Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Modelling fluvial responses to episodic sediment supply regimes in mountain streams Müller (Mueller), Josef Tobias 2019

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2019_may_mueller_josef.pdf [ 65.7MB ]
Metadata
JSON: 24-1.0377728.json
JSON-LD: 24-1.0377728-ld.json
RDF/XML (Pretty): 24-1.0377728-rdf.xml
RDF/JSON: 24-1.0377728-rdf.json
Turtle: 24-1.0377728-turtle.txt
N-Triples: 24-1.0377728-rdf-ntriples.txt
Original Record: 24-1.0377728-source.json
Full Text
24-1.0377728-fulltext.txt
Citation
24-1.0377728.ris

Full Text

Modelling fluvial responses to episodic sediment supplyregimes in mountain streamsbyJosef Tobias Mu¨ller (Mueller)Diplom-Geograph, Geographisches Institut, RheinischeFriedrich-Wilhelms-Universita¨t Bonn, 2011A THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFDoctor of PhilosophyinTHE FACULTY OF GRADUATE AND POSTDOCTORALSTUDIES(Geography)The University of British Columbia(Vancouver)March 2019c© Josef Tobias Mu¨ller (Mueller), 2019The following individuals certify that they have read, and recommend to the Fac-ulty of Graduate and Postdoctoral Studies for acceptance, the dissertation entitled:Modelling fluvial responses to episodic sediment supply regimes in mountain streamssubmitted by Josef Tobias Mu¨ller (Mueller) in partial fulfillment of the requirementsfor the degree of Doctor of Philosophy in GeographyExamining Committee:Marwan A. HassanSupervisorErkan IstanbulluogluSupervisory Committee MemberIan McKendrySupervisory Committee MemberBrett EatonUniversity ExaminerRoger BeckieUniversity ExamineriiAbstractLarge, episodically occurring sediment supply events may temporarily dominate channel mor-phology and sediment transport in mountain streams. Field studies of channel response tothese events are challenging to undertake, as a long data record is needed to reasonably as-sess a system’s state of response in the context of episodic supply. Greater confidence in theobserved state of response of a system can be achieved with flume experiments where fluvialresponse can be observed in detail after episodic events are introduced in a controlled fashion.Yet, the amount of work necessary to carry out these experiments is large, which limits thenumber of experimental conditions that can be studied, and thus their utility for addressingapplied problems of channel adjustment.To overcome this limitation, I developed the 1-D morphometric sediment transport modelBESMo, which allows large numbers of simulations to be run in batches, generating ensembleresults. This model was used to recreate results from flume experiments, after which the exper-imental conditions were extended to include a broader range of simulated pulse frequencies,magnitudes, and grain size compositions. It was shown that the sequencing of pulse eventsof different magnitudes has only a short term effect on the slope and grain size response ofthe channel. Furthermore, thresholds were identified that allow for the categorization of flu-vial response to episodic sediment supply regimes into one of (a) constant-feed-like, or (b)pulse-dominated.The practical utility of BESMo for studying fluvial response to large sediment supplyevents was demonstrated through the study of potential geomorphic effects following theremoval of a dam in the Carmel River, California, USA. This showed the advantage of BESMofor simulating many different future scenarios, as stochasticity could be explicitly includedthrough varied hydrographs. This allowed results to be interpreted in light of the uncertaintyin future flood occurrence. Finally, to overcome data limitations on surface grain size distri-butions, I developed machine-learning based methods to detect grain size distributions fromimages. Collectively, this work has advanced our understanding and ability to characterisedownstream channel response to episodic supply events, and to better obtain data needed forthis characterisation.iiiLay SummaryThe objective of this thesis is to understand how mountain streams adjust to sediment fromepisodic events, such as landslides. This is difficult to study in the field or lab, where long ob-servation times are needed to capture a river’s response to multiple events. To overcome thislimitation, I developed BESMo, a versatile model for studying channel adjustment to sedimentadditions. I found that in the long term, the ordering of events is not important for the adjust-ment of the river. Furthermore, rivers adjusts to episodic events in different modes, based onwhether events occur more (or less) frequently than the river can rework them. I then used BE-SMo to simulate how the Carmel River adjusts to the removal of a dam, helping to determinethe best options for river management in the future. Finally, I developed a machine-learningbased method to detect grain size distributions from images.ivPrefaceThis thesis is original work completed by J. Tobias Mu¨ller. Guidance was given by the super-visory committee. This thesis includes one manuscript, and two complementary chapters thatare planned to be submitted for publication as individual manuscripts. Chapter 2 is publishedin Earth Surface Dynamics (Mu¨ller and Hassan, 2018) with co-author Marwan A. Hassan, whosupervised the work. Chapter 3 and Chapter 4 are complementary Chapters.The work in Chapter 3 is based on a project done for the Monterey Peninsula Water Man-agement District (MPWMD) and the California American Water Company (CalAM) in coop-eration with Balance Hydrologics and AECOM, resulting in a separate report that was sub-mitted for review in November 2018 and is being prepared for publication. Shawn Chartrandand Kealie Pretzlav helped with gathering the data, and Leonora King helped editing the text.Mu¨ller, T. and M. A. Hassan (2018), Fluvial response to changes in the magnitude and fre-quency of sediment supply in a 1-D model. Earth Surface Dynamics, 6(4), 10411057.vContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiLay Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xList of Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiiList of Acronyms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xivAcknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvDedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvii1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Systems approach to episodic events . . . . . . . . . . . . . . . . . . . . . . . . 41.2 System understanding from numerical models . . . . . . . . . . . . . . . . . . . 72 Fluvial response to changes in the magnitude and frequency of sediment supplyin a 1-D model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142.3.1 Model setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152.3.2 Model calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192.3.3 Event sequencing simulations . . . . . . . . . . . . . . . . . . . . . . . . 222.3.4 Equilibrium simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242.4.1 Model calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24vi2.4.2 Event sequencing simulations . . . . . . . . . . . . . . . . . . . . . . . . 252.4.3 Equilibrium simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . 272.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292.5.1 Extension of flume results with the numerical model . . . . . . . . . . . 292.5.2 Development of timescales from the equilibrium simulations . . . . . . 302.5.3 Interpretation of the equilibrium simulations . . . . . . . . . . . . . . . 342.5.4 Implications of the equilibrium simulations . . . . . . . . . . . . . . . . 352.6 Summary and conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403 Simulation of sediment pulse effects on Carmel River after the Los Padres Damremoval, Monterey County, California, USA . . . . . . . . . . . . . . . . . . . . . . . 433.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443.3 Study area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463.4 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483.4.1 Model boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . 523.4.2 Node initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533.4.3 Hydrology submodel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553.4.4 Sediment supply . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 693.5.1 Overview of bed elevation changes along the channel . . . . . . . . . . 743.5.2 Overview of surface grain size changes along the channel . . . . . . . . 763.5.3 Comparison to MEI simulations . . . . . . . . . . . . . . . . . . . . . . . 773.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 793.6.1 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 793.6.2 Placing simulation results within the Carmel River context . . . . . . . 823.6.3 Comparison to other dam removal projects . . . . . . . . . . . . . . . . 843.6.4 Reflection on the use of BESMo . . . . . . . . . . . . . . . . . . . . . . . 863.7 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 874 Identifying surface grain size distributions from images using computer visionand machine learning methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 894.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 894.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 904.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 944.3.1 Bed surface datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 944.3.2 Identifying individual grains: StoneID . . . . . . . . . . . . . . . . . . . 954.3.3 Matching GSDs to colour images: DistID . . . . . . . . . . . . . . . . . 984.3.4 CNN to map grain areas to colour images: U-Net+StoneID . . . . . . . 99vii4.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1004.4.1 StoneID . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1004.4.2 DistID . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1014.4.3 U-Net+StoneID . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1024.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1024.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1045 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1105.1 Summary of contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1115.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118A Sensitivity analysis plots for BESMo in episodic supply experiments . . . . . . . . 130B Los Padres Dam removal supplementary material . . . . . . . . . . . . . . . . . . . 136B.1 Model testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136B.2 Elevation data availability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141B.3 Sediment transport data availability . . . . . . . . . . . . . . . . . . . . . . . . . 144B.4 Rating Curves Mainstem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147B.5 Rating Curves Tributaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149B.6 Node Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153B.7 Grain Size Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155B.8 Reach Averaged Crosssections . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156B.9 Reach Flow Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160B.10 Median Sediment Supply and Reservoir Depletion Curves . . . . . . . . . . . . 164B.11 Detailed Results of the focus scenarios . . . . . . . . . . . . . . . . . . . . . . . . 167B.11.1 No Action Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . 167B.11.2 Historical Supply Simulation Results . . . . . . . . . . . . . . . . . . . . 175B.11.3 Pulsed Supply Simulation Results . . . . . . . . . . . . . . . . . . . . . 181B.11.4 Uncontrolled Release Simulation Results . . . . . . . . . . . . . . . . . . 188B.12 Median Results of focus Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . 195B.12.1 Average Hydrographs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195B.12.2 Dry Hydrographs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198B.13 Median Results of all Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201C DistID and U-Net model parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 213C.1 Employed ResNet50 architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . 213C.2 Employed VGG19 architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214C.3 Employed UNet architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217viiiList of Tables2.1 Overview of runs in the flume experiments. . . . . . . . . . . . . . . . . . . . . 212.2 Sequencing of events in runs that were simulated as permutations of the origi-nal flume experiment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.3 List of forcing and reacting parameters as well as timescales in our simulations. 372.4 Application of the fluvial evacuation time to the rapids reach in East Creek. . . 383.1 Complete list of model parameters for No Action simulation and changes ap-plied to Historical, Pulse, and Uncontrolled Supply simulations. . . . . . . . . 493.2 Carmel River main stem discharge ratio from 4748 days of overlapping meandaily flow data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573.3 Flood frequency table for the Robles del Rio USGS gauge using mean dailydischarge on the day of yearly peak flow. . . . . . . . . . . . . . . . . . . . . . . 583.4 Peaking factor for flood classes from the historical peak flows. . . . . . . . . . . 593.5 Probability of number of floods per year and ordered peak mean daily flowratios between the floods within one year in relation to the highest mean dailyflow at the Robles del Rio USGS gauge . . . . . . . . . . . . . . . . . . . . . . . 603.6 Overview of sediment feed scenarios by rating curve type. . . . . . . . . . . . . 663.7 Exponential decay curves for the three simulated scenarios. . . . . . . . . . . . 663.8 Values of E∗, the ratio between sediment volume eroded in the first year af-ter the management action begins with the first flood event in each time series,calculated from median sediment supply values after 1 simulation year for bed-load material. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 854.1 Overview of the grain detection methods used in this chapter. . . . . . . . . . . 944.2 Overview of colour thresholds per size class in the experiments, some devel-oped from HSV data, some from RGB data. . . . . . . . . . . . . . . . . . . . . . 974.3 Difference in reported accuracy and loss from model configurations. . . . . . . 101ixList of Figures1.1 Landslide providing episodic sediment supply into a stream in Hinton, Alberta 11.2 Event effectiveness as a function of frequency and magnitude . . . . . . . . . . 61.3 Temporal features of disturbance reaction with the example of sediment storagein a streambed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62.1 Flowchart stating the main components of the model and the flow of informa-tion between them. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152.2 (a) Variation of the grain size distribution between model runs. (b) Combina-tions of pulse magnitude and pulse period (or recurrence interval) used in theequilibrium model runs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232.3 Comparison of (a) slope, (b) mean surface Dsg, (c) surface Ds90, and (d) sedi-ment transport rate between the numerical simulation and the flume experiments 262.4 Comparison of slope and mean surface grain size Dsg from runs with differentevent sequencing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282.5 (a) Slope and (b) armouring ratio for the last 400 h of 2 out of 40 experimentsusing σ = 1.6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292.6 (a) Distribution of the ratios of slope during the last pulse to the constant-feedslope for all 40 runs with σ = 1.6. (b) Mean slope ratios for all runs grouped bywidth of GSD (σ). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302.7 (a) Mean slope ratios in non-dimensional timescale. (b) Relative armouringratio in non-dimensional timescale. . . . . . . . . . . . . . . . . . . . . . . . . . 333.1 Map of simulation reaches. The exact location of each node can be found inAppendix Table B.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473.2 Long profile of the full run simulation domain with annotation showing theposition of the fixed-elevation boundary condition. . . . . . . . . . . . . . . . . 523.3 Components of the calculation of the flood time series with the hydrology sub-model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 563.4 Number of floods per water year from the historical record at the Robles delRio USGS gage. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59x3.5 Relative mean daily flow as time series in five flow classes from the averagesof all events in the historical record. Timing of floods in the calendar year inrelation to mean daily peak flood magnitude. . . . . . . . . . . . . . . . . . . . 613.6 Cumulative discharge for 1000 randomly generated hydrographs after 10, 30,and 60 years, sorted by 10-year cumulative discharge. Subsection of 300 runsthat were simulated for each project simulation with BESMo. . . . . . . . . . . 613.7 Exponential decay curves modelled after Marmot dam removal data. . . . . . 683.8 Comparison of projected bed elevation change from the 2017 initial profile forthe wet hydrologic condition and the Historical, Pulse and Uncontrolled sedi-ment supply simulations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 703.9 Comparison of projected change of the geometric mean grain size of the bedsurface Dg for the wet hydrologic condition and the Historical, Pulse and Un-controlled sediment supply simulations. . . . . . . . . . . . . . . . . . . . . . . 713.10 Comparison of projected change of the geometric mean grain size of the bedsurface D90 for the wet hydrologic condition and the Historical, Pulse and Un-controlled sediment supply simulations. . . . . . . . . . . . . . . . . . . . . . . 723.11 Comparison of projected bed elevation change simulated by BESMo and re-ported by MEI (2002) for the roughly equivalent condition of removing SanClemente Dam and no bedload bypass at Los Padres Dam. . . . . . . . . . . . . 784.1 Identification of individual stones from colour thresholded images with StoneID. 964.2 Using classified stone areas from U-Net with the StoneID method and directpixel count GSD prediction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1004.3 Results of StoneID for YW data cropped from the original image. . . . . . . . . 1054.4 Comparison of StoneID derived GSDs to manual data. . . . . . . . . . . . . . . 1064.5 Comparison between mean sizes of D90, D85, D65, D50, D25, and D16 fromStoneID and manual data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1064.6 Results of StoneID for AM data cropped from the original image. . . . . . . . . 1074.7 Comparison of StoneID and DistID (VGG19,HSV) derived GSDs to manual data.1084.8 Comparison between mean sizes of D90, D85, D65, D50, D25, and D16 fromStoneID and DistID (VGG19,HSV) to data from manual grid-based counts. . . 1084.9 Comparison of StoneID and U-Net derived grain area predictions. . . . . . . . 109xiList of SymbolsSymbol DefinitionARelp Armouring ratio at end of last pulseARconst Armouring ratio at end of constant feed runDsg Surface geometric mean grain size [m]Dsubg Subsurface geometric mean grain size [m]Ds50 Surface 50th percentile grain size [m]Ds84 Surface 84th percentile grain size [m]Ds90 Surface 90th percentile grain size [m]D f g Feed geometric mean grain size [m]D f 90 Feed 90th percentile grain size [m]Dag Armoured geometric mean grain size [m]Da90 Armoured 90th percentile grain size [m]f Ii Proportion of ith grain size class exchanged betweenthe surface and the subsurfaceFi Surface frequency of ith grain size classFr Froude numberFpulse Pulse frequency [1/s]g Gravity [m/s2]GSD f luv Channel surface grain size distributionGSDpulse Pulsed sediment feed grain size distributionh Water depth [m]ks Roughness height [m]lr Channel length [m]La Active layer thickness [m]Mpulse Pulse magnitude [m3]nk Roughness height coefficientna Scale of bed height fluctuationpbi Bedload transport rate fraction of ith grain size classqb Bedload transport rate [m2/s]qbi Fractional bedload transport rate [m2/s]Qw Water discharge [m3/s]S f Friction slope [m/m]S0 Bed slope [m/m]Smlp Mean channel slope at end of last pulse [m/m]xiiSconst Slope at the end of constant feed run [m/mt Time [s]Tpp Pulse period [s]Tf e Fluvial evacuation time [s]Tar Fluvial armouring time [s]Tsim Duration of simulation [s]u∗ Dimensionless shear velocityU f luv Fluvial export velocity [m/s]Upulse Virtual pulse velocity [m/s]wr Channel width [m]x Downstream distance [m]αr Manning-Strickler coefficientα Active layer exchange ratioηb Bed surface elevation [m]λ Bed porosityρ Water density [kg/m3]σ Wideness of generated grain size distributionτb Boundary shear stress [Pa]τrm Reference shear stress [Pa]xiiiList of AcronymsAcronym DescriptionAM Dataset of flume images collected by Alex MitchellAECOM Architecture, Engineering, Consulting, Operations, andMaintenance (engineering company)BESMo Bedload Scenario Model (modelling software developed for thisthesis)BOMC Bed Of Many ColoursCNN Convolutional Neural Network (type of machine learningarchitecture)CSUMB California State University, Monterey Bay (educationalinstitution)CRRDR Carmel River Reroute and Dam Removal (project)DistID Distribution Identification (Method developed in Chapter 4)GIS Geographic Information System (software for spatial analysis)GSD Grain size distributionHEC-RAS Hydraulic Engineering Center - River Analysis System, USArmy Corps (modelling software)HEC-SSP Hydraulic Engineering Center - (modelling software)HSV Hue, Saturation, Value (digital colour representation)LPD Los Padres Dam (structure)MPWMD Monterey Peninsula Water Management District (public agency)MEI Mussetter Engineering Inc. (engineering company)PRMS Precipitation-Runoff Modeling System (modelling software)ResNet50 Residuals Network, size 50 (type of CNN)RGB Red, Green, Blue (digital colour representation)SRH-1D Sedimentation and River Hydraulics - One Dimension, USBureau of Reclamation (modelling software)StoneID Stone Identification (Method developed in Chapter 4)UAV Unmanned Aerial Vehicle (remote controlled aircraft, commonlyknown as a drone)URS United Research Services (engineering company, now part ofAECOM)USGS United States Geological Survey (public agency)VGG19 Visual Geometry Group, size 19 (type of CNN)YW Dataset of flume images and GSDs collected by Yinlue WangxivAcknowledgementsI first want to thank my family for their love and moral support. Even though I moved faraway to write this thesis, my parents, sisters, and brothers-in-law could always cheered meup with calls and videos of my nieces and nephews. I want to thank all my friends back inGermany, who showed in visits that 6 years abroad never caused our friendships to change.This is especially true for my lifelong friends Felix M., Martin H., and Matthias. Urs, FelixS., Steph, and crew kept me on my heels, thank you for that. Christoph Disch, along with allmy other friends from Bonn, provided moral support over the blackest of coffees. Thank youThomas Hoffmann for getting the first stones rolling.The journey through the PhD program would have been immensely harder without myfriends and colleagues in Vancouver. Thank you Leah and Bodhi for cheering me up manytimes. Leonora King and David Reid are the best housemates to have (especially while finish-ing a thesis) and helped greatly with proof-reading. The labgroup was an amazing supportin discussing all things from raw ideas to the tiniest details of plots. Thank you Shawn foroffering me the chance to work on a great project.Thank you Marwan for being an amazing supervisor, whose door was always open toprovide advice, encouragement, and support. Thank you for being patient even though Iovershot all deadlines, thank you for catching my worst mistakes, and thank you for trustingin my abilities. Thank you Erkan and Ian for supporting me in this endeavour.Regarding the second chapter, I thank Maria Elgueta and Claudia von Flotow for their col-laboration in conducting the flume experiments, Conor McDowell for comments on an earlydraft of this work, and Eric Leinberger for improving the figures for publication. I am gratefulxvfor suggestions by Gary Parker and Chenge An that improved this chapter, editing by Kim-berly Hill, and reviews by three anonymous referees. Carles Ferrer-Boix helped greatly in thedevelopment of an early version of BESMo. Regarding the third chapter, I thank Shawn Char-trand and Kealie Pretzlav for sourcing some of the input data for the model and providingsupport for writing the text. Final editing support from Leonora King improved this chaptergreatly. I thank Alex Michell and Yinlue Wang for providing data of their flume experimentsthat was used to develop the methods in the fourth chapter, and Matteo Saletti for providingvaluable insights on the stone identification approach.Computational resources for both the numerical simulations and the machine learningtraining were provided by WestGrid1 and Compute Canada2.I thank the University of British Columbia for supporting me with a Four Year Fellowshipand the Stiftung fu¨r Kanadastudien for providing a travel grant that helped me to conduct theflume experiments.1www.westgrid.ca, last access: Jan 1st, 20192www.computecanada.ca, last access: Jan 1st, 2019xviDedicationFu¨r meine Eltern Ellen und Karl-Josef. Danke fu¨r die unendliche Unterstu¨tzung.xviiChapter 1IntroductionFigure 1.1: Landslide providing episodic sediment supply into a stream in Hin-ton, Alberta. Credit: Marwan A. Hassan.Rivers are commonly affected by large and episodic sediment inputs, a result of both naturallyoccurring landslides (Korup, 2013), and human activity, such as mining (Nelson and Church,2012) or dam removals (Grant and Lewis, 2015). These large sediment supply events mighttemporarily dominate sediment transport and storage regimes within river channels (Hassan1et al., 2005, 2008a). Hassan et al. (2018) found that especially in formerly glaciated mountainenvironments the coupling of fluvial channels to hillslopes is widespread, leading to a highpotential for colluvial input into the channel. Consequently, these episodic supply eventslead to downstream impacts that can impact fish habitat (May and Lisle, 2012) and humaninfrastructure (Kondolf and Pie´gay, 2011).The range and magnitude of natural and anthropogenic supply events, in combinationwith the complexity of geomorphic mountain channel processes (Wohl, 2013), make mountainriver form and behaviour particularly challenging to manage and predict. This complexity ne-cessitates a better understanding of how large sediment supply events affect the morphologyand sediment transport dynamics of mountain streams. Despite over a century of research intosediment transport, we continue to lack understanding of many aspects of channel responseto changing sediment supply (Hassan and Zimmermann, 2012).Sediment supply plays an important role in shaping the morphology and the sedimenttexture of gravel bed rivers from the reach to the catchment scale (Hassan et al., 2008a; Pfeifferet al., 2017). In the context of sediment transport, rivers are often conceptualized to be in oneof two states: (1) supply limited, or (2) transport limited (Montgomery and Buffington, 1997).When the sediment supply rate exceeds the transport capacity of a stream, storage of mate-rial increases and the system is considered transport limited. If the transport capacity of thestream is higher than the sediment supply rate, stored sediment is depleted and the channelis considered supply limited. However, this simple conceptualization is rendered more com-plex when storage dynamics are considered, which temper a channel’s response to externalsediment supply (Lisle and Church, 2002). Finally, sediment supply plays an important role inchanges in the bed surface grain sizes and the development of bed surface structures, whichcan reduce the sediment transport capacity by stabilizing the bed surface (Dietrich et al., 1989;Church et al., 1998).To better understand the impact of changes in frequency, magnitude, and grain size ofsediment supply events on a fluvial channel, simplified process interactions can be studiedin both flume experiments (e.g. Cui et al., 2003; Elgueta-Astaburuaga and Hassan, 2017; Johnsonet al., 2015; Sklar et al., 2009) and numerical models (e.g. Cui and Parker, 2005; An et al., 2017a).2While both approaches have provided significant insight into the complex process interac-tions in the fluvial reworking of supply events, these studies simulated channel response toa relatively limited range of event frequencies and magnitudes. For this thesis the 1-D mor-phometric model BESMo (’Bedload scenario model’) was developed, which allows for manydifferent configurations of sediment supply and hydrology conditions to be evaluated con-currently. BESMo has the advantage of incorporating computational methods that allow forthe consideration of uncertainty and stochasticity in a channel’s response to episodic sedimentsupply.Using the developed model, this thesis investigates the following general questions:1. How do changes in the frequency and magnitude of sediment pulses impact gravel bedrivers at the reach scale?2. How do different sediment supply regimes in combination with different hydrologicconditions impact the river downstream?Associated to these questions and motivated by a lack of data in answering the earlier ques-tions, a methodological problem was identified:3. What method can improve the identification of grain size distributions in flume experi-ments?After a general introduction to the research approach and the significance of numericalmodels, Chapter 2 describes the development of BESMo and how it is used to expand uponsupply conditions examined in flume experiments, which provides insight into research ques-tion 1 (above). The second research question is explored in Chapter 3 using a dam removalcase study on the Carmel River in California. For this chapter, different sediment supply con-ditions were simulated in BESMo with hundreds of different simulated hydrographs, whichwere generated using a stochastic approach based on historical flood records. Due to theimportant role that armouring and channel bed features play in the bed surface response toepisodic sediment supply, a need for better grain size characterisation in flume experimentswas identified. To address this issue (stated in the third research question), an automated sta-31.1. Systems approach to episodic eventstistical approach for identifying both grain size distributions and individual grains in imagestaken of flume surfaces was developed, described in Chapter 4.The following section provides an overview of the role episodic events play in geomorphicsystems. This leads into a section on the advantages of numerical models to explore this rolein greater detail, providing grounds for the use of numerical models to answer the researchquestions of this thesis.1.1 Systems approach to episodic eventsChorley and Kennedy (1971) suggest a system hierarchy to structure the understanding of nat-ural processes. One of the theoretical frameworks suggested is to order processes and formsinto ‘cascading systems’. Within these systems, components are characterized by chained en-ergy or mass throughputs, and therefore the output of one subsystem is the input of the next.This concept is used for studies of sediment dynamics in mountain catchments in the form ofsediment budgets, in which stored and transported sediment volumes are quantified for allrelevant processes (Slaymaker, 2003). This puts the spatial hierarchy of a sediment cascade inthe temporal context of time spent in storage or transport and allows a picture of catchmentscale sediment dynamics to be derived. The spatial components of the system are homoge-neous areas, distinguishable by material, morphometry or type of active processes. These canbe derived from field mapping or GIS analysis, as for example done by Schrott et al. (2003) andHoffmann et al. (2013).Morphometric indices such as curvature, slope, or catchment area can be used to delin-eate system components in the landscape. Montgomery and Foufoula-Georgiou (1993) describea distinctive relation between the morphometric indices ‘local slope’ and ‘contributing area’,which, when applied to their study catchments, depict an inflection point that indicates achange from debris-flow-dominated channels to alluvial channels. Montgomery (1999) furtherinvestigate the spatial and temporal aspects of form-process relationships to develop a ‘pro-cess domain’ concept. He argues that spatial variability in process activity must be consideredto understand temporal patterns of system change and forcing of system reactions to distur-bance. An application of this concept to a formerly glaciated catchment in British Columbia is41.1. Systems approach to episodic eventsgiven by Brardinoni and Hassan (2006). In their study area, the topographic signatures do notmatch currently active processes. The authors ascribe this disequilibrium to the relict glacialforms, whose legacy is more dominant than the morphologic activity of contemporary pro-cesses. Hence, the application of the process domain concept must be viewed in the context ofthe individual history and configuration of the system studied.After characterising the spatial extent of the sediment cascade, knowledge of the tempo-ral patterns of linkage functionality between the components of the geomorphic system isneeded to understand sediment transport dynamics. If the action of a process is episodic, theeasiest way to assess volumes of transported sediment is by building a relationship betweenthe frequency and magnitude of events. In debris flow dominated catchments in particular,information about event frequency is of high importance in the study of sediment transport(Dietrich and Dunne, 1978).Both fluvial and hillslope domains are subject to the action of highly episodic processes.Wolman and Miller (1960) introduced the term geomorphic effectiveness, which is calculatedas the product of frequency and magnitude of events. They found that most work is doneby moderately sized events, as the events highest in magnitude are low in frequency (see Fig-ure 1.2). River channels are shaped by the sequenced action of floods with varying magnitudesand frequencies, but fluvial sediment transport is in most cases simulated as an linear func-tion of discharge. Even though the frequency-magnitude concept could allow for the calcula-tion of a long-term average impact, the relationship between flood magnitudes and sedimenttransport is more complex. For example, flows of similar peak magnitude show different ge-omorphic effectiveness in shaping morphology or transporting sediment due to hydrographcharacteristics (Hassan et al., 2006a, 2014). Guthrie and Evans (2007) analyse geomorphic ef-fectiveness of landslides in terms of work (material transported), persistence (residence time)and formation (morphological impact) and find moderately sized events to be most efficient.They also find the persistence time of the impact to be the most meaningful measure of theeffectiveness of events.The effect of episodic supply events on the fluvial system can be viewed under a distur-bance concept. Bull (1991) describes a hypothetical temporal succession of disturbance events51.1. Systems approach to episodic eventsApplied stressMagnitudeFrequencyWorkFigure 1.2: Event effectiveness as a function of frequency and magnitude. After:Wolman and Miller (1960).with the example of sediment storage in a stream-bed (see Figure 1.3). Episodic disturbanceevents of sediment contribution and sediment export result in an increase or decrease of thestream-bed elevation, respectively. This reaction is not instantaneous, but delayed by a reac-tion time (Ra) in which no adjustment takes place. The system processes the disturbance overthe relaxation time (Rx), which together with the reaction time constitutes the total responsetime (Rt). If there is no further disturbance, the system status remains unchanged over a timeof persistence (Ps). Sequenced disturbance events might lead to overlap of reaction times andrelaxation times and thus create a complex pattern of system changes. Due to the complex-ity of geomorphic systems, it might be impossible to discern these temporal phases of systemreaction in field studies.TimeSedimentstorageEvent 1Event 2Event 3Reaction RelaxationResponseReactionResponseRelaxationPersistenceReactionFigure 1.3: Temporal features of disturbance reaction with the example of sedi-ment storage in a streambed. After: Bull (1991).61.2. System understanding from numerical models1.2 System understanding from numerical modelsNumerical models are used to combine the understanding of multiple processes in a frame-work that enables us to study process interactions that represent complex real world systems(Phillips, 2003). Use of these models can serve to meet practical objectives linking cause andconsequence, but can also allow for exploration of new research questions and experimentaldesigns (Escauriaza et al., 2017). Wilcock and Iverson (2003) state that while prediction with-out a model is just speculation, using models without making predictions can be helpful forreviewing our understanding how processes interact and expose knowledge gaps.Numerical process modelling has a long tradition in fluvial geomorphology (Coulthardand Van De Wiel, 2013) and is based on the development of process understanding throughmeasuring process rates. Dietrich et al. (2003) give an overview of mechanistic mathematicalmodels of separate morphological processes, which if combined in a framework and solvedtogether builds a landscape evolution model. These models open up the opportunity to predictbehaviour of rivers from the reach to the landscape scale, for instance by linking hillslopeswith fluvial domains (Tucker and Bras, 1998). Recently, modelling has focussed on the effectsof environmental change (Van De Wiel et al., 2011), necessitating the inclusion of processes thatwere not traditionally a focus in geomorphology, such as the effect of vegetation on landscapeevolution (Istanbulluoglu, 2005).The development of combined process models is rooted in the previously described ’sys-tems approach’. Numerical models in geomorphology must be based on an understanding ofthe process interactions at play, which can be illustrated in the form of a sediment cascade. Ifa process that affects the studied phenomena is not represented numerically, the prediction ofthe model will be (within the accuracy that can be expected from a simplified model) false andpotentially misleading. In general, more complex problems require more complex models, asthe effect of more processes and/or conditions must be considered (Van De Wiel et al., 2011).Model predictions must be critically evaluated as a matter of course. Real geomorphicsystems are thermodynamically open systems, where we cannot know all energetic exchangesand processes at play within the boundary of a particular study. Furthermore, the history of71.2. System understanding from numerical modelsthe modelled system might be more important for the result than the mechanics of the processinteractions, which is typical for chaotic systems (Phillips, 2003). This makes finding realisticapproximations of complex open systems difficult, which would still be true even if we hadperfect formal descriptions of the individual contemporary processes (Beven, 2002).A way to approximate uncertainty is to link the range of model results to the uncertaintyin input conditions by running many simulations ’Monte-Carlo style’, where the separate sim-ulations begin with varied initial conditions (Willgoose et al., 2003). This approach is used hereand is a main advantage of BESMo over previous models, due to its capability of running manysimulations at once on a computing cluster and easily compiling batch results. The model wascalibrated with results from flume experiments and then used to explore the effect of variedevent frequency, magnitude and grain size composition (see Chapter 2). In this way, the modelwas used to expand our theoretical understanding of the impact of hillslope processes on thefluvial system. The capability of the model to explore different management options in a riveris demonstrated in Chapter 3 for a dam removal case. This case study brought challenges oflow data availability and unknown future hydrology: while these difficulties might appear topresent a problematic situation in which to apply a model, a modelling effort here is the bestway to assess the sensitivity of the system to future conditions (Oreskes et al., 1994).8Chapter 2Fluvial response to changes in themagnitude and frequency of sedimentsupply in a 1-D model2.1 SummaryIn steep headwater reaches, episodic mass movements can deliver large volumes of sedimentto fluvial channels. If these inputs of sediment occur with a high frequency and magnitude,the capacity of the stream to rework the supplied material can be exceeded for a significantamount of time. To study the equilibrium conditions in a channel following different episodicsediment supply regimes (defined by grain size distribution, frequency, and magnitude ofevents), we simulate sediment transport through an idealized reach with our numerical 1-Dmodel “BESMo” (Bedload Scenario Model). The model performs well in replicating flume ex-periments of a similar scope (where sediment was fed constantly, in one, two, or four pulses)and allowed the exploration of alternative event sequences. We show that in these experi-ments, the order of events is not important in the long term, as the channel quickly recoverseven from high magnitude events. In longer equilibrium simulations, we imposed differentsupply regimes on a channel, which after some time leads to an adjustment of slope, grain92.2. Introductionsize, and sediment transport that is in equilibrium with the respective forcing conditions. Weobserve two modes of channel adjustment to episodic sediment supply. (1) High-frequencysupply regimes lead to equilibrium slopes and armouring ratios that are like conditions inconstant-feed simulations. In these cases, the period between pulses is shorter than a “fluvialevacuation time”, which we approximate as the time it takes to export a pulse of sedimentunder average transport conditions. (2) In low-frequency regimes the pulse period (i.e., recur-rence interval) exceeds the “fluvial evacuation time”, leading to higher armouring ratios dueto the longer exposure of the bed surface to flow. If the grain size distribution of the bed is fineand armouring weak, the model predicts a decrease in the average channel slope. The ratiobetween the “fluvial evacuation time” and the pulse period constitutes a threshold that canhelp to quantify how a system responds to episodic disturbances.2.2 IntroductionMass movements in mountainous regions often deliver sediment directly to the stream net-work, resulting in coupled conditions that can trigger immediate channel responses duringrelatively large delivery events. Notably, delivery events can reset the local channel profileand govern construction and maintenance of channel bed architecture downstream of deliv-ery points. The local response rate and trajectory following a delivery event is a function ofthe prevailing watershed flow regime, the magnitude of the delivery event, gradients in chan-nel width (Ferrer-Boix et al., 2016), and in some instances the concentrated activity of aquaticspecies such as salmon (Hassan et al., 2008b). While lowland river systems have been the fo-cus of a substantial body of research, less work has been carried out within steep mountainstreams, particularly concerning mountain channel responses to changes in flow or sedimentsupply regimes.In mountain streams, large, episodic inputs may temporarily dominate channel processesand morphology, significantly altering sediment transport and storage within the stream chan-nel (Hassan et al., 2005, 2008a). Lisle et al. (1997) conducted flume experiments and numericalmodelling that showed that sediment pulses are mainly reworked in situ, in contrast to adownstream translation in the form of a sediment wave. This finding is supported by Lisle102.2. Introductionet al. (2001), where little evidence of sediment waves was found in the field. Lisle and Church(2002) suggested that a stream channel responds to changes in the sediment supply by alteringboth storage and sediment transport rates. They describe that after a sediment pulse occurs,a first phase with low armouring rates allows for high transport rates in reworking the in-troduced material, corresponding to supply limited conditions. This is followed by a secondphase in which armouring develops and transport rates decrease, corresponding to transportlimited conditions. If the fluvial capacity is too low to evacuate the mass or grain size ofmaterial in the current hydrological regime, lag sediment can remain in the channel and dom-inate the local morphology for a long time (e.g., Benda et al., 2005; Brummer and Montgomery,2006). Patterns of cyclic behaviour, associated with the rapid input of sediment from exter-nal sources, have been described in a number of field observations (e.g., Roberts and Church,1986; Madej and Ozaki, 1996; Madej, 1999, 2001; Miller and Benda, 2000; Hoffman and Gabet, 2007;Hassan et al., 2008a) and some experimental studies (e.g., Cui et al., 2003; Sklar et al., 2009; Ven-ditti et al., 2010; von Flotow, 2013; Elgueta, 2014; Ferrer-Boix and Hassan, 2014, 2015; Johnson et al.,2015; Elgueta-Astaburuaga and Hassan, 2017; An et al., 2017a). These studies observed fining ofthe bed surface, higher mobility and thus increased transport rates following such episodicsediment supply events. A reverse trend for coarsening and stabilizing of the bed was alsonoted as the supply was exhausted and a decrease in sediment transport rates followed (e.g.,Dietrich et al., 1989; Church et al., 1998; Hassan and Church, 2000; Nelson et al., 2009). Cui andParker (2005) support these findings using numerical modelling and further point out thatabrasion can play an important role in the reworking of sediment pulses. Field observations(e.g., Benda, 1990; Pryor et al., 2011), and flume experiments (e.g., Pryor et al., 2011; Luzi, 2014)also document cycles of aggradation and degradation due to changes in the sediment supply.The observations discussed above suggest that changes in the sediment supply rate may leadto significant changes in bed elevation, bed surface texture, channel stability, and bed mor-phology. An analytical model developed by Blom et al. (2017) showed that the local channelgeometry and surface grain size composition is mainly governed by long-term mean sedimentsupply rates and not by short-term changes in supply conditions. In contrast to cyclic sedi-ment supply, cyclic hydrographs were found to mainly affect sediment transport rates and112.2. Introductionhave a lesser impact on bed surface texture and channel morphology (Parker et al., 2007). Wongand Parker (2006) reported that cyclic hydrographs cause a part of the channel bed to undergocyclic aggradation and degradation forming a hydrograph boundary layer. Cyclic sedimentsupply causes a similar effect which is termed the sedimentograph boundary layer (An et al.,2017b).These findings imply that the morphological impact of the sediment pulse is most preva-lent at the point of entrance, while the downstream portion mostly conveys the subsequentlyeroded material. The time needed for channel adjustments to occur after a large sedimentinput event depends on the amount and the texture of the delivered material. Brummer andMontgomery (2006) reported that 2 years after the supply event, the most mobile fractions (e.g.,the fine fractions) were evacuated by a series of moderate floods while the largest grain sizesremained in the channel as lag deposits because flows were below their flow competence. Fur-ther, they showed that selective transport of sediment led to the development of an armourlayer after only a few flow events. This armour layer protects the supplied material in the bedsubsurface, increases bed stability, and causes lateral erosion and channel widening.We expect the sediment transport rate in a channel to reach a long-term balance betweenerosional and depositional forces, even though there can be periodic changes in the short term.This state is defined as a “dynamic equilibrium” following Ahnert (1994). If the external forc-ing on the system changes, the channel will be in a transient state of adjustment towards anew dynamic equilibrium. Little attention has been directed to the question of what effecta change in the frequency of sediment supply events may have on the response of alluvialstreams. Brunsden and Thornes (1979) proposed that if the frequency (i.e., recurrence interval)of disturbing events is shorter than the time necessary for a system to adjust to new bound-ary conditions (“relaxation time”), then transience will dominate the system and it may neverachieve equilibrium (Brunsden, 1980). Wolman and Miller (1960) suggest that mountain chan-nels experiencing direct inputs of sediment are good examples of such systems, where form isdefined by extreme events rather than events of intermediate magnitude and frequency. Theconcept of this so-called “temporal sensitivity” was later elaborated on by Thomas (2001) andBrunsden (2001), although since their studies little attempt has been made in fluvial geomor-122.2. Introductionphology to address this issue in practice. Bull (1991) applies the theory of Brunsden and Thornes(1979) to the impact of a hypothetical temporal succession of disturbance events on sedimentstorage, which can either increase or decrease the stream-bed elevation. The system processesthe disturbance over the relaxation time, which together with a potential reaction time con-stitutes the total response time. If there is no further disturbance, the system status remainsunchanged over a time of persistence. The concept from Bull (1991) is based on a system’strend towards a dynamic equilibrium between the forcing by, and the reworking of, distur-bances. Howard (1982) concluded that if episodic inputs occur with a frequency that matchesthe inverse of the relaxation time, the output of the system will remain in a constant equilib-rium with the average value of the forcing. Flume based insights about equilibrium conditionsand timescales of adjustment to changes in sediment supply rates are discussed in some stud-ies (e.g., Elgueta-Astaburuaga and Hassan, 2017; Pryor et al., 2011), but response times are onlyquantified in few cases (e.g., Podolak and Wilcock, 2013).The paragraphs above illustrate how the frequency at which events occur may be funda-mental in defining the response of a fluvial system to a change in boundary conditions. There-fore, it appears that event frequency should be a central aspect in investigations regarding theeffect of episodic sediment supply on streams. Consequently, our understanding of involvedprocesses remains incomplete. For example, it is uncertain whether the freshly delivered sedi-ment that buries and is transferred over the bed surface is simply removed by the subsequentfloods or whether there is some exchange between armoured and structured bed and the fineand mobile deposits. Furthermore, Hassan and Zimmermann (2012) asserted that it is impor-tant to study how quickly internal changes in grain size, channel morphology, and sedimentstorage occur when the stream shifts between cycles of aggradation and degradation.Our main research objective is to describe the impact of episodic sediment supply on chan-nel bed evolution in simulations using a 1-D morphodynamic numerical model for a bed ofmultiple grain sizes. We use the model to recreate conditions from experiments conducted atthe Mountain Channel Hydraulic Experimental Laboratory, University of British Columbia,where a set of experiments were carried out to examine the impacts of episodic sediment sup-ply on bed surface evolution and channel adjustment of a gravel bed stream (von Flotow, 2013;132.3. MethodsElgueta, 2014; Elgueta-Astaburuaga and Hassan, 2017; Elgueta-Astaburuaga, 2018). Although theseexperiments provide detailed information on channel adjustment to changes in the sedimentsupply regime, they are limited in terms of the number of experiments and the range of scenar-ios that could be conducted. The performance of the model is tested against the experimentalresults obtained in the laboratory and then used to further explore controls and responsesof the fluvial system to changes in flow and sediment supply regimes. The specific researchquestions addressed in this study are as follows:1. Can the numerical model recreate the channel response that was observed in flume ex-periments of similar scope?2. Does the sequencing of supply events play a role in the reaction of a gravel-bed stream,when several events of specified magnitudes occur in a different order?3. How will different combinations of episodic sediment supply, obtained by varying theirmagnitude and frequency, impact channel evolution of a gravel-bed stream?2.3 MethodsWe applied the 1-D morphodynamic model BESMo (Bedload Scenario Model) to calculate ca-pacity based sediment transport under different sediment supply regimes. We chose valuesfor model parameters to match the flume experiments as closely as possible and used mea-surements of sediment transport rate, surface grain size distribution, and slope to calibrate themodel. Matching our research questions, we then conducted two types of simulations:1. In “event sequencing simulations”, we simulated different permutations of events tounderstand the role event succession plays in long-term channel response.2. In “equilibrium simulations”, we used the same model setup and imposed different, butwithin each run regular, supply event frequencies. These simulations were run until weachieved a recurrent pattern in slope and grain size adjustment, allowing us to identifyhow the channel adjusts to the supply regime in the long term.142.3. Methodsfp pFigure 2.1: Flowchart stating the main components of the model and the flowof information between them. The temporal loop is advanced as the newelevation affects the slope in the flow model. The components are colouredas follows: blue – flow related, dark yellow – sediment volume related,peach – particle size related, and green – geometry related.2.3.1 Model setupThe structure of the model is similar to other models designed to reproduce and interpretdata from flume experiments (e.g., Cui and Parker, 2005; Wong and Parker, 2006; Ferrer-Boixand Hassan, 2014; An et al., 2017a). Figure 2.1 gives an overview of the implemented modelcomponents and their basic interaction. The model can be subdivided into a “hydraulic part”and a “sediment part”, both of which are subject to “external forcing” that varies in accordingto the modelling scenarios.We set up the modelling environment to run on a Compute Canada research cluster, which152.3. Methodsallows us to simulate many different input conditions in parallel and compare results quickly.We use a backwater flow model as suggested by Cui et al. (2006), implementing a thresholdFroude number (Fr) to switch conditions between supercritical and subcritical flow:Fr =√Q2wgw2r h3. (2.1)The Froude number (Fr) is calculated as a function of discharge (Qw), gravity (g), channelwidth (wr) and water depth (h) (Eq. 2.1). The threshold Fr = 0.9 simplifies the calculation offlow conditions, allowing us to spatially iterate through the nodes only once from downstreamto upstream. In the case of subcritical flow, the water depth is solved locally as a function ofdownstream friction slope (Sf) and bed slope (S0) (Eq. 2.2a):dhdx=S0 − Sf1− Fr2 for subcritical: Fr < 0.9 (2.2a)h =(n2Q2wSf)3/10for supercritical: Fr ≥ 0.9. (2.2b)Water depth under supercritical flow conditions is calculated locally assuming steady uniformflow using the Manning–Strickler formulation (Eqs. 2.2b and 2.3), where αr is a coefficient of8.1 (Parker, 1991) and roughness height (ks):n =k1/6sαrg1/2. (2.3)ks is calculated using the constants nk and Ds90, the surface grain size for which 90 % of thesurface is finer:ks = nkDs90. (2.4)In the case of steady and uniform flow the bed slope S0 is equal to the friction slope (Sf) withbed elevation (ηb) at downstream position x:Sf =−dηbdx= S0. (2.5)162.3. MethodsIf the solution of the water depth with Eq. (2.2a) is numerically unstable on the current nodedistribution, the model subdivides the channel into more nodes and reiterates the subdivisionuntil a stable backwater curve is found. This approach does not properly represent the locationof hydraulic jumps (Cui et al., 2006), which should not be a problem as we average conditionsover a node spacing of at least one channel width. Boundary shear stress (τb) is then calculatedwith the depth-slope product:τb = ρghSf, (2.6)where ρ is the water density. τb is then converted to the shear velocity u∗, which is used in thesediment routing component:u∗ =√τb/ρ. (2.7)The volumetric unit bedload transport rate per size class qbi is calculated using the sedimenttransport function provided by Wilcock and Crowe (2003). The change in bed elevation ∂ηb foreach node x per time step t follows from the Exner equation of mass conservation:(1− λ)∂ηb∂t= −∂qb∂x, (2.8)where λ is bed porosity. The volumetric bedload transport rate per unit width is given as qband is calculated per grain size class i in the sediment mixture of n size classes:qb =n∑i=1qbi. (2.9)The model incorporates subsurface stratigraphy using the active layer concept (Parker, 2008),which gives the Hirano equation:(1− λ)[∂∂t(LaFi)− f Ii ∂La∂t]= −(∂qbi∂x− f Ii ∂qb∂x), (2.10)where La is the active layer thickness, Fi is the surface frequency of the ith grain size class, f Iiis the ith grain size class proportion exchanged between the surface and the subsurface, andqbi = pbiqb is the volumetric unit bedload transport rate of the ith grain size class where pbi172.3. Methodsrepresents the ith fraction of the bedload transport rate. The active layer thickness is calculatedas La = naDs90, with the parameter na, representing the scale of bed fluctuations.The grain size distribution of the sediment flux between the active layer and the substrateis either calculated from the subsurface texture when the bed degrades, or from a linear com-bination of surface and bedload grain size distributions when the bed aggrades (Hoey andFerguson, 1994):f Ii =fi for∂ηb∂t < 0αFi + (1− α)pbi for ∂ηb∂t > 0,(2.11)where fi is the fraction of the ith grain size class in the subsurface, and α is a constant. Thevertical stratigraphy is stored in 10 cm high layers following Viparelli et al. (2010). By keepingtrack of grain size distributions within the surface and subsurface layers, the model can pre-serve the history of phases of erosion or aggradation. This allows emergent properties such asarmouring layers to occur. To study the combined effect that the active layer thickness factorna and the active layer exchange ratio α have on the model results, we executed sensitivityruns shown in the Supplement to this paper (Supplement Figs. S1–S4).The Exner equation (Eq. 2.8) in combination with the expression for the friction slope inEq. (2.2) and the sediment transport function by Wilcock and Crowe (2003) form an non-linearadvection–diffusion system that allows the calculation of bed elevation as a function of spaceand time (An et al., 2017a). An upwind scheme was used for the numerical discretization. Themodel needs an initial bed profile and an initial value of the surface grain size to be solvable.Sediment boundary conditions are given by the sediment feed rate and grain size distributionson the inlet, and a fixed bed elevation at the outlet of the simulated reach. The flow boundarycondition is a water surface height of 0.1 m over normal flow at the outlet. The bedload trans-port function is used to calculate transport rates for each channel cross section. During themodel run, changes of the sediment transport rate are dependent on changes in the sedimentsupply, channel slope, and surface grain size distribution.182.3. Methods2.3.2 Model calibrationWe used data from flume experiments to calibrate the model. The objective of the flume ex-periments was to measure the adjustment of an alluvial steep channel to different frequenciesof sediment supply. The experiments were carried out in a water recirculating flume whichis 18 m long, 1 m wide, and 1 m deep in the Mountain Channel Hydraulic Experimental Lab-oratory at the University of British Columbia. Here we will provide a brief summary of theflume setup and experimental design, for more details see von Flotow (2013), Elgueta (2014),Ferrer-Boix and Hassan (2015), and Elgueta-Astaburuaga and Hassan (2017).The experiment consisted of seven 40 h long runs with different sediment supply frequen-cies, while keeping the total sediment input the same at 300 kg per run. The experiments wererun continuously, i.e., the bed surface at the end of run 1 was the starting condition for run2 and so on. For all runs, flow was held constant so that the sediment feed regime could bestudied with no changes in flow regime (Table 2.1). The difference between the runs was thespreading of the supply over a changing input frequency, which was either constant, in onepulse, in two pulses, or in four pulses. The bed was fixed in the first 1 m downstream sectionof the flume head box with stones equivalent to about Ds84 of the experimental bed material.In the remainder of the flume the bed initially consisted of 0.1 m of loose material with particlesizes ranging from 0.5 to 64 mm with a Ds50 of 5.64 mm, matching downscaled (by a factor of 3)conditions of a study reach in East Creek, British Columbia, Canada. The flume slope was setto 0.022 m m−1. Measurements include water depth, water surface slope, water velocity, bedsurface slope, bed surface particle size distribution, bed elevation, sediment transport rate,and bedload texture. Measurements of the water surface elevation were conducted through-out the experiment using a mechanical point gauge with 0.001 m precision. Photos were usedto manually sample bed surface grain size distributions and the bed elevation was recordedwith a green laser scanner at a 2 mm resolution. The bed surface scans were used to measurethe bed surface slope along the thalweg, i.e., the line of lowest elevation along the flume. Flowvelocity measurements were conducted using an ADV profiler. The grain size and count ofparticles exiting the flume were recorded with a camera and a light table at the outlet of flume(Zimmermann et al., 2008). The transport rate measurements were done at 30 Hz and validated192.3. Methodsafter the experiments by total exported weight.Using the model described, we simulated a 12 m long and 1 m wide channel in 13 down-stream nodes each spaced 1 m apart. The model was set to calculate sediment transport forall nodes in time steps of 10 s. All simulations used a constant water discharge of 0.065 m3 s−1and a geometric mean grain size of 5.64 mm for both initial bed and sediment feed (full distri-bution shown in Fig. 2.2a). We chose to use a normally distributed approximation of the flumegrain size with a width of σ = 1.6 to be consistent with distributions used for the equilibriumsimulations. This distribution and the original flume GSD are statistically the same. The initialchannel slope was 0.022 m m−1, also matching the parameters from the flume experiments.We calibrated the model by visually reducing the difference between measured and simu-lated values of bed slope (S), surface grain size parameters (Dsg, Ds90), and transport rate (qb).In the calibration runs more importance was given to recreating S, Dsg, and qb than to a goodmatch in Ds90. We first increased the reference Shields stress (τ∗rm) in the Wilcock and Croweformula to roughly match simulated and measured qb. Afterwards, we varied the grain ex-change ratio (α) (Eq. 2.11) and the coefficient of the active layer thickness (na). As we achieveda good visual match between simulation and flume measurements, we did not see the need tocalibrate more parameters.202.3.MethodsTable 2.1: Overview of runs in the flume experiments. All runs were 40 h long. The texture of the initial bed mixtureand the sediment feed were identical. Plots of resulting slope, Dsg, Ds90, and sediment transport are shown inFig. 2.3. The symbols in column two signify the feed regime where 0 represents no feed, C represents constantfeed, and 1, 2, and 4 represent the number of pulses in 40 h.Flume Symbol Feed Mean water depth Bed slope after Water surface slope Feed rate Feed Pulse Pulserun regime after 40 h (m) 40 h (m m−1) at 40 h (m m−1) (g ms−1) duration (min) magnitude (kg) period (h)R1 0 None NA 0.017 NA 0 – – –R2 C Constant 0.073 0.016 0.017 2.0833 – – Const.R3 1 One pulse 0.080 0.018 0.019 83.3 60 300 40 hR4 4 Four pulses 0.083 0.020 0.020 83.3 30 75 10 hR5 2 Two pulses 0.072 0.022 0.020 83.3 15 150 20 hR6 C Constant 0.075 0.022 0.020 2.0833 – – Const.R7 0 None 0.073 0.022 0.020 0 – – –Initial slope 2.2 %; mixture Ds50 = 5.64 mm; mixture Ds90 = 11.2 mm, Qw = 65 L s−1; duration of each run = 40 h; totalfeed= 300 kg per run. NA represents not available.212.3. MethodsTable 2.2: Sequencing of events in runs that were simulated as permutations ofthe original flume experiment. Each of the seven periods was 40 h long andeach run lasted 280 h in total. There was no sediment input during the nofeed runs (0). Within all other period types, 300 kg of sediment was fed over40 h, either constantly (C) or in pulses (one, two, or four events). Besidesrecreating the original flume sequence (OF), we simulated two runs wherethe pulsed events occur either in order from many pulses to few (MtF) orfrom few to many (FtM). To explore if the system could rebound from theimpact of a certain pulse phase during a constant-feed phase, we createdtwo additional runs where this was the case (cMtF and cFtM), which led toa 600 kg higher total sediment feed.Simulation run Event sequence symbol Mass fed total (kg)OF Original flume 0 C 1 4 2 C 0 1500MtF Many to few 0 C 4 2 1 C 0 1500FtM Few to many 0 C 1 2 4 C 0 1500cMtF C-buffered: many to few C 4 C 2 C 1 C 2100cFtM C-buffered: few to many C 1 C 2 C 4 C 21002.3.3 Event sequencing simulationsWe explored the role that the sequencing of the pulse events could have on the flume studyby simulating the “original flume” event sequence and comparing the result to alternativesequencing of events (see Table 2.2). The alternative event sequences are using the same pulsedistributions (four pulses, two pulses, or one pulse over 40 h), but the pulse order is eitherfrom “few to many” (FtM) (i.e., one pulse, then two pulses, then four pulses) or from “manyto few” (MtF) (i.e., four pulses, then two pulses, then one pulse) per 40 h phase. To allowthe system more time to recover from pulse events, we simulated two more cases where eachpulsed phase is buffered from the next one by a 40 h constant-feed phase. These runs are called“c-buffered: many to few” (cMtF) and “c-buffered: few to many” (cFtM).2.3.4 Equilibrium simulationsIn our final set of experiments, we kept the frequency and magnitude of pulse events constantto achieve equilibrium slope and grain size conditions. The use of numerical modelling allowsfor the comparison of many simulations with differing grain size distributions, pulse frequen-cies, and pulse magnitudes. We expect a channel under episodic sediment supply to adjust222.3. Methodssynchronously to the frequency of external forcing events. The added sediment volume froma supply event will increase the channel slope at first. After the supply of sediment ends andmaterial is removed from the channel, the slope will decrease, and the surface grain size willbegin to reflect the sediment starved conditions. As the long-term sediment input equals thelong-term sediment output, the channel will eventually achieve a condition where the capac-ity to erode material (through increased slope in conjunction with changes in the surface grainsize) equals the depositional forcing (i.e., long-term sediment input) of the supply regime. Inthis state the adjustment of channel slope and grain size to each sediment input event will re-turn to the same values after every pulse. All runs achieved this equilibrium condition within20 000 simulation hours.00.20.40.60.81Fraction finerσ = 0.05σ = 0.1σ = 0.2σ = 0.4σ = 0.7σ = 1σ = 1.3σ = 1.6σ = 1.8Flumeexperiments10010010110210310110–1 102Pulse period Tpp (h)Grainsize (mm)(b)(a)100 10110–1 102Pulsemagnitude Mpulse (kg)Numerical simulationFlume experimentsFigure 2.2: (a) Variation of the grain size distribution between model runs (lines)and values from the flume experiments (circles). All distributions have ageometric mean grain size (Dsg) of 5.64 mm. The different σ values rep-resent the width of the distribution, calculated for normal distributions inphi-scaled sediment sizes. The data for the flume experiments are roughlymatched with σ = 1.6. (b) Combinations of pulse magnitude and pulseperiod (or recurrence interval) used in the equilibrium model runs. As thetotal mass of supplied material is the same for all simulations, runs withhigh pulse frequencies (i.e., low pulse periods Tpp) are smaller in magni-tude. Four of the combinations match the flume runs (circles).To find the equilibrium slope resulting from different sediment supply regimes, we simu-lated different combinations of sediment supply frequency (Fpulse) and magnitude (Mpulse) fornine different grain size distributions (Fig. 2.2a). All distributions have the same mean grain232.4. Resultssize of 5.64 mm, but differ in the width of the distribution by the standard deviation (σ), whichwas chosen for a phi-scaled, normally distributed sample. While σ = 0.05 represents a nearlyuniform sediment mixture, σ = 1.6 roughly matches the grain size distribution of the flumeexperiments. We used 11 grain size classes in the simulations and set the initial GSD to the feedGSD. Under the hydraulic conditions applied all grain sizes are initially mobile, which mightchange during the simulations due to the effect of armouring and changes in bed slope. Fig-ure 2.2b shows the combinations of frequency and magnitude used in this study. Each modelrun delivered the same input mass over the simulation time (150 000 kg over 20 000 h). Wethen distributed this total mass over different pulse frequencies, with four of the combinationsmatching the flume experiments. Each pulse was 10 min in length. The lowest frequency waschosen to be one pulse every 400 h (pulse period time: Tpp = 400 h), and the highest frequencywas constant feed (one 10 min long pulse every 10 min). We selected a range of 40 pulse fre-quencies for which the whole number of cycles summed to 20 000 h. In total we executed 360simulations, 9 different σ with 40 frequencies each.2.4 Results2.4.1 Model calibrationElgueta-Astaburuaga and Hassan (2017) describe the flume results we used for model calibrationin detail. We will give a short summary of the findings here. An initial run without sedimentfeed over an unstructured bed showed a high sediment output while the bed armoured. Thisrun was similar in output grain sizes and grain mobility to the run with one sediment pulseover a structured bed. This shows that active restructuring of the bed occurred in both ofthese runs. On the other end of the spectrum, the constant feed and four pulse runs weresimilar in their low sediment output and showed different grain mobility. This implies that thesystem reacts differently at a threshold frequency somewhere between two pulses per 40 h andfour pulses per 40 h, which Elgueta-Astaburuaga and Hassan (2017) interpret as the relaxationtime of the bed to a pulse event.We used these experiments to test the ability of the numerical model to recreate the flume242.4. Resultsresults in discharge, pulse frequency, pulse magnitude, slope, and grain sizes. In the exper-iment, mean grain size (Dsg) and slope (S) were measured over a central 2 m long section toavoid a bias of bed surface measurements due to inflow and outflow conditions. The mea-surement intervals varied between every 1 and every 20 h (depending on the pulse interval).The values for the numerical simulation are averaged over the whole reach and were recordedevery 10 min during the simulation time. We achieved a best match in slope S, Dsg, Ds90,and the transport rate qb by increasing the reference Shields stress τ∗rm in the Wilcock andCrowe formula by a factor of 2. Wilcock (2001) suggests taking the same approach of increas-ing the threshold shear stress to match a sediment transport calculation to field data. Otherresearchers use this method to calibrate models to field and flume data (Gary Parker, personalcommunication, 2018; Chartrand et al., 2015).Figure 2.3 shows the comparison between flume measurements and the calibrated modelresults with the grain exchange ratio α = 0.45 (Eq. 2.11) and the active layer thickness factorna = 2. The sensitivity of the model to changes in α and na is shown in Supplement Figs. S1–S4. Due to the long interval between measurements in the flume experiments, some short-termslope responses to individual sediment pulses might be hidden (e.g., after hours 80 and 180 inFig. 2.3a). The model underpredicts both the slope and the mean surface grain size (Dsg) (seeFig. 2.3b) in the first 60 h, while the coarse grain size fractions (Ds90) and average transportrate are over predicted for the first 30 h (see Fig. 2.3c and d). This might be due to imperfectinitial conditions or boundary effects in the flume experiments. For the rest of the simulationboth slope and Dsg show good agreement with the flume results. D90 is overpredicted in themodel, but this is seen as a minor issue because the simulated transport rates mainly dependon Dsg and the slope. The transport rate in the simulation lags behind the light table data,which might be due to our numerical implementation of diffusion. As the model matches theaverage transport rates well, we did not see the need to improve the temporal agreement.2.4.2 Event sequencing simulationsAfter obtaining a good match between the model and the flume data, we simulated alternativeevent sequences as described in Table 2.2. Figure 2.4a shows the adjustment of slope in runs252.4. Results1.61.41.822.22.42.6·10−2Slope(m m )D sg (mm)D s90 (mm)Transport rate (     )(a)Numerical simulationFlume experiments681012141618(b)24262830323436(c)0 40 80 120 160 200 240 28010210–2100No feed Constant feed One pulse Four pulses Two pulses Constant feed No feedTime (h)g ms(d)-1Figure 2.3: Comparison of (a) slope, (b) mean surface Dsg, (c) surface Ds90, and(d) sediment transport rate between the numerical simulation and the flumeexperiments.262.4. Resultsthat preserved the same sediment feed volume and had the same duration as the flume exper-iments (OF, MtF, and FtM), but the frequency of events is ordered differently. At the end ofthe simulations, all runs approach the same slope value of 0.022 m m−1, which shows that themain factor determining the long-term slope is the total volume of sediment fed. The sequenc-ing of events seems to play a role in the slope adjustment over the short term, here about80 h after the events. On this short timescale, large pulses increase the slope quickly, whilethe smaller, more frequent pulses lead to a more gradual adjustment of slope. Figure 2.4bshows the runs where pulse phases were buffered by constant-feed phases (cMtF and cFtM),increasing the total sediment feed by 600 kg. This did not change the pattern of adjustment sig-nificantly compared to the earlier runs, as the constant-feed phases only prolonged the effectof the previous pulse phase.The effect of event sequencing on the channel response in Dsg is shown in Fig. 2.4c forall runs in Table 2.2. The order of events only has a weak impact on patterns of adjustmentin Dsg, as maximum grain size conditions are reached within 20–30 h. Afterwards no furtheradjustment occurs until the introduction of fine material with the next pulse lowers the surfacegrain size again. This means that armouring of the channel surface happens quickly in relationto the time between pulse events. The subsurface is made of the same grain size distributionas the sediment feed, so its mean size is 5.56 mm. Therefore, we can infer that an armouringratio (Dsg/Dsubg) of about 2.2 was reached within 20 h and then increased towards 2.7 over thefollowing 260 h. A finer, less armoured surface at the time of each supply event is followedby a coarsening of the surface as the finer grain sizes are more mobile, and thus more easilyevacuated in the time without feed between pulses.2.4.3 Equilibrium simulationsOur second set of simulations explored the equilibrium conditions that are reached under dif-ferent supply regimes. As we will explain in the discussion, the effect of the supply regime canbe constrained to changes in slope (S) and the surface grain size distribution (GSDfluv), whichin the following will be characterized by the armouring ratio between surface and subsurfacemean grain size (Dsg/Dsubg). After different times in the simulations, these parameters reach272.4. Results22.41.62.8·10−2Slope(m m   )D sg (mm)(b)1.622.42.8·10−2(a)0 40 80 120 160 200 240 2806810121416Time (h)(c)FtM - few to manyMtF - many to fewcFtM - buered few to manycMtF - buered many to fewOF - original flume-1Slope(m m   )-1Figure 2.4: Comparison of slope and mean surface grain size Dsg from runs withdifferent event sequencing (see Table 2.2) (σ = 1.6, factor of 2 increase ofτ∗rm). Event sequences are (a) rearranged pulse phases to the flume experi-ments and (b) a setup where the pulsed phases are buffered with constant-feed phases. Panel (c) shows Dsg for the runs in both (a) and (b).a time-independent periodic adjustment that is illustrated in Fig. 2.5 for the last 400 h of twosimulations. Figure 2.6a shows box plots of the distribution of slope values during the lastpulse of all 40 runs with σ = 1.6. The presented normalized slopes (S/Sconst) indicate howthe slope for each pulse frequency compares to the constant-feed slope of runs with the samegrain size distribution. Figure 2.6b shows the change in normalized mean slope (Smlp/Sconst)with pulse frequency, which corresponds to the red lines in Fig. 2.6a and is our main indicatorfor the equilibrium state of the channel slope. Each line represents a different width σ of theGSD over 40 runs with increasing Tpp.282.5. Discussion·10−2SlopeD/Dsg subg3.23.43.63.844.2(a)19600 19700 19800 19900 200002.833.23.43.6Time (h)(b)Pulse period = 20 hMeanPulse period = 40 hMeanFigure 2.5: (a) Slope and (b) armouring ratio for the last 400 h of 2 out of 40 exper-iments using σ = 1.6. A run with a low event frequency is shown in blue,and a run with a high event frequency is shown in red. In these exampleruns the mean slope in equilibrium is higher for the run with longer pulseperiods.2.5 Discussion2.5.1 Extension of flume results with the numerical modelThe numerical model shows good agreement with the temporal response of mean surfacegrain size and slope from the flume experiments. As the initial bed grain size distributionswere well mixed, the good match in mean surface grain size also implies a good match in thearmouring ratios. A series of runs with an alternative sequencing of events showed that, whilethe adjustment of mean surface grain size was not sensitive to the order of the pulsed phases,the evolution of slope differed considerably. In the cases where the first pulsed phase consistedof one large magnitude event (FtM and cFtM), the slope increased quickly and the followinghigher-frequency, lower magnitude pulse phases did not modify the system considerably. Incontrast, cases where multiple smaller event phases occurred first (MtF and cMtF), the slopeincreased more gradually. All runs ended at about the same slope after the 280 h simulation292.5. DiscussionS mlp /Sconst103 104 105 106(s)Tpp (h)S/S const(b) (a)0.33 1.7 4 13.3 26.7 50 80 125 200 3120.811.21.41.61.822.22.40.70.80.911.11.21.3σ = 0.05σ = 0.1σ = 0.2σ = 0.4σ = 0.7σ = 1σ = 1.3σ = 1.6σ = 1.8Tpp26.7 h4 h 50 h13.3 h 80 h125 h 200 hFigure 2.6: (a) Distribution of the ratios of slope during the last pulse to theconstant-feed slope for all 40 runs with σ = 1.6. We normalized the val-ues of the slope in the last pulse with the slope of the constant-feed run ofthe same grain size distribution width (σ), allowing for the comparison ofthe equilibrium slopes between runs with different σ. The red lines repre-sent the normalized mean slopes (Smlp/Sconst), which we chose as the mainindicator for the equilibrium state of the channel slope. Note that data fromlonger pulse periods Tpp will contain more data points for the box plots, asthere are more slope values recorded during the longer time between pulses(sampling every 10 min). The red crosses are outliers in the distribution ofslope values, illustrating the extreme slope values during the time right af-ter the pulse was introduced into the channel. (b) Mean slope ratios for allruns grouped by width of GSD (σ).time, which implies that while the low frequency, large magnitude events strongly alter thechannel in the short term, the sequencing of events does not play an important role in thelong run. The constant-feed-buffered runs (cFtM and cMtF) show similar behaviour to theirunbuffered counterparts, as the constant-feed phases preserve the bed state of the previouspulse phase. The main driver of the slope adjustment is the total sediment feed, which isconsistent with findings by Blom et al. (2017).2.5.2 Development of timescales from the equilibrium simulationsThe definition of a sediment supply regime can be based on different aspects of sediment inputinto a stream, either from outside or within the channel. For simplicity, we restrict the defini-tion of a sediment supply regime to the input of material into the fluvial system from outsidethe active channel. Sediment supply from storage close to the channel can be viewed as exter-302.5. Discussionnal supply if it only occurs episodically (e.g., less than yearly) in large flooding events. Thisview allows us to describe the sediment supply regime by a combination of frequency, mag-nitude, and grain size distribution of sediment supply events over a multi-event time frame(as in Benda and Dunne, 1997a). The time frame must be long enough to contain enough sed-iment supply events to allow the stream bed to adjust to the external forcing by changing itsinternal configuration of sediment storage and bed structuring. Even though a natural chan-nel might never reach an equilibrium to a certain sediment supply regime, it might produceregular patterns of transient adjustment to the supply events.In our case, the forcing on the system is a combination of pulse frequency (Fpulse), pulsemagnitude (Mpulse), pulse grain size distribution (GSDpulse), and water discharge (Qw). Due tothe simple geometry and the lack of bedforms in a 1-D numerical model, the fluvial reaction tothe forcing is restricted to the bedload transport rate (qb), channel slope (S), and channel grainsize distribution (GSDfluv) (see Table 2.3). Pulse frequency and magnitude can be combinedin the virtual pulse velocity (Upulse), with the magnitude normalized by reach width (wr) andlength (lr):Upulse = Fpulse · Mpulse/(wr · lr). (2.12)A characteristic pulse period time (Tpp) for a sediment supply regime is the inverse of thepulse frequency:Tpp = 1/Fpulse. (2.13)We define a reach averaged fluvial export velocity Ufluv = qb/lr in order to compare the sed-iment export of material to the pulsed sediment input. Considering a multi-event time frameTsim( Tpp) for each simulation, we can assume that the fluvial system will adjust to theexternal forcing over time, leading to an adjustment of the reach averaged fluvial sedimenttransport to match the external forcing of the virtual pulse velocity:Tsim  Tpp : Ufluv ≈ Upulse. (2.14)The time it takes to achieve an equilibrium state is highly dependent on the initial condi-312.5. Discussiontions of the simulations. Instead of using a process rate threshold to find an equilibrium time,we can infer a simulation-time independent relation between the supply regime (i.e., Tpp) andthe state of the system in equilibrium (i.e., Ufluv, S, and GSDfluv). This way we only have to runsimulations for long enough to verify equilibrium with the respective supply regime (in ourcase 20 000 h), and then observe properties of the channel at the very end of the simulation,even though equilibrium might have been achieved earlier.When comparing the system state in equilibrium between many different supply regimes,while keeping the initial fluvial reworking capacity constant (mainly geometry and Qw), wecan identify how the equilibrium conditions change under different supply regimes. Dueto the fixed channel geometry in the 1-D model, only the bed slope (S) and the grain sizedistribution (GSDfluv) can adjust to the change in the supply regime.To better compare the temporal adjustment to sediment pulses of different magnitudes andfrequencies, we non-dimensionalized the pulse period Tpp with a fluvial evacuation time Tfe:Tfe =lrwrDfgUfluv(DfgDag)2. (2.15)This timescale is a representation of how long it would take to remove a layer of sediment aslong and as wide as the flume (lr and wr) with the thickness of the median feed grain size Dfg,under the average transport rate Ufluv, multiplied by an estimate of the ratio of feed grain sizeto armoured grain size (Dfg/Dag)2, which can be interpreted as an inverse of the degree ofpotential bed armouring similar to Ds90/Ds50 (Recking, 2012). We developed Eq. (2.15) visuallyby matching the inflection of lines in Fig. 2.7a to Tpp/Tfe = 1.Figure 2.7a shows the same data as Fig. 2.6b, but in non-dimensionalized time Tpp/Tfe. Aratio of Tpp/Tfe < 1 can be interpreted as a condition in which the pulsed input of materialoccurs faster than the fluvial removal of a Dfg thick theoretical layer of material under averagetransport conditions modified by armouring. At a ratio of Tpp/Tfe > 1, the sediment is fed intime steps longer than the time that the fluvial system needs to remove said theoretical layer.Figure 2.7b shows the normalized armouring ratio (ARelp/ARconst), which was obtainedby dividing the armouring ratio ARelp = Dsg/Dsubg at the end of the last pulse for each simu-322.5. DiscussionARelp /Ar constTpp / TarS mlp /Sconst10–110–2 100 101 102Tpp / T10–110–2 100 101 102ARelp ≈ ARcons t ARelp > ARcons tSmlp ≈ Scons t Smlp > Scons tSmlp < Scons tσ = 0.05σ = 0.1σ = 0.2σ = 0.4σ = 0.7σ = 1σ = 1.3σ = 1.6σ = 1.8σ = 0.05σ = 0.1σ = 0.2σ = 0.4σ = 0.7σ = 1σ = 1.3σ = 1.6σ = 1.80.70.80.911.11.21.30.9511.051.11.151.21.25fe(b)(a)Figure 2.7: (a) Mean slope ratios in non-dimensional timescale. Tpp/Tfe < 1:model runs with a high-frequency sediment supply show similar equilib-rium slopes and armouring ratios as conditions of constant sediment feed(Smlp ≈ Sconst). Tpp/Tfe > 1: if the GSD is narrow (σ < 0.4) or the pulse pe-riod is not much longer than the fluvial evacuation time, we observe lowerequilibrium slopes than in constant-feed runs (Smlp < Sconst). Runs witheither very low frequency of supply events or wide GSD (σ ≥ 0.4) showequilibrium slopes that are higher than the respective constant-feed runs(Smlp > Sconst). (b) Relative armouring ratio in non-dimensional timescale.Tpp/Tar > 1: low frequency of supply events leads to an increase in armour-ing ratio compared to constant-feed runs (ARelp > ARconst), especially forwide GSDs (σ ≥ 0.4).lation with the armouring ratio at the end of the corresponding constant-feed run. Similarly tothe fluvial evacuation time that we used to non-dimensionalize the slope adjustment, we usedan armouring time Tar to non-dimensionalize the grain size adjustment:Tar =lrwrD f 90Ufluv0.5(D f 90Da90)2. (2.16)This timescale represents how long it would take to remove a layer of sediment as long and aswide as the flume (lr and wr) with the thickness of the supplied D f 90, under the reach averagedtransport rate Ufluv = qb/lr, multiplied by an estimate of the ratio of the sediment supply Df90to the Da90 of an armoured bed. We developed Eq. (2.16) visually by matching the inflectionof lines in Fig. (2.7b) to Tpp/Tar = 1.332.5. Discussion2.5.3 Interpretation of the equilibrium simulationsThe condition of Tpp/Tfe = 1 constitutes a threshold in the slope adjustment to pulsed sed-iment supply. Pulse periods shorter than the fluvial evacuation time (Tpp/Tfe < 1) lead toequilibrium slopes similar to the constant-feed equilibrium slopes (Smlp ≈ Sconst). In the caseof narrow GSDs (σ < 0.4), simulations with pulse periods longer than the fluvial evacuationtime (Tpp/Tfe > 1) show an up to 20 % lower slope than in the constant-feed equivalent run(Smlp < Sconst). In contrast, runs with either very low frequency of supply events or wide GSD(σ ≥ 0.4) show equilibrium slopes that are up to 30 % higher than the respective constant-feedruns (Smlp > Sconst). Simulations with a wider range of material (σ > 1) show a drop inSmlp/Sconst right at Tpp/Tfe = 1, but then an increase in Smlp/Sconst at longer pulse distances.We interpret conditions of lower Smlp to be less armoured, as they coincide with lower valuesof ARelp/ARconst as shown in Fig. 2.7b. The simulations showing an intermittent decrease ofSmlp seem to be in a state in which material can be efficiently exported from the system with-out having intense armouring limiting the slope adjustment. These large pulses increase theslope rapidly, leading to high shear velocities which causes high transport rates.We interpret the cause for increasing slope and armouring ratios for long pulse periods inthe following way. A longer time between pulses (Tpp > Tar) causes more intense armour-ing, which shows that the channel bed is starved of sediment between pulses, as finer grainfractions are removed from the surface. This leads to the development of an armouring layer,which restricts the incision into lower deposits, initially limiting the sediment output of thesystem (Ufluv < Upulse). This imbalance between input and output of material leads to an in-creased sediment storage over time, which due to the fixed geometry of the channel leads toan increase in slope. This can increase the shear velocity and in return leads to higher sedi-ment output rates. This response loop between armouring and slope adjustment will continueuntil the sediment output matches the long-term sediment supply (Ufluv ≈ Upulse). Note thatdue to the restricted geometry of our model, slope and grain size are the main parametersin the channel that can change in response to the sediment supply regime. It is possible thatother morphological adjustments (e.g., channel width) could compensate for the transport ratedisequilibrium in a similar fashion.342.5. DiscussionThe non-dimensionalized presentation of the simulation results shows two distinct modesof adjustment of the fluvial system to episodic sediment supply regimes: (1) “constant-feed-like” behaviour in runs where supply events are of high frequency and low magnitude. Underthis kind of forcing the equilibrium slopes and armouring ratios are similar to equilibriumconditions of runs with constant sediment feed. (2) “Pulse-dominated” behaviour occurred inruns where sediment was fed in low frequency and high magnitude events.An interesting finding is that when Tpp > Tfe, an increase in slope only occurs in simu-lations that have grain size distributions wide enough to allow armouring to occur. In thesecases, we could use the timescale of Tar to determine if the armouring would be significantenough to prevent a decrease of the equilibrium slope. Even though armouring develops veryquickly in both our simulations and the flume experiments, its long-term persistence in thetime between sediment pulses is what governs the channel response. Hence, the grain sizedistribution in a series of supply events can be more important for the channel response in thelong term than the frequencies and magnitudes of the individual events themselves.It is notable that the channel response at Tpp = Tfe does not change abruptly, but thesystem response slowly tilts to either pulse-dominated on the one end, or constant-feed-likeon the other end of the spectrum. While all constant-feed-like channels (for each σ) have verysimilar equilibrium properties, all pulse-dominated channels are different in both slope andarmouring ratios.2.5.4 Implications of the equilibrium simulationsApplying the threshold of Tpp/Tfe = 1 between constant-feed-like and pulse-dominated sup-ply regimes to the flume experiments is inconclusive. The unity of Tpp = Tfe ≈ 6 h lies be-tween the constant-feed runs and the highest frequency runs (four pulses with Tpp = 10 h),which means that we have no experimental constant-feed-like pulsed regime where Tpp < Tfe.Elgueta-Astaburuaga and Hassan (2017) found that the four-pulse phase caused a sedimenttransport response that was similar to the constant-feed runs, implying that Tfe would liebetween 10 h and 20 h. The flume experiments were not executed long enough to reach equi-librium, which complicates the attribution of a specific channel response to a specific forcing.352.5. DiscussionBesides recreating the 280 h flume experiment in the model, the 360 equilibrium simulationsinclude four configurations that repeat the constant-feed and three pulse periods for 20 000 h.These four configurations only reached the equilibrium after about 12 000 h of simulated time,which is very long in comparison to the conditions of the flume experiments where each sup-ply regime only lasted 40 h.362.5.DiscussionTable 2.3: List of forcing and reacting parameters, and timescales in our simulations with abbreviations and their di-mension (T represents time and L represents length).Forcing parameters Reacting parameters TimescalesPulse frequency Fpulse (1 T−1) Fluvial export velocity Ufluv (L T−1) Simulation time Tsim (T)Pulse magnitude Mpulse (L3) Channel slope S (L L−1) Pulse period Tpp (T)Pulse grain size distr. GSDpulse (L) Surface grain size distr. GSDfluv (L)Water discharge Qw (L3 T−1) Derived from simulations:Derived parameters: Fluvial evacuation time Tfe (T)Virtual pulse velocity Upulse (L T−1) Armouring time Tar (T)372.5.DiscussionTable 2.4: Application of the fluvial evacuation time to the rapids reach in East Creek. The system is assumed to be inequilibrium with a matching fluvial transport rate and long-term sediment supply rate. The time of active fluvialtransport is estimated to be 100 h yr−1. We approximated the long-term fluvial transport rate Ufluv with three yearsof data from a sediment trap below the reach. We assumed that the subsurface grain size measurements reflect theaverage supply GSD and the surface grain size measurements represent the long-term average state of armouringin the reach. Besides the original “East Creek” data and the two “threshold” pulse frequency fits, we assumed fourmore scenarios with doubled armoured grain sizes or doubled fluvial transport rates to give a rough estimate oferror bounds.Fluvial parameters Supply regime TimescalesUfluv(= Upulse) Da50 Da90 Mpulse Fpulse Tfe Tar Tpp/Tfe Tpp/Tar(m3 yr−1) (mm) (mm) (m3) (1 yr−1) (yrs) (yrs)East Creek 0.75 57 150 0.75 1 2.23 3.58 0.45 0.28Threshold Tfe 0.75 57 150 1.67 0.45a 2.23 3.58 1 0.62– with 2×Ufluv 1.5 57 150 1.67 0.90a 1.11 1.79 1 0.62– with 2× Da50 and Da90 0.75 114 300 0.42 1.80a 0.56 0.89 1 0.62Threshold Tar 0.75 57 150 2.68 0.28b 2.23 3.58 1.61 1– with 2×Ufluv 1.5 57 150 2.68 0.56b 1.11 1.79 1.61 1– with 2× Da50 and Da90 0.75 114 300 0.67 1.12b 0.56 0.89 1.61 1For all calculations: supply Df90 = 90 mm, supply Df50 = 32 mm, channel width = 2.5 m, channel length = 72 m.Some frequencies calculated to match. a Tpp = Tfe and b Tpp = Tar382.5. DiscussionThe numerical model was calibrated with only one set of flume experiments. As our studymainly focusses on comparing different simulation results from the same model, the appli-cability of our results to other flume studies or field cases is uncertain. However, we areconfident that the numerical model is an adequate tool to gain insight on the effect of episodicsediment supply on fluvial channels in a general sense. For example if there was an inaccuracyin the calculation of the shear velocity, it would affect all model runs and thus be counterbal-anced by relating the changed resulting slope to the constant-feed slope (Smlp/Sconst), and thechanged armouring ratio to the constant-feed armour ratio (ARelp/ARconst).We tested this by executing two additional batches of simulations with a 25 % decrease anda 25 % increase in the total mass fed respectively. If the experimental design would stronglyaffect the threshold between constant-feed-like and pulse-dominated conditions, we wouldexpect these simulations to plot differently than the data in Fig. 2.7b. But as shown in Sup-plement Fig. S5 for the case of σ = 1.6, the change in the slope ratio Smlp/Sconst is relativelyinsensitive to the total feed volume. We also expect this to be the case when changing thegrain size distributions to include large particles that are initially immobile with the applieddischarge. In such simulations, the slope would increase to a point where these initially im-mobile grain sizes become mobile at a very high equilibrium slope. As we compare all resultsto the corresponding constant-feed equilibrium slope of the same grain size distribution, theseconditions might collapse on the existing data as well. However, it is possible that the channelparameters would become extreme in a way that the empirically derived transport functionby Wilcock and Crowe would no longer be realistically applicable.As we developed the fluvial evacuation time Tfe and the armouring timescale Tar purelyfrom observations in numerical simulations, their usefulness remains to be proven in the field.If such a threshold behaviour between episodic sediment supply event frequency and the flu-vial adjustment of a channel exists, it should be possible to find signatures in channel morphol-ogy, sediment storage volume, or channel slope when comparing streams subject to differentpulse periods and with different fluvial transport capacities. It is possible to use Eq. (2.15)with information about the long-term sediment supply volume, grain size supply, and aver-age channel dimensions to calculate Tfe and thus infer the matching threshold pulse period392.6. Summary and conclusionswhere Tpp = Tfe. If the long-term sediment supply occurs in more frequent events than thisthreshold, the system can be assumed to experience constant-feed-like sediment supply. Ifthe supply frequency is lower, we would expect to find a morphological signature of a pulse-dominated supply regime.To provide an example application of the developed timescales, we applied Eqs. (2.15) and(2.16) to data from East Creek, which is a small creek in the Fraser watershed close to Vancou-ver in BC, Canada. The “rapids” channel section of this creek was used to design the flumeexperiments as a 1 to 6 Froude scaled model. Table 2.4 shows the resulting timescales in sevendifferent scenarios that use the reported values for sediment supply, grain size distribution,and channel dimensions from Cienciala and Hassan (2013) and Papangelakis and Hassan (2016).In East Creek, we assume that of all the supplied sediment is contributed by annual eventswith a magnitude that matches the annual fluvial transport. As the calculated fluvial evacu-ation time Tfe is 2.23 years, this scenario implies that the system would behave in a constant-feed-like manner, which would only change if the supply events were more than 2.23 yearsapart on average (i.e., Tpp = 2.23 years), as shown in the “threshold Tfe” calculation. Due tothe high uncertainty in our assumption that the measured values represent equilibrium con-ditions, we calculated two more scenarios with double transport rates (2×Ufluv) and doublearmoured grain size (2× Da50 and Da90). The last three calculations in Table 2.4 show whichpulse frequency is needed to match the armouring timescale (Tar). While we do not know ifEast Creek is in equilibrium with the sediment supply regime and the measurements used donot reflect long-term conditions, these calculations can still give a rough idea of whether a sys-tem is constant-feed-like in our classification of channel response to episodic supply regimes.2.6 Summary and conclusionsWe characterized an episodic sediment supply regime in terms of event frequency, magni-tude, and supplied grain size distribution. To test the effect that different episodic sedimentsupply regimes can have on the morphology of a mountain stream, we developed a numericalmodel to recreate and extend simulations from flume experiments. The model performs wellin recreating the flume experiments in both slope and grain size distributions (GSD), which402.6. Summary and conclusionsare the two variables that represent morphological adjustment in our model. Channel widthis fixed and bedforms are assumed to be absent, even though bedforms did occur in the flumeexperiments.To understand the extent to which event succession plays a role in the flume experiments,we simulated alternative pulse successions of large-to-small events (i.e., infrequent-to-frequent)and small-to-large events (i.e., frequent-to-infrequent), while keeping the total sediment vol-ume feed the same. These simulations show that different pulse frequency sequences have nostrong effect on the long-term slope and GSD of the bed surface. In the short term large pulseevents can dominate the channel response causing an abrupt increase in slope, while the effectof subsequent smaller events is subdued as the channel is still adjusted to the large pulse. Ifsmaller events dominate at first, the channel adjusts more gradually.In our second set of simulations, we imposed different episodic sediment supply regimeswith the same total sediment supply volume on the same initial channel geometries with con-stant discharge. While being kept constant within a run, the episodic supply regimes differedin event frequencies, magnitudes, and GSD. We simulated 40 different event frequencies forwhich the sum of event magnitudes matched an overall equal total sediment supply. All 40pulse configurations were calculated for 9 GSD that differed in the width of the distribution σaround the same geometric mean grain size. The channels adjusted to the episodic sedimentsupply until they reached an equilibrium state in which each successive pulse led to the sameslope and grain size adjustment. We compared this state between runs and found a distinctiveregime change when the time between pulses (Tpp) became lower than a fluvial evacuationtime (Tfe), which we developed as a measure of the time it takes to remove a Ds50 thick layerfrom the channel surface under average transport conditions, modified by a measure of po-tential armouring (see Eq. 2.15).The condition of Tpp < Tfe causes a constant-feed-like sediment supply regime, as themodel runs show similar slopes and surface grain size distributions as constant-feed runsof the same GSD. When Tpp > Tfe the sediment supply regime becomes pulse-dominated.Under these conditions, we observed a lower relative slope in cases where the GSD is narrow(σ < 0.4), as the long time between pulses in combination with a low armouring potential412.6. Summary and conclusionsallows more erosion in the reach, ultimately lowering the equilibrium slope. If the GSD iswide enough to allow armouring (σ ≥ 0.4), a stronger armouring layer can develop duringthe periods of selective transport of smaller grain sizes and bedload starvation. This limits theminimum slope and increases sediment storage (and thus slope) in the long term.The application of the episodic supply regime classification to data from East Creek showsthat the threshold to a pulse-dominated regime lies at the fluvial evacuation time of roughly2.2 years. This creek probably receives sediment at a lower interval, which indicates a pulse-dominated regime. The armouring timescale lies around 3.5 years, indicating that if the long-term sediment supply was introduced over event frequencies between 2.2 and 3.5 years, itwould be removed most efficiently and result in a lower slope.Steeper channels than East Creek could show both a lower fluvial evacuation time (dueto higher slope, smaller channel area) and a lower pulse frequency (more landslide domi-nated), which could make these channels more likely to be pulse-dominated. Further studyof field cases is needed to strengthen the case for our classification of channel response typesto episodic supply regimes. In natural rivers, there are further modes of adjustment that thesystem can undergo after receiving sediment pulses, for example changes in bed forms or thestorage of excess material in sediment bodies along the channel. Still, the condition when achannel is receiving more material per pulse than what can be exported in the same time frame(i.e., the ratio of Tpp/Tfe is above 1), should be observable in natural rivers as irregularities inchannel long profiles due to increased sediment storage. In our model, we only supplied grainsizes that were transportable by the imposed flow conditions. In field streams it can be thatthe biggest clasts (e.g., boulders) are only transportable by extreme flow events, which wouldfurther increase the slope of reaches with high sediment supply.42Chapter 3Simulation of sediment pulse effects onCarmel River after the Los Padres Damremoval, Monterey County, California,USA3.1 SummaryWe used a one-dimensional bedload transport model (BESMo) model to study the effect of theproposed removal of the Los Padres Dam (LPD) on the Carmel River in Monterey Bay, Cal-ifornia, USA. Based on multiple sediment management options currently being considered,we developed general sediment supply scenarios that represent conditions of continued sed-iment retention, the re-establishment of background sediment supply, or the exposure of thereservoir sediment to erosion. To explicitly include the stochastic nature of floods in their ef-fect on sediment transport, we developed a method to generate synthetic hydrographs thatare based on the frequency, magnitude, shape, and timing of flood events in the historicalrecord. All scenarios that restore sediment feed below LPD cause large parts of the channelto aggrade. We found that establishing a long-time sediment feed (e.g. in the ’Pulse supply’433.2. Introductionscenario through managing how material is sluiced out of the reservoir) is most effective inestablishing a channel bed that has finer bed material than current conditions. In contrast,the ’Uncontrolled supply’ scenario with rapid erosion of the reservoir fines the bed for about10 years, after which the bed surface becomes significantly coarser through armouring. Overthe 60 year timeframe, the style of dam removal might be not be geomorphically significant.The establishment of connectivity between the upstream reaches and the main stem is whatgoverns the system response between 10-60 years after the Los Padres dam is removed. Thisechoes the previously recognized importance of the upstream areas for the supply of fine sed-iment on the channel morphology of the Carmel River.3.2 IntroductionIn the United States, a large number of dams were constructed between the late 1950s and thelate 1970s (Graf , 1999), trapping sediment and decreasing sediment yield from the mountainsto the coast by 20% (Kondolf and Pie´gay, 2011). Older dams pose a contemporary safety risk dueto surpassing the design age, particularly in areas like California that are at risk of earthquakes.Further, dams can have negative ecological impacts as they change the flow and sedimentregimes and can hinder fish passage (Grant, 2001; Grant and Lewis, 2015; Major et al., 2017).While there exist a range of sediment management options to restore downstream riverfunctions, (Kondolf et al., 2014), the removal of dams is shown to be particularly effective inefforts to restore biotic diversity due to unregulated flow regimes and the re-establishment ofdiverse channel features (Bednarek, 2001). Additionally, the removal of dams that block fishpassage opens up upstream habitat (Grant and Lewis, 2015). Recent removals of large dams(> 30 m) on the US west coast were undertaken at the Elwha river with both the Elwha dam(2012) and the Glines Canyon dam (2014) (East et al., 2015), and the Condit dam on the WhiteSalmon River in 2011 (Wilcox et al., 2014). The geomorphological downstream effects of damremoval vary due to the different reservoir characteristics and removal approaches, yet thesediment release from the reservoir is commonly coupled with an increase in downstreamsediment transport and increased deposition (Major et al., 2017). Giving a recent overviewof dam removal studies, Foley et al. (2017) state that the geomorphic response to removals is443.2. Introductiontypically fast, and that both geomorphic and ecological connectivity are restored quickly.This study focuses on the Carmel River where the San Clemente Dam (SCD) was removedin 2015 (Harrison et al., 2018), and where further upstream the Los Padres Dam (LPD) is beingconsidered for removal (MPWMD, 2016), motivating this study. The geomorphology of theCarmel River is still adjusting to the removal of SCD and to a partial rerouting of the mainriver channel that was done to prevent erosion of former reservoir sediments.The dam removal project of LPD is highly complex. The potential geomorphic responseof the river to the upcoming LPD removal will be superimposed onto the ongoing river ad-justment to the SCD removal. Additionally, the river is impacted by a wide range of naturaldisturbance events such as wildfire and landslide activity which have historically contributedsignificant sediment to the river and are known to impact sediment transport dynamics. Fi-nally, the climatic regime of the area leads to highly variable precipitation patterns which arelikely to change in unpredictable ways due to climate change over the coming decades. Col-lectively, the co-occurrence of complexity across so many variables and scales requires novelapproaches to integrating uncertainty and stochasticity into predictions of river morphologyadjustment.Numerical modelling has been shown to be useful in assessing potential downstream im-pacts of dam removal (Cui and Wilcox, 2008; De Rego, 2018), providing information for man-agers to understand long-term effects of dam removal (Foley et al., 2017). To simulate the effectsof possible management scenarios on the Carmel River, we used the numerical 1-D model BE-SMo (Bedload Scenario Model; Chapter 2 of this thesis), which was developed to study thesediment transport and geomorphological effect of large sediment pulses in fluvial systems,making it highly suitable for this task of simulating a dam removal. The model integratesapproaches from other studies in calculating both hydraulic and sediment dynamics based oninput discharge and sediment supply conditions (e.g., Cui and Parker, 2005; Wong and Parker,2006; Ferrer-Boix and Hassan, 2014; An et al., 2017a), and is similar to other models used tosimulate the effects of dam removal (e.g. Viparelli et al., 2011; Cui et al., 2006).However, past modeling efforts have been computationally limited and cannot adequatelyintegrate the range of complex process at work in the Carmel River. Integrating stochasticity453.3. Study areaand uncertainty is particularly important in forecasting dam removal impacts due to inter-relations between multiple poorly constrained fluvial, climatic, and hydrological processes.The BESMo model employed in this study is a novel modelling approach that overcomes thelimitations of prior models by facilitating computation of multiple stochastic sedimentologi-cal and hydrological scenarios to be run in parallel to investigate uncertainty in river response.Unlike methods of traditional platforms such as the US Army Corps Hydraulic EngineeringCenter River Analysis System (HEC-RAS) or the US Bureau of Reclamations Sedimentationand River Hydraulics - One Dimension (SRH-1D), this allows us to simulate hundreds of dif-ferent hydrographs per scenario and thus to explore the effects of uncertainty in future hydrol-ogy.In this context, the objectives of this chapter are twofold. First, we want to provide apractical assessment of the response of the Carmel River to four different management optionsto inform managers and decision makers in charge of the dam removal. The second objective isto situate BESMo within the broader context of modelling dam removals. We assess the utilityof BESMo for capturing the complexity and uncertainty of hydrological and sedimentologicalvariables that impact river response to dam removal, and we consider the insight that can begained from this computational approach.3.3 Study areaThe Carmel River Watershed is approximately 660 km2, originating in the Santa Lucia Moun-tains and terminating at the Carmel Lagoon into Carmel Bay, just south of the town of Carmel-by-the-Sea (see Figure 3.1). Following decades of decoupling of the channel from upstreamsediment supply, the elevation of the river bed has lowered and the bed surface has coarsened(Kondolf , 1982; GMA, 2008). Furthermore, along the lowermost 24.5 km of the mainstreamCarmel River the banks were altered by installation of a variety of materials to decrease theprobability of bank erosion during flood events (Hampson, 2018).The 32 m high San Clemente Dam was build 1921 for water storage and removed in 2015 inthe Carmel River Reroute and Dam Removal (CRRDR) project, as the dam was deemed seis-mically unsafe, threatening 1,500 homes and other buildings on the floodplains of the lower463.3. Study areaCarmel River (Harrison et al., 2018). Furthermore, Carmel River provides a crucial anchor habi-tat for the threatened steelhead trout species Oncorhynchus mykiss, of which the juveniles hadlimited access upstream until the dam was removed (Harrison et al., 2018). During the CRRDRproject, the Carmel River was rerouted into the lower end of its tributary, San Clemente Creek,to prevent the erosion of former reservoir deposits, which were additionally protected witha diversion dike. The re-route channel itself was constructed as a step-pool reach fixed withlarge boulders and excess sediment was added on the banks to be redistributed in floods.Sources: Esri, USGS, NOAA±0 2 4 6 8 101 KilometersLegendCarmel RIverTributariesLos Padres reservoirReachAbove Los PadresLos Padres to Cachagua CreekCachagua Creek to San Clemente DamSan Clemente Dam to Tularcitos CreekTularcitos Creek to Las Garzas CreekLas Garzas Creek to Robinson Canyon RoadRobinson Canyon Road to Potrero CreekPotrero Creek to LagoonLagoonReach 3BReach 4AReach 5Reach 4BReach 3AReach 2Reach 1BReach 1ALagoon Potrero Creek Robinson Canyon Creek Las Garzas CreekHitchcock Creek Tularcitos CreekSan Clemente CreekPine Creek Cachagua Creek© OpenStreetMap (and)contributors, CC-BY-SA,Sources: Esri, USGS, NOAAFigure 3.1: Map of simulation reaches. The exact location of each node can befound in Appendix Table B.2Los Padres Dam is located approximately 40 river km upstream from the river mouth andwas constructed in 1948 and 1949 as an earth fill barrier 45 m high. The contributing watershedabove Los Padres Dam is approximately 114 km2, with a mean annual precipitation of 993 mm(AECOM, 2017). Much of the upper Carmel River watershed has been burned in previous473.4. Methodswildfires that introduced episodic sediment supply in the years 1977, 1999, 2008 and 2016.The Marble Cone fire in 1977 was the most significant supply event, transporting 727,800 m3of sediment into the reservoir. The initial water storage capacity was 3.9 Mio m3, which overtime has been reduced through sediment accumulation to about 48% of usable storage. Thefuture sediment accumulation is estimated to be between 12,300-24,600 m3 per year (MPWMD,2016). The reservoir sediment storage was estimated to be 1,550,000 m3 (AECOM, 2017), ofwhich roughly 75% is transportable as bedload. We estimate the background sediment feedfrom upstream of the reservoir to follow the same size distribution as the material stored inthe reservoir, giving a mean estimate of sand and gravel transport rate of about 13,400 m3yr−1.3.4 MethodsWe simulated the response of Carmel River to different sediment supply scenarios usingBESMo. We calculated bedload sediment transport in a 1-D channel with a node spacing of500 m with the transport function developed by Wilcock and Crowe (2003). We introduced anunerodible boundary condition at the former SCD site to prevent unrealistic erosion (see Sec-tion 3.4.1) and reduced the node spacing in the re-route channel to between 30 m and 150 mto test the model by comparing the simulated profile adjustment with surveys done over athree-year period (see Appendix B.1).We used the best available data to initialize elevation and grain size values at each node(see Section 3.4.2). For the full 60-year simulations, we generated synthetic hydrographs (seeSection 3.4.3) and explored the effects of sediment supply defined in four different manage-ment scenarios that are presented in Section 3.4.4. We present a summary of model inputparameters in Table 3.1, which for many parameters are the same across all four sedimentsupply simulations.483.4.MethodsTable 3.1: Complete list of model parameters for No Action simulation and changes applied to Historical, Pulse, andUncontrolled Supply simulations.Parameter No Action Historical Supply Pulse Supply Uncontrolled Sup-plySimulation SpatialBoundariesLos Padres Dam to Carmel Lagoon Same as No Action Same as No Action Same as No ActionSimulation TimePeriod60 years (2017-2077) Same as No Action Same as No Action Same as No ActionNode Distribution 500 m 500 m, ∼100 m inthe San ClementereachSame as HistoricalSupplySame as HistoricalSupplyRiverbed SedimentLayers100 Layers: 1 active layer and 99 subsur-face layers; surface layer depth ranges from0.5 to 1 mSame as No Action Same as No Action Same as No ActionSediment TransportMechanicsWilcock and Crowe (2003) Same as No Action Same as No Action Same as No ActionModel Time Step Variable, between 5 s and 1 min Same as No Action Same as No Action Same as No ActionHydrology Random annual hydrographs and number offloods based on MPWMD1 discharge data ofentire basin, internal boundary conditions attributary confluencesSame as No Action Same as No Action Same as No ActionSediment TransportPeaking FactorHydrograph peaking factor applied to days offlood peak ranging from 3 to 1.6Same as No Action Same as No Action Same as No ActionUpstream FlowBoundary Condi-tionHydrographs interpolated from Robles delRio gauge to Los Padres Reservoir assumingno flood peak attenuationSame as No Action Same as No Action Same as No Action1Monterey Peninsula Water Management District493.4.MethodsDownstream La-goon BoundaryCondition0.85 m water surface elevation (URS, 2013) Same as No Action Same as No Action Same as No ActionSediment SupplyBoundary Condi-tionsMain stem and major tributary sediment sup-ply rating curves for chronic conditions withboundary conditions at tributariesSame as No Ac-tion, with addedsediment supplyequal to 13,400m3yr−1 inferredfrom long-termsedimentation rateincluding Marble-Cone fireSediment supplyat Los Padres dambased on RC 1-6through a culvert.Tributaries same asin No Action.Sediment supplyat Los Padres dambased on Exp 1-3.Tributaries same asin No Action.River Bed SurfaceSediment SizesData sourced from MEI (2002), 2015 CSUMBdata courtesy of Doug SmithSame as No Ac-tion, added TetraTech (2015) datato represent boul-der steps in SanClemente ReachSame as HistoricalSupplySame as HistoricalSupplyRiver Bed Subsur-face Sediment SizesSubsurface grain sizes set to distribution re-ported as MEI80K, MEI (2002), subsurfacemaximum grain size set to between 512 and2048 mm to control bed erosion based on trialruns completed Oct-Nov 2017Same as No Action Same as No Action Same as No ActionReservoir evacua-tion curveN/A N/A N/A Modelled afterMarmot Dam re-moval, See Table3-6Number of 60-yearhydrographs300 60-year hydrographs Same as No Action Same as No Action Same as No ActionRiver Bed Longitu-dinal ProfileProfile constructed using Whitson Engineers(2017), URS (2013), and NED19 datasets; re-sulting profile used in simulationSame as No Action Same as No Action Same as No Action503.4.MethodsRiver Cross Sec-tional GeometrySourced from Normandeau (2016) and sup-plemented with URS (2013) data when notavailableSame as No Action Same as No Action Same as No Action513.4. MethodsFigure 3.2: Long profile of the full run simulation domain with annotation show-ing the position of the fixed-elevation boundary condition.3.4.1 Model boundary conditionsThe simulations required specification of several boundary conditions: (a) the riverbed ele-vation at the downstream-most node at the Pacific Ocean (which we assume is fixed), (b) thebedload sediment supply rate at the upstream-most node at Los Padres Dam and from eachof the major tributaries, and (c) the water flow rate at the upstream-most node at Los PadresDam and from each of the major tributaries. Additionally, test simulations revealed that themodel calculates unrealistic riverbed erosion within the re-route reach at the San ClementeDam Removal project site. Construction conditions at the upstream end of the re-route reachintroduced a relatively fixed river profile condition at the transition from the former reservoirdeposit to the reroute section. This was accomplished by capping shallowly occurring bedrockat this location (depth below bed surface is approximately 1.5 m) with steel rebar and concrete.We emulate this constructed condition in the model with an internal fixed elevation boundarycondition by splitting the model domain into two parts: (1) the upper part from Los PadresDam to the beginning of the reroute upstream of the former San Clemente Dam and (2) thelower part from the reroute to the mouth of Carmel River (see Figure 3.2). This adjustmentprevented both erosion and aggradation at this point in the channel and sediment was fullyconveyed through this node. However, the local grain size distribution could adjust accordingto the composition of the upstream bedload supply.523.4. Methods3.4.2 Node initializationTo begin each simulation, the 87 model nodes along the river long profile (see Figures 3.1 and3.2) were initialized with values of elevation and grain size distribution. A detailed list of thedata used for each node is provided in Appendix Table B.2.Elevation profileThe initial elevation profile was based on two different channel surveys and DEM data, as thesurveys are available only for the lower parts of the river (see Appendix B.2). We calculated theelevation for nodes in the upper Carmel River from a 3m-resolution digital elevation model.Grain size distributionsFor each model node, we specified three different initial Grain Size Distributions (GSDs): (1)initial active layer GSD, (2) initial subsurface GSD, and (3) maximum subsurface GSD. Thegrain sizes used for this purpose were collected from several different sources: MEI (2002),URS (2013) and Chow et al. (2016) 2. Through initial model runs we determined that modelresults are sensitive to the initial configuration of the surface and subsurface grain size distri-butions. This is unsurprising for two reasons. First, the bed surface GSD is used to estimatechannel roughness through the 90th-percentile grain size (D90), which affects the calculatedcross-sectional average velocity. This in turn can impact the calculation of water depth andthe associated cross-sectional average shear stress. Second, mass balance calculations for ad-justment of the grain size fractions present on the channel bed surface are dependent on thethickness of the active layer. The active layer concept simplifies bedload transport as a two-layer system: grains in transport within the active layer and immobile subsurface grains. Theinterface between the active layer and the subsurface is an exchange surface for grain sizefractions.The active layer thickness is typically parametrized as the local D90 grain size multipliedby a constant which ranges from 1 to 2 (we use a value of 2 following Parker, 2008). As a result,the D90 affects both the bed surface roughness and the depth of the bed which participates2from here forward referred to as the CSUMB data, provided to us by Douglas Smith533.4. Methodsin bedload transport. Relatively small D90 grain sizes led to unrealistically large depths ofbed erosion along the mainstem Carmel River reach downstream of the Narrows within trialsimulations. As a result, we have carefully identified a defensible, but not demonstrable (dueto a lack of field data), initialization of grain sizes for the No Action Simulation. Our choicesgenerally reflect field observations of bed surface grain sizes where data is not available, mostnotably within the vicinity of the so-called Steinbeck pool, and construction specifications ofthe bed surface for the San Clemente Dam Removal step-pool reach. In general, all channelevolution models are sensitive to the initialization of bed surface and subsurface grain sizedistributions. Unfortunately, almost all studies like the present one lack actual informationto minimize uncertainty with respect to this model input because it is impractical to samplethe bed to minimize this uncertainty (i.e. Church et al., 1987; Rice and Church, 1996; Bunte andAbt, 2001), or data is collected for reasons that go beyond channel evolution modelling andtherefore concessions are made in order to collect a diversity of data rather than data for oneparticular purpose, and subsurface data is rarely collected, as in the present case. Grain sizespecifications for each model node are provided in Appendix Table B.2 and the distributionsare presented in Figure B.19.Initial active layer GSD The active layer GSD is important mainly at the beginning of the sim-ulations, as it describes the transportable size classes directly exposed at the channel surface.For the initialization of this layer, we used the active-layer data specified within the URS simu-lations, except for the lowest reaches where the coarser CSUMB data is a better representationof current condition. The CSUMB data also mitigated unrealistically large simulated channelbed erosion depths in the spin-up runs.Initial subsurface GSD The subsurface GSD lies directly beneath the active layer and the sizeclasses get incorporated into the active layer if the channel erodes. We deemed the URS sub-surface data too fine for the initial subsurface GSD, as the channel eroded significantly in thespin-up runs. The MEI subsurface GSD data mitigated simulated erosion and was deemed abetter representation of subsurface conditions. Note, however, that both datasets are derived543.4. Methodsfrom surface population estimates because we lack measured subsurface GSDs.Maximum subsurface GSD Whereas the URS model (URS, 2013) assumed the subsurface GSDto be present for the whole depth of the subsurface (virtual depths of multiple meters), wefound that to prevent unrealistic erosion directly at the dam site and in the lower reaches, wehad to introduce layers of coarser maximum grain sizes below the first few subsurface layers.We generated these by removing all smaller size classes in these layers and only specified thepresence of a uniform, large grain size (either 512 mm or 256 mm). We do not have field datato support the presence of these more erosion resistant layers, but the approach yields modelresults which are on average consistent with previous model results of the spatial patterns oferosion and deposition (URS, 2013), and more importantly with observed conditions at theSan Clemente Dam removal project site (see Appendix B.1). We specify the depth at whichthis layer begins and which size we use in Appendix Table B.2.3.4.3 Hydrology submodelTo generate synthetic hydrographs, we relied on statistical data derived from both historicalflow records and simulated tributary discharge provided by MPWMD, who generated thedata from watershed scale hydrologic modelling with PRMS (Markstrom et al., 2015). Below, wespecify the methods used to randomly select an annual peak flow magnitude from historicaldistributions, as well as the number of peak flows for each year of the 60-year simulationperiod. An overview of the different components of this calculation is given in Figure 3.3. Itis important to note that BESMo simulation uses the same randomly constructed hydrologictime series for each sediment supply scenario.Reach and tributary discharge We segmented the Carmel River into 5 main reaches followingAECOM (2017). We then further subdivided reaches 1, 3, and 4 to better represent the effect oftributaries as we recalculated the hydrology for each sub reach. A map of the reach locations,catchment areas, tributaries, and the position of model nodes is shown in Figure 3.1. For eachsimulation, we generated a random hydrograph of mean daily flow for the reference reach that553.4. MethodsFigure 3.3: Components of the calculation of the flood time series with the hy-drology submodel. Blue: Input data from historical records and watershedmodelling. Green: statistical data per flood class. Orange: steps to generateflood time seriesincluded the Robles del Rio (RR) USGS gauge (Reach 3A). We then calculated the dischargefor all other reaches by multiplying the D3A reach discharge by the historical discharge ratioof each of the other simulation reaches. The discharge ratios for both the main stem (used formodelled sediment transport) and the tributaries (used for tributary sediment feed from ratingcurves) are listed in Table 3.2 and were calculated as averages from 4748 days of overlappingrecords provided by the MPWMD as part of ongoing watershed scale hydrologic modellingwith PRMS (Markstrom et al., 2015). The overlapping period of records extends from October1st, 2001 to September 30th, 2014. For each simulation reach, the estimated hydraulics for eachsequential daily discharge was simulated in a backwater flow calculation.563.4. MethodsTable 3.2: Carmel River main stem discharge ratio from 4748 days of overlappingmean daily flow data (Oct 1st, 2001-Sept 30th,2014). The flow data is fromhistorical record where available, and otherwise modeled.Main Stem ID Reach DischargeRatio toRRStandarddeviationBelow Los Padres ’BL’ D1A 0.66 0.11Below Los Padres + CachaguaCreek’BL+CA’ D1B 0.70 0.11Sleepy Hollow Weir ’SHW’ D2 0.93 0.08Robles del Rio Gauge ’RR’ D3A 1.00 0.00Don Juan ’DJ’ D3B 1.05 0.11Near Carmel Gauge ’NC’ D4A 1.06 0.18Highway 1 ’HWY1’ D4B 1.02 0.21TributariesCachagua Creek ’CA trib’ D1B 0.04 0.02San Clemente Creek ’CL trib’ D2 0.12 0.03Tularcitos Creek ’TU trib’ D2 0.01 0.01Las Garcas Creek ’GA trib’ D3B 0.07 0.03Robinson Canyon Creek ’RC trib’ D4A 0.01 0.01Potrero Creek ’PO trib’ D4A 0.01 0.01Hitchcock Creek ’HI trib’ D3A 0.01 0.00Random hydrographs with flood events We generated random hydrographs for the 60-year anal-ysis period to simulate plausible future hydrologic conditions. We did so by assuming thatfuture conditions will be statistically similar to the historical record of flood magnitude andfrequency, flood duration, and number of floods per year. As a result, we extracted the follow-ing information from the historical records to develop the hydrographs: (1) annual peak flows,(2) number of floods per year, (3) flood duration, and (4) timing of flood events within a year.With this information, we first defined a flood event as any flow above a threshold of 3 m3s−1mean daily flow for at least two consecutive days (flood duration) in the historical record. Wethen group all flood events into 5 classes by the maximum mean daily flow, with roughly aminimum of 10-20 events in each of the higher flood classes for the 60-year simulation period(Table 3.4). After this step, we used this catalogue of historical flood events to develop random60-year daily simulation hydrographs.573.4. MethodsTable 3.3: Flood frequency table for the Robles del Rio USGS gauge using meandaily discharge on the day of yearly peak flow.Expecteddischarge(m3s−1)Exceedancechance (%)Lower confidenceinterval (5%)(m3s−1)Higher confidence inter-val (95%) (m3s−1)1282.76 0.2 2156.36 710.75986.28 0.5 1631.56 573.05787.14 1 1279.66 474.49608.41 2 968.32 381.48405.14 5 620.98 268.27275.71 10 407.83 190.97168.11 20 237.79 121.7759.07 50 78.76 44.5317.83 80 24.67 12.768.88 90 13.16 5.994.79 95 7.69 3.061.33 99 2.66 0.78Annual peak flow, number of floods per year and flood magnitudes For each full simulation, werandomly generated a 60-year record of daily flows by simultaneously carrying out three mod-elling steps: (1) randomly choosing an annual mean daily peak flow magnitude, (2) determin-ing peaking factors (see below) to apply to the annual mean daily peaks, and (3) determiningthe number of floods for each simulation year. For each year in a simulation, we randomlyselected from among the flood frequency classes previously calculated with HEC-SSP and de-scribed in AECOM (2017, Section 2.6.3 Flood Frequency Analysis). However, instead of usingthe instantaneous annual peak flows to prepare the flood frequency statistics, we calculatedthe frequency and magnitudes of the mean daily flows for each specific day correspondingto an instantaneous flood peak within the historical record (Table 3.3). This is consistent withthe modelling approach of past local studies (MEI, 2002; URS, 2013, e.g.). However, unlikeprevious studies, we applied peaking factors: the ratio between mean daily flow and the in-stantaneous peak flow at the day of highest flood discharge (Table 3.4). This captures thenon-linear relationship between instantaneous discharge and the rate of bedload transport,which would be lost by only using mean daily flow values.583.4. MethodsTable 3.4: Peaking factor for flood classes from the historical peak flows.Flood Class: 1 2 3 4 5Mean daily flow 3-7 m3s−17-21 m3s−121-42.5 m3s−142.5-85 m3s−1>85 m3s−1# in historicalrecord63 43 22 25 11Peaking factor 1.6 1.6 2 3 3Figure 3.4: Number of floods per water year from the historical record at the Rob-les del Rio USGS gage.We simultaneously decided how many floods occur in each simulation year by using as-sociated probabilities for the historical period. With the historical RR gauge data, we calcu-lated flood occurrence probabilities for between 0-8 floods per year based on a flood dischargethreshold of 3 m3s−1 (Table 3.5 and Figure 3.4). We then randomly sampled from this dis-tribution to determine the number of floods for each simulation year. The choice of floodmagnitude and the number of floods for each simulation year was carried out independentof each other so that the simulations are not strictly constrained by the hydrologic characterof the historical record. This method of developing the hydrologic records for the simulationperiod means we are not overly restricting our analysis to assumptions of stationarity, despitereliance on the historical discharge records. Next, we determine how intra-annual flood peakscompare if more than 2 floods are randomly chosen for a simulation year.We analysed the historical data to determine how intra-annual flood peaks varied from593.4. MethodsTable 3.5: Probability of number of floods per year and ordered peak mean dailyflow ratios between the floods within one year in relation to the highest meandaily flow at the Robles del Rio USGS gauge. The cumulative probabilitycolumn indicates that most years in the historical record have two floods peryear.Number of floodsper yearOccurrences inrecordCumulativeprobabilityAverage maximumdaily mean flow ratios0 3 0.05 01 14 0.23 12 19 0.31 1, 0.363 10 0.75 1, 0.42, 0.194 5 0.84 1, 0.42, 0.22, 0.125 4 0.90 1, 0.59, 0.30, 0.23, 0.146 3 0.95 1, 0.66, 0.24, 0.13, 0.09, 0.098 3 1.00 1, 0.52, 0.36, 0.29, 0.20, 0.10, 0.08, 0.07event to event, depending on how many floods occurred in a given year. The results areshown in the far right-hand column of Table 3.5, presented as the ratio of the highest meandaily flow at the RR gage to the flood peak associated mean daily flow. This data is importantfor a few reasons. First, it is a critical link for construction of the random annual hydrographsfor the 60-year simulation period because the statistical analysis of annual peak flows onlyrecognizes the maximum flow event for each year and does not contain information on morethan one flood within the same water year. Second, like the data in Table 3.3 and Table 3.4,information on average intra-annual flood variability permits us to remain consistent with thestatistical nature of historical floods, while supporting a stochastic approach.Flood duration and timing within each year We assumed the duration of each flood event to bemainly dependent on the peak flow magnitude. To attribute a flood duration to each floodin the randomly constructed records, we calculated average hydrographs within the historicalrecord at the RR gauge for the five flood categories in Table 3.4. The average hydrographsare shown in Figure 3.5 (left panel) where floods have average durations of approximately10 to 50 days or more, based on flow conditions prior to the peaks. We assign the timing ofeach flood event within a year based on the most likely flood day-of-year observed from the603.4. Methods-50 0 50 100 150Time in days00.10.20.30.40.50.60.70.80.91Magnitude normalized> 85 cms42-85 cms21-42 cms7-21 cms3-7 cms50 100 150 200 250 300 350Day of calendar year50100150200250Peak flow (cms)> 85 cms42-85 cms21-42 cms7-21 cms3-7 cmsFigure 3.5: Left: Relative mean daily flow as time series in five flow classes fromthe averages of all events in the historical record. The flood classes in thelegend are in mean daily flow at peak day. Right: Timing of floods in thecalendar year in relation to mean daily peak flood magnitude.0 200 400 600 800 1000Run number012345678Cumulative discharge (cms)104after 10 yearsafter 30 yearsafter 60 years0 50 100 150 200 250 300Run number012345678Cumulative discharge (cms)104after 10 yearsafter 30 yearsafter 60 yearsFigure 3.6: Left: Cumulative discharge for 1000 randomly generated hydro-graphs after 10, 30, and 60 years, sorted by 10-year cumulative discharge.Right: Subsection of 300 runs that were simulated for each project simula-tion with BESMo. Runs 1-100 represent the 100 wettest cases of first 10-yearcumulative discharge (10th percentile and lower), runs 101-200 representaverage conditions (45th-55th percentile), and runs 201-300 represent dryconditions in the first 10 years (90th percentile and higher).historical record (Figure 3.5, right panel) for the associated hydrologic category. In the eventof overlapping flood events in time, we move the overlapping event to center around anotherlikely flood day.613.4. MethodsHydrograph use in BESMo We used the described approach to generate a population of 1,000synthetic hydrographs that statistically match the historical record. It is important to note thatwe do not use climate projections to develop future records of possible daily discharge, butit is feasible to do so by changing the underlying flood frequency-magnitude relation. Forthis study we are interested in the effect that very dry or very wet conditions have on theCarmel River after the potential dam removal. To create a subset of hydrographs to simulate,we ranked the resulting discharge time series from driest to wettest based on cumulative dis-charge during the first 10 years, during which we expect the channel adjustment to be mostresponsive. We then chose the 100 wettest, the 100 driest, and 100 average3 simulations fromthe 1,000 randomly constructed hydrographs (Figure 3.6). Because a given hydrograph is clas-sified based only on the first 10 years of discharge, annual cumulative discharge is still highlyvariable after 30 and 60 simulation years.An advantage of using the sorted collection of hydrographs is that our simulation resultscan be linked to statistical tendencies of the range of hydrologic conditions (Wilcock, 2001).Based on the flood statistics, the 100 wet-year hydrographs have discharges higher than what90% of the historical record indicates. Similarly, the average hydrologic conditions are de-signed to represent hydrographs that produce discharges within 45% and 55% of what thehistorical record indicates. The dry hydrologic conditions have discharges lower than 90% ofwhat the historical record shows.Hydraulics BESMo simulates river flow with the normal flow approximation and the back-water solution to the momentum equation(Cui et al., 2006). It does not directly incorporatethe effect of a non-uniform cross-sectional channel shape on bed shear stress. However, wedo account for such conditions in calculation of the water surface profile with the backwatersolution. This permits us to better represent the effect of high flows on the average channelbed shear stresses at each model node. To accomplish this, we capture the non-uniformity ofcross-sectional shape through data previously reported which relates water depth and flowarea to discharge, averaged for each model reach (URS, 2013). With this information, we cal-350 simulations on either side of the median.623.4. Methodsculate an approximated water surface width based on water mass conservation (i.e. discharge= cross-sectional average velocity times flow area) and a relationship between discharge andflow velocity from MEI (2003) (see Appendix B.9).3.4.4 Sediment supplyWe describe available sediment transport data for both the Carmel River and its tributariesin Appendix B.3. Based on the general management options for the Los Padres dam andreservoir, we developed the following four sediment supply scenarios:No Action simulation The No Action project alternative assumes no action is taken at LosPadres dam or within the reservoir. This means coarse material would continue to accumulatein the reservoir, and the only sediment to bypass the dam is that which is carried in suspensionover the top of the dam during floods. As a result, the only bedload sediment supply toreaches downstream of the former San Clemente Dam in this scenario is that from tributariesdownstream of Cachagua Creek. This simulation is the baseline simulation for comparativepurposes.Historical Supply simulation The historical sediment supply simulation assumes that someactions are taken to restore historical sediment supply in the Carmel River. This means thatbedload-sized sediment supplied from the watershed contributing to the Los Padres reservoirwould once again contribute to the sediment budget for the watershed downstream of theDam. The simulation does not account for any of the bedload sediment presently stored withinthe reservoir deposit. This simulation can serve as a baseline end member for the Carmel Riverwatershed under approximated unmodified sediment yields. However, the effects of resumedhistorical supply to the watershed downstream of Los Padres is a function of present-day riverconditions downstream of the Dam, which are the result of a mixture of human-driven impactsrelated to dam construction and river/floodplain modification. This simulation can also serveas a representation of controlled dam removal alternatives, after the existing reservoir deposithas been stabilized, removed, or otherwise is no longer a factor.633.4. MethodsAECOM (2017) report a total dry unit weight of 1,831,850 tons reservoir sediment4, whichincludes 26% of silt- and clay-sized particles. Including the Marble-Cone fire in 1977, this de-posit of sediment corresponds to a long-term average sedimentation rate of 18,100 m3yr−1. Abedload rating curve was calculated by scaling down the sedimentation rate of 18,100 m3yr−1by 74%, resulting in 13,400 m3yr−1 of bedload sediment. This rating curve was applied toflood events with flows higher than 3 m3s−1. Simulation hydrology used for the HistoricalSupply is identical to that used for the No Action simulations.Pulse Supply simulation The Pulse Supply simulation assumes that some actions are takento manage the introduction of pulses of bedload from the Los Padres reservoir deposits intothe Carmel River, which may include a sluicing tunnel, or dredging. A preferred means ofsediment relocation has not been identified, but the Pulse Supply simulation is intended tomimic the behaviour of sediment introduced below Los Padres Dam in any of these scenarios.The Pulse Supply simulation is designed to evaluate probable downstream responses un-der a range of conditions which emulate how passive sediment transfer may affect down-stream reaches with the introduction of sediment pulses. The Reservoir Alternatives memo(AECOM, 2017) highlights four different sediment management alternatives for Los PadresReservoir: (a) excavate, truck and dump, (b) sluice tunnel, (c) bypass tunnel, (d) some combi-nation of these three approaches. Our implementation most closely reflects the sluice tunnel,as we use flow-weighted sediment rating curves for a flow range up to5 140 m3s−1. We furtherimpose a minimum flow of 8.5 m3s−1 for sediment transport to start, which represents theclosing of the sluice gate to improve reservoir fill times.Sediment delivered through the sluice structure to downstream reaches is sourced fromthe Los Padres reservoir deposits. Additionally, upstream sediment supply replenishes thereservoir deposit in the same way the sediment feed was calculated in the Historical scenario(flow weighted). The rating curves are of the form Qs = aQbw, with Qs as sediment feed in4This figure was updated and revised in 2018, but because estimated reservoir capacity changedby less than one percent, model simulations were not re-run as they would likely have a similarlynegligible effect.5The Report indicates an approximate maximum operational flow of just over 140 m3s−1 for ahorseshoe-shaped sluice tunnel of approximately 4.5 m in width.643.4. Methodstons per day and Qw as discharge in ft3s−1. The relationship is based on empirical data fromtributaries of the Carmel River, suggesting coefficient values for a ranging from 0.0002 - 0.6,and exponent values for b ranging from 1.2 - 3.6. We assume that sediment supply duringthe simulations to the structure is at or near the respective capacity of the daily simulatedflow, mimicking open-channel conditions. The grain size distribution of the pulsed sedimentsupply is a mixture of the sand and coarser fractions presently within the reservoir.We use the same 300 60-year hydrographs as previously outlined for simulations of wet,average and dry conditions. We present 6 different rating curves across the 3 different hydro-logic conditions (18 different pulse-like sediment supply scenarios) in Table 3.6. One benefit ofusing flow-weighted rating curves is that sediment pulse size is scaled by flow, and thereforethe hydrologic time series remains one of the primary control parameters of the simulations.Plots of sediment yield and reservoir evacuation over time for all hydrologic conditions arepresented in Appendix B.10.653.4.MethodsTable 3.6: Overview of sediment feed scenarios by rating curve type. We use the pre-calculated hydrographs to predictboth the storage depletion time and the median sediment supply that would follow from each sediment feedscenario. RC 1 to 6: Take effect only at minimum discharge of 8.5 m3s−1 and limit flow to a maximum of 144 m3s−1.Flow over the maximum is delayed to subsequent days until the flow volume of the simulated flood was conveyedthrough the structure.Sediment feed type Median timeto depletion (years)Median timeto 50% of storage (years)Median sediment supplyin first 10 years (m3yr−1)ID Formula wet average dry wet average dry wet average dryRC 1 Qs = 0.35 ∗Q1.5w,inst 34.91 never never 7.42 never never 73,300 10,100 500RC 2 Qs = 0.50 ∗Q1.5w,inst 6.87 never never 4.82 37.88 54.89 84,100 14,400 800RC 3 Qs = 0.05 ∗Q1.75w,inst 58.84 never never 8.29 never never 72,600 8000 400RC 4 Qs = 0.15 ∗Q1.75w,inst 3.96 28.43 34.92 3.18 17.93 26.92 87,800 24,000 1100RC 5 Qs = 0.025 ∗Q2w,inst3.36 21.90 29.90 2.92 11.43 23.98 87,800 32,300 1200RC 6 Qs = 0.05 ∗Q2w,inst 2.97 14.92 24.87 2.87 6.87 18.99 87,800 46,200 1800Table 3.7: Exponential decay curves for the three simulated scenarios.ID Formula descriptionExp 1 S = −6 ∗ log Qw,cum + 150 Low storage decayExp 2 S = −9.3 ∗ log Qw,cum + 178.59 USGS best fit for Marmot damExp 3 S = −10 ∗ log Qw,cum + 180 High storage decay663.4. MethodsUncontrolled Supply simulation The Uncontrolled Supply simulation assumes that the LosPadres dam is removed without taking steps to manage the subsequent erosion of the reservoirdeposits. Our approach of simulating this case is based on observations by Major et al. (2012),who describe that the sediment storage of the Marmot reservoir decreased in an exponentialfashion after the removal of the dam. This indicates that the initial sediment supply is veryhigh but decreases quickly. Grant and Lewis (2015) found this observation to be valid for mul-tiple dam removal cases. As the Marmot reservoir is similar in both reservoir size and particlesize, we assumed the Los Padres reservoir would deplete in a similar fashion. We designedthree potential decay curves (Exp 1 to 3) which envelope the natural decay rates reported byMajor et al. (2012). The curve Exp 2 matches the data from Marmot dam, while Exp 1 and Exp 3represent higher and lower storage decay rates, respectively (see Table 3.7 and Figure 3.7). Therating curves are of the form:S = −a ∗ log Qw,cum + bwhere S represents storage (expressed in percent remaining), Qw,cum is the cumulative dis-charge (in m3) since the removal of the dam, and a and b are empirically determined.We assumed that the material from all zones would leave the reservoir well mixed and ex-cluded all grain sizes < 1 mm. A background sediment feed rate of 13,400 m3yr−1 is added tothe sediment export calculated from the decay curve (spread over the simulated hydrograph).This is different from the Pulse Supply scenario, in which the background feed replenishedthe reservoir sediment storage. In contrast to the Pulse Supply scenarios, we did not impose aminimum discharge to erode the reservoir deposits. We imposed the same 300 60-year hydro-graphs as in all other simulations. The remaining model specifications are the same as in thePulse Supply scenarios and summarized in Table 3.1.In general, the simulation of uncontrolled sediment release in form of exponential decaycurves leads to:1. Large volumes of the reservoir sediment are eroded early even during small floods, asthe decay curve is not depending on flood magnitude.2. Following this, the erosion of sediment in the reservoir will decrease substantially within673.4. MethodsFigure 3.7: Exponential decay curves modelled after Marmot dam removal data.The ’USGS best fit’ matches our ’Exp 2’ scenario, while the ’Exp 1’ matchesa visually fit ’low scenario’ and ’Exp 3’ matches a visually determined ’highscenario’.10-20 years.3. A lot of material will never be eroded, as the decay curves approach a set minimum fillpercentage asymptotically. The volume of material left, represents sediment out of reachof the stream, between 2% and 40% of the initial volume in this simulation).Plots of sediment yield and reservoir evacuation over time for all hydrologic conditionsare presented in Appendix B.10.Tributary sediment input Sediment input from each of the tributaries is calculated using thebedload sediment rating curves given in Appendices B.4 and B.5, and is introduced at thenode closest to the confluence. Because only relatively large flood events are simulated (flowsgreater than 3 m3s−1 at the Robles del Rio gage), episodic bedload rating curves are used tocalculate tributary sediment inputs.683.5. Results3.5 ResultsWe focus our review and comparison on those results which occurred in the wet hydrographsimulations. These simulations represent more frequent and larger floods for the first 10 sim-ulation years than estimated for 90% of the historical record (see Section 3.4.3). These wetconditions are most relevant to project planning, as they represent a worst-case in terms ofsediment transport activity. The associated comparative plots for the dry and average hydro-logic conditions are described in detail in Appendix B.11 and overview figures are presentedin Appendix B.12. To further focus the discussion, we only present results of one sedimentsupply curve each for the ’Pulse supply’ and the ’Exponential decay’ scenarios (RC 4 andExp 2, respectively), which were designed to represent the conservative assumptions for thesemanagement options.We assess management scenario impacts on channel geomorphology through changes tobed elevation6, mean surface grain size Dg and the 90th-percentile fractional grain size D90from Los Padres Dam to the Pacific Ocean. In general, all four simulations lead to an increaseof channel bed elevations through net bedload deposition from the former San Clemente Damto the Pacific Ocean over the 60-year simulation time period (Figure 3.8). The most rapid ratesof deposition occur from years 1 through 10, and diminish thereafter through year 60, indicat-ing that the depositional rate is proportional to flood magnitude early in the simulation timeperiod. Upstream of the former San Clemente Dam, the primary response to resumption ofLos Padres (and upstream) sediment supply is net deposition. In contrast, the No Action sim-ulation results in up to two meters of further bed erosion up until the profile approaches theformer San Clemente reservoir site. Projected bed elevation responses over the entire modeldomain are associated with coarsening of the bed surface relative to initial conditions (Fig-ure 3.9 and Figure 3.10). Coarsening is most pronounced under the No Action simulation anddecreases with the additional sediment supplied from the Los Padres reservoir storage andupper contributing watershed. This outcome suggests that the gravel and coarser bedload6Elevation change is calculated by projecting the deposited sediment volumes equally across theaverage reach-based cross-sectional shape (see Section 3.6.1 and Appendix B.8). For the case of bederosion, we calculate elevation change based on a rectangular channel.693.5. ResultsFigure 3.8: Comparison of projected bed elevation change from the 2017 initialprofile for the wet hydrologic condition and the Historical, Pulse and Un-controlled sediment supply simulations. Shaded regions capture the 25th-75th percentile responses across the 100 simulations for the wet condition.Results shown for simulation year 1, 10, 30 and 60.content from the watershed area upstream of the former San Clemente Dam is large relative tothe fractional content of bedload supplied along the lower mainstem and is therefore impor-tant in setting the ultimate texture of the riverbed surface. Furthermore, the bedload supplysourced from the Los Padres reservoir storage and the upstream contributing watershed isimportant in terms of moderating the overall coarsening response. This result has clear im-plications for expectations around future steelhead habitat conditions related to actions at LosPadres Dam.703.5. ResultsFigure 3.9: Comparison of projected change of the geometric mean grain size ofthe bed surface Dg for the wet hydrologic condition and the Historical,Pulse and Uncontrolled sediment supply simulations. Shaded regions cap-ture the 25th-75th percentile responses across the 100 simulations for thewet condition. Results shown for simulation year 1, 10, 30 and 60.713.5. ResultsFigure 3.10: Comparison of projected change of the geometric mean grain sizeof the bed surface D90 for the wet hydrologic condition and the Histori-cal, Pulse and Uncontrolled sediment supply simulations. Shaded regionscapture the 25th-75th percentile responses across the 100 simulations forthe wet condition. Results shown for simulation year 1, 10, 30 and 60.723.5. ResultsThe range of depositional depths between the 25th and 75th percentiles is characteristicallylarge for year 1, decreases at most locations by year 10, and continues to decrease through years30 and 60 from the Narrows to the Pacific Ocean (Figure 3.8). Reduction in the projected rangeof depositional depths through time from Los Padres Dam to the Pacific Ocean suggests that:• The net effect of the 100 random wet hydrographs diminishes in time; and• Projected bed elevations across the 100 random wet hydrographs evolve to a narrow setof response trajectories spatially, regardless of the supply simulation.These two results imply a reasonable degree of confidence for the spatial and temporal trendsof projected bed elevation change under the four different sediment supply simulations. How-ever, the magnitudes of projected bed elevation change in particular are limited by (a) the basicmethod used to translate projected channel bed volume changes to bed elevations, (b) the 1-DBESMo model build, and (c) the available topographic data. We provide more discussion re-lated to limitations (b) in Section 3.6.1.There is one notable exception to the response generalizations made in the previous para-graphs. The evolution of projected bed elevations within the vicinity of the Tularcitos Creekconfluence is indicated by ranges of elevation values between the 25th and 75th percentilesthat are larger than the median values. This result suggests that projected average bed el-evations along the mainstem Carmel River around Tularcitos Creek are sensitive to the se-quence of future floods under the simulated wet conditions. Consideration of this resultwith projections for the evolution of the D90 grain size provides some insight for the cou-pled bed elevation-grain size response around the Tularcitos Creek confluence (Figure 3.8 andFigure 3.10). The magnitude of deposition at the former San Clemente Dam sets the averagelongitudinal bed slope along the mainstem Carmel River, which in turn leads into the Tu-larcitos Creek confluence reach. Larger magnitudes of deposition lead to larger average bedslopes, and hence higher bedload transport rates. Higher bedload transport rates drive thedownstream advance of a relatively coarse D90 grain size response, which makes it as far asthe Tularcitos Creek confluence (Figure 3.10). The downstream shape of the D90 grain sizeresponse indicates that the coarse fraction of the bed surface has a relatively large amount of733.5. Resultsspatial variability in the vicinity of Tularcitos Creek, with projected spatial changes of severalhundred millimetres over about 1.5 km of river length. However, variation in the downstreamextent of the D90 grain size response for any given supply simulation upstream of TularcitosCreek spans about 1 km or more of river length. The spatial domain of this D90 grain sizevariation correlates with the upstream zone of relatively large variations in the bed elevationprojections. This implies that the range of bed elevation responses in the vicinity of TularcitosCreek is conditioned by the magnitude of bedload deposition at the former San Clemente Damduring the first 10 projection years, coupled with the particular and associated downstreamadvance of coarse grain size fractions, which serves to limit future bed elevation adjustmentsas the simulations proceed beyond year 10. This highlights that field-based monitoring of bedelevation and bed surface texture response upstream of Tularcitos Creek may provide the dataneeded to make informed decisions regarding the likely trajectory of channel responses at thislocation.3.5.1 Overview of bed elevation changes along the channelProjected bedload sediment deposition magnitude differs by location and across the four sed-iment supply simulations. Deposition is greatest just downstream of Los Padres Dam of thethree Los Padres supply simulations, ranging from 5-6 m with the Uncontrolled simulationprojecting the most deposition and the Historical Supply simulation projecting the smallestamount (Figure 3.8). The projected magnitude of deposition just downstream of Los PadresDam is plausible because topographic data local to the Dam suggests about 6 m of net bed el-evation decline since the Dam was constructed in 1949 (Appendix B.11.2). By contrast, the NoAction simulation projects 0.3-0.6 m of additional bed elevation decreases near the Dam overthe 60-year model time period. It is thus expected that simulations which resume downstreambedload supply from areas upstream of Los Padres Dam (i.e. sediment stored in the reservoiror from the upstream contributing basin) would lead to recovery of local average channel bedelevations because steeper bed slopes are needed to facilitate transport of increased bedloadsupply.The magnitude of deposition varies spatially, with an upper bound given by the Uncon-743.5. Resultstrolled Supply simulation and a lower bound given with the No Action simulation, whicheven shows net bed erosion of up to 1.5 m. Projected deposition reaches between 1.5 and2.4 m as the former San Clemente reservoir area is approached. This magnitude of depositionseems reasonable given that the former reservoir pool drove upstream sediment deposition,which led to significant average longitudinal bed slope reduction close to the former reservoirpool. The Carmel River Reroute and Dam Removal (CRRDR) project locked in the effect ofthe reservoir pool into the post-construction local channel profile (San Clemente Creek fixednode; Figure 3.2). As a result, steeper bed slopes leading into the former reservoir pool areaare required in order to transport the increased upstream bedload supply, particularly for thethree supply simulations which pass bedload sediment downstream of Los Padres Dam.The CRRDR project also introduced segments of bed slope through the former San ClementeCreek arm of the San Clemente Dam reservoir pool which were considerably different fromand larger than those bedslopes downstream of the former dam site. As a result, solutions ofthe BESMo model for the wet hydrologic conditions lead to deposition magnitudes between2.1 and 3.7 m at the former San Clemente Dam site. This magnitude of projected depositionand the downstream advance of deposition beyond the former dam site acts to smooth outabrupt profile changes and facilitate bedload transport rates which converge to similar magni-tudes across locations of profile change. As discussed above, this model result is relevant forthe Tularcitos Creek confluence area because steepening of the bed profile downstream of theformer San Clemente Dam leads to a range of possible average bed elevation responses in thisarea.Downstream of Hitchcock Creek and as far downstream as beyond Garza Creek (station18.3 km), all four supply simulations are projected to yield little bed elevation changes andfluctuate around the initial bed elevation. This suggests that this roughly 4.6 km of mainstemCarmel River is already adjusted to a bedload transport capacity (i.e. under 2017 profile andtexture conditions) that can accommodate upstream supply increases across a majority of thegrain size classes. This relatively high transport capacity lessens moving toward The Narrowsat station 13.7 km. Approaching The Narrows, all four upstream supply simulations projectroughly 1.5 m of deposition by year 10, and this magnitude of deposition continues through753.5. Resultsyear 60 of the simulations. The deposition response at The Narrows may be attributed tothe lateral confinement of the channel, such that bed elevation is the primary mechanism tofacilitate transport of increased bedload supply. Just downstream of The Narrows, the NoAction simulation exhibits between 0 and 1 m of net deposition, whereas the three scenarioswith higher sediment supply converge to an approximate net deposition magnitude of 1.2 m.Moving farther downstream toward the mouth at the Pacific Ocean, the four supply simu-lations diverge slightly, but yield between 1.2 and 1.8 m of net deposition at year 60. Notablyhowever, a majority of the projected deposition along the lowermost 11 km of the mainstemCarmel River occurs by year 10 of the simulations. Therefore, resuming bedload supply fromcontributing areas upstream of the former San Clemente Dam and Los Padres Dam coupledwith relatively large floods early within the simulations drives rapid movement of bedloadthrough the system to the lowermost mainstem reaches, resulting in deposition there. Underthe average and dry hydrologic conditions, the downstream delivery of bedload supply to thelowermost mainstem Carmel River is significantly delayed compared to the wet conditions,with the overall depositional pattern present by simulation year 30 (as shown by figures inAppendix B.12).3.5.2 Overview of surface grain size changes along the channelUnder the wet hydrologic conditions, all four sediment supply simulations project increasesto the Dg and D90 grain sizes from Los Padres Dam to the Pacific Ocean by year 10 of thesimulations (Figure 3.9 and Figure 3.10). Coarsening over time is due to high relative transportrates of the finer size classes within the supply and bed surface distributions. Within thisoverall coarsening trend, however, maintenance of fining of the bed surface is projected as apossibility. At the end of simulation year 1, some of the wet hydrographs for the Los Padressupply simulations result in no or little change in Dg and D90 conditions compared to theinitial bed surface, except through the CRRDR project reach. Directly below the Los Padresdam, the Uncontrolled Supply scenario shows a potential fining of the bed reaching about12 km downstream. This may be attributed to the relatively large supply of fine material fromdam removal. By year 5 (not shown), bed surface texture conditions begin evolving to coarser763.5. Resultsstates. We observe a similar fining response for average and dry hydrologic conditions (figuresprovided in Appendix B.12). Persistence of the grain size fining responses ranges from 1 to 60years, depending on location, and is modulated by the co-evolution of the longitudinal bedprofile and bed surface texture. When the time scale of bed surface texture evolution occursat rates comparable to bed profile adjustments, the bed texture can maintain a finer texturecompared to conditions which drive the profile to adjust more rapidly. Under more rapidprofile adjustment, texture change cannot keep pace and is subsequently reset by youngerepisodes of deposition and sediment sorting.The lowermost 10.7 km of mainstem Carmel River follow the more rapid topographicprofile evolution trajectory across all hydrologic conditions, and as a result end up gener-ally more coarse than initial conditions. Moving upstream, texture conditions for the reachbetween Tularcitos and Las Garzas Creek typically trends to the initial texture states, with atendency to smooth out spatial jumps in the initial bed texture configuration. At the formerSan Clemente Dam and downstream of Los Padres Dam, bed texture conditions evolve to sig-nificantly coarser conditions across all three hydrologic and sediment supply simulations. Acoarser texture represents a coupled response with relatively large depths of sediment deposi-tion at both locations. In other words, steeper bed slopes are generally maintained by coarserbed surfaces.3.5.3 Comparison to MEI simulationsMEI (2002) conducted simulations to inform about effects of the San Clemente Dam removaland also reported early fining in response to the release of bedload sediment stored within theSan Clemente reservoir pool, followed by recovery toward, and in some cases coarser thaninitial bed surface grain size conditions. In particular, MEI reported that early simulationsperiods characterized by relatively dry conditions show the strongest downstream fining re-sponse.Figure 3.11 illustrates a comparison of longitudinal bed elevation profiles reported by: (a)MEI (2002) for the 1985-750C simulation, and (b) the present study for the No Action simula-tion at Los Padres Dam. The comparison is generally favourable for two specific reaches. First,773.5. ResultsFigure 3.11: Comparison of projected bed elevation change simulated by BESMoand reported by MEI (2002) for the roughly equivalent condition of re-moving San Clemente Dam and no bedload bypass at Los Padres Dam.The MEI (2002) condition represents their 1985-750C simulation results(data from Table 7.8 therein), which are plotted at the mid-point withinthe study sub-reaches. The 1985-750C simulation represents a reasonablebasis of comparison to the No Action simulation at Los Padres Dam un-der the wet hydrologic condition because both simulations have elevatedrates of sediment supply to the mainstem Carmel River downstream of theformer San Clemente Dam (see Appendix D.23 of MEI, 2002)both studies suggest a tendency for little to no net bed elevation change within the vicinity ofthe Tularcitos Creek confluence. Second, both studies suggest a net depositional condition forthe lowermost 15 km of the mainstem, extending roughly from the Narrows to the Highway1 bridge. Compared to the HEC-6T results, BESMo generally projects more deposition alongthe lowermost 11 km of the mainstem. Depositional differences between the two projectionsalong the lower mainstem range from 0.3 to 0.8 m, excluding conditions at the downstreamboundary. Although there is general consistency between the models in the vicinity of the Tu-larcitos Creek confluence, the BESMo model projects up to a meter of bed elevation responsevariability between the 25th and 75th percentile values.The projected profile consistency between the four different upstream supply simulationsas well as that reported by MEI (2002) highlights that deposition should be anticipated andcould range upwards of 2 m on average 60 years into the future. The timing of the depositionalsignal depends on the sequence and magnitude of floods (Figure 3.8). The variability projectedby BESMo is due to model sensitivity related to the timing and sequencing of future floods, aswell as the associated sorting of bed surface sediments which tends to reinforce the persistenceof early bed slope responses. Flooding observed during the winter of 2017 along Paso Honda783.6. DiscussionRoad (Monterey Herald, January 9, 2017) suggests a depositional trajectory may be evolvingdownstream of the Tularcitos Creek confluence given the relatively large magnitude of floodsthat occurred from January-February 2017 (Harrison et al., 2018).3.6 DiscussionIn this section, we first discuss BESMo model limitations and how these might affect the sim-ulation results. This is followed by an interpretation of the results in the Carmel River context.We then compare our results to other dam removal projects and a give a reflection on the useof BESMo for this task.3.6.1 LimitationsEvaluation of BESMo results must be understood through model construction limitations re-garding the fixed channel geometry, the bed elevation change calculation, and the accuracy ofinput data for the model.Fixed channel geometry BESMo does not account for the partitioning of streamflows betweenthe main channel and adjacent floodplain areas. This limitation was addressed by using fieldobservations of discharge and average flow velocity to capture the effect that increasing cross-sectional flow area has on the structure of flow velocity within the main channel. This isimportant because flow velocity is a key parameter used to estimate the rate of bedload trans-port. BESMo also does not account for lateral channel migration, nor widening of the channelat any model node. Both limitations are moderated to some degree by the common occurrenceof channel bank protection as well as bedrock along the mainstem Carmel River from CarmelValley Village to the mouth at the Pacific Ocean. The primary challenge that this introducesrelated to projection of channel conditions is that a lack of channel migration or wideningmeans that local sources of sediment are not represented in the model. This introduces un-known short-term effects into model projections roughly at a 5-year time scale. Since effectswill be local in spatial scale, likely at the level of one to two model nodes, we do not expectlocal widening effects to change the larger-scale spatial trends of the results reported here.793.6. DiscussionAdditionally, due to the 1-D construction of BESMo, model results do not provide reliableprojections of how flooding conditions may change in association with any particular set ofresults. Flooding could be evaluated with model runs within HEC-RAS using projected bedelevation and surface texture conditions generated by BESMo.The enhanced deposition projected by BESMo in comparison with the results presented byMEI (2002) (for the Narrows to Highway 1) might also be attributed to the 1-D model construc-tion, specifically as the model does not directly represent overbank flows. The 1-D model buildof BESMo means that cross-sectionally averaged downstream velocities at sediment transport-ing flows are larger compared to MEI (2002). A higher average velocity would lead to com-paratively lower bed elevations due to increased bedload transport capacities. However, itappears the BESMo No Action simulation evolves to bed surface grain sizes which may beseveral factors larger than those for the MEI 1985-750C simulations. Larger grain sizes willlead to comparatively lower average velocities due to higher particle drag, which in generalwill promote deposition of the larger grain sizes in transport. As a result, the larger mag-nitude of projected deposition simulated by BESMo may be due to at least two contributingand coupled factors related to the 1-D model build, but other factors related to differencesin model builds may also be important. Nevertheless, in the context of the lower mainstemCarmel River and the present study, we suggest that BESMo may over-project the magnitudeof deposition by approximately 0.3 m.With respect to the present discussion, it is pertinent to ask the following question: Howdoes the BESMo model build introduce uncertainty with respect to drawing conclusions re-garding the nature of future channel adjustments under the four different sediment supplysimulations? This question is particularly relevant within the context of understanding theBESMo results with respect to considerations specific to the Carmel River watershed.Elevation change calculations The fixed channel width in BESMo further means that any depo-sition (or erosion from the channel bed) of sediment leads to a linear increase (or decrease) inchannel elevation at each modelling node. Because our focus is to project the most probableriver profile response over the simulation period given the model configuration and set-up803.6. Discussionand input data relative to a specific response for any particular year, this simplification hasonly a small impact on the simulated sediment budget. It does, however, lead to unrealisticincreases or decreases in channel elevations if the stored sediment volume changes. There-fore, we focus on interpreting the projected absolute channel storage changes at each modelnode, and secondarily discuss how projected storage changes may translate to actual changesof channel bed elevation. We approximate how storage may translate to elevation changes ateach model node by proportionally distributing storage change projections based on the av-erage reach-based cross-sectional shape (see Appendix B.8) for the case of deposition. For thecase of bed erosion, we attribute the volume change equally across the average reach-basedcross-sectional shape.Accuracy of input data Last, it is important to note that all results presented and discussedhere are a reflection of the Carmel River BESMo build for this project, along with the modelconfiguration, setup and input data. There are many uncertainties with regard to actual fieldconditions and how they are accounted for with the input data. First, channel profile databetween the former San Clemente Dam and Los Padres Dam is based on the USGS NationalElevation Dataset (1/9 arc-second) and, as a result, might not resolve the channel bottom ele-vation. Second, the present model has been developed with the best available grain size data.However, bed surface grain size census data is spatially limited with respect to the modeldomain, and subsurface data is largely lacking. Therefore, we recommend model results beinterpreted with respect to general spatial trends across the simulations, as opposed to resultsat a particular location and point in time. Third, sediment transport rating curves were de-veloped using data largely collected in the 1980s and do not reflect recent shifts in sedimentsupply.With these limitations in mind, results from modelling of the four different sediment sup-ply simulations show clear spatial trends. Temporal trends, on the other hand, are directlyrelated to the timing and magnitude of larger floods within the 60-year simulation time pe-riod. Consistency of spatial trends between the four supply simulations suggests that resultspresented here can be used to plan for expected outcomes related to sediment management813.6. Discussionactions at Los Padres Dam.3.6.2 Placing simulation results within the Carmel River contextPrior to 2015 and the removal of the San Clemente Dam, channel bed elevation and bed sur-face grain size conditions along the mainstem Carmel River were governed by the combinedeffects of: (1) constructing the San Clemente Dam (1921) and Los Padres Dam (1949), whichled to a significant reduction in the supply of bedload sediment from the upper watershed tothe downstream mainstem Carmel River; (2) in-stream gravel mining along the middle main-stem in the 1950s and 60s, which amplified the effects from dam construction because bedloadavailable for downstream transport was further reduced; and (3) channel bank armouringalong many reaches of the mainstream Carmel River along the lowermost 24.5 km of river,which decreased bedload supply available from lateral channel migration (Paola et al., 1999) orcross-section enlargement as a result of bank erosion.The reduction to bedload supply since 1921 has led to widespread lowering of river bedelevations to varying magnitudes from Los Padres Dam to the Pacific Ocean as well as a gen-eral coarsening of the bed surface over the same reach (Kondolf , 1982; GMA, 2008). Giventhese past actions in the watershed, we expect that any resumption of bedload supply fromthe upper watershed will drive increasing average bed elevations over time, and possibly a re-duction in the bed surface coarseness, depending on the grain size distribution of the supplyand the riparian vegetation conditions (Kondolf and Curry, 1986). However, river flows havechanged since river flows were first known to have been diverted to support local agriculturein 1771. This (Gudde, 2010) means that the response of the Carmel River mainstem to actionstaken at Los Padres Dam today or in the near future will not necessarily occur in a way thatdrives river conditions to pre-existing states. The mainstem river is in some ways irrevocablychanged, and therefore the present simulations can help to build understanding of how themainstem Carmel River may respond to bedload-focused actions at Los Padres.The three supply simulations which pass bedload to the mainstem Carmel River down-stream of Los Padres Dam (Historical, Pulse, and Uncontrolled supply) show a consistentbed elevation response from Hitchcock Creek to the mouth at the Pacific Ocean. This find-823.6. Discussioning mostly holds across dry, average and wet hydrologic conditions and at year 60 there isclear trend of between 1.2 to 2 m of net sediment deposition along the lowermost 9 km ofthe mainstem, with a peak in net deposition of 1.5 m just upstream of The Narrows. Whilethis finding fits with the expectation of bed aggradation following the long period of lowersediment supply, it represents an unquantified risk of increased flooding. We recommendthat future studies should carefully evaluate this risk using results reported herein. Interest-ingly, a net depositional response in these locations also brings potential benefit to channelmorphology and natural riverine function because rising bed elevations will more frequentlyactivate side and alternate channels and will lead to natural construction of in-channel habi-tat elements and features. The potential benefits will be locally and randomly accentuated asrising bed elevations will also lead to a temporal spike in wood contributions from channelbanks due to increased mortality with a rising riparian water table. Along developed rivercorridors it is common for potential negative impacts to be mirrored by potential positive im-pacts. Going forward this counterpoint should be carefully evaluated with respect to local andfeasible mitigating actions that can minimize or otherwise remove the expected risk. Similarconsiderations should be given to the mainstem Carmel River downstream of the TularcitosCreek confluence through Carmel Valley Village, where projected conditions are particularlysensitive to the timing and sequencing of future large floods.All four sediment supply simulations suggest further evolution of conditions through theCarmel River Reroute and Dam Removal project reach. The primary projected response is awidespread increase in average bed elevations. Bed surface grain sizes are also projected toshow a strong coarsening trend. The three sediment supply simulations which pass sedimentdownstream of Los Padres Dam are projected to drive significant local bed elevation gains,ranging from near to 6 m at the Dam to about a meter downstream of the Cachagua Creekconfluence. Deposition of this magnitude will trigger a complete resetting of the river corridor.Corridor resetting at this level will also likely result in the delivery of significant quantities oflarge wood to the Carmel River Reroute and Dam Removal project reach, and possibly beyond.Wood delivery to the Dam removal project reach will likely benefit to physical habitat as woodcan anchor development of diverse channel patterns and local morphologic conditions, as833.6. Discussionwell as instream and overbank habitat elements and features. Potential benefits are likely tobe proportional to the magnitude of sediment supply at Los Padres Dam. Risks to furtherdownstream reaches are anticipated to be moderated by an intact riparian corridor betweenthe former San Clemente Dam site and the Tularcitos Creek confluence.3.6.3 Comparison to other dam removal projectsTo get a better idea about the relative geomorphic impact of the post-dam removal sedimentsupply, Major et al. (2017) suggest to use V∗, a ratio between the volume of stored sedimentto the background sediment load. To gauge the impact of geomorphic change in the first yearafter the dam removal, Grant and Lewis (2015) suggest to use E∗, which represents a ratio be-tween the volume eroded in the first year after removal and the background sediment flux.To use these parameters in context of our simulation results for the Carmel River, it is impor-tant to consider the relative impact of the Marbel Cone fire to a potential sediment release bydam removal, as it delivered nearly half of the stored sediment in the reservoir. If sedimentpulses the size of the Marbel Cone fire are used as examples of naturally occurring events, thenthe removal of LPD would cause a sediment pulse at maximum just twice the size of naturalevents.For the Los Padres reservoir, V∗ is 79.1 if the the Marbel Cone fire event is included in thenatural sediment transport conditions, and 167.7 if it is discounted. Major et al. (2017) describedam removals where V∗ > 20 as having a high geomorphic impact on downstream reaches,meaning that downstream channel adjustment to the event might take multiple years, po-tentially decades. Following this statement, a management option that restricts the sedimentvolume released by the dam removal to a maximum of 392,200 m3 would cause moderate ge-omorphic response, meaning that the channel morphology would potentially recover within afew years. In light of the fact that the Marbel Cone fire event alone of 727,800 m3 exceeds thismagnitude, the release of at least 50% of the currently stored reservoir sediment might cause ahigh geomorphic impact on Carmel River, but it would be within the magnitude of a naturallyoccurring sediment supply events.The value of E∗ is shown to be correlated well with the transport distance of coarse sed-843.6. DiscussionTable 3.8: Values of E∗, the ratio between sediment volume eroded in the first yearafter the management action begins with the first flood event in each timeseries, calculated from median sediment supply values after 1 simulationyear for bedload material.E∗Sediment supply scenario wet average dryPulse Supply RC 1 2.71 0.28 0.14RC 2 3.87 0.4 0.2RC 3 2.17 0.2 0.09RC 4 6.52 0.6 0.28RC 5 8.78 0.7 0.32RC 6 12.54 1 0.46Uncontrolled Sup-plyExp 1 12.2 11.45 8.43Exp 2 18.85 17.66 13.06Exp 3 20.25 18.97 14.03iment after the dam removal (Major et al., 2017). We present this value for the Pulse Supplyand Uncontrolled Supply scenarios in Table 3.8. Mechanically, the main difference betweenthe Pulse Supply and the Uncontrolled Supply scenarios is the dependence of sediment feedrates on the flow record. In the Pulse Supply management option, sediment is only releasedin flow events larger than 3 m3s−1, which leads to low supply in the first year for the aver-age and dry hydologic scenarios. On the other hand, the Uncontrolled Supply scenario showshigher values of E∗ even in the average and dry hydrologic scenarios, as the sediment feedrate is dependent on the cumulative discharge through the reservoir. These calculations showthat the sediment release from LPD would have the largest impact on downstream channelgeomorphology in the Uncontrolled Supply scenarios, matching the spatial trends shown foryear 1 in Figures 3.8 to 3.10. However, Figure 3.10 shows that the Pulse Supply scenario RC 4has a farther-reaching impact on the surface grain size D90 than the Uncontrolled Supply sce-nario Exp 2, indicating that while E∗ might be a good indicator for geomorphic effects of adam removal in the short term, other effects than the magnitude of introduced sediment willdominate the geomorphic response on longer timescales.The values of V∗ and E∗ help to place the potential LPD sediment release scenarios within853.6. Discussionthe context of estimates reported by other studies and summarized by Major et al. (2017). Boththe initial sediment reservoir volume and the ratio V∗ for LPD are comparable to Condit dam.The Uncontrolled Supply simulations all show at least 10 times lower values of E∗ for the wethydrographs than the data reported for Condit dam, attributable to its much finer and moreeasily erodible sediment Wilcock (2001). Values for both E∗ and V∗ are comparable to the GlinesCanyon removal, which was about 10 times larger in initial sediment volume (16 Mio m3) andwas removed in several stages East et al. (2015).An interesting observation from our simulations is that the re-establishment of historicalsupply conditions without additional release of reservoir sediments already causes an aggra-dational response and a significant change in surface grain sizes of the mainstem Carmel River.It seems that the magnitude of the short-term release of reservoir sediments has a compara-bly small effect in comparison to the general addition of fine sediments from the upper CarmelRiver and its tributaries. Over the 60 year timeframe, the style of dam removal might be not begeomorphically significant. The establishment of connectivity between the upstream reachesand the main stem is what governs the system response between 10-60 years after the LosPadres dam is removed. This echoes the previously recognized importance of the upstreamareas for the supply of fine sediment on the channel morphology of the Carmel River (Kondolf ,1982).3.6.4 Reflection on the use of BESMoDespite the limitations discussed earlier, BESMo is an appropriate tool for the task of esti-mating the geomorphic adjustment to the presented management options on LPD, and offerssignificant advantages over prior modelling work. Previous modelling efforts generally un-derestimated the time it took for significant geomorphic change to occur after dam removals(Foley et al., 2017). As BESMo is based on similar transport relations as these models, we mightalso expect to underestimate the fluvial response time until geomorphic change occurs. Yet, asour focus is the long term adjustment of Carmel River and the relative impact of the four man-agement alternatives, we see this as a minor issue. Further, the employment of stochasticallybased hydrographs in BESMo is an important improvement over other models in regards of863.7. Concluding remarkspredicting response times and magnitudes, as we explicitly include the effects of early largefloods in the wet hydrograph scenarios.In addition to methodological limitations, a significant factor of uncertainty is due to thelimited availability of data with which to initialize the model. However, all models wouldinclude this limitation. BESMo was developed to explicitly cope with the variation of param-eters, which in this study was the uncertainty of future hydrology and the sediment supplyvariations from the dam management options. In our view we could further improve the pre-sented results by varying more input parameters. This would have provided better insightinto the effect of data limitations, and offered a way to give specific advice to planners aboutwhich parameters should be collected in the future to improve the understanding of dam re-moval effects on the Carmel River.3.7 Concluding remarksThis study was conducted to gain insight into potential effects of different management ac-tions taken on the Los Padres Dam and reservoir. It is yet to be decided if the dam will beremoved or if other actions on the dam and reservoir are more feasible. For our purposes, fourscenarios of sediment management were specified as: (1) ’No Action’ would be taken, whichwould lead to continued accumulation of sediment in the reservoir and no sediment supplyimmediately downstream of LPD; (2) ’Historical supply’ would be re-established, routing allmaterial entering the reservoir downstream of LPD, preventing further accumulation in thereservoir; (3) ’Pulse supply’ represents a controlled removal of LPD, flushing reservoir sed-iment downstream through a sluice gate which would be opened under certain flow condi-tions; and (4) ’Uncontrolled supply’ represents the full removal of LPD without any measuresundertaken to prevent rapid erosion of the reservoir deposits.The Carmel River watershed offers a complex setting for a dam removal in multiple as-pects: The climate is dry, but floods constituting large sediment transport events occur regu-larly. Additionally, the river was altered significantly first by the construction of two dams inthe 1920s and 1940s, and the removal of San Clemente Dam in 2015. The ensuing low sedi-ment availability and related decrease in fine material was further pronounced by efforts to873.7. Concluding remarksstabilize the banks in the lower river. Furthermore, fires historically supplied large volumesof sediment into the channel, with the exceptional Marble Cone fire in 1977 constituting about50% of all material stored behind LPD. This means that an uncontrolled release of all materialout of the reservoir would only constitute twice the volume of that natural event.In adapting BESMo to the Los Padres Dam and Reservoir Alternatives Study, consider-able effort was undertaken to build defensible model inputs. There is little available elevationand grain size data for the watershed upstream of the former San Clemente Dam in particu-lar. While the sediment supply conditions were developed based on the given managementoptions, the uncertainty in the occurrence of significant floods shortly after a potential damremoval necessitated an approach of creating synthetic hydrographs based on historical dis-charge records. This allowed us to explicitly test wet conditions with early floods, which haveshown to be associated with situations in which the sediment supply scenarios diverge mostsignificantly. In the longer time frame of 60 years, the supply scenarios converge to the con-ditions of the Historical Supply scenario, which most closely resembles natural conditions.However, the response of channel geomorphology to actions taken at the dam site can havefurther complex effects that are hard to predict, especially on a long time-scale (Foley et al.,2017).Our findings show that the removal of the dam can move the Carmel River sediment dy-namics closer to natural conditions, primarily due to the re-establishment of connectivity ofthe upper watershed to the main stem, causing a fining of the channel bed surface. In compar-ison to the use of BESMo to explore the effect of event frequency and magnitude on channelmorphology in the previous chapter, we found that the frequent pulses in the Pulse Supplyalternative are more effective in keeping the surface grain size distribution relatively fine formore than a decade. Given enough time for armouring to occur, for example after an uncon-trolled dam removal was reworked, the channel surface coarsens within a few years. Thesefindings are significant for the decision of which actions should be taken on LPD, as both pat-terns of channel aggradation/degradation and bed surface fining are important for both floodmitigation and the preservation of fish habitat.88Chapter 4Identifying surface grain sizedistributions from images usingcomputer vision and machine learningmethodsDuring the previous chapters, the modelling efforts with BESMo demonstrated the need forimproved input data of surface grain size distributions. Higher temporal resolution of surfacegrain size data would have allowed for a more detailed model calibration using flume exper-iments in Chapter 2. A better spatial coverage of initial surface grain size distributions forthe Carmel River would have reduced the uncertainty in the modelling of the dam removalalternatives in Chapter 3. This chapter presents newly developed methods that can overcomethese data limitations in future work. The chronologic nature of how this work was executedled to these methods being presented at the end of the thesis.4.1 SummaryInsufficient temporal and spatial coverage of grain size distributions in field data limits stud-ies of channel response to episodic events in estimating grain size distributions of sediment894.2. Introductionsupply, channel surface adjustment, and sediment mobility. This lack in data might be over-come by collecting images of channel surfaces and automatically classifying grain sizes withcomputer vision methods. Colour based semi-automatic methods to identify grain size dis-tributions from images are currently employed in the flume lab at UBC, but have shown tobe unreliable due to paint chipping off stones and irregular lighting conditions. Furthermore,changing conditions between experiments (such as used camera type) make it necessary torecalibrate colour thresholds for grain size class identification. This chapter presents threedifferent methods to identify grain size distributions from images of coloured grains. Firstly,an algorithm to identify individual stones is developed (referred to as StoneID), which canbe used to study spatial stone distributions in the bed and to calculate GSDs of images. Sec-ondly, the outputs of this algorithm are used to train a neural network to determine GSD(referred to as DistID). Thirdly, individual stone locations from StoneID are used to explorethe performance of the U-Net machine learning method to find stone locations (referred to asU-Net+StoneID).These methods were developed on two datasets of images showing grains coloured by sizeclass, with the objective of reducing the manual work needed for the analysis of both the loca-tions of coarse grains and grain size distributions. Machine learning methods are less sensitiveto variations in colouring of stones and lighting conditions when trained with datasets that re-flect this range in conditions and that are manually validated. Even though the used datasetswere not fully validated manually, the employed methods show good results in matching theavailable data. Especially the larger grain size classes such as D90 were estimated within 2 mmby the DistID and the StoneID methods. The U-Net model was successfully used to classifystone areas within the accuracy of the training data, which shows potential in automaticallyidentifying individual stones if combined with the StoneID method. More manually validatedtraining data is needed if these methods are to be applied to more datasets.4.2 IntroductionThe bed surface grain size distribution (GSD) in rivers is the most important factor used toestimate sediment mobility (Church and Ferguson, 2015). The distribution of grain sizes also904.2. Introductionprovides information about channel structuring (Church et al., 1998; Zimmermann et al., 2010)and sediment supply from upstream of the study reach, hillslopes, and tributaries (Rice andChurch, 1998), and is therefore an important factor in assessing aquatic habitat and channelmorphology (Kondolf and Wolman, 1993). Both numerical and analytical sediment transportmodels simplify GSDs with statistical parameters such as the median grain size (D50), or the90th percentile (D90). These parameters have been shown to be useful tools for describingfluvial processes such as the onset of meandering (MacKenzie and Eaton, 2017) or the devel-opment of channel bed armouring which reduces particle entrainment and transport (Hassanet al., 2006b).A popular method to collect GSDs in the field was suggested by Wolman (1954), whichincludes a grid-based sampling of individual stones by hand. To obtain a statistically satisfac-tory sample, approximately 400 stones must be individually measured. Detailed informationon fine grain sizes (< 8 mm) is not possible to obtain with this approach (Rice and Church, 1996,1998). A similar method of grid-bound sampling can be employed using images of channelbed surfaces. However, if stone sizes are determined from visual measurements, only thevisible surface axis can be estimated, which leads to error as stones may be partly obscureddue to imbrication (Church et al., 1987). Sthly et al. (2017) used the semi-automatic softwareBASEGrain (Detert and Weitbrecht, 2013) to measure the stone b-axis from images and com-pare them to manual measurements. They report a ratio between the manual b-axis lengthto the image-based b-axis length that is comparable to the ratio between the manual b-axislength and the sieve size. This means that both methods might yield similar results, but bothhave different systematic errors: the sieving-method measures a combination of b and c-axis,while the visual method does not capture the stone surface completely. It is likely that thesesystematic errors vary with particle shape.Another way of estimating grain size parameters in the field is by using surface roughnessderived from terrestrial laser scanning (Heritage and Milan, 2009). Recently, advances weremade in deriving grain size parameters from images taken with UAVs (Unmanned AerialVehicles) (Carrivick and Smith, 2018; Detert et al., 2018; Woodget et al., 2018).In a laboratory setting, stones can be coloured by their respective size fraction (termed ’Bed914.2. IntroductionOf Many Colours’, BOMC), which permits the inference of a sieve-measured stone size withthe visual occurrence of a stone. A BOMC allows for the visual identification of the flume bedsurface composition, for example by converting images to a visual point count (Wilcock andMcArdell, 1993).While these field and flume methods to acquire grain size distributions are reliable, theeffort required to collect statistically robust samples (or to colour all grains in an experiments)is great, especially when sediment is coarser than sand or fine gravel (Church and Kellerhals,1978). The analysis of experimental flume results and numerical modelling in Chapter 2 showsthat the response of a fluvial channel to sediment supply is highly dependent on the historyof surface grain size, and therefore many successive measurements are needed. Furthermore,the case study presented in Chapter 3 demonstrates that a lack of grain size data is a majorsource of uncertainty for predicting the response of rivers to changes in sediment supply. Itis therefore of great importance to develop methods which can be used to collect statisticallyrobust samples in both field and lab settings to accurately characterise bed surface sedimenttexture. Furthermore, in a laboratory setting it is important that these methods are rapid andnon-intrusive, so that grain sizes can be reliably measured during an experiment.There are a number of studies with the objective to automate grain size detection fromimages, following two main approaches: (1) use image segmentation to find and measureindividual stones; or (2) extract statistical image properties and find correlations to respectiveGSDs. Graham et al. (2005) developed an image segmentation based photo-sieving method forautomated stone detection and grain size estimates based on greyscale images. The methodthen segments individual stones by enhancing stone interstices and extracts stone surfaceswith a watershed-algorithm. The b-axis of stones are then measured by automatically fittingstone areas with ellipses. A similar approach is used in the free software BASEGrain, wheregrains are separated by two automatically determined thresholds in greyscale images and thenmeasured geometrically (Detert and Weitbrecht, 2013). Chung and Chang (2013) followed thesame approach, but use machine learning to find appropriate greyscale-thresholds. The biasof computer vision based image segmentation methods and a general discussion of limitationsof automated grain size detection is given by Graham et al. (2010). An example for the second924.2. Introductionclass of methods based on statistical image properties is the software Cobblecam developed byWarrick et al. (2009). Their approach uses correlations between spatial scales of image contrastto known GSDs, and is stated to be less sensitive to lighting conditions than methods thatsegment and measure individual stones. Buscombe (2013) suggests that linking these kindsof image statistics to GSDs can overcome some limitations of photo-sieving methods. Whenapplied to coloured sediments, all mentioned greyscale-based approaches are problematic, asthe contrast between dark shaded areas in grain interstices, and bright areas on top of grainsis weak.Convolutional Neural Networks (CNNs) are a class of machine learning methods that canbe a powerful tool for computer vision, especially for image classification (Witten et al., 2016).While classical computer vision algorithms rely on pre-defined filters to extract features ofimages, CNNs ’learn’ filters by matching input images to a classified output. The network ofneurons is flexible enough to by itself learn the importance of image properties such as spatialscales of image contrast as used in the autocorrelation method mentioned earlier.In this study, different machine learning algorithms are trained to detect grain size distri-butions from BOMC images that were collected during experiments at the Mountain ChannelHydraulic Experimental Laboratory at the University of British Columbia. Data from two ex-periments were used: (1) Yinlue Wang’s experiment (referred to as YW) was designed to studythe development of transverse ribs under different constant discharges; (2) Alex Mitchell’s ex-periment (referred to as AM) studied sediment transport conditions and morphology changesunder symmetrical hydrographs.In this chapter three methods for image based detection of grain size data are presentedand evaluated: Firstly, an algorithm to identify individual stones is developed (hereafter re-ferred to as StoneID), which can be used to study spatial stone distributions in the bed and tocalculate GSDs of images. Secondly, the outputs of this algorithm are used to train a neuralnetwork to determine GSD (hereafter referred to as DistID). Thirdly, individual stone locationsfrom StoneID are used to explore the performance of the U-Net machine learning method tofind stone locations (hereafter referred to as U-Net+StoneID). These methods were developedwith a large collection of bed surface images from flume experiments in combination with934.3. Methodsmanually acquired grain size distributions. The three approaches complement each other:StoneID is used to define both GSDs and individual stone locations, which can then be usedto generate data with which the machine learning methods can be trained reliably.4.3 MethodsAn overview of the methods used in this chapter is given in Table 4.1. While the fully manualgrid-based stone count was used for validation, the other approaches require different degreesof manual adjustment when applied to new datasets.4.3.1 Bed surface datasetsThe StoneID method was developed on nine BOMC images, each covering a 5 m long and0.5 m wide area of a flume. GSDs for the images were sampled manually from 100 pointsTable 4.1: Overview of the grain detection methods used in this chapter.Method Description Required manual work perdatasetGrid-basedstone countStones are counted by colouralong a spatial grid and an area-based grain size frequency is de-rived.Manual sampling of at least 100locations on a regular grid.StoneID Algorithm to identify individualstone locations and sizes fromcolour thresholded images.Definition of colour thresholds,min. and max. expected diame-ters and stone areas.DistID Machine learning method tomatch GSDs to images.None after training. Retrainingneeded if applied to images thatare not similar to the trainingdataset.U-Net Machine learning method to ex-tract coloured stone areas fromimages.Same as DistID.U-Net+StoneIDUse of StoneID on the stone ar-eas detected with U-Net.Same as DistID. Additionallymin. and max. expected diam-eters and stone areas.944.3. Methodswith a grid-based colour count. The dataset (referred to as YW) was collected by Yinlue Wang.To develop DistID, a dataset of 1522 images was used. These images were acquired duringeight flume experiments conducted by Alex Mitchell (dataset referred to as AM). Each imageshows a 320 mm by 320 mm section along the centre of a 1 m wide and 12 m long flume. Theimages were either 1 m apart covering 8 m of the flume length (448 images), or six images weretaken adjacent to each other covering a 2 m central section of the flume length (1074 images).The grain size mixtures used for these experiments are coloured by sieved grain size class inhalf-phi intervals.4.3.2 Identifying individual grains: StoneIDDue to the colouring of the stones, an estimate of the surface coverage of each grain size classcan be based on the digital colour values in the images. As each stone does not display uni-form colouring, this method requires the definition of pixel value ranges for each grain colourby which the image can be classified. Colour values are typically stored as RGB values, split-ting the data into bands of red (R), green (G), and blue (B) colour information. While RGBrepresents different colours (e.g. yellow) as combinations of these three individual channels,testing showed that the HSV representation showed better results for all colours except black.This is due to the clearer segmentation of colours in HSV, where the colour is represented ashue (H), the intensity of the colour as saturation (S), and the light-dark shading as value (V).The colour, mean size, and HSV threshold values for each class is listed in Table 4.2While the colour-based mapping of each pixel in the image to a grain size class might besufficient to get area estimates for each size class, errors are introduced through both colourchipping and colour variations e.g. from reflections. Further, an area estimate based on pixelcounts is not sufficient to identify stone locations. To overcome these limitations, additionalparameters are used to extract individual stone locations in images. The method was imple-mented in MATLAB. The algorithm iterates through each size class, starting with the largestsize class. The approach is similar to the one described by Graham et al. (2005), but seperatesthe stone identification by colour class. A flowchart of the method is given with Figure 4.1.While finding locations of individual stones with StoneID works automatically for a set954.3. Methods1. Filter noise from imageHSV colour thresholdsRepeat for each colourIndividual stone areasStone recognitionsFinal stone recognitionsOriginal image Filtered image2. Threshold colours by HSV or RGB3. Find stone boundariesfrom continuuous areas4. Dilate stone boundaries(closing half-moon shapes)5. Split merged stones withinverse watershed algor.Masked light green stones6. Merge stone recognitionswhere areas are too close to be individual stones7. Reject stone recognitionssmaller than the minimumexpected area8. Split stone recognitionslarger than the maximumexpected area9. Remove identified stoneareas from image,continue with next colourFigure 4.1: Identification of individual stones from colour thresholded imageswith StoneID. After the image data is prepared in steps 1 and 2, steps 3to 9 are repeated for each colour class. The example image is from the AMdataset and the steps are illustrated with results for the light green size class.964.3. MethodsTable 4.2: Overview of colour thresholds per size class in the experiments, somedeveloped from HSV data, some from RGB data. Each colour threshold hadto be reconfigured for each set of experiments due to changes in lighting andgrain size classes. AM images do not contain white stones.HSV values Size thresholdsColour Meansize(mm)Hue SaturationValue minArea(mm2)maxArea(mm2)minDiam.(mm)light blue AM 4.8 0.59-0.650.27-1.000.00-1.008 100 1light blue YW 4.8 0.60-0.740.00-1.000.00-1.008 100 1dark green AM 6.75 0.47-0.570.25-1.000.00-0.6515 200 2dark green YW 6.75 0.31-0.470.00-1.000.00-1.0015 200 2yellow AM 9.6 0.11-0.190.00-1.000.00-1.0022 350 3yellow YW 9.6 0.09-0.160.32-1.000.58-1.0022 350 3red AM 13.6 0.90-1.000.16-1.000.00-1.00100 1100 8red YW 13.6 0.98-0.020.28-1.000.49-1.00100 1100 8light green AM 27 0.24-0.470.00-1.000.00-1.00300 2100 15light green YW 27 0.18-0.310.00-1.000.46-1.00300 2100 15white YW 38.5 0.00-1.000.00-0.300.77-1.00600 4000 35RGB valuesRed Green Blueblack (wet)YW19 0-57 0-43 0-122 160 1600 10black (dry) YW 19 0-115 0-104 72-164 160 1600 10black (all) AM 19 0-72 0-73 0-156 160 1600 10of images, the colour value parameters have to be reconfigured for new lighting conditionsor changed camera properties, and new thresholds have to be found when new grain sizeclasses are introduced. Besides the identification of individual stones, this method permits an974.3. Methodsestimation of the GSD of an image based on the surface areas of each stone class.4.3.3 Matching GSDs to colour images: DistIDA Convolutional Neural Network (CNN) is a machine learning type which uses convolutionalfilters, computational operations iterating over an image. The input image undergoes manyconvolutional filter layers with a successive reduction in the filter extent, while the number offilter layers increases. The output of the convolutional part of the network is fed into a ’fullyconnected’ layer of neurons, which undergoes further reduction until the specified outputshape is reached (e.g. six grain size classes). The layers are in sequence connected by neurons,which have an input, an output and a weight which can be ’trained’. This training adjusts theweights of millions of neurons iteratively until the network ’learns’ to match the image inputto the output (in this case the GSD).The DistID model was developed in python with the machine learning library Keras. Boththe ResNet50 (He et al., 2016) and the VGG19 (Simonyan and Zisserman, 2014) architectures wereused, as these are reported to perform well in image classification tasks, for example in the ’Im-ageNet’ challenge, a benchmark dataset for visual recognition (Witten et al., 2016). ResNet50features more weights than VGG19, but is trained more quickly due to a ’shortcut’ connectionthat skips individually training multiple layers. Both the ResNet50 and VGG19 were initial-ized with weights that were trained for the ImageNet challenge. Even though the pre-trainedweights were developed for the task of identifying 1000 different objects in images, the weightsare expected to represent generally usable filters such as edge detectors. Testing confirmed thatusing a pre-trained model yields a better training accuracy and a lower mean squared errorthan a random model initialization (see Table 4.3). While propagating from an image input toa classified output, the layers of the neural network represent a decreasing amount of image-locational information (by systematically sub-sampling the layer, e.g. with max-pooling) andan increasing amount of information that has proven to be helpful to determine the trainedoutput. In the last ’fully-connected’ layers of both the ResNet50 and the VGG19 models, eachindividual neuron is connected with all neurons of the previous and the next layer. This meansthat there is no image-locational dependence of the neurons any more. This ’fully-connected’984.3. Methodspart of the network is easily fit to any output shape (e.g. the number of grain size classes) andis not initialized with pre-trained weights. The final ’full-connected’ configuration chosen wasfound by experimenting with the number, sizes, and activation functions of layers. A detailedlist of the model layers used for this study is given in Appendix C.The AM dataset was used to develop DistID. As the pre-trained versions of both ResNet50and VGG19 used input image shapes of 300 by 300 pixels in 3 channels, the images weremedian-filtered, rescaled to a resolution of 1 pixel per mm, and cropped to 300 by 300 pixels.The colour channels were converted to HSV values for most of the training runs (as specifiedin Table 4.3). The matching GSDs for each image in the AM dataset were generated withStoneID, while manual GSD were created for 9 images using a grid-based colour count with150 samples per image.The model was trained with 1,215 images and the performance tested with 307 images thatwere not used in the training steps. This assures that over-fitting the model with the trainingdata can be detected, where the model learns aspects specific to the training data, but notthe general task. The training data was augmented by flipping the images horizontally andvertically, artificially increasing the number of images for training by a factor of four.4.3.4 CNN to map grain areas to colour images: U-Net+StoneIDBesides the mapping of images to a limited number of classes, some machine learning meth-ods can be trained to achieve spatial predictions between input images and classified outputimages. One example is U-Net (Ronneberger et al., 2015), where the input image is mapped toa fully connected layer (similarly to VGG19), which is then mapped back to the spatial scaleof the input image. The training data for this method was generated with the StoneID methoddescribed earlier.The U-Net model was trained on a subset of the AM dataset of 500 images, split into 450training and 50 validation images. Due to the U-Net architecture requirements, the imageswere rescaled to 512 by 512 pixels. Details on the model configuration can be found in Ap-pendix C. The output of the U-Net model can be used both as input to the individual stoneidentification (step 3, Figure 4.1) in the StoneID method and to directly estimate GSDs from994.4. ResultsFigure 4.2: Using classified stone areas from U-Net with the StoneID methodand direct pixel count GSD prediction. The surface images show StoneID-generated training data and not U-Net generated output.pixel counts (see Figure 4.2). This combination of methods is in the following referred to asU-Net+StoneID.4.4 Results4.4.1 StoneIDAn example of the image input for StoneID is given with Figure 4.3a, showing part of animage in the YW dataset. Figure 4.3b shows the 2,426 stones which were identified in theircorresponding colour classes. The residual areas of the image where no stones were found isshown in Figure 4.3c. Over all nine images, a total of 72,209 stones were identified, on averagecovering 58.5% of the total image area. The undetected 41.5% in area are composed by a mixof (1) grain sizes that were too small to be included in the method, (2) grains that partly orfully lost colour, and (3) grains that were submerged in pools changing the apparent colour.The StoneID method was used to derive GSDs from the surface coverage of identifiedstones. The resulting distributions for sizes ≥ 5 mm are shown in Figure 4.4 with the corre-sponding manually validated GSDs. The StoneID GSDs match the manual data well for thecoarser fractions, which is illustrated by the mean of nine GSDs shown in Figure 4.5. The finerfractions are under represented, which is a limitation of this method.1004.4. ResultsTable 4.3: Difference in reported accuracy and loss from model configurations.All models were run in Keras on a Compute Canada cluster using a TeslaP100 with 12GB of memory. Training was done for 50 epochs, where oneepoch is an iteration through all training images. The most reliable accu-racy estimation is for the test data performance (bold), as good training dataperformance can be due to overfitting.Training data performance Test data performanceModeltypeInitialweightsColours Accuracy MeansquarederrorAccuracy MeansquarederrorVGG19 ImageNet HSV 0.87 0.0079 0.80 0.0115VGG19 ImageNet RGB 0.88 0.0076 0.79 0.0101ResNet50 ImageNet HSV 0.86 0.0085 0.67 0.0195ResNet50 ImageNet RGB 0.87 0.0080 0.66 0.0193ResNet50 random HSV 0.65 0.0220 0.58 0.0240A sample for the performance of StoneID with the AM dataset is shown in Figure 4.6,where panel (a) shows the original image, (b) shows the cropped stone locations, (c) shows theresidual area, and (d) shows the GSD derived from stone areas.4.4.2 DistIDThe AM dataset of 1,215 images was used to develop a neural network to learn the predictionof GSDs from images. The resulting performance reported by Keras is shown in Table 4.3for the models used, with details provided in Appendix C. All models were trained for 50epochs (full iterations through the dataset) to make performance comparable. The VGG19was most accurate (80%) with HSV colour images when using pre-trained ImageNet weights.The ResNet50 models all showed lower accuracy, but the computational time was about 40%lower.The accuracy describes the match of the model output GSD in relation to the input GSD,with discrepancies being measured with the mean squared error. Due to the large amount ofimages, the input GSDs were created with StoneID, which means that the model is not trainedagainst perfect data from the beginning. Nonetheless, the good performance of 80% of theVGG model indicates that image patterns were found that correlate well with the input GSDs1014.5. Discussionduring the training of the model.To test the performance in more detail, 9 images were manually validated with 150 grid-based colour counts each. The comparison of resulting GSDs from manual counts, StoneIDinput data, and resulting DistID are shown in Figure 4.7. Surprisingly, on this dataset DistIDperforms better than StoneID, with both methods being within 1 and 2 mm of the manual data(see Figure 4.8).4.4.3 U-Net+StoneIDA comparison of a StoneID classification with a prediction from U-Net is shown in Figure 4.9.Notably, this specific image was not part of the training dataset, meaning that the shownU-Net prediction (Figure 4.9d) was generated independently from the StoneID classification(Figure 4.9b). The U-Net model performs well in recreating the grain size class based areaestimates from StoneID. Differences between the predictions occur mainly on the borders ofthe grains, as illustrated by an image of the difference of predictions in Figure 4.9c. The U-Netmodel was not trained to separate grains directly, which is why the StoneID methods (fromstep 5 on, Figure 4.1) have to be employed on the U-Net output classes to separate mergedstones and extract individual stone locations, as outlined in Figure 4.2.4.5 DiscussionThe three developed methods (StoneID, DistID, and U-Net+StoneID) offer new ways of esti-mating GSDs from bed surface images of coloured stones. StoneID and U-Net+StoneID furthercan be used to locate individual stones and derive properties on the clustering of large stones,a descriptive parameter that is currently used in a study on the formation of transverse ribs influme experiments.StoneID works well for large grain sizes, matching the manually estimated D65, D85, andD90 in the YW dataset within 0.15 mm (see Figure 4.5), and in the AM dataset within 2 mm (seeFigure 4.8). For these comparisons the GSDs were created from material ≥ 5 mm, as the indi-vidual identification of grains was unsuccessful for smaller material. These small grain sizesoverlapped in colour space with the larger grains in red and black, preventing the lumping1024.5. Discussionof fines into one class. StoneID misses some of the larger stones that lost colour due to chip-ping or showed unexpected colour values due to changes in lighting conditions. Generally,the method shows more false negatives than false positives in the colour classification, whichintroduces a systematic error that is hard to quantify. Further, some stones are identified asbeing smaller than they actually are, sometimes splitting them in multiple identifications. Thisis a clear limitation of this method and some manual validation could be done.The DistID method was trained with GSDs generated with StoneID, meaning that the ac-curacy of the trained model depends on the accuracy of StoneID, which likely underestimatedthe frequency of smaller grain size classes. Due to the large amount of images, manual datawas not created for the full set of images. However, manual validation against nine samplesshows that the method performs reasonably well (see Figure 4.8). The VGG19 architectureused for DistID has the potential to be improved by customizing the network architectureand creating model ensembles, which will improve the predictive performance of the method.While the training of the current DistID implementation required a large amount of data, theretraining on other datasets can be quicker and less data intensive as the model already learnedthe basic relation of BOMC image data to GSDs. DistID can be expected to be advantageousover the other methods in situations where the bed surface contains material which is toosmall for a grid-based stone count (i.e. < 8 mm). In these cases, several bulk samples of sed-iment can be sieved to obtain the GSDs to be used in combination with surface images of thesampled bed surface areas as training data for DistID.The initial grain area estimates of StoneID can be done with a U-Net model instead, whichmight offer the advantage not to require the respecification of colour thresholds when themethod is applied to a different experimental setup. This advantage has to be further tested,as in this study U-Net+StoneID was only trained and applied to one dataset. The performanceof the U-Net based grain area classification can be improved greatly if a dataset of about 100images would be manually validated. This would allow to train the model without missingstone predictions in the training dataset.While all discussed methods were developed on coloured beds, they can be extended touncoloured beds. While the loss of colour information will reduce the accuracy of the meth-1034.6. Conclusionsods, the use of elevation data might offset this loss. Elevation data could not be utilized inthis study, as the available laser-scanner DEMs are of low resolution (2 mm) and contain manyartefacts. For all current and future experiments a newer method of generating DEMs is em-ployed, which uses overlapping high resolution images. These new DEMs can improve themachine learning methods by switching the third channel in the HSV colour space to containthe DEM values. This approach also can be used for real-time UAV-based grain size detection.4.6 ConclusionsMachine learning methods offer a great potential to automate data collection from bed surfaceimages. All methods perform reasonably well, particularly with respect to larger grain sizes.For some studies the inability to identify small grain sizes might be of less importance, asthese sizes are also missed during stone counts in the field (Rice and Church, 1996). In settingswhere the bed surface contains significant amounts of sand and fine gravel, DistID could betrained using GSDs derived from sieving, avoiding the problem of image based segmentationof small grains. The addition of training data is expected to enhance the performance of DistIDand U-Net+StoneID. This lack of data limited the results in the presented study.1044.6. ConclusionsFigure 4.3: (a) Median-filtered input image for the stone recognition withStoneID. (b) Resulting image cropped by areas where 2426 stones wereidentified. (c) Residual area where no stones were identified shown incolour with grey background.1054.6. Conclusions0 10 20 30 40grain size (mm)00.20.40.60.81cumulative frequencyD50: 11.4 mmD50: 14.1 mmD90: 24.5 mmD90: 25.3 mmManualStoneID0 10 20 30 40grain size (mm)00.20.40.60.81cumulative frequencyD50: 10.9 mmD50: 14.7 mmD90: 23.8 mmD90: 25.1 mmManualStoneID0 10 20 30 40grain size (mm)00.20.40.60.81cumulative frequencyD50: 14.3 mmD50: 15 mmD90: 24.9 mmD90: 25.1 mmManualStoneID0 10 20 30 40grain size (mm)00.20.40.60.81cumulative frequencyD50: 13.2 mmD50: 15.3 mmD90: 24.4 mmD90: 25.4 mmManualStoneID0 10 20 30 40grain size (mm)00.20.40.60.81cumulative frequencyD50: 13.4 mmD50: 15.7 mmD90: 25.1 mmD90: 25.5 mmManualStoneID0 10 20 30 40grain size (mm)00.20.40.60.81cumulative frequencyD50: 14.5 mmD50: 15.8 mmD90: 26.5 mmD90: 25.5 mmManualStoneID0 10 20 30 40grain size (mm)00.20.40.60.81cumulative frequencyD50: 15.7 mmD50: 16.2 mmD90: 26.1 mmD90: 25.5 mmManualStoneID0 10 20 30 40grain size (mm)00.20.40.60.81cumulative frequencyD50: 16.2 mmD50: 16.2 mmD90: 26 mmD90: 25.4 mmManualStoneID0 10 20 30 40grain size (mm)00.20.40.60.81cumulative frequencyD50: 15.6 mmD50: 16.2 mmD90: 25.7 mmD90: 25.5 mmManualStoneIDFigure 4.4: Comparison of StoneID derived GSDs to manual data. Coloured cir-cles show the colouring of the stones.-Figure 4.5: Comparison between mean sizes of D90, D85, D65, D50, D25, and D16from StoneID and manual data.1064.6. ConclusionsFigure 4.6: (a) Median-filtered input image for the stone recognition withStoneID. (b) Resulting image cropped by areas where stones were identi-fied. (c) Residual area where no stones were identified shown in colourwith grey background. (d) Derived GSD from stone area.1074.6. Conclusions0 5 10 15 20 25 30grain size (mm)00.20.40.60.81cumulative frequencyD50: 11 mmD50: 12.5 mmD50: 13.7 mmD90: 22 mmD90: 23.5 mmD90: 23.5 mmManualStoneIDDistID0 5 10 15 20 25 30grain size (mm)00.20.40.60.81cumulative frequencyD50: 8.7 mmD50: 12.5 mmD50: 12 mmD90: 21.3 mmD90: 23.5 mmD90: 21.7 mmManualStoneIDDistID0 5 10 15 20 25 30grain size (mm)00.20.40.60.81cumulative frequencyD50: 14.1 mmD50: 12.5 mmD50: 13 mmD90: 23.6 mmD90: 23.5 mmD90: 23.2 mmManualStoneIDDistID0 5 10 15 20 25 30grain size (mm)00.20.40.60.81cumulative frequencyD50: 9.4 mmD50: 12.5 mmD50: 11.8 mmD90: 22.1 mmD90: 23.5 mmD90: 22.4 mmManualStoneIDDistID0 5 10 15 20 25 30grain size (mm)00.20.40.60.81cumulative frequencyD50: 9.2 mmD50: 12.5 mmD50: 11.1 mmD90: 19.1 mmD90: 23.5 mmD90: 21.1 mmManualStoneIDDistID0 5 10 15 20 25 30grain size (mm)00.20.40.60.81cumulative frequencyD50: 12 mmD50: 12.5 mmD50: 12.8 mmD90: 23.4 mmD90: 23.5 mmD90: 23.5 mmManualStoneIDDistID0 5 10 15 20 25 30grain size (mm)00.20.40.60.81cumulative frequencyD50: 11.2 mmD50: 12.5 mmD50: 13.2 mmD90: 22.2 mmD90: 23.5 mmD90: 23.3 mmManualStoneIDDistID0 5 10 15 20 25 30grain size (mm)00.20.40.60.81cumulative frequencyD50: 9.1 mmD50: 12.5 mmD50: 12.2 mmD90: 20.9 mmD90: 23.5 mmD90: 21.9 mmManualStoneIDDistID0 5 10 15 20 25 30grain size (mm)00.20.40.60.81cumulative frequencyD50: 14.1 mmD50: 12.5 mmD50: 12.9 mmD90: 23.6 mmD90: 23.5 mmD90: 23.3 mmManualStoneIDDistIDFigure 4.7: Comparison of StoneID and DistID (VGG19, HSV) derived GSDs tomanual data.Figure 4.8: Comparison between mean sizes of D90, D85, D65, D50, D25, and D16from StoneID and DistID (VGG19,HSV) to data from manual grid-basedcounts.1084.6. ConclusionsFigure 4.9: Comparison of StoneID and U-Net derived grain area predictions fora 13 by 13 cm large image sub-sample. (a) Original image. (b) StoneIDgrain classifications based on colour values and expected grain geometry.(c) Image of difference between U-Net and StoneID classifications. (d) U-Net classifications coloured by the respective grain size classes. The U-Netoutput resolution is reduced by about 40% in comparison to the StoneIDmethod.109Chapter 5Concluding remarksThe objective of this thesis is to build our understanding of how mountain rivers respond toepisodic sediment supply. Field studies of channel response to these events are challenging toundertake, as a long data record is needed to reasonably assess a system’s state of responsein the context of episodic supply. A fluvial system might be undergoing transient adjustmentto previous, potentially unknown, episodic events, or it might be approaching a steady statewhere no significant adjustment occurs. Greater confidence in the observed state of responseof a system can be achieved with flume experiments where the fluvial response can be ob-served in detail after episodic events are introduced in a controlled fashion. Yet, the amountof work necessary to carry out these experiments is large, which limits the number of experi-mental conditions that can be studied, and thus their utility for addressing applied problemsof channel adjustment. To overcome this limitation, the 1-D morphometric sediment trans-port model BESMo was developed, which allows large numbers of simulations to be run inbatches, generating ensemble results. Chapter 2 presents the identification of thresholds inthe fluvial response to episodic sediment supply, including the application to a field case. Thegeneral utility of BESMo for studying fluvial response to large sediment supply events is pre-sented in Chapter 3 with the study of potential geomorphic effects following the removal of adam in the Carmel River. Finally, to overcome data limitations on surface grain size distribu-tions, machine-learning based methods were developed to detect grain size distributions fromimages, as described in Chapter 4.1105.1. Summary of contributionsA more detailed list of contributions made in this thesis is presented in the next section,followed by a discussion of potential avenues for future work.5.1 Summary of contributionsThe 1-D morphometric sediment transport model BESMo (Bedload Scenario Model) wasdeveloped and calibrated to recreate data from flume experiments.As presented in Chapter 2, BESMo was developed to simulate the effect that different com-binations of event frequencies, magnitudes, and grain size compositions have on a simplifiedfluvial channel. As described earlier, the study of episodic events is limited by the range ofconditions that can be recreated and understood in flume and field based studies. Thus thecalibration of BESMo with one set of flume experiments allowed for a broader range of sim-ulated conditions to be assessed (320 instead of 4, an increase by a factor of 80) and for theobservation time to be extended by a factor of 500 per simulation (20,000 hours instead of 40hours each).When a channel is impacted by multiple episodic supply events of different magnitudes,the sequencing of these events has only a short term effect on the slope and grain size re-sponse of the fluvial channel.BESMo was used to recreate channel adjustments observed during flume experiments, show-ing that if a channel is impacted by multiple episodic supply events of different magnitudes,the sequencing of these events has only a short term effect on the slope and grain size responseof the channel. If a large pulse occurs early in the time series, it causes an abrupt increase inchannel slope and the effect of subsequent smaller pulses is subdued. If a series of small pulsesoccurs first, the slope responds more gradually and the effect of a subsequent large pulse issubdued. Both of these hypothetical time series lead to similar channel conditions in the longterm, indicating that the total volume of sediment supply is the governing factor of channelequilibrium conditions. As the grain size response is quicker than the response in channelslope, the event sequencing has a small effect on the surface grain size distributions.The fluvial response to episodic sediment supply regimes is either constant-feed-like or1115.1. Summary of contributionspulse-dominated, which can be determined by ratios of the event return period (Tpp) to thefluvial evacuation time (Tfe) and the fluvial armouring time (Tfa).In a second set of simulations, the range of simulated pulse frequencies, magnitudes, andgrain size compositions were extended, representing different sediment supply regimes. Dur-ing the simulations these conditions caused the fluvial channel to adjust its slope, with somecombinations of input conditions resulting in similar slopes to the constant-feed simulations,while others led to significantly higher slopes. These two response types can be separated bya threshold ratio of the timescales of sediment supply to the timescale of the fluvial response.When the return period of sediment pulses (Tpp) is shorter than the fluvial evacuation time(Tfe), channel slopes are similar to constant-feed simulations. When Tpp > Tfe, channel slopesare significantly higher, indicating pulse-dominated conditions. A similar effect is observedfor the armouring ratio of the channel, which shows a threshold in armouring response whenTpp is longer than the fluvial armouring time (Tfa). These thresholds allow for the categoriza-tion of fluvial response to episodic sediment supply regimes into one of (a) constant-feed-like,or (b) pulse-dominated. This distinction is useful for predicting the long-term behaviour offluvial channels which experience irregular supply events.Chapter 2 included a sample of such an application for the case of East Creek, showingthat the threshold to a pulse-dominated regime lies at the fluvial evacuation time of roughly2.2 years. This creek probably receives sediment at a lower interval, which indicates a pulse-dominated regime. The armouring timescale lies around 3.5 years, indicating that if the long-term sediment supply was introduced over event frequencies between 2.2 and 3.5 years, itwould be removed most efficiently and result in a lower slope.BESMo is an effective tool for applied problems of channel adjustment, and was used topredict channel response to the removal of the Los Padres Dam, on the Carmel River, Cali-fornia. In general, all scenarios show that the release of sediment from the dam will causethe river to aggrade. The bed will fine during phases of sediment release, but coarsen withinseveral years after the reservoir is depleted of sediment.As presented in Chapter 3, BESMo was used to study the geomorphic response of the Carmel1125.1. Summary of contributionsRiver to the removal of the Los Padres Dam for 60 years in the future. The simulations indi-cate that increased sediment supply from the reservoir can return the Carmel River sedimentdynamics closer to pre-dam conditions, primarily due to the re-establishment of connectivitybetween the upper watershed and the main stem. Specifically, the predicted return to his-torical sediment connectivity will lead to channel bed aggradation and fining of the channelbed surface downstream of the dam. While these changes potentially increase fish habitat,widespread aggradation poses an increased flood risk.Different sediment supply scenarios were developed to represent management optionsconsidered for the dam. Simulations in which sediment was released from the Los Padresreservoir in smaller, more frequent pulses proved to be more effective for establishing a finersurface texture which could last more than a decade. On the other hand, in simulations wherethe sediment release was less frequent and larger in magnitude, for example following an un-controlled dam removal, the fine surface texture was only established intermittently. As smallgrain sizes were not replenished, their removal from the channel bed led to the developmentof channel armouring within a few years of low sediment supply.Ensemble simulations of sediment transport are a useful approach for understanding howuncertainty in future conditions affects the system in the long term (>10 years).Determining the response of the Carmel River to the Los Padres dam removal is a complexproblem, as both current conditions and future changes on the river are not fully understood.As detailed prediction of such a system is nearly impossible, the application of a 1-D modellike BESMo helps to explore the response of the river with simplified process interactions.Thus, the comparison between scenarios of sediment release from the Los Padres reservoiroffers insight into large scale patterns of river response which are generally more meaningfulfor managers than highly localized predictions.BESMo offers significant advantages over prior modelling work by allowing for the explo-ration and quantification of uncertainty in input parameters. The uncertainty in the occurrenceof significant floods shortly after a potential dam removal necessitated an approach of creatingsynthetic hydrographs based on historical discharge records. This allowed for the generation1135.1. Summary of contributionsof flow series representing historically ”wet” conditions, which have shown to be associatedwith situations where the sediment supply scenarios diverge most significantly. Includingthese uncertainties in the presentation of simulation data is of great practical importance forcommunicating complex modelling results. Despite the limitations of 1-D models, BESMohas shown to be an appropriate tool for understanding the future geomorphic response of theCarmel River to potential management options at Los Padres Dam.The detection of grain size distributions from images can be greatly improved with ma-chine learning-based models.The lack of grain size distributions in field data for future studies of channel response toepisodic events might be overcome by collecting drone based imagery. Currently employedsemi-automatic methods to identify grain size distributions from images in the lab are unreli-able due to paint chipping off stones and irregular lighting conditions. Furthermore, changingconditions between experiments (such as used camera type) make it necessary to recalibratecolour thresholds for grain size class identification.Three different methods to identify grain size distributions from images of coloured grainswere presented in Chapter 4. These methods were developed on two datasets of images show-ing grains coloured by size class, with the objective of reducing the manual work needed forthe analysis of both the locations of coarse grains and grain size distributions. Machine learn-ing methods can be less sensitive to variations in colouring of stones and lighting conditionswhen trained with datasets that reflect this range in conditions and that are manually vali-dated. Even though the used datasets were not fully validated, the employed methods showedgood results in matching the available data. Especially the larger grain size classes such asD90 were estimated within 2 mm by the DistID and the StoneID methods. The U-Net modelwas successfully used to classify stone areas within the accuracy of the training data, whichpromises great potential in automatically identifying individual stones if combined with theStoneID method. More manually validated training data is needed if these methods are to beapplied to more datasets.1145.2. Future work5.2 Future workNetwork-scale model: BESMo has proven useful for recreating sediment transport conditionsranging from those at the small scale (flume experiment) to the large scale (Carmel River catch-ment). In the latter case, only the main-stem of the river was simulated while tributary sedi-ment supply was derived from rating curves. In other settings, the explicit inclusion of tribu-taries in the simulation domain might be necessary to understand the network-scale impactsof episodic sediment supply in a catchment. BESMo could be used in such a study to investi-gate the geomorphic effects that changing episodic sediment supply frequency and magnitudehas in a channel network.Variation of hydrologic conditions in accordance to climate change scenarios: A wide range of syn-thetic hydrographs were generated for the Los Padres dam removal study. These dischargetime series were based on static climatic conditions of flood frequency and magnitude cap-tured by a single stream gauge. It is likely that climate change will impact flood frequency-magnitude relations, which could be represented in BESMo and used to understand the sensi-tivity of fluvial sediment transport to the impact of climate change.Further applications of BESMo: Studies simulating sediment transport are never based on per-fect input data, as field data characterising key processes is challenging, or in some cases, notpossible to collect in adequate detail or over appropriate timescales (e.g. subsurface grain sizesor flow conditions). As stated before, BESMo offers a way to approach the uncertainty of realworld systems by relying on ensembles of simulations.While any modelling study of bedload sediment transport could benefit of this approach,a possible application related to episodic sediment supply is the study of appropriate gravelaugmentation magnitudes and frequencies. For example, simulations of this type could helpfinding which sediment supply conditions are most effective, for improveing fish habitat whilereducing problems with ’overloading’ the channel with sediment and causing aggradation.115Application of developed time scales: The application of the fluvial evacuation time (Tfe) andthe fluvial armouring time (Tfa) to more field cases is necessary to better understand the effectof episodic sediment supply on mountain streams. Such studies could investigate morpho-logic change at tributary junctions where sediment is supplied to the main channel from de-bris flows. An important aspect of a suitable field site would be the availability of records ofepisodic supply from which frequency-magnitude relations could be derived.Improvement of machine learning based grain size identification More high quality validation datawill help to improve the developed machine learning methods significantly, as these modelsare primarily limited by the quality of the input training data. This method should also be ap-plicable for unpainted, or natural sediment: While the loss of colour information will reducethe accuracy of the models, incorporating other co-variables, such as elevation data, mightoffset this loss. Elevation data could not be utilized in this thesis, as the available laser-scannerDEMs are of low resolution (2 mm per pixel) in comparison to the image data and containsmany artefacts. For all current and future experiments a newer method of generating DEMsis employed, which uses overlapping high resolution images. These new DEMs can improvethe machine learning methods by switching the third channel in the HSV colour space to con-tain the DEM values. Preserving three channels is preferable, as this assures that pre-trainedmodels can be used, which are shown to require less training data as some aspects of objectdetection are already established. Alternatively, a CNN (e.g. VGG19 or ResNet50) could befully retrained on 4 channels (3 colour channel channels plus DEM values).Automated grain size estimates from remotely sensed images: The stone identification from un-coloured sediment will open up the possibility to apply the machine learning methods onimages captured remotely, such as with UAVs (Unmanned Aerial Vehicles), allowing for rapidgrain size scanning of large areas. This can be achieved after collecting new validation data,as the method can not be directly translated from the flume setting due to differences in e.g.image resolution, DEM resolution, inconsistent lighting, and stone geometry. A reliable UAV-based method to identify grain size distribution can be tied back to the numerical modelling116work with BESMo, as this would help to generate larger input datasets in a fast and cost effec-tive way.117BibliographyAECOM (2017), Los Padres Dam and Reservoir Alternatives and Sediment ManagementStudy Draft Preparation Technical Memorandum-Deliverable for Task 1. Consulting reportprepared for Monterey Peninsula Water Management District in association with HDR En-gineering and Stillwater Sciences.Ahnert, F. (1994), Equilibrium, scale and inheritance in geomorphology, Geomorphology, 11(2),125–140, doi:10.1016/0169-555x(94)90077-9.An, C., Y. Cui, X. Fu, and G. Parker (2017a), Gravel-bed river evolution in earthquake-proneregions subject to cycled hydrographs and repeated sediment pulses, Earth Surface Processesand Landforms, 42(14), 2426–2438, doi:10.1002/esp.4195.An, C., X. Fu, G. Wang, and G. Parker (2017b), Effect of grain sorting on gravel bed riverevolution subject to cycled hydrographs: Bed load sheets and breakdown of the hydrographboundary layer, Journal of Geophysical Research: Earth Surface, 122(8), 1513–1533, doi:10.1002/2016jf003994.Ayers, W. Y. (1995), Water Resources Data, California, Water Year 1994. Volume 2. Pacific SlopeBasins from Arroyo Grande to Oregon State Line except Central Valley., Tech. rep., doi:10.3133/wdrca942.Balance Hydrologics, Inc. (2001), Unpublished Sediment Transport Measurements, CarmelRiver, Water Year 2001.Bednarek, A. T. (2001), Undamming rivers: a review of the ecological impacts of dam removal,Environmental management, 27(6), 803–814.Benda, L. (1990), The influence of debris flows on channels and valley floors in the oregoncoast range, usa, Earth Surface Processes and Landforms, 15(5), 457–466, doi:10.1002/esp.3290150508.Benda, L., and T. Dunne (1997a), Stochastic forcing of sediment supply to channel networksfrom landsliding and debris flow, Water Resources Research, 33(12), 2849–2863, doi:10.1029/97WR02388.Benda, L., M. A. Hassan, M. Church, and C. L. May (2005), Geomorphology of steepland head-waters: The transition from hillslopes to channels, Journal of the American Water ResourcesAssociation, 41(4), 835–851.118Beven, K. (2002), Towards a coherent philosophy for modelling the environment, in Proceedingsof the Royal Society of London A: Mathematical, Physical and Engineering Sciences, vol. 458, pp.2465–2484, The Royal Society, The Royal Society, doi:10.1098/rspa.2002.0986.Blom, A., L. Arkesteijn, V. Chavarrı´as, and E. Viparelli (2017), The equilibrium alluvial riverunder variable flow and its channel-forming discharge, Journal of Geophysical Research: EarthSurface, 122(10), 1924–1948, doi:10.1002/2017jf004213.Brardinoni, F., and M. A. Hassan (2006), Glacial erosion, evolution of river long profiles,and the organization of process domains in mountain drainage basins of coastal britishcolumbia, Journal of Geophysical Research, 111(F1), doi:10.1029/2005JF000358.Brummer, C. J., and D. R. Montgomery (2006), Influence of coarse lag formation on the me-chanics of sediment pulse dispersion in a mountain stream, Squire Creek, North Cascades,Washington, United States, Water Resources Research, 42(7), doi:10.1029/2005wr004776.Brunsden, D. (1980), Applicable model of long term landform evolution, Zeitschr. Geomorph.NF, Suppl. Bd., 36, 16–26.Brunsden, D. (2001), A critical assessment of the sensitivity concept in geomorphology, Catena,42(2), 99–123, doi:10.1016/s0341-8162(00)00134-x.Brunsden, D., and J. B. Thornes (1979), Landscape sensitivity and change, Transactions of theInstitute of British Geographers, 4(4), 463–484, doi:10.2307/622210.Bull, W. B. (1991), Geomorphic responses to climatic change, 326 pp. pp., Oxford University Press,doi:10.2307/215397.Bunte, K., and S. R. Abt (2001), Sampling surface and subsurface particle-size distributionsin wadable gravel-and cobble-bed streams for analyses in sediment transport, hydraulics,and streambed monitoring, Gen. Tech. Rep. RMRS-GTR-74. Fort Collins, CO: US Departmentof Agriculture, Forest Service, Rocky Mountain Research Station. 428 p., 74.Buscombe, D. (2013), Transferable wavelet method for grain-size distribution from imagesof sediment surfaces and thin sections, and other natural granular patterns, Sedimentology,60(7), 1709–1732, doi:10.1111/sed.12049.Carrivick, J. L., and M. W. Smith (2018), Fluvial and aquatic applications of structure frommotion photogrammetry and unmanned aerial vehicle/drone technology, Wiley Interdisci-plinary Reviews: Water, 6(1), e1328, doi:10.1002/wat2.1328.Chartrand, S. M., M. A. Hassan, and V. Radic´ (2015), Pool-riffle sedimentation and surfacetexture trends in a gravel bed stream, Water Resources Research, 51(11), 8704–8728, doi:10.1002/2015wr017840.Chorley, R. J., and B. A. Kennedy (1971), Physical geography: a systems approach, Prentice-HallLondon.Chow, K., L. Luna, A. Delforge, and D. Smith (2016), 2015 Pre-San Clemente Dam RemovalMorphological Monitoring of the Carmel River Channel in Monterey County, California.119Chung, C.-H., and F.-J. Chang (2013), A refined automated grain sizing method for estimatingriver-bed grain size distribution of digital images, Journal of hydrology, 486, 224–233.Church, M., and R. Ferguson (2015), Morphodynamics: Rivers beyond steady state, WaterResources Research, 51(4), 1883–1897, doi:10.1002/2014wr016862.Church, M., and R. Kellerhals (1978), On the statistics of grain size variation along a gravelriver, Canadian Journal of Earth Sciences, 15(7), 1151–1160, doi:10.1139/e78-121.Church, M., M. A. Hassan, and J. F. Wolcott (1998), Stabilizing self-organized structures ingravel-bed stream channels: Field and experimental observations, Water Resources Research,34(11), 3169–3179, doi:10.1029/98wr00484.Church, M. A., D. G. McLean, and J. F. Wolcott (1987), River bed gravels: sampling and analy-sis, in Sediment transport in gravel-bed rivers, edited by C. R. Thorne, J. C. Bathurst, and R. D.Hey, pp. 43–78, Wiley.Cienciala, P., and M. A. Hassan (2013), Linking spatial patterns of bed surface texture, bedmobility, and channel hydraulics in a mountain stream to potential spawning substrate forsmall resident trout, Geomorphology, 197, 96–107, doi:10.1016/j.geomorph.2013.04.041.Coulthard, T., and M. Van De Wiel (2013), 9.34 Numerical Modeling in FluvialGeomorphology, in Treatise on Geomorphology, pp. 694–710, Elsevier, doi:10.1016/b978-0-12-374739-6.00261-x.Cui, Y., and G. Parker (2005), Numerical model of sediment pulses and sediment-supplydisturbances in mountain rivers, Journal of Hydraulic Engineering, 131(8), 646–656, doi:10.1061/(asce)0733-9429(2005)131:8(646).Cui, Y., and A. Wilcox (2008), Development and application of numerical models of sedimenttransport associated with dam removal, Chapter, 23, 995–1020.Cui, Y., G. Parker, T. E. Lisle, J. Gott, M. E. Hansler-Ball, J. E. Pizzuto, N. E. Allmendinger,and J. M. Reed (2003), Sediment pulses in mountain rivers: 1. experiments, Water ResourcesResearch, 39(9), doi:10.1029/2002wr001803.Cui, Y., G. Parker, C. Braudrick, W. E. Dietrich, and B. Cluer (2006), Dam removal expressassessment models (dream). part 1: model development and validation, Journal of HydraulicResearch, 44(3), 291–307, doi:10.1080/00221686.2006.9521683.Curry, R., and G. Kondolf (1983), Sediment transport and channel stability, carmel river, cali-fornia, Unpublished report to MPWMD.De Rego, K. G. (2018), Decadal-scale evolution of elwha river downstream of glines canyondam: perspectives from numerical modeling, Ph.D. thesis, University of British Columbia.Detert, M., and V. Weitbrecht (2013), User guide to gravelometric image analysis by basegrain,Advances in science and research, pp. 1789–1795.Detert, M., L. Kadinski, and V. Weitbrecht (2018), On the way to airborne gravelometry basedon 3d spatial data derived from images, International journal of sediment research, 33(1), 84–92,doi:10.1016/j.ijsrc.2018.02.001.120Dibblee, T., and J. Minch (2007), Geologic map of the bickmore canyon quadrangle, montereyand san benito counties.Dietrich, W. E., and T. Dunne (1978), Sediment budget for a small catchment in mountainousterrain, Z. Geomorphol, 29, 191–206.Dietrich, W. E., J. W. Kirchner, H. Ikeda, and F. Iseya (1989), Sediment supply and the de-velopment of the coarse surface layer in gravel-bedded rivers, Nature, 340(6230), 215–217,doi:10.1038/340215a0.Dietrich, W. E., D. G. Bellugi, L. S. Sklar, J. D. Stock, A. M. Heimsath, and J. J. Roering (2003),Geomorphic transport laws for predicting landscape form and dynamics, Prediction in geo-morphology, 135, 103–132, doi:10.1029/135gm09.East, A. E., G. R. Pess, J. A. Bountry, C. S. Magirl, A. C. Ritchie, J. B. Logan, T. J. Randle, M. C.Mastin, J. T. Minear, J. J. Duda, et al. (2015), Large-scale dam removal on the elwha river,washington, usa: River channel and floodplain geomorphic change, Geomorphology, 228,765–786, doi:10.1016/j.geomorph.2014.08.028.Elgueta, M. A. (2014), Channel adjustment of a gravel-bed stream under episodic sedimentsupply regimes, Master’s thesis, University of British Columbia, doi:10.14288/1.0165847.Elgueta-Astaburuaga, M. A. (2018), Effects of episodic sediment supply on channel adjustmentof an experimental gravel bed, PhD thesis, University of British Columbia, doi:10.14288/1.0364558.Elgueta-Astaburuaga, M. A., and M. A. Hassan (2017), Experiment on temporal variation ofbed load transport in response to changes in sediment supply in streams, Water ResourcesResearch, 53(1), 763–778, doi:10.1002/2016wr019460.Escauriaza, C., C. Paola, and V. R. Voller (2017), Computational models of flow, sedimenttransport and morphodynamics in rivers, in Gravel-Bed Rivers, pp. 1–31, John Wiley & Sons,Ltd Chichester, UK, doi:10.1002/9781118971437.ch1.Ferrer-Boix, C., and M. A. Hassan (2014), Influence of the sediment supply texture on mor-phological adjustments in gravel-bed rivers, Water Resources Research, 50(11), 8868–8890,doi:10.1002/2013wr015117.Ferrer-Boix, C., and M. A. Hassan (2015), Channel adjustments to a succession of waterpulses in gravel bed rivers, Water Resources Research, 51(11), 8773–8790, doi:10.1002/2015wr017664.Ferrer-Boix, C., S. M. Chartrand, M. A. Hassan, J. P. Martı´n-Vide, and G. Parker (2016), On howspatial variations of channel width influence river profile curvature, Geophysical ResearchLetters, 43(12), 6313–6323, doi:10.1002/2016gl069824.Foley, M. M., J. Bellmore, J. E. O’Connor, J. Duda, A. E. East, G. Grant, C. Anderson, J. A.Bountry, M. J. Collins, P. J. Connolly, et al. (2017), Dam removal: Listening in, Water ResourcesResearch, 53(7), 5229–5246.Freeman, L., M. Webster, and M. F. Friebel (1996), Pacific Slope Basins from Arroyo Grande toOregon State Line except Central Valley. US Geological Survey Water-Data Report CA-96-2.121GMA (2008), 2007 Carmel River Surveys. Consulting report prepared for the Monterey Penin-sula Water Management Agency by Graham Matthews and Associates.Graf, W. L. (1999), Dam nation: A geographic census of american dams and their large-scalehydrologic impacts, Water resources research, 35(4), 1305–1311.Graham, D. J., I. Reid, and S. P. Rice (2005), Automated sizing of coarse-grained sed-iments: image-processing procedures, Mathematical Geology, 37(1), 1–28, doi:10.1007/s11004-005-8745-x.Graham, D. J., A.-J. Rollet, H. Pie´gay, and S. P. Rice (2010), Maximizing the accuracy of image-based surface sediment sampling techniques, Water Resources Research, 46(2), doi:10.1029/2008wr006940.Grant, G. (2001), Dam removal: Panacea or pandora for rivers?, Hydrological Processes, 15(8),1531–1532, doi:10.1002/hyp.473.Grant, G. E., and S. L. Lewis (2015), The remains of the dam: What have we learned from15 years of us dam removals?, in Engineering Geology for Society and Territory-Volume 3, pp.31–35, Springer.Gudde, E. G. (2010), California place names: the origin and etymology of current geographical names,Univ of California Press.Guthrie, R., and S. Evans (2007), Work, persistence, and formative events: the geomorphicimpact of landslides, Geomorphology, 88(3), 266–275.Hampson, L. (2018), Unpublished data of bank protection occurrence, extent and material typealong the mainstem Carmel River. Microsoft Excel spreadsheet. .Harrison, L. R., A. E. East, D. P. Smith, J. B. Logan, R. M. Bond, C. L. Nicol, T. H. Williams,D. A. Boughton, K. Chow, and L. Luna (2018), River response to large-dam removal in amediterranean hydroclimatic setting: Carmel river, california, usa, Earth Surface Processesand Landforms.Hassan, M. A., and M. Church (2000), Experiments on surface structure and partial sedi-ment transport on a gravel bed, Water Resources Research, 36(7), 1885–1895, doi:10.1029/2000wr900055.Hassan, M. A., and A. E. Zimmermann (2012), Channel response and recovery to changes insediment supply, in Gravel-Bed Rivers, pp. 464–473, John Wiley & Sons, Ltd, doi:10.1002/9781119952497.ch33.Hassan, M. A., M. Church, T. E. Lisle, F. Brardinoni, L. Benda, and G. E. Grant (2005), Sedimenttransport and channel morphology of small, forested streams, Journal of the American WaterResources Association, 41(4), 853–876, doi:10.1111/j.1752-1688.2005.tb03774.x.Hassan, M. A., R. Egozi, and G. Parker (2006a), Experiments on the effect of hydrographcharacteristics on vertical grain sorting in gravel bed rivers, Water Resources Research, 42(9),doi:10.1029/2005wr004707.122Hassan, M. A., R. Egozi, and G. Parker (2006b), Experiments on the effect of hydrographcharacteristics on vertical grain sorting in gravel bed rivers, Water Resources Research, 42(9),doi:10.1029/2005wr004707.Hassan, M. A., B. J. Smith, D. L. Hogan, D. S. Luzi, A. E. Zimmermann, and B. C. Eaton (2008a),18 sediment storage and transport in coarse bed streams: scale considerations, in Gravel-BedRivers VI: From Process Understanding to River Restoration, edited by P.-H. Habersack, H. andM. Rinaldi, pp. 473–496, Elsevier, doi:10.1016/s0928-2025(07)11137-8.Hassan, M. A., A. S. Gottesfeld, D. R. Montgomery, J. F. Tunnicliffe, G. K. C. Clarke, G. Wynn,H. Jones-Cox, R. Poirier, E. MacIsaac, H. Herunter, and S. J. Macdonald (2008b), Salmon-driven bed load transport and bed morphology in mountain streams, Geophysical ResearchLetters, 35(4), doi:10.1029/2007gl032997.Hassan, M. A., D. Brayshaw, Y. Alila, and E. Andrews (2014), Effective discharge in small for-merly glaciated mountain streams of british columbia: Limitations and implications, WaterResources Research, 50(5), 4440–4458, doi:10.1002/2013wr014529.Hassan, M. A., S. Bird, D. Reid, C. Ferrer-Boix, D. Hogan, F. Brardinoni, and S. Chartrand(2018), Variable hillslope-channel coupling and channel characteristics of forested mountainstreams in glaciated landscapes, Earth Surface Processes and Landforms, doi:10.1002/esp.4527.He, K., X. Zhang, S. Ren, and J. Sun (2016), Deep residual learning for image recognition,in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778,doi:10.1109/cvpr.2016.90.Heritage, G. L., and D. J. Milan (2009), Terrestrial laser scanning of grain roughness in a gravel-bed river, Geomorphology, 113(1-2), 4–11, doi:10.1016/j.geomorph.2009.03.021.Hoey, T. B., and R. Ferguson (1994), Numerical simulation of downstream fining by selec-tive transport in gravel bed rivers: Model development and illustration, Water ResourcesResearch, 30(7), 2251–2260, doi:10.1029/94wr00556.Hoffman, D. F., and E. J. Gabet (2007), Effects of sediment pulses on channel morphologyin a gravel-bed river, Geological Society of America Bulletin, 119(1-2), 116–125, doi:10.1130/b25982.1.Hoffmann, T., T. Mller, E. A. Johnson, and Y. E. Martin (2013), Postglacial adjustment of steep,low-order drainage basins, canadian rocky mountains, Journal of Geophysical Research: EarthSurface, 118(4), 17p., doi:10.1002/2013jf002846.Howard, A. D. (1982), Equilibrium and time scales in geomorphology: Application to sand-bed alluvial streams, Earth Surface Processes and Landforms, 7(4), 303–325, doi:10.1002/esp.3290070403.Istanbulluoglu, E. (2005), Vegetation-modulated landscape evolution: Effects of vegetationon landscape processes, drainage density, and topography, Journal of Geophysical Research,110(F2), doi:10.1029/2004jf000249.123Johnson, J. P. L., A. C. Aronovitz, and W. Kim (2015), Coarser and rougher: Effects of finegravel pulses on experimental step-pool channel morphodynamics, Geophysical Research Let-ters, 42(20), 8432–8440, doi:10.1002/2015gl066097.Kondolf, G. (1982), Recent channel instability and historic channel changes of the carmel river,monterey county, california [masters thesis], Santa Cruz: University of California Santa Cruz.Kondolf, G., and H. Pie´gay (2011), Geomorphology and society, in Handbook of Geomorphology,edited by K. Gregory, pp. 105–117, SAGE Publications, London.Kondolf, G. M., and R. R. Curry (1986), Channel erosion along the carmel river, montereycounty, california, Earth Surface Processes and Landforms, 11(3), 307–319.Kondolf, G. M., and M. G. Wolman (1993), The sizes of salmonid spawning gravels, WaterResources Research, 29(7), 2275–2285, doi:10.1029/93wr00402.Kondolf, G. M., Y. Gao, G. W. Annandale, G. L. Morris, E. Jiang, J. Zhang, Y. Cao, P. Carling,K. Fu, Q. Guo, R. Hotchkiss, C. Peteuil, T. Sumi, H.-W. Wang, Z. Wang, Z. Wei, B. Wu, C. Wu,and C. T. Yang (2014), Sustainable sediment management in reservoirs and regulated rivers:Experiences from five continents, Earths Future, 2(5), 256–280, doi:10.1002/2013ef000184.Korup, O. (2013), 9.15 Landslides in the Fluvial System, pp. 244–259, doi:10.1016/b978-0-12-374739-6.00240-2.Lisle, T. E., and M. Church (2002), Sediment transport-storage relations for degrading, gravelbed channels, Water Resources Research, 38(11), 1–1–1–14, doi:10.1029/2001wr001086.Lisle, T. E., J. E. Pizzuto, H. Ikeda, F. Iseya, and Y. Kodama (1997), Evolution of a sedimentwave in an experimental channel, Water Resources Research, 33(8), 1971–1981, doi:10.1029/97wr01180.Lisle, T. E., Y. Cui, G. Parker, J. E. Pizzuto, and A. M. Dodd (2001), The dominance of disper-sion in the evolution of bed material waves in gravel-bed rivers, Earth Surface Processes andLandforms, 26(13), 1409–1420, doi:10.1002/esp.300.Luzi, D. S. (2014), Sediment transport and morphological response of a semi-alluvial channel:insights from a froude scaled laboratory model, PhD thesis, University of British Columbia,doi:10.14288/1.0165910.MacKenzie, L. G., and B. C. Eaton (2017), Large grains matter: contrasting bed stability andmorphodynamics during two nearly identical experiments, Earth Surface Processes and Land-forms, 42(8), 1287–1295, doi:10.1002/esp.4122.Madej, M. A. (1999), Temporal and spatial variability in thalweg profiles of a gravel-bed river, Earth Surface Processes and Landforms, 24(12), 1153–1169, doi:10.1002/(SICI)1096-9837(199911)24:12<1153::AID-ESP41>3.0.CO;2-8.Madej, M. A. (2001), Development of channel organization and roughness following sedimentpulses in single-thread, gravel bed rivers, Water Resources Research, 37(8), 2259–2272, doi:10.1029/2001wr000229.124Madej, M. A., and V. Ozaki (1996), Channel response to sediment wave propagation and move-ment, Redwood Creek, California, USA, Earth Surface Processes and Landforms, 21(10), 911–927, doi:10.1002/(sici)1096-9837(199610)21:10<911::aid-esp621>3.0.co;2-1.Major, J. J., J. E. OConnor, C. J. Podolak, M. K. Keith, G. E. Grant, K. R. Spicer, S. Pittman, H. M.Bragg, J. R. Wallick, D. Q. Tanner, et al. (2012), Geomorphic response of the Sandy River, Oregon,to removal of Marmot Dam, US Department of the Interior, US Geological Survey.Major, J. J., A. E. East, J. E. O’Connor, G. E. Grant, A. C. Wilcox, C. S. Magirl, M. J. Collins,and D. D. Tullos (2017), Geomorphic responses to dam removal in the united states–a two-decade perspective, D. Tsutsumi, D. and JB Laronne (eds), Gravel-Bed Rivers: Processes andDisasters. New York, NY: John Wiley and Sons, pp. 355–383, doi:10.1002/9781118971437.ch13.Markham, K., J. Palmer, M. F. Friebel, and L. F. Trujillo (1992), Pacific Slope Basins from ArroyoGrande to Oregon State Line except Central Valley. US Geological Survey Water-Data ReportCA-92-2.Markstrom, S. L., R. S. Regan, L. E. Hay, R. J. Viger, R. M. Webb, R. A. Payn, and J. H. La-Fontaine (2015), PRMS-IV, the precipitation-runoff modeling system, version 4, US Geologi-cal Survey Techniques and Methods, (6-B7).May, C. L., and T. E. Lisle (2012), River profile controls on channel morphology, debris flowdisturbance, and the spatial extent of salmonids in steep mountain streams, Journal of Geo-physical Research, 117, doi:10.1029/2011JF002324.MEI (2002), Carmel River Sediment-Transport Study, Monterey County, California. Report pre-pared for California Department of Water Resources.MEI (2003), San Clemente Reservoir and Carmel River Sediment-Transport Modeling to Evalu-ate Potential Impacts of Dam Retrofit Options. Submitted to American Water Works ServiceCompany. .Miller, D. J., and L. E. Benda (2000), Effects of punctuated sediment supply on valley-floorlandforms and sediment transport, Geological Society of America Bulletin, 112(12), 1814–1824,doi:10.1130/0016-7606(2000)112<1814:eopsso>2.0.co;2.Montgomery, D. R. (1999), Process domains and the river continuum, JAWRA Journal of theAmerican Water Resources Association, 35(2), 397–410.Montgomery, D. R., and J. M. Buffington (1997), Channel-reach morphology in mountaindrainage basins, Geological Society of America Bulletin, 109(5), 596–611.Montgomery, D. R., and E. Foufoula-Georgiou (1993), Channel network source representationusing digital elevation models, Water Resources Research, 29(12), 3925–3934.MPWD (1986), Unpublished sediment transport measurements, carmel river, water years 1984- 1986.MPWMD (2016), Request for Proposals Los Padres Dam and Reservoir Alternatives and Sedi-ment Management Study.125Nelson, A. D., and M. Church (2012), Placer mining along the fraser river, british columbia:the geomorphic impact, Bulletin, 124(7-8), 1212–1228.Nelson, P. A., J. G. Venditti, W. E. Dietrich, J. W. Kirchner, H. Ikeda, F. Iseya, and L. S. Sklar(2009), Response of bed surface patchiness to reductions in sediment supply, Journal of Geo-physical Research: Earth Surface, 114(F2), doi:10.1029/2008jf001144.Normandeau (2016), Carmel river cross-section surveys.Oreskes, N., K. Shrader-Frechette, and K. Belitz (1994), Verification, validation, and confirma-tion of numerical models in the earth sciences, Science, 263(5147), 641–646.Paola, C., G. Parker, D. C. Mohrig, and K. X. Whipple (1999), The influence of transport fluc-tuations on spatially averaged topography on a sandy, braided fluvial fan, in NumericalExperiments in Stratigraphy: Recent Advances in Stratigraphic and Sedimentologic Computer Sim-ulations, edited by J. Harbaugh, W. Watney, and E. Rankey, pp. 211–218, Special Publicationsof SEPM.Papangelakis, E., and M. A. Hassan (2016), The role of channel morphology on the mobilityand dispersion of bed sediment in a small gravel-bed stream, Earth Surface Processes andLandforms, 41(15), 2191–2206, doi:10.1002/esp.3980.Parker, G. (1991), Selective sorting and abrasion of river gravel. i: Theory, Journal of HydraulicEngineering, 117(2), 131–147, doi:10.1061/(asce)0733-9429(1991)117:2(131).Parker, G. (2008), Transport of gravel and sediment mixtures, in Sedimentation Engineering, vol.110, pp. 165–251, American Society of Civil Engineers, doi:10.1061/9780784408148.ch03.Parker, G., M. Hassan, and P. Wilcock (2007), 10 adjustment of the bed surface size distributionof gravel-bed rivers in response to cycled hydrographs, in Gravel-Bed Rivers VI: From Pro-cess Understanding to River Restoration, pp. 241–285, Elsevier, doi:10.1016/s0928-2025(07)11127-5.Pfeiffer, A. M., N. J. Finnegan, and J. K. Willenbring (2017), Sediment supply controls equi-librium channel geometry in gravel rivers, Proceedings of the National Academy of Sciences,114(13), 3346–3351, doi:10.1073/pnas.1612907114.Phillips, J. D. (2003), Sources of nonlinearity and complexity in geomorphic systems, Progressin Physical Geography, 27(1), 1–23.Podolak, C. J. P., and P. R. Wilcock (2013), Experimental study of the response of a gravelstreambed to increased sediment supply, Earth Surface Processes and Landforms, 38(14), 1748–1764, doi:10.1002/esp.3468.Pryor, B. S., T. Lisle, D. S. Montoya, and S. Hilton (2011), Transport and storage of bed materialin a gravel-bed channel during episodes of aggradation and degradation: a field and flumestudy, Earth Surface Processes and Landforms, 36(15), 2028–2041, doi:10.1002/esp.2224.Recking, A. (2012), Influence of sediment supply on mountain streams bedload transport, Ge-omorphology, 175-176, 139–150, doi:10.1016/j.geomorph.2012.07.005.126Rice, S., and M. Church (1996), Sampling surficial fluvial gravels; the precision of size distri-bution percentile sediments, Journal of Sedimentary Research, 66(3), 654–665.Rice, S., and M. Church (1998), Grain size along two gravel-bed rivers: statistical variation,spatial pattern and sedimentary links, Earth Surface Processes and Landforms, 23(4), 345–363,doi:10.1002/(sici)1096-9837(199804)23:4<345::aid-esp850>3.0.co;2-b.Roberts, R. G., and M. Church (1986), The sediment budget in severely disturbed watersheds,queen charlotte ranges, british columbia, Canadian Journal of Forest Research, 16(5), 1092–1106, doi:10.1139/x86-189.Ronneberger, O., P. Fischer, and T. Brox (2015), U-net: Convolutional networks for biomedicalimage segmentation, in International Conference on Medical image computing and computer-assisted intervention, pp. 234–241, Springer, Springer International Publishing, doi:10.1007/978-3-319-24574-4_28.Schrott, L., G. Hufschmidt, M. Hankammer, T. Hoffmann, and R. Dikau (2003), Spatial distri-bution of sediment storage types and quantification of valley fill deposits in an alpine basin,reintal, bavarian alps, germany, Geomorphology, 55(1), 45–63.Simonyan, K., and A. Zisserman (2014), Very deep convolutional networks for large-scale im-age recognition, arXiv preprint arXiv:1409.1556.Sklar, L. S., J. Fadde, J. G. Venditti, P. Nelson, M. A. Wydzga, Y. Cui, and W. E. Dietrich (2009),Translation and dispersion of sediment pulses in flume experiments simulating gravel aug-mentation below dams, Water Resources Research, 45(8), doi:10.1029/2008wr007346.Slaymaker, O. (2003), The sediment budget as conceptual framework and management tool,in The Interactions between Sediments and Water, Developments in Hydrobiology, vol. 169, editedby B. Kronvang, pp. 71–82, Springer Netherlands, doi:10.1007/978-94-017-3366-3_12.Sthly, S., H. Friedrich, and M. Detert (2017), Size ratio of fluvial grains’ intermediate axes as-sessed by image processing and square-hole sieving, Journal of Hydraulic Engineering, 143(6),06017,005, doi:10.1061/(asce)hy.1943-7900.0001286.Tetra Tech (2015), Final Design Report for Channel Restoration Design for the Carmel RiverReroute and San Clemente Dam Removal (CRRDR) project.Thomas, M. F. (2001), Landscape sensitivity in time and spacean introduction, Catena, 42(2-4),83–98, doi:10.1016/s0341-8162(00)00133-8.Tucker, G. E., and R. L. Bras (1998), Hillslope processes, drainage density, and landscape mor-phology, Water Resources Research, 34(10), 2751–2764, doi:10.1029/98wr01474.URS (2013), Carmel River Reroute & San Clemente Dam Removal Project: Draft Subtask 4.2.2River Flood & Sediment Transport Modeling Technical Memorandum. Consulting reportprepared for the State Coastal Conservancy and California American Water.Van De Wiel, M. J., T. J. Coulthard, M. G. Macklin, and J. Lewin (2011), Modelling theresponse of river systems to environmental change: progress, problems and prospectsfor palaeo-environmental reconstructions, Earth-Science Reviews, 104(1-3), 167–185, doi:10.1016/j.earscirev.2010.10.004.127Venditti, J. G., W. E. Dietrich, P. A. Nelson, M. A. Wydzga, J. Fadde, and L. Sklar (2010), Effectof sediment pulse grain size on sediment transport rates and bed mobility in gravel bedrivers, Journal of Geophysical Research: Earth Surface, 115(F3), doi:10.1029/2009jf001418.Viparelli, E., O. E. Sequeiros, A. Cantelli, P. R. Wilcock, and G. Parker (2010), River morphody-namics with creation/consumption of grain size stratigraphy 2: numerical model, Journal ofHydraulic Research, 48(6), 727–741, doi:10.1080/00221686.2010.526759.Viparelli, E., D. Gaeuman, P. Wilcock, and G. Parker (2011), A model to predict the evolutionof a gravel bed river under an imposed cyclic hydrograph and its application to the trinityriver, Water Resources Research, 47(2).von Flotow, C. (2013), Temporal adjustments of a streambed following an episodic sedimentsupply regime, Master’s thesis, University of British Columbia, doi:10.14288/1.007421.Warrick, J. A., D. M. Rubin, P. Ruggiero, J. N. Harney, A. E. Draut, and D. Buscombe (2009),Cobble cam: grain-size measurements of sand to boulder from digital photographs andautocorrelation analyses, Earth Surface Processes and Landforms, 34(13), 1811–1821, doi:10.1002/esp.1877.Whitson Engineers (2017), Carmel River Thalweg Survey.Wilcock, P. R. (2001), Toward a practical method for estimating sediment-transport rates ingravel-bed rivers, Earth Surface Processes and Landforms, 26(13), 1395–1408, doi:10.1002/esp.301.Wilcock, P. R., and J. C. Crowe (2003), Surface-based transport model for mixed-size sediment,Journal of Hydraulic Engineering, 129(2), 120–128, doi:10.1061/(asce)0733-9429(2003)129:2(120).Wilcock, P. R., and R. M. Iverson (2003), Prediction in geomorphology, 3–11 pp., Wiley OnlineLibrary, doi:10.1029/135gm01.Wilcock, P. R., and B. W. McArdell (1993), Surface-based fractional transport rates: Mobiliza-tion thresholds and partial transport of a sand-gravel sediment, Water Resources Research,29(4), 1297–1312.Wilcox, A. C., J. E. O’Connor, and J. J. Major (2014), Rapid reservoir erosion, hyperconcentratedflow, and downstream deposition triggered by breaching of 38 m tall condit dam, whitesalmon river, washington, Journal of Geophysical Research: Earth Surface, 119(6), 1376–1394,doi:10.1002/2013jf003073.Willgoose, G. R., G. R. Hancock, and G. Kuczera (2003), A framework for the quantitativetesting of landform evolution models, Prediction in geomorphology, 135, 195–216, doi:10.1029/135gm14.Witten, I. H., E. Frank, M. A. Hall, and C. J. Pal (2016), Data Mining: Practical machine learningtools and techniques, Morgan Kaufmann.Wohl, E. (2013), Mountain rivers revisited, vol. 19, John Wiley & Sons.1285.2. Future workWolman, M. G. (1954), A method of sampling coarse river-bed material, EOS, TransactionsAmerican Geophysical Union, 35(6), 951–956.Wolman, M. G., and J. P. Miller (1960), Magnitude and frequency of forces in geomorphicprocesses, The Journal of Geology, 68(1), 54–74, doi:10.1086/626637.Wong, M., and G. Parker (2006), One-dimensional modeling of bed evolution in a gravel bedriver subject to a cycled flood hydrograph, Journal of Geophysical Research: Earth Surface,111(F3), n/a–n/a, doi:10.1029/2006jf000478.Woodget, A., C. Fyffe, and P. Carbonneau (2018), From manned to unmanned aircraft: Adapt-ing airborne particle size mapping methodologies to the characteristics of suas and sfm,Earth Surface Processes and Landforms, 43(4), 857–870, doi:10.1002/esp.4285.Zimmermann, A., M. Church, and M. A. Hassan (2010), Step-pool stability: Testing thejammed state hypothesis, Journal of Geophysical Research: Earth Surface (2003–2012), 115(F2).Zimmermann, A. E., M. Church, and M. A. Hassan (2008), Video-based gravel transport mea-surements with a flume mounted light table, Earth Surface Processes and Landforms, 33(14),2285–2296.129Appendix ASensitivity analysis plots for BESMo inepisodic supply experiments130APPENDIX A. SENSITIVITY ANALYSIS PLOTS FOR BESMO IN EPISODIC SUPPLYEXPERIMENTS0.0150.020.0250.03Slope (m/m)0.0150.020.0250.03Slope (m/m)0.0150.020.0250.03Slope (m/m)0.0150.020.0250.03Slope (m/m)Flume experimentsChosen parameter (α = 0.45) α = 0.35 α = 0.4 α = 0.45 α = 0.5Boundary of sensitivity runs(a) na = 1(b) na = 1.5(c) na = 2(d) na = 2.50 40 80 120 160 200 240 280Time (hours)No feed Constant feed One pulse Four pulses Two pulses Constant feed No feedFigure A.1: Sensitivity of modelled slope to active layer thickness factor na andactive layer exchange ratio α in the ’Original flume’ (OF) event sequence.131APPENDIX A. SENSITIVITY ANALYSIS PLOTS FOR BESMO IN EPISODIC SUPPLYEXPERIMENTSFlume experimentsChosen parameter (α = 0.45) α = 0.35 α = 0.4 α = 0.45 α = 0.5Boundary of sensitivity runsSurface Dg (mm)681012141618Surface Dg (mm)681012141618Surface Dg (mm)681012141618Surface Dg (mm)6810121416180 40 80 120 160 200 240 280(a) na = 1(b) na = 1.5(c) na = 2(d) na = 2.5Time (hours)No feed Constant feed One pulse Four pulses Two pulses Constant feed No feedFigure A.2: Sensitivity of modelled Surface Dg to active layer thickness factor naand active layer exchange ratio α in the ’Original flume’ (OF) event se-quence.132APPENDIX A. SENSITIVITY ANALYSIS PLOTS FOR BESMO IN EPISODIC SUPPLYEXPERIMENTSFlume experimentsChosen parameter (α = 0.45) α = 0.35 α = 0.4 α = 0.45 α = 0.5Boundary of sensitivity runsNo feed Constant feed One pulse Four pulses Two pulses Constant feed No feed0 40 80 120 160 200 240 280Surface D90 (mm)24262830323436Surface D90 (mm)24262830323436Surface D90 (mm)24262830323436Surface D90 (mm)24262830323436(a) na = 1(b) na = 1.5(c) na = 2(d) na = 2.5Time (hours)Figure A.3: Sensitivity of modelled Surface D90 to active layer thickness factorna and active layer exchange ratio α in the ’Original flume’ (OF) event se-quence.133APPENDIX A. SENSITIVITY ANALYSIS PLOTS FOR BESMO IN EPISODIC SUPPLYEXPERIMENTSFlume experimentsChosen parameter (α = 0.45) α = 0.35 α = 0.4 α = 0.45 α = 0.5Boundary of sensitivity runsTransport rate ( ) 10210–2100g msTransport rate ( ) 10210–2100g msTransport rate ( ) 10210–2100g msTransport rate ( )102100g ms0 40 80 120 160 200 240 280Time (hours)(a) na = 1(b) na = 1.5(c) na = 2(d) na = 2.5No feed Constant feed One pulse Four pulses Two pulses Constant feed No feedFigure A.4: Sensitivity of modelled transport rate to active layer thickness factorna and active layer exchange ratio α in the ’Original flume’ (OF) event se-quence.134APPENDIX A. SENSITIVITY ANALYSIS PLOTS FOR BESMO IN EPISODIC SUPPLYEXPERIMENTS10−2 10−1 100 101 1020.70.80.911.11.21.3Tpp/T f eS mlp/Sconst125% Mass feed, σ = 1.6100% Mass feed, σ = 1.675% Mass feed, σ = 1.6Figure A.5: Effect of a 25% increased and 25% decreased total sediment feed onthe equilibrium slope in the non-dimensionalized time scale. All simula-tions were executed with σ = 1.6.135Appendix BLos Padres Dam removalsupplementary materialB.1 Model testingThe following demonstrates the key model evaluation steps and tests directed by the technicalreview committee (TRC) and evaluated by the project team. Model iterations were pursued toproduce results that are more suitable for informing the project alternative assessment processand so the results of those iterative tests are included in this report.Prior to executing the No Action Simulation, the TRC requested a comparison of longitudi-nal profile adjustment within the San Clemente project reach (model station 99,015 to 105,000feet upstream of the Pacific Ocean), for the period water years 2015- 2017 (WY2015-17). Com-parison was made between BESMo and observations of channel elevation change at selectchannel locations within the San Clemente project reach. The comparison period was sim-ulated within BESMo for the hydrograph shown in Figure B.1, subject to the peaking factoradjustments. The measured longitudinal profiles for the project reach were collected in early2014 and in 2017 and were generously provided to us by Doug Smith of CSUMB.We modified the Carmel River build of BESMo for this comparison in three ways. First,we implemented a local model build with more closely spaced nodes within the project reachin order to provide for a better comparison between model and observation. The inset plot136B.1. Model testingFigure B.1: Hydrograph used in the San Clemente comparison. Hydrograph rep-resents mean daily flow reported for the RR gage by the USGS.of Figure B.2 illustrates the nodes used for the San Clemente comparative simulation, and theinitial model longitudinal profile for the project reach was taken from the final San Clementeproject design (Tetra Tech, 2015). Second and after a few trial runs to evaluate model fidelitywithin the relatively steep and coarse combined flow steppool segments of the San Clementeproject reach, we lowered the volumetric content of fine channel substrate to 33% (from 50%)and accordingly increased the coarse mixture content to 66% (from 50%). Figure B.3 illus-trates the GSD used in the BESMo simulation for the former reservoir deposits upstream ofthe reroute reach, and for the steeper constructed step-pool channel segments. Last, due to theshort distance between nodes, we reduced the model time step to 1 second during periods ofbedload transport, otherwise we used a time step of 5 minutes. This change in the model wastested and confirmed to not affect the simulated results. No other changes were made to theCarmel River build of BESMo.The San Clemente project reach comparative results are shown in Figure B.4 and Figure B.5.Figure B.4 shows the longitudinal profiles for BESMo and observation (labeled as post withinthe legend) in WY2014 and WY2017, separately. The WY2017 profiles represent the adjustmentthrough the project reach due primarily to the floods of WY2017 (Figure B.5). A few relevantobservations can be made. First, differences between BESMo and the WY2017 observationsgenerally range from less than 0.5 to 1.0 meters, for most of the project reach, except at the SanClemente Creek tributary node and the very downstream end of the project reach. Second,137B.1. Model testingFigure B.2: BESMo simulation nodes for the reroute corridor at the former SanClemente Reservoir. The detailed view shows the higher-resolution ofnodes within the reroute channel.Figure B.3: Grain size distributions of both the fine upstream deposits (MEI, 2002)and the coarse steps (Tetra Tech, 2015).the general spatial pattern of erosion and deposition between BESMo and the WY2017 obser-vations is similar. There is notable agreement within the upper former reservoir deposits andthrough the middle part of the steeper step-pool channel segments. This is a bit surprisinggiven the simplicity of how BESMo represents actual physical conditions and processes thatgive rise to river profiles.Third, differences between BESMo and WY2017 observations flip in trend around the SanClemente Creek tributary node. Last, BESMo simulates up to 1.5 feet of deposition by WY2017within the downstream end of the combined flow reach, upstream of the former San ClementeDam. Simulated deposition is largely a result of bed erosion simulated upstream from thedownstream most channel segment. Despite these differences, BESMo does a reasonably good138B.1. Model testingFigure B.4: 4 Comparison between the measured and simulated reroute channellong profiles over the comparative period of WY2014-17.Figure B.5: Elevation difference between the measured and simulated reroutechannel long profiles over three years.job of capturing net erosion within the former reservoir deposit, as well as the reach averagebed slope downstream of San Clemente Creek to just upstream of the former San ClementeDam. As discussed above, results are sensitive to the initial GSDs for the surface and sub-139B.1. Model testingsurface sediments. Given this sensitivity, it is encouraging that BESMo compares with obser-vations as well as it does because a 3-year simulation period with very little net adjustmentduring the first two years is a difficult basis for comparison between a simplified numeri-cal model, and real-world conditions within a steep, and spatially complex channel segmentwithin which sediment transport processes depart substantially from the empirical Wilcock-Crowe transport function.140B.2. Elevation data availabilityB.2 Elevation data availabilityWe use three different sources of channel elevation data to construct the initial simulationlongitudinal profile for the BESMo simulations (Figure B.6)• Whitson Engineers survey of the main stem from the Lagoon to river station approxi-mately 24 km upstream of the Lagoon (Whitson Engineers, 2017);• URS HEC-RAS model build from the river station at 24 km to approximately 32 km forthe San Clemente Dam removal scenario (URS, 2013);• USGS National Elevation Dataset, spatial resolution 1/9th arc-second (NED19) fromhead of former San Clemente Dam reservoir to Los Padres Dam at river station ap-proximately 42 km. This NED19 dataset was compiled by the authors of this report, asdetailed channel geometry surveys are not available for the channel between the formerreservoir deposit of San Clemente Dam and the Los Padres Dam. To convert the digitalelevation model data to a channel long profile, we averaged elevation values recordedalong the channel within a circular 100m area around the model nodes (Figure B.7).141B.2. Elevation data availabilityFigure B.6: Channel profile from the three data sources. Both Whitson Engineers(2017) and URS (2013) lack coverage of the upper section of the river. Weused the NED19 dataset to fill the holes of the other data and generate afull profile of the study reach.142B.2. Elevation data availabilitySources: Esri, HERE, DeLorme, Intermap, increment P Corp., GEBCO, USGS, FAO, NPS,NRCAN, GeoBase, IGN, Kadaster NL, Ordnance Survey, Esri Japan, METI, Esri China (HongKong), swisstopo, MapmyIndia, © OpenStreetMap contributors, and the GIS User CommunityExtraction of Elevation from NED19 DEM±0 330 660 990 1,320165 MetersLegendStreamline provided by AECOM500m model nodes100m Buffer to average DEM elevationStreamline from NED19 DEMElevation points from NED19DEMFigure B.7: Extraction of model node elevation from the NED19 (1/9th arc seconddigital elevation model) below the Los Padres Reservoir. For each 500mmodel node along the streamline, we averaged the elevation from NED19within a 100m buffer along the streamline.143B.3. Sediment transport data availabilityB.3 Sediment transport data availabilitySediment transport data has been collected in the Carmel River watershed sporadically from1981 to 2001 in the mainstem and tributary channels. Sediment availability and transportrates can vary considerably both temporally and spatially in the watershed. Due to extremeepisodic events such as fires and landslides, large pulses of sediment can significantly increasethe sediment transport rates. We have collated the available bedload transport data, for bothbedload and suspended sediment, and updated previous rating curves for each tributary andmainstem location.Each rating curve was estimated from available sediment transport datasourced from the USGS Water-Data Reports (Markham et al., 1992; Ayers, 1995; Freeman et al.,1996), Monterey Peninsula Water District field office data archives (MPWD, 1986), (Curry andKondolf , 1983), and unpublished data collected by Balance Hydrologics for the San ClementeDam Removal efforts (Balance Hydrologics, Inc., 2001). Each rating curve was estimated usingthe best-fit power law when possible. When applicable, more representative rating curveswere developed manually or with outliers not included.Carmel River Mainstem Sediment SupplyThe Carmel River at Via Mallorca has the most abundant sediment transport dataset. Thesite has been a USGS stream gage since 1962 making this site an ideal location to collectpaired sediment-flow measurements. Data has also been collected at Schulte Bridge, RobinsonCanyon Road, and Robles del Rio. Appendix B.4 presents the sediment rating curves for eachmainstem location with available data.Carmel River Watershed Tributary Sediment SupplyA considerable portion of the total sediment load in the Carmel River watershed comes fromthe tributaries. There are seven major tributaries in the Carmel River watershed that have beenhistorically monitored: Tularcitos Creek, Cachagua Creek, San Clemente Creek, Las GarzasCreek, Robinson Canyon Creek, Potrero Creek, and Hitchcock Creek. The hydrologic and sed-iment contribution from each tributary to the mainstem Carmel River varies, dependent uponthe mean annual rainfall in the contributing watershed and the underlying geology and as-144B.3. Sediment transport data availabilitysociated sediment production processes. For example, the underlying geology in the Tularci-tos Creek watershed is largely sourced from the easily erodible Santa Margarita sandstone,which introduces an abundance of sand. Despite a low mean annual precipitation, the Tu-larcitos Creek watershed supplies a lot of sand, considerably changing the sediment characterdownstream of the Tularcitos Creek confluence. Table B.1 summarizes the differences in meanannual precipitation and geology in each of the main tributaries.Table B.1: Summary of Carmel River watershed major tributaries; Mean annualprecipitation, watershed size, and geology.Tributary WatershedSize(km2)Mean An-nual Per-cipitation(mm)1Geology2Tularcitos Creek 90.6 546 San Margarita Sandstone, MioceneMarine clastic shale, sandstone, andconglomerateCachagua Creek 74.5 805 Mixed Miocene marine sandstoneand Mesozoic granitic rocksSan ClementeCreek25.1 950 Mesozoic granitic rocks, with someMesozoic metasedimentary rocksLas Garcas Creek 21.2 724 Mesozoic granitic rocks, with someMiocene unnamed sedimentaryredbedsRobinson CanyonCreek8.7 564 Miocene unnamed sedimentaryredbeds and marine sandstonePotrero Creek 9.3 582 Miocene Monterey Formation shale,and Quaternary landslide and allu-vial gravel, sand, and silt/clayHitchcock Creek 7.4 635 Mesozoic granitic rocks, primarilygranodiorite, some Miocene marinerocks1 Estimated using Monterey County Isohyetal lines of average annual Rainfall in inches, published May14,2014, accessed April 30, 20182 Geologic information sourced from Geologic maps of various quadrangles scale 1:24,000, Dibblee andMinch (2007)Pine Creek, located to the south of San Clemente Creek, is another major tributary inthe Carmel River watershed. Because access to Pine Creek is difficult, there is limited sedi-ment transport and hydrologic data available and so rating curves were not developed. Ap-145B.3. Sediment transport data availabilitypendix B.5 presents the sediment rating curves for each tributary with available data.Bedloaddata for Tularcitos Creek suggests a two-phase rating curve with one rating curve for flowsless than 0.4 m3s−1 and another for flows 0.4 m3s−1 and greater. Although the precise rea-son for this two-phase rating curve cannot be conclusively determined without further study,we hypothesize sediment availability is a function of bank geometry and erodibility changeswith height. The data used to develop the San Clemente Creek rating curves were collected attwo different locations on San Clemente Creek; Curry and Kondolf (1983) collected samples atthe inlet to the former San Clemente Reservoir and Balance Hydrologics staff collected sam-ples approximately 3.2 km upstream. For the bedload rating curve, 4 of the 5 available pointswere collected by Balance Hydrologics at the upper site. The one measurement presented inCurry and Kondolf (1983) is consistent with the Balance Hydrologics rating curve. No sedimenttransport data was collected in Hitchcock Canyon Creek, but Curry and Kondolf (1983) reportsediment rating curves, which are presented here.146B.4. Rating Curves MainstemB.4 Rating Curves MainstemFigure B.8: Rating curves for Carmel River at Via Mallorca. The site is co-locatedwith USGS gage 11143250.Figure B.9: Rating curves for Carmel River at Schulte Bridge.147B.4. Rating Curves MainstemFigure B.10: Rating curves for Carmel River at Robinson Canyon Road.Figure B.11: Rating curves for Carmel River at Robles del Rio.148B.5. Rating Curves TributariesB.5 Rating Curves TributariesFigure B.12: Rating curves for Tularcitos Creek at Sleepy Hollow.Figure B.13: Rating curves for Cachagua Creek at Princess Camp.149B.5. Rating Curves TributariesFigure B.14: Rating curves for San Clemente Creek.Figure B.15: Rating curves for Las Garzas Creek.150B.5. Rating Curves TributariesFigure B.16: Rating curves for Robinson Canyon Creek at Robinson CanyonRoad.Figure B.17: Rating curves for Potrero Creek.151B.5. Rating Curves TributariesFigure B.18: Rating curves for Hitchcock Canyon Creek.152B.6. Node InitializationB.6 Node InitializationTable B.2: Input parameters for the full runs by simulation node. The first twonodes (ID 1 and 2) were added to provide a smoother boundary conditionfrom high-magnitude sediment input events from the former Los Padresreservoir. Reaches change where the hydrology changes, as different back-water curves are applied for each reach. The depth of maximum subsur-face grain size is different between spinup runs (indicated by *) and the fulllength runs (trial runs and alternatives).NodeIDElevation(ft)Distancefrommouth(ft)ReachIDHydrologyStationIDInitial SurfaceGrain Size Distri-bution IDExternalSedimentFeed IDInitial Subsur-face Grain SizeDistribution IDDepth ofmaximumsubsur-face grainsize (ft)Maximumsubsur-facegrain size(mm)Northingin mEastingin m1 965.8 D1A BL URS Active 99577 DAM MEI80k 2 512 4027565.6 619425.22 942.4 D1A BL URS Active 99577 DAM MEI80k 2 512 4027565.6 619425.2Analysis domain begins here / End of extended boundary3 918.2 137795 D1A BL URS Active 99577 MEI80k 2 512 4027565.6 619425.24 893.8 136155 D1A BL URS Active 99577 MEI80k 2 512 4027565.6 619425.25 869.7 134514 D1A BL URS Active 99577 MEI80k 2 512 4027960.2 619666.06 845.7 132874 D1A BL URS Active 99577 MEI80k 2 512 4028370.0 619661.47 824.9 131234 D1A BL URS Active 99577 MEI80k 2 512 4028759.9 619932.98 803.2 129593 D1A BL URS Active 99577 MEI80k 2 512 4029219.6 620077.59 782.8 127953 D1B BL+CA URS Active 99577 CA trib MEI80k 2 512 4029530.6 619874.610 764.2 126312 D1B BL+CA URS Active 99577 MEI80k 2 512 4029937.5 619602.611 746.8 124672 D1B BL+CA URS Active 99577 MEI80k 2 512 4030061.2 619150.912 728.2 123031 D1B BL+CA URS Active 99577 MEI80k 2 512 4030208.8 618749.913 706.6 121391 D1B BL+CA URS Active 99577 MEI80k 3.3 512 4029996.8 618384.114 683.4 119751 D1B BL+CA URS Active 99577 MEI80k 3.3 512 4030020.0 617994.815 661.0 118110 D1B BL+CA URS Active 99577 MEI80k 6.6 512 4030016.6 617549.916 637.8 116470 D1B BL+CA URS Active 99577 MEI80k 6.6 512 4030036.9 617120.417 616.1 114829 D1B BL+CA URS Active 99577 MEI80k 6.6 512 4030038.3 616799.418 596.9 113189 D1B BL+CA URS Active 99577 MEI80k 6.6 512 4030199.7 616385.919 577.3 111549 D1B BL+CA URS Active 99577 MEI80k 6.6 512 4030648.3 616343.620 563.4 109908 D1B BL+CA URS Active 99577 MEI80k 6.6 512 4030822.7 615908.021 544.2 108268 D1B BL+CA URS Active 99577 MEI80k 6.6 512 4031028.3 615498.222 535.4 106627 D1B BL+CA URS Active 99577 MEI80k 6.6 512 4031493.4 615454.323 527.5 104987 D1B BL+CA URS Active 99577 MEI80k 6.6 512 4031958.7 615578.624 518.8 103346 D1B BL+CA URS Active 99577 MEI80k 6.6 512 4032127.8 615309.225 495.4 101706 D1B SHW URS Active 99577 CL trib MEI80k 0 512 4032509.4 615416.126 461.7 100066 D1B SHW URS Active 99577 MEI80k 2 512 4032914.0 615645.3Spinup run begins here27 441.1 98425 D2 SHW URS Active 99577 MEI80k 2 / 1.6* 512 4033245.0 615791.528 431.7 96785 D2 SHW URS Active 99577 MEI80k 2 / 1.6* 512 4033476.2 616148.829 414.7 95144 D2 SHW URS Active 99577 MEI80k 16 / 3.3* 512 4033861.6 615860.330 400.1 93504 D2 SHW URS Active 99577 MEI80k 16 / 3.3* 512 4034068.1 615509.631 386.1 91864 D2 SHW URS Active 86400 MEI80k 16 / 3.3* 512 4034088.1 615093.632 372.3 90223 D2 SHW URS Active 86400 MEI80k 16 / 3.3* 512 4034499.7 615008.233 359.2 88583 D2 SHW URS Active 86400 MEI80k 16 / 3.3* 512 4034973.5 614973.534 346.3 86942 D2 SHW URS Active 86400 MEI80k 16 / 3.3* 512 4035415.8 615145.135 333.7 85302 D2 SHW URS Active 86400 MEI80k 16 / 3.3* 512 4035843.5 615212.8153B.6. Node Initialization36 320.3 83661 D2 SHW URS Active 81900 MEI80k 16 / 3.3* 512 4036292.7 615198.337 306.9 82021 D3A RR URS Active 81900 TU trib MEI80k 16 / 3.3* 512 4036724.7 615244.538 296.0 80381 D3A RR URS Active 81900 MEI80k 16 / 3.3* 512 4036986.4 614882.939 285.4 78740 D3A RR URS Active 81900 MEI80k 16 / 3.3* 512 4037387.0 614597.140 275.6 77100 D3A RR URS Active 81900 MEI80k 16 / 3.3* 512 4037322.1 614196.041 266.3 75459 D3A RR URS Active 71400 MEI80k 16 / 3.3* 512 4037335.6 613709.442 257.3 73819 D3A RR URS Active 71400 HI trib MEI60k 16 / 3.3* 512 4037513.1 613262.643 248.2 72178 D3A RR URS Active 71400 MEI60k 16 / 3.3* 512 4037602.1 612785.944 239.0 70538 D3A RR URS Active 71400 MEI60k 16 / 3.3* 512 4038088.3 612720.345 229.4 68898 D3A RR URS Active 71400 MEI60k 16 / 3.3* 512 4038397.6 612388.646 219.6 67257 D3A RR URS Active 71400 MEI60k 16 / 3.3* 512 4038838.0 612173.047 209.9 65617 D3B RR URS Active 62100 MEI60k 16 / 3.3* 512 4039263.2 611918.748 201.1 63976 D3B DJ URS Active 62100 GA trib MEI60k 16 / 3.3* 512 4039648.7 611620.349 191.5 62336 D3B DJ URS Active 62100 MEI60k 16 / 3.3* 512 4039862.7 611191.050 182.2 60696 D3B DJ URS Active 62100 MEI60k 16 / 3.3* 512 4040236.7 610871.251 174.7 59055 D3B DJ URS Active 62100 MEI60k 16 / 3.3* 512 4040716.9 610760.452 167.3 57415 D3B DJ URS Active 62100 MEI60k 16 / 3.3* 512 4041132.5 610516.853 160.4 55774 D3B DJ URS Active 62100 MEI60k 16 / 3.3* 512 4041325.0 610056.354 152.3 54134 D3B DJ URS Active 62100 MEI60k 16 / 3.3* 512 4041488.4 609606.355 144.3 52493 D3B DJ URS Active 42400 MEI60k 16 / 3.3* 512 4041704.0 609182.556 135.4 50853 D3B DJ URS Active 42400 MEI60k 16 / 3.3* 512 4041888.8 608735.357 127.9 49213 D3B DJ URS Active 42400 MEI40k 16 / 3.3* 512 4041868.4 608249.058 121.2 47572 D4A DJ URS Active 42400 MEI40k 16 / 3.3* 512 4042226.1 607903.659 116.0 45932 D4A NC URS Active 42400 MEI40k 16 / 3.3* 512 4042518.6 607510.660 111.3 44291 D4A NC URS Active 42400 MEI40k 16 / 3.3* 512 4042482.2 607063.561 106.2 42651 D4A NC URS Active 42400 MEI40k 16 / 3.3* 512 4042275.7 606614.262 101.2 41010 D4A NC URS Active 42400 RC trib MEI40k 16 / 3.3* 512 4042241.9 606178.463 95.8 39370 D4A NC URS Active 42400 MEI40k 16 / 3.3* 512 4042645.5 605892.464 90.3 37730 D4A NC URS Active 35900 MEI40k 16 / 3.3* 512 4042857.4 605458.865 85.4 36089 D4A NC URS Active 35900 MEI40k 16 / 3.3* 512 4042717.4 604998.966 79.7 34449 D4A NC URS Active 35900 MEI40k 16 / 3.3* 512 4042805.1 604643.267 74.1 32808 D4A NC URS Active 35900 MEI40k 16 / 3.3* 512 4043196.8 604349.368 69.9 31168 D4A NC URS Active 29200 MEI40k 16 / 3.3* 512 4043122.1 603964.069 64.3 29528 D4A NC URS Active 29200 MEI40k 16 / 3.3* 512 4042844.6 603586.770 60.1 27887 D4A NC URS Active 29200 MEI40k 16 / 3.3* 512 4043033.6 603136.271 55.2 26247 D4A NC URS Active 29200 MEI40k 16 / 3.3* 512 4043023.9 602667.572 50.6 24606 D4A NC URS Active 29200 MEI40k 16 / 3.3* 512 4043321.2 602282.673 47.1 22966 D4A NC URS Active 29200 MEI40k 16 / 3.3* 512 4043728.2 602073.174 42.6 21325 D4A NC SC MEI40k 2 512 4043944.1 601631.675 38.2 19685 D4A NC SC PO trib MEI40k 2 512 4044089.8 601214.976 35.5 18045 D4A NC SC MEI40k 2 512 4044248.6 600782.777 32.5 16404 D4A NC SC MEI40k 2 512 4044302.9 600296.278 27.5 14764 D4B NC CRO MEI40k 2 512 4044024.2 599955.079 25.8 13123 D4B HWY1 CRO MEI40k 2 512 4044148.5 599604.580 22.0 11483 D4B HWY1 CRO MEI40k 2 512 4044210.6 599137.181 17.8 9843 D4B HWY1 CRO MEI40k 2 512 4043944.3 598734.082 14.0 8202 D4B HWY1 CRO MEI40k 2 512 4043978.5 598244.683 10.6 6562 D4B HWY1 CRO MEI40k 2 512 4043874.1 597757.584 8.3 4921 D4B HWY1 CRO MEI40k 2 512 4043999.9 597316.485 4.9 3281 D4B HWY1 CRO MEI40k 2 512 4044321.8 596935.886 2.1 1640 D4B HWY1 CRO MEI40k 2 512 4044370.3 596514.687 0.0 0 D4B HWY1 CRO MEI40k 2 512 4044017.2 596177.6154B.7. Grain Size DistributionsB.7 Grain Size DistributionsFigure B.19: Grain size distributions used for the node initialization. SC and CROwere provided by Douglas Smith, who reported them in the report ’2015Pre-San Clemente Dam Removal Morphological Monitoring of the CarmelRiver Channel in Monterey County, California’. The three MEI40k, 60kand 80k grain size distributions were compiled from plots the MEI 2002San Clemente dam modelling report. The URS Active grain size distribu-tions were compiled from the URS modelling report. The ’only256mm’and ’only512mm’ grain size distributions were used to artificially preventerosion of deeper subsurface layers.155B.8. Reach Averaged CrosssectionsB.8 Reach Averaged CrosssectionsFigure B.20: Averaged crosssection for reach D5Figure B.21: Averaged crosssection for reach D4B156B.8. Reach Averaged CrosssectionsFigure B.22: Averaged crosssection for reach D4AFigure B.23: Averaged crosssection for reach D3B157B.8. Reach Averaged CrosssectionsFigure B.24: Averaged crosssection for reach D3AFigure B.25: Averaged crosssection for reach D2158B.8. Reach Averaged CrosssectionsFigure B.26: Averaged crosssection for reach D1B159B.9. Reach Flow CalculationsB.9 Reach Flow CalculationsFigure B.27: Flow velocity from discharge, based on data from MEI (2003)Figure B.28: Averaged crosssection for reach D5160B.9. Reach Flow CalculationsFigure B.29: Averaged crosssection for reach D4BFigure B.30: Averaged crosssection for reach D4A161B.9. Reach Flow CalculationsFigure B.31: Averaged crosssection for reach D3BFigure B.32: Averaged crosssection for reach D3A162B.9. Reach Flow CalculationsFigure B.33: Averaged crosssection for reach D2Figure B.34: Averaged crosssection for reach D1B163B.10. Median Sediment Supply and Reservoir Depletion CurvesB.10 Median Sediment Supply and Reservoir DepletionCurves164B.10. Median Sediment Supply and Reservoir Depletion CurvesFigure B.35: Uncontrolled Supply (Exp1-3) scenarios. Each line represents themedian of 100 simulations.165B.10. Median Sediment Supply and Reservoir Depletion CurvesFigure B.36: Median reservoir depletion in the Pulse Supply (RC1-6) and Uncon-trolled Supply (Exp1-3) scenarios. Each line represents the median of 100simulations.166B.11. Detailed Results of the focus scenariosB.11 Detailed Results of the focus scenariosB.11.1 No Action Simulation ResultsWe review the No Action project alternative results for the five model reaches running fromLos Padres Dam to the Pacific Ocean. Results are summarized in Figure B.37 through Fig-ure B.41 for the average hydrologic condition. Within each Figure the top plot illustratesresults for the wet hydrologic condition, the middle plot for the average condition and thebottom plot for the dry condition. Each profile-type line plot illustrates results for 100 simula-tions at three different simulation times: 10-, 30- and 60-years. To highlight the most probableresponse trajectory for the simulations, we also plot the median response for the 100 simula-tions at each simulation time; the median responses are shown as the thicker lines in the plots.We focus on the following figures:• Figure B.37 shows the simulated change in channel bed elevation;• Figure B.38 shows the simulated change of the bed surface median grain size (Dg);• Figure B.39 shows the simulated change of the bed surface 90th-percentile grain size(D90);• Figure B.40 shows the simulated change of the longitudinal channel bed slope; and• Figure B.41 shows the simulated unit bedload sediment transport rate.Downstream of Los Padres (42-32 km): As expected, reaches downstream of Los PadresDam up to channel station 35 km are projected to degrade for the wet, average and dry hy-drologic conditions. The median simulation projection of channel bed degradation at year 60ranges from 0 to roughly -1.5 m relative to initial channel bed elevations (circa 2017) acrossall three hydrologic conditions. The location of the most severe degradation along this seg-ment appears to follow an existing spatial trend of downstream increasing bed slope (see Fig-ure B.40). Some of the eroded bed material is transported downstream and deposited betweenstations 32 km and 35 km, where the downstream trend of bed slopes is presently decreasing.167B.11. Detailed Results of the focus scenariosFigure B.37: Change in elevation compared to the initial elevation profile from2017. While BESMo reports only change in sediment volume, we trans-lated the volume change to an elevation change by using the averagedcrosssection profiles shown in Appendix B.8 and converting the sedimentstorage from cross sectional area to depth using the same ratios shown inAppendix B.9 for flow area to flow depth. In the case of erosion, we as-sumed a rectangular cross section with constant channel width. Top: high,Middle: average, bottom: low cumulative discharge in the first 10 years.Per subplot the profiles of 100 simulations is shown in 10, 30, and 60-yeartime slices. The three solid lines in each subplot signify the median condi-tion at each node for each of the three time slices; all other lines representdata from individual model runs.In all cases, timing of the largest projected changes to bed elevation occur during the simula-tion period of the largest floods. Both the Dg and the D90 are projected to increase from 15-168B.11. Detailed Results of the focus scenariosFigure B.38: Mean surface grain size (Dg) for top: high, Middle: average, bottom:low cumulative discharge in the first 10 years. Per subplot the profiles of100 simulations is shown in 10, 30, and 60-year time slices. The three solidlines in each subplot signify the median condition at each node for each ofthe three time slices; all other lines represent data from individual modelruns.to 40% within this section of the channel. Transport rates here are comparably low due to therelatively low discharge and low bedload sediment supply.Former San Clemente Dam project reach (32.5-30.5 km): The simulations project twodominant responses within the former San Clemente project reach: channel bed degradationwithin the former reservoir deposits, and extending about 1000 m upstream, and downstreamprogressive channel bed aggradation from approximately the San Clemente Creek confluence169B.11. Detailed Results of the focus scenariosFigure B.39: Coarse fractions of the surface grain size expressed as 90th percentilesize (D90) for top: high, Middle: average, bottom: low cumulative dis-charge in the first 10 years. Per subplot the profiles of 100 simulations isshown in 10, 30, and 60-year time slices. The three solid lines in each sub-plot signify the median condition at each node for each of the three timeslices; all other lines represent data from individual model runs.to the former dam site. The median simulation projection of channel bed degradation at year60 ranges from 0 to about -1.5 m relative to initial channel bed elevations (circa 2017) acrossall three hydrologic conditions. The location of the most severe degradation response occursjust upstream of the reroute reach. Notably, the dry scenario produces the largest magnitudedegradation response because a majority of the bedload transporting flows occur later in thesimulation. The median simulation projection of channel bed aggradation at year 60 ranges is170B.11. Detailed Results of the focus scenariosFigure B.40: Slope profile for top: high, Middle: average, bottom: low cumulativedischarge in the first 10 years. Per subplot the profiles of 100 simulationsis shown in 10, 30, and 60-year time slices. The three solid lines in eachsubplot signify the median condition at each node for each of the threetime slices; all other lines represent data from individual model runs.comparatively similar to the degradation case, ranging from 0 to about +1.5 m relative to initialchannel bed elevations (circa 2017) across all three hydrologic conditions. The location of themost severe aggradation response is coincident with the former dam location node, or withina few nodes downstream. Also, the largest magnitude aggradation is produced under the dryscenario. Erosion within the former reservoir deposit is a function of the relatively fine bedsurface GSD there, and projected deposition downstream suggests that the model is evolv-171B.11. Detailed Results of the focus scenariosFigure B.41: Averaged transport rate between time slices (Year 10: data from 0-10years, Year 30: data from 10-30 years, Year 60: data from 30-60 years). Fortop: high, Middle: average, bottom: low cumulative discharge in the first10 years. Per subplot the profiles of 100 simulations is shown in 10, 30,and 60-year time slices. The three solid lines in each subplot signify themedian condition at each node for each of the three time slices; all otherlines represent data from individual model runs.ing to an overall flatter profile through the San Clemente project reach. This is a reasonableprojected model result because the San Clemente project reach is a significant downstreamperturbation to the longitudinal slope trend. Erosion within the upstream segment and depo-sition downstream produces an overall significant coarsening of the Dg and D90 bed surfacegrain sizes. Coarsening of the Dg ranges from 10s to many 100s of mm coarser than initial172B.11. Detailed Results of the focus scenariosconditions, and coarsening of the D90 ranges from 10s of mm up to roughly 1500 mm morecoarse than initial. This suggests that the projected spatial pattern of changes to bed elevationand bed slope may be limited beyond the 60-year model simulation period.Between former San Clemente Dam and Tularcitos creek (30.5-25.5 km): The simula-tions indicate that aggradational responses simulated within the downstream part of the SanClemente project reach progressively diminish moving downstream through this reach. As aresult, this reach is a buffer for upstream changes and therefore shows a large range of tempo-ral and spatial variation across the three hydrologic conditions and the 300 simulations. Ele-vation responses range from upwards of +3.7 m relative to initial channel bed elevations (circa2017) just downstream of the former dam to approximately +1.5 m by year 60. As with thetwo upstream reaches, the timing of bed elevation response is governed by the sequencing offlood events. However, unlike the two upstream reaches, the simulated range and magnitudeof elevation responses is similar across the hydrologic conditions. All simulations indicatethat the Dg and D90 bed surface grain sizes will coarsen over time, and that the downstreampropagation of the D90 response is stronger relative to the Dg.Tularcitos to Robinson Canyon Creek (25.5-13 km): The simulations indicate a generaltendency for channel bed aggradation downstream of station 20.5 km, with little net changesuggested for locations upstream of this station. Aggradation through the lower 8 km of thisreach may be associated with downstream occurrence of the Narrows, where the depositionalsignal begins to steadily diminish moving downstream. The Narrows could trigger local de-position with resumption of bedload supply from the watershed in between Los Padres andSan Clemente, and this response could then propagate upstream. Across the three hydro-logic conditions, net deposition ranges from 0 to upwards of 2 m by year 60 relative to initialchannel bed elevations (circa 2017). The Dg and D90 bed surface grain sizes show generalcoarsening trends across the three hydrologic conditions, but the magnitude of coarsening di-minishes in the downstream direction for the 60-year simulation. The D90 shows variation inthe downstream extent of coarsening relative to the Dg, suggesting that spatial propagationof the coarsening response is dependent on the sequencing of flood events. Coarsening of theDg ranges from roughly 20 to 100 mm coarser than initial conditions over the 60-year simu-173B.11. Detailed Results of the focus scenarioslation, whereas coarsening of the D90 ranges upwards to about 200 mm coarser than initialconditions.Robinson Canyon Creek to Outlet (13-0 km): The lower most reach of the simulation do-main shows two general spatial and temporal trends with respect to bed elevation: little netchange of bed topography immediately downstream of the Narrows and aggradation alongthe lower most 9 km of the river. Across the three hydrologic conditions, aggradation by year60 ranges from +0.6 to +1.5 m relative to initial channel bed elevations (circa 2017), dimin-ishing to 0 at the downstream most model node at the Pacific Ocean. Coarsening of the Dgand D90 bed surface grain sizes continue through this reach, and steadily diminishes movingdownstream toward the ocean.Synthesis For the No Action project simulation at Los Padres Dam, persistence of low relativesediment supply downstream of Los Padres Dam is simulated to drive further channel beddegradation to roughly channel station 35 km. Downstream of this station, BESMo projectsthe most significant spatial gradients in channel response likely due to constructed channelconditions at the San Clemente project reach. Strong profile adjustment here suggests thatconstructed conditions are not in steady-state with upstream projected supplies of water andbedload sediment for the No Action project simulation. In general, the former reservoir de-posit area is a location of channel bed erosion, with deposition downstream of San ClementeCreek up and to the former San Clemente Dam site.The reach from the former dam site to Tularcitos Creek is a response transition reach, hav-ing a somewhat wide range in the magnitude and spatial extent of aggradation. As a result,the evolution of future conditions from the former San Clemente Dam site to Tularcitos Creekare particularly sensitive to the sequencing and magnitude of future floods. A general aggra-dation response of +1.5 m relative to initial channel bed elevations is simulated from GarzasCreek to the Narrows, followed by little to no net bed elevation change downstream of theNarrows to station 9 km. The lower most 9 km have a consistent aggradation response rang-ing from +0.6 to +1.5 m relative to initial channel bed elevations.The channel bed surface is projected to coarsen throughout the simulation reach, from Los174B.11. Detailed Results of the focus scenariosPadres Dam to the Pacific Ocean, despite the reintroduction of bedload supply from in be-tween Los Padres and the former San Clemente Dam to downstream reaches. This simulationresponse highlights that the magnitude and gradation of the reintroduced bedload supply isinsufficient to limit general bed coarsening, which is substantial over much of the model do-main (i.e. factor 2 increase of the Dg at a minimum). This result for the No Action projectsimulation is not encouraging for steelhead or resident trout. The sequencing of floods gov-erns the timing of profile adjustment from Los Padres to the Pacific Ocean, particularly fromthe former San Clemente Dam site to Tularcitos Creek, with earlier floods driving early change,and later floods driving later change. However, the general magnitude of profile response isindependent of large flood timing of the 60- year simulation. These set of results raise theexpectation that resumption of bedload supply from the area upstream of Los Padres Damwill have spatial patterns of response similar to those for the No Action project simulation,downstream of the former San Clemente Dam site, however the magnitudes may be morepronounced.B.11.2 Historical Supply Simulation ResultsThe results of the Historical Supply simulations are shown in the following figures:• Figure B.42 shows the simulated change in channel bed elevation;• Figure B.43 shows the simulated change of the bed surface median grain size (Dg);• Figure B.44 shows the simulated change of the bed surface 90th-percentile grain size(D90);• Figure B.45 shows the simulated change of the longitudinal channel bed slope; and• Figure B.46 shows the simulated unit bedload sediment transport rate.Resumption of the estimated long-term average natural sediment supply of 13,400 m3/yrto reaches downstream of Los Padres Dam result in the projection of nearly 4.6 m of aggra-dation just downstream of the Dam. Using historical topographic data, we estimate that ap-proximately 6 m of net bed erosion has occurred downstream of the dam since construction in175B.11. Detailed Results of the focus scenariosFigure B.42: Change in elevation compared to the initial elevation profile from2017. While BESMo reports only change in sediment volume, we trans-lated the volume change to an elevation change by using the averagedcrosssection profiles shown in Appendix B.8 and converting the sedimentstorage from cross sectional area to depth using the same ratios shown inAppendix B.9 for flow area to flow depth. In the case of erosion, we as-sumed a rectangular cross section with constant channel width. Top: high,Middle: average, bottom: low cumulative discharge in the first 10 years.Per subplot the profiles of 100 simulations is shown in 10, 30, and 60-yeartime slices. The three solid lines in each subplot signify the median condi-tion at each node for each of the three time slices; all other lines representdata from individual model runs.1949. As a result, simulated bedload deposition downstream of the dam of up to 4.6 m overthe 60-year simulation time period is plausible. The aggradation response at the dam lessens176B.11. Detailed Results of the focus scenariosFigure B.43: Mean surface grain size (Dg) for top: high, Middle: average, bottom:low cumulative discharge in the first 10 years. Per subplot the profiles of100 simulations is shown in 10, 30, and 60-year time slices. The three solidlines in each subplot signify the median condition at each node for each ofthe three time slices; all other lines represent data from individual modelruns.downstream as expected, but increases up to an estimated 1.5 m as the former San Clementereservoir backwater zone and deposit is approached. This response also makes sense becausethe average bed slope through this formerly reservoir affected region is flatter than the pre-dam condition.Careful inspection of Figure B.42 also reveals that episodes of bed erosion are simulatedwithin the first 10 years of the simulation within the former reservoir affected region and177B.11. Detailed Results of the focus scenariosFigure B.44: Coarse fractions of the surface grain size expressed as 90th percentilesize (D90) for top: high, Middle: average, bottom: low cumulative dis-charge in the first 10 years. Per subplot the profiles of 100 simulations isshown in 10, 30, and 60-year time slices. The three solid lines in each sub-plot signify the median condition at each node for each of the three timeslices; all other lines represent data from individual model runs.upstream for primarily average and dry hydrologic conditions. We assume this is due torelatively low local sediment supply as a result of deposition downstream of Los Padres Dam(Figure B.46). As the deposition downstream of the dam continues in time, bedload supplyincreases and all hydrologic conditions tend toward a similar spatial pattern of bed profileresponse in between Los Padres Dam and the upstream end of the San Clemente project reach.Downstream of the former San Clemente Dam all three hydrologic conditions project up178B.11. Detailed Results of the focus scenariosFigure B.45: Slope profile for top: high, Middle: average, bottom: low cumulativedischarge in the first 10 years. Per subplot the profiles of 100 simulationsis shown in 10, 30, and 60-year time slices. The three solid lines in eachsubplot signify the median condition at each node for each of the threetime slices; all other lines represent data from individual model runs.to 2.1 m of bed deposition, with relaxation of net bed aggradation toward Hitchcock Creek.This result is similar to the result discussed and presented above for the No Action updatedsimulation. The increased local sediment budget is driven by redistribution of material fromthe reroute reach at San Clemente Creek.Downstream of Hitchcock Creek the Historical Supply simulations project a varied re-sponse of a net elevation change between -0.6 and +0.3 m by the end of the 60-year simulation179B.11. Detailed Results of the focus scenariosFigure B.46: Averaged transport rate between time slices (Year 10: data from 0-10years, Year 30: data from 10-30 years, Year 60: data from 30-60 years). Fortop: high, Middle: average, bottom: low cumulative discharge in the first10 years. Per subplot the profiles of 100 simulations is shown in 10, 30,and 60-year time slices. The three solid lines in each subplot signify themedian condition at each node for each of the three time slices; all otherlines represent data from individual model runs.time period. This response is consistent across all three hydrologic conditions, and the profileis about 0.3 to 0.6 m higher than in the No Action simulation for the whole lower part of theriver. Comparability of the aggradational response and magnitude along the lower 23 km ofthe Carmel River across the three hydrologic conditions for the Historical Supply conditionssuggests deposition is likely to occur and independently of hydrology, given sufficient time180B.11. Detailed Results of the focus scenariosfor the channel to respond. A similar conclusion was drawn for the No Action simulation.Downstream of Los Padres Dam, the channel bed is projected to coarsen. With additionalsediment supplied from upstream, we would expect increased sediment mobility and bedmixing, eventually winnowing the finer grain sizes from both the surface and shallow sub-surface. The trajectory of the surface D90 is dependent upon the hydrologic conditions of thefirst 10 years of the simulation, with wet years producing a coarser bed, likely as a result ofhigh mobility and bed mixing. Results from the end of the 60-year simulation run times aresimilar between the hydrologic scenarios, with the coarsest bed just downstream of the LosPadres Dam and near the former San Clemente Dam site. In each of the hydrologic condi-tions, a coarser D90 has prograded downstream to Las Garzas Creek in wet conditions, andjust upstream of Las Garzas Creek in average and dry conditions.Compared to the No Action simulation, the geometric mean grain size (Dg), has more vari-ability in the lower reaches across the 60-year simulation time frame, suggesting an overallhigher sediment mobility throughout the simulation period in the lower reaches. The Histori-cal Supply simulation also produces a mean surface grain size similar to the model input grainsize distribution (indicated on the plots by the dashed cyan line, hidden behind the simulationresults), and is finer than the final Dg for the No Action simulation (dashed black line). Thisresult suggests that resumption of the Historical Supply to the river downstream of Los Padresmay have tangible benefits for steelhead habitat conditions within the simulation time periodof 60 years. It further implies that the Historical Supply is sufficient to prevent further bedsurface coarsening along the lower river downstream of the former San Clemente Dam, rela-tive to initial conditions. This was not the case for the No Action simulation results where weconcluded that the additional bedload sediment supply sourced in between Los Padres Damand the former San Clemente Dam, as represented in the model, was insufficient to preventnet bed surface coarsening along the lower Carmel River over the simulation time period.B.11.3 Pulsed Supply Simulation ResultsTo simplify the discussion, we will only describe the results for the pulse scenario using ratingcurve RC4. The results of the other curves are presented as median values in Appendix B.13.181B.11. Detailed Results of the focus scenariosFollowing the structure from the previous results, we review the Pulsed Supply simulation inthe following figures:• Figure B.47 shows the simulated change in channel bed elevation;• Figure B.48 shows the simulated change of the bed surface median grain size (Dg);• Figure B.49 shows the simulated change of the bed surface 90th-percentile grain size(D90);• Figure B.50 shows the simulated change of the longitudinal channel bed slope; and• Figure B.51 shows the simulated unit bedload sediment transport rate.The sediment supply from rating curve RC4 to reaches downstream of Los Padres Damcauses up to 5.8 m of aggradation just downstream of the Dam (see Figure B.47). This isabout 1.2 m more than in the Historical scenario and within the range of measured historicalelevation profiles below the dam. The aggradation response at the dam lessens downstreambut increases up to an estimated 2.4 m (Historical supply 1.5 m) as the former San Clementereservoir backwater zone and deposit is reached.The channel between Los Padres Dam and the upstream end of the San Clemente projectreach shows a 0.3-0.9 m higher deposition compared with the Historical Supply simulation,with an exception between Pine creek and Cachagua creek where both simulations show noelevation change.Downstream of the former San Clemente Dam all three hydrologic conditions result inup to 2.7 m of sediment deposition, which steadily decreases until Tularcitos Creek, where thePulsed Supply scenario shows about 0.5 m less deposition than the Historical Supply scenario.At the request of the TRC, we apply a 50% increase in boundary shear stress resistance to pre-vent perhaps unrealistic erosion because of incomplete bed surface and subsurface grain sizedistribution data. After 60 years, the Pulsed Supply scenario RC4 does not show significanterosion or deposition within 3 km upstream and downstream of Garzas Creek, which agreeswith the Historical Supply scenario. Further downstream of this reach and up to the mouth of182B.11. Detailed Results of the focus scenariosFigure B.47: Change in elevation compared to the initial elevation profile from2017. While BESMo reports only change in sediment volume, we trans-lated the volume change to an elevation change by using the averagedcrosssection profiles shown in Appendix B.8 and converting the sedimentstorage from cross sectional area to depth using the same ratios shown inAppendix B.9 for flow area to flow depth. In the case of erosion, we as-sumed a rectangular cross section with constant channel width. Top: high,Middle: average, bottom: low cumulative discharge in the first 10 years.Per subplot the profiles of 100 simulations is shown in 10, 30, and 60-yeartime slices. The three solid lines in each subplot signify the median condi-tion at each node for each of the three time slices; all other lines representdata from individual model runs.the river we observe relatively consistent deposition of 1.5 m of sediment after 60 simulationyears, which is about 0.3 m more deposition than in the Historical Supply scenario.183B.11. Detailed Results of the focus scenariosFigure B.48: Mean surface grain size (Dg) for top: high, Middle: average, bottom:low cumulative discharge in the first 10 years. Per subplot the profiles of100 simulations is shown in 10, 30, and 60-year time slices. The three solidlines in each subplot signify the median condition at each node for each ofthe three time slices; all other lines represent data from individual modelruns.Comparability of the aggradational response and magnitude along the lower 23 km of theCarmel River across the three hydrologic conditions for the Pulsed Supply conditions sug-gests deposition is likely to occur and independently of hydrology, given enough time for thechannel to respond. A similar conclusion was drawn for the No Action and Historical Supplysimulations.The median grain size adjustments (Dg, Figure B.48), the Pulsed Supply scenario shows184B.11. Detailed Results of the focus scenariosFigure B.49: Coarse fractions of the surface grain size expressed as 90th percentilesize (D90) for top: high, Middle: average, bottom: low cumulative dis-charge in the first 10 years. Per subplot the profiles of 100 simulations isshown in 10, 30, and 60-year time slices. The three solid lines in each sub-plot signify the median condition at each node for each of the three timeslices; all other lines represent data from individual model runs.similar conditions as the Historical scenario. The biggest difference occurs below the LosPadres reservoir, where the sediment feed introduces more fine material. The coarse frac-tions signified by the 90th-percentile grain sizes (D90, Figure B.49) show a distinct coarseningin response to the increased sediment supply. This might be caused by the increased mobilityof the finer material, as the transport function used in BESMo will cause higher transport ratesfor higher contents of sand in the bed.185B.11. Detailed Results of the focus scenariosFigure B.50: Slope profile for top: high, Middle: average, bottom: low cumulativedischarge in the first 10 years. Per subplot the profiles of 100 simulationsis shown in 10, 30, and 60-year time slices. The three solid lines in eachsubplot signify the median condition at each node for each of the threetime slices; all other lines represent data from individual model runs.In conclusion, the elevation profile modestly aggrades in the Pulsed Supply scenario RC4compared to the Historical Supply scenario, which follows the large increase in sediment sup-ply due to the inclusion of the reservoir deposits (over 60 years Historical Supply: 1,087,900m3, Pulsed Supply: 1,667,700 m3). Fining of the bed close to Los Padres reservoir is causedby the increased supply of relatively fine fractions. It is notable that the surface grain sizedoes not seem to be significantly changed between the two scenarios, even though the aver-186B.11. Detailed Results of the focus scenariosFigure B.51: Averaged transport rate between time slices (Year 10: data from 0-10years, Year 30: data from 10-30 years, Year 60: data from 30-60 years). Fortop: high, Middle: average, bottom: low cumulative discharge in the first10 years. Per subplot the profiles of 100 simulations is shown in 10, 30,and 60-year time slices. The three solid lines in each subplot signify themedian condition at each node for each of the three time slices; all otherlines represent data from individual model runs.age transport rates throughout the river (Figure B.51) increases by often a factor of two. Wewant to note that the Pulse scenario RC3, which evacuates reservoir sediment more slowly,projects a finer bed grain size distribution (Dg 10-40% lower). This implies that a more con-stant sediment feed from the Los Padres reservoir is more effective in fining the bed than asteeper, pulse-like rating curve.187B.11. Detailed Results of the focus scenariosOur findings suggest that the Pulsed sediment supply may have tangible benefits for steel-head habitat conditions within the simulation time period of 60 years. These simulations fur-ther show that the rate of supply of sediment from the reservoir has a strong impact on thebed surface composition along the whole river.B.11.4 Uncontrolled Release Simulation ResultsTo simplify the discussion, we will only describe the results for the pulse scenario using the ex-ponential decay curve Exp2. The results of the other curves are presented as median values inAppendix B.13. Following the structure from the previous results, we review the UncontrolledRelease simulation in the following figures:• Figure B.52 shows the simulated change in channel bed elevation;• Figure B.53 shows the simulated change of the bed surface median grain size (Dg);• Figure B.54 shows the simulated change of the bed surface 90th-percentile grain size(D90);• Figure B.55 shows the simulated change of the longitudinal channel bed slope; and• Figure B.56 shows the simulated unit bedload sediment transport rate.The sediment supply from the Uncontrolled Release scenario Exp 2 to reaches downstreamof Los Padres Dam causes up to 6.7 m of aggradation just downstream of the Dam (see Fig-ure B.52). This is about 2 m more than in the Historical scenario. The aggradation response atthe dam lessens downstream but increases up to an estimated 2.7 m (Hist. 2 m) as the formerSan Clemente reservoir backwater zone and deposit is approached. The channel between LosPadres Dam and the upstream end of the San Clemente project reach shows a 0.6-1.8 m higherdeposition than in the Historical scenario, which is about double than in the Pulsed supplyscenario RC4.Downstream of the former San Clemente Dam all three hydrologic conditions project upto 4 m of sediment deposition, which steadily decreases until Hitchcock Creek, with the un-controlled release scenario showing about 0.6 m more deposition than the Historical scenario.188B.11. Detailed Results of the focus scenariosFigure B.52: Change in elevation compared to the initial elevation profile from2017. While BESMo reports only change in sediment volume, we trans-lated the volume change to an elevation change by using the averagedcrosssection profiles shown in Appendix B.8 and converting the sedimentstorage from cross sectional area to depth using the same ratios shown inAppendix B.9 for flow area to flow depth. In the case of erosion, we as-sumed a rectangular cross section with constant channel width. Top: high,Middle: average, bottom: low cumulative discharge in the first 10 years.Per subplot the profiles of 100 simulations is shown in 10, 30, and 60-yeartime slices. The three solid lines in each subplot signify the median condi-tion at each node for each of the three time slices; all other lines representdata from individual model runs.Per request of the TRC, in this section of the river we apply a 50increase in boundary shearstress resistance to prevent an erosional signal which occurred in early runs of the Historical189B.11. Detailed Results of the focus scenariosFigure B.53: Mean surface grain size (Dg) for top: high, Middle: average, bottom:low cumulative discharge in the first 10 years. Per subplot the profiles of100 simulations is shown in 10, 30, and 60-year time slices. The three solidlines in each subplot signify the median condition at each node for each ofthe three time slices; all other lines represent data from individual modelruns.scenario.Within 3 km upstream and downstream of Garzas Creek the Uncontrolled Release after60 years shows between 0.6 m of erosion under wet hydrographs, or between 0.6 m erosionand 0.6 m deposition under average and dry hydrographs, which agrees with the Historicalscenario. Further downstream of this reach and up to the mouth of the river we observerelatively consistent deposition of 1.5 m of sediment after 60 simulation years. The trend is190B.11. Detailed Results of the focus scenariosFigure B.54: Coarse fractions of the surface grain size expressed as 90th percentilesize (D90) for top: high, Middle: average, bottom: low cumulative dis-charge in the first 10 years. Per subplot the profiles of 100 simulations isshown in 10, 30, and 60-year time slices. The three solid lines in each sub-plot signify the median condition at each node for each of the three timeslices; all other lines represent data from individual model runs.about 0.3 m more elevation change than in the Historical scenario and matches data from thePulsed Supply scenarios.The comparable aggradational response along the lower 23 km of the Carmel River acrossthe three hydrologic conditions for both the Uncontrolled Release and the Pulsed Supply sim-ulations suggests that the deposition is likely to occur independently of hydrology if givenenough time for the channel to respond. A similar conclusion was drawn for the No Action191B.11. Detailed Results of the focus scenariosFigure B.55: Slope profile for top: high, Middle: average, bottom: low cumulativedischarge in the first 10 years. Per subplot the profiles of 100 simulationsis shown in 10, 30, and 60-year time slices. The three solid lines in eachsubplot signify the median condition at each node for each of the threetime slices; all other lines represent data from individual model runs.and Historical Supply simulations.In regards of median grain size adjustments (Dg, Figure B.53), the Uncontrolled Releasescenario shows similar conditions as both the Pulsed Supply and the Historical scenarios,except that under the dry hydrograph more fine material leaves the reservoir and reduces theaverage grain size below the Los Padres dam in the early phase of the simulations (10 yearlines). The coarse fractions signified by the 90th-percentile grain sizes (D90, Figure B.54) show192B.11. Detailed Results of the focus scenariosFigure B.56: Averaged transport rate between time slices (Year 10: data from 0-10years, Year 30: data from 10-30 years, Year 60: data from 30-60 years). Fortop: high, Middle: average, bottom: low cumulative discharge in the first10 years. Per subplot the profiles of 100 simulations is shown in 10, 30,and 60-year time slices. The three solid lines in each subplot signify themedian condition at each node for each of the three time slices; all otherlines represent data from individual model runs.a similar response, also mainly in the first 10 years of the simulations. At the 30 and 60 yearmarks the bed surface is generally coarser than in the Historical scenario, which is due tothe increased mobility of the bed due to the finer material supplied by the reservoir, as thetransport function used in BESMo will cause higher transport rates for higher contents of sandin the bed.193B.11. Detailed Results of the focus scenariosIn conclusion, the elevation profile increases moderately in the Uncontrolled Release sce-nario Exp 2 in comparison to the Historical scenario, which stands in contrast to the largeincrease in sediment supply due to the inclusion of the reservoir deposits (over 60 years His-torical: 1,087,900 m3, Exp 2: 1,541,800 m3). Fining of the bed close to Los Padres reservoir iscaused by the increased supply of relatively fine fractions. In comparison to the Pulsed Supplyscenarios, this fine sediment is introduced mainly within the first 10-20 years of the simula-tions. It is notable that the surface grain size does not seem to be significantly changed betweenthe scenarios, even though the average transport rates throughout the river (Figure B.56) in-creases by often a factor of two.In comparison to the Pulsed Supply scenarios, the Uncontrolled Release scenarios showmore change in grain size and bed surface elevation within the first years of the simulations.Generally, in the Pulsed Supply scenarios the sediment feed from the reservoir is spread overa longer time frame, which leads to a more continuous supply of fine material into the upperCarmel River. Due to the quick decrease in supply rates in the Uncontrolled Release scenarios,the bed surface coarsens stronger.194B.12. Median Results of focus ScenariosB.12 Median Results of focus ScenariosB.12.1 Average HydrographsFigure B.57: Comparison of projected bed elevation change from the 2017 initialprofile for the average hydrologic condition and the Historical, Pulse andUncontrolled sediment supply simulations. Shaded regions capture the25th-75th percentile responses across the 100 simulations for the averagecondition. Results shown for simulation year 1, 10, 30 and 60.195B.12. Median Results of focus ScenariosFigure B.58: Comparison of projected change of the geometric mean grain size ofthe bed surface Dg for the average hydrologic condition and the Histori-cal, Pulse and Uncontrolled sediment supply simulations. Shaded regionscapture the 25th-75th percentile responses across the 100 simulations forthe average condition. Results shown for simulation year 1, 10, 30 and 60.196B.12. Median Results of focus ScenariosFigure B.59: Comparison of projected change of the geometric mean grain size ofthe bed surface D90 for the average hydrologic condition and the Histori-cal, Pulse and Uncontrolled sediment supply simulations. Shaded regionscapture the 25th-75th percentile responses across the 100 simulations forthe average condition. Results shown for simulation year 1, 10, 30 and 60.197B.12. Median Results of focus ScenariosB.12.2 Dry HydrographsFigure B.60: Comparison of projected bed elevation change from the 2017 initialprofile for the dry hydrologic condition and the Historical, Pulse and Un-controlled sediment supply simulations. Shaded regions capture the 25th-75th percentile responses across the 100 simulations for the dry condition.Results shown for simulation year 1, 10, 30 and 60.198B.12. Median Results of focus ScenariosFigure B.61: Comparison of projected change of the geometric mean grain sizeof the bed surface Dg for the dry hydrologic condition and the Histori-cal, Pulse and Uncontrolled sediment supply simulations. Shaded regionscapture the 25th-75th percentile responses across the 100 simulations forthe dry condition. Results shown for simulation year 1, 10, 30 and 60.199B.12. Median Results of focus ScenariosFigure B.62: Comparison of projected change of the geometric mean grain sizeof the bed surface D90 for the dry hydrologic condition and the Histori-cal, Pulse and Uncontrolled sediment supply simulations. Shaded regionscapture the 25th-75th percentile responses across the 100 simulations forthe dry condition. Results shown for simulation year 1, 10, 30 and 60.200B.13. Median Results of all ScenariosB.13 Median Results of all ScenariosElevation changeFigure B.63: Median elevation change in the Pulse Supply (RC1-6) and Uncon-trolled Supply (Exp1-3) scenarios under the ’Dry’ hydrologic regime. Eachline represents the median of 100 simulations.201B.13. Median Results of all ScenariosFigure B.64: Median elevation change in the Pulse Supply (RC1-6) and Uncon-trolled Supply (Exp1-3) scenarios under the ’Average’ hydrologic regime.Each line represents the median of 100 simulations.202B.13. Median Results of all ScenariosFigure B.65: Median elevation change in the Pulse Supply (RC1-6) and Uncon-trolled Supply (Exp1-3) scenarios under the ’Wet’ hydrologic regime. Eachline represents the median of 100 simulations.203B.13. Median Results of all ScenariosSurface D90Figure B.66: Median surface D90 in the Pulse Supply (RC1-6) and UncontrolledSupply (Exp1-3) scenarios under the ’Dry’ hydrologic regime. Each linerepresents the median of 100 simulations.204B.13. Median Results of all ScenariosFigure B.67: Median surface D90 in the Pulse Supply (RC1-6) and UncontrolledSupply (Exp1-3) scenarios under the ’Average’ hydrologic regime. Eachline represents the median of 100 simulations.205B.13. Median Results of all ScenariosFigure B.68: Median surface D90 in the Pulse Supply (RC1-6) and UncontrolledSupply (Exp1-3) scenarios under the ’Wet’ hydrologic regime. Each linerepresents the median of 100 simulations.206B.13. Median Results of all ScenariosSurface DgFigure B.69: Median surface Dg in the Pulse Supply (RC1-6) and UncontrolledSupply (Exp1-3) scenarios under the ’Dry’ hydrologic regime. Each linerepresents the median of 100 simulations.207B.13. Median Results of all ScenariosFigure B.70: Median surface Dg in the Pulse Supply (RC1-6) and UncontrolledSupply (Exp1-3) scenarios under the ’Average’ hydrologic regime. Eachline represents the median of 100 simulations.208B.13. Median Results of all ScenariosFigure B.71: Median surface Dg in the Pulse Supply (RC1-6) and UncontrolledSupply (Exp1-3) scenarios under the ’Wet’ hydrologic regime. Each linerepresents the median of 100 simulations.209B.13. Median Results of all ScenariosSlopeFigure B.72: Median slope in the Pulse Supply (RC1-6) and Uncontrolled Supply(Exp1-3) scenarios under the ’Dry’ hydrologic regime. Each line repre-sents the median of 100 simulations.210B.13. Median Results of all ScenariosFigure B.73: Median slope in the Pulse Supply (RC1-6) and Uncontrolled Supply(Exp1-3) scenarios under the ’Average’ hydrologic regime. Each line rep-resents the median of 100 simulations.211B.13. Median Results of all ScenariosFigure B.74: Median slope in the Pulse Supply (RC1-6) and Uncontrolled Supply(Exp1-3) scenarios under the ’Wet’ hydrologic regime. Each line repre-sents the median of 100 simulations.212Appendix CDistID and U-Net model parametersC.1 Employed ResNet50 architecture______________________________________________________________________________________Layer (type) Output Shape Param # Connected to======================================================================================### 48 layers omitted from the Keras implementation for ResNet50 ####______________________________________________________________________________________activation_49 (Activation) (None, 10, 10, 2048) 0 add_16[0][0]______________________________________________________________________________________flatten_1 (Flatten) (None, 204800) 0 activation_49[0][0]______________________________________________________________________________________dense_1 (Dense) (None, 512) 104858112 flatten_1[0][0]______________________________________________________________________________________dropout_1 (Dropout) (None, 512) 0 dense_1[0][0]______________________________________________________________________________________dense_2 (Dense) (None, 256) 131328 dropout_1[0][0]______________________________________________________________________________________dropout_2 (Dropout) (None, 256) 0 dense_2[0][0]213C.2. Employed VGG19 architecture______________________________________________________________________________________dense_3 (Dense) (None, 128) 32896 dropout_2[0][0]______________________________________________________________________________________dense_4 (Dense) (None, 10) 1290 dense_3[0][0]======================================================================================Total params: 128,611,338Trainable params: 128,558,218Non-trainable params: 53,120______________________________________________________________________________________optimizer: stochastic gradient descentoptimizers.SGD(lr=0.0003, momentum=0.9, nesterov=True, decay=1e-7)loss: mean absolute errorC.2 Employed VGG19 architecture_________________________________________________________________Layer (type) Output Shape Param #=================================================================input_1 (InputLayer) (None, 300, 300, 3) 0_________________________________________________________________block1_conv1 (Conv2D) (None, 300, 300, 64) 1792_________________________________________________________________block1_conv2 (Conv2D) (None, 300, 300, 64) 36928_________________________________________________________________block1_pool (MaxPooling2D) (None, 150, 150, 64) 0_________________________________________________________________block2_conv1 (Conv2D) (None, 150, 150, 128) 73856_________________________________________________________________214C.2. Employed VGG19 architectureblock2_conv2 (Conv2D) (None, 150, 150, 128) 147584_________________________________________________________________block2_pool (MaxPooling2D) (None, 75, 75, 128) 0_________________________________________________________________block3_conv1 (Conv2D) (None, 75, 75, 256) 295168_________________________________________________________________block3_conv2 (Conv2D) (None, 75, 75, 256) 590080_________________________________________________________________block3_conv3 (Conv2D) (None, 75, 75, 256) 590080_________________________________________________________________block3_conv4 (Conv2D) (None, 75, 75, 256) 590080_________________________________________________________________block3_pool (MaxPooling2D) (None, 37, 37, 256) 0_________________________________________________________________block4_conv1 (Conv2D) (None, 37, 37, 512) 1180160_________________________________________________________________block4_conv2 (Conv2D) (None, 37, 37, 512) 2359808_________________________________________________________________block4_conv3 (Conv2D) (None, 37, 37, 512) 2359808_________________________________________________________________block4_conv4 (Conv2D) (None, 37, 37, 512) 2359808_________________________________________________________________block4_pool (MaxPooling2D) (None, 18, 18, 512) 0_________________________________________________________________block5_conv1 (Conv2D) (None, 18, 18, 512) 2359808_________________________________________________________________block5_conv2 (Conv2D) (None, 18, 18, 512) 2359808_________________________________________________________________block5_conv3 (Conv2D) (None, 18, 18, 512) 2359808215C.2. Employed VGG19 architecture_________________________________________________________________block5_conv4 (Conv2D) (None, 18, 18, 512) 2359808_________________________________________________________________block5_pool (MaxPooling2D) (None, 9, 9, 512) 0_________________________________________________________________flatten_1 (Flatten) (None, 41472) 0_________________________________________________________________dense_1 (Dense) (None, 512) 21234176_________________________________________________________________dropout_1 (Dropout) (None, 512) 0_________________________________________________________________dense_2 (Dense) (None, 256) 131328_________________________________________________________________dropout_2 (Dropout) (None, 256) 0_________________________________________________________________dense_3 (Dense) (None, 128) 32896_________________________________________________________________dense_4 (Dense) (None, 10) 1290=================================================================Total params: 41,424,074Trainable params: 41,424,074Non-trainable params: 0_________________________________________________________________optimizer: stochastic gradient descentoptimizers.SGD(lr=0.0003, momentum=0.9, nesterov=True, decay=1e-7)loss: mean absolute error216C.3. Employed UNet architectureC.3 Employed UNet architecture_______________________________________________________________________________________Layer (type) Output Shape Param # Connected to=======================================================================================input_2 (InputLayer) (None, 512, 512, 3) 0_______________________________________________________________________________________lambda_2 (Lambda) (None, 512, 512, 3) 0 input_2[0][0]_______________________________________________________________________________________conv2d_20 (Conv2D) (None, 512, 512, 16) 448 lambda_2[0][0]_______________________________________________________________________________________dropout_10 (Dropout) (None, 512, 512, 16) 0 conv2d_20[0][0]_______________________________________________________________________________________conv2d_21 (Conv2D) (None, 512, 512, 16) 2320 dropout_10[0][0]_______________________________________________________________________________________max_pooling2d_5 (MaxPooling2D) (None, 256, 256, 16) 0 conv2d_21[0][0]_______________________________________________________________________________________conv2d_22 (Conv2D) (None, 256, 256, 32) 4640 max_pooling2d_5[0][0]_______________________________________________________________________________________dropout_11 (Dropout) (None, 256, 256, 32) 0 conv2d_22[0][0]_______________________________________________________________________________________conv2d_23 (Conv2D) (None, 256, 256, 32) 9248 dropout_11[0][0]_______________________________________________________________________________________max_pooling2d_6 (MaxPooling2D) (None, 128, 128, 32) 0 conv2d_23[0][0]_______________________________________________________________________________________conv2d_24 (Conv2D) (None, 128, 128, 64) 18496 max_pooling2d_6[0][0]_______________________________________________________________________________________dropout_12 (Dropout) (None, 128, 128, 64) 0 conv2d_24[0][0]_______________________________________________________________________________________217C.3. Employed UNet architectureconv2d_25 (Conv2D) (None, 128, 128, 64) 36928 dropout_12[0][0]_______________________________________________________________________________________max_pooling2d_7 (MaxPooling2D) (None, 64, 64, 64) 0 conv2d_25[0][0]_______________________________________________________________________________________conv2d_26 (Conv2D) (None, 64, 64, 128) 73856 max_pooling2d_7[0][0]_______________________________________________________________________________________dropout_13 (Dropout) (None, 64, 64, 128) 0 conv2d_26[0][0]_______________________________________________________________________________________conv2d_27 (Conv2D) (None, 64, 64, 128) 147584 dropout_13[0][0]_______________________________________________________________________________________max_pooling2d_8 (MaxPooling2D) (None, 32, 32, 128) 0 conv2d_27[0][0]_______________________________________________________________________________________conv2d_28 (Conv2D) (None, 32, 32, 256) 295168 max_pooling2d_8[0][0]_______________________________________________________________________________________dropout_14 (Dropout) (None, 32, 32, 256) 0 conv2d_28[0][0]_______________________________________________________________________________________conv2d_29 (Conv2D) (None, 32, 32, 256) 590080 dropout_14[0][0]_______________________________________________________________________________________conv2d_transpose_5 (Conv2DTrans (None, 64, 64, 128) 131200 conv2d_29[0][0]_______________________________________________________________________________________concatenate_5 (Concatenate) (None, 64, 64, 256) 0 conv2d_transpose_5[0][0]conv2d_27[0][0]_______________________________________________________________________________________conv2d_30 (Conv2D) (None, 64, 64, 128) 295040 concatenate_5[0][0]_______________________________________________________________________________________dropout_15 (Dropout) (None, 64, 64, 128) 0 conv2d_30[0][0]_______________________________________________________________________________________conv2d_31 (Conv2D) (None, 64, 64, 128) 147584 dropout_15[0][0]_______________________________________________________________________________________218C.3. Employed UNet architectureconv2d_transpose_6 (Conv2DTrans (None, 128, 128, 64) 32832 conv2d_31[0][0]_______________________________________________________________________________________concatenate_6 (Concatenate) (None, 128, 128, 128 0 conv2d_transpose_6[0][0]conv2d_25[0][0]_______________________________________________________________________________________conv2d_32 (Conv2D) (None, 128, 128, 64) 73792 concatenate_6[0][0]_______________________________________________________________________________________dropout_16 (Dropout) (None, 128, 128, 64) 0 conv2d_32[0][0]_______________________________________________________________________________________conv2d_33 (Conv2D) (None, 128, 128, 64) 36928 dropout_16[0][0]_______________________________________________________________________________________conv2d_transpose_7 (Conv2DTrans (None, 256, 256, 32) 8224 conv2d_33[0][0]_______________________________________________________________________________________concatenate_7 (Concatenate) (None, 256, 256, 64) 0 conv2d_transpose_7[0][0]conv2d_23[0][0]_______________________________________________________________________________________conv2d_34 (Conv2D) (None, 256, 256, 32) 18464 concatenate_7[0][0]_______________________________________________________________________________________dropout_17 (Dropout) (None, 256, 256, 32) 0 conv2d_34[0][0]_______________________________________________________________________________________conv2d_35 (Conv2D) (None, 256, 256, 32) 9248 dropout_17[0][0]_______________________________________________________________________________________conv2d_transpose_8 (Conv2DTrans (None, 512, 512, 16) 2064 conv2d_35[0][0]_______________________________________________________________________________________concatenate_8 (Concatenate) (None, 512, 512, 32) 0 conv2d_transpose_8[0][0]conv2d_21[0][0]_______________________________________________________________________________________conv2d_36 (Conv2D) (None, 512, 512, 16) 4624 concatenate_8[0][0]_______________________________________________________________________________________219dropout_18 (Dropout) (None, 512, 512, 16) 0 conv2d_36[0][0]_______________________________________________________________________________________conv2d_37 (Conv2D) (None, 512, 512, 16) 2320 dropout_18[0][0]_______________________________________________________________________________________conv2d_38 (Conv2D) (None, 512, 512, 6) 102 conv2d_37[0][0]=======================================================================================Total params: 1,941,190Trainable params: 1,941,190Non-trainable params: 0_______________________________________________________________________________________batch_size=5,epochs=50,optimizer=’adam’,loss=’binary_crossentropy’220

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0377728/manifest

Comment

Related Items