A rapid and objective characterization ofchannel morphology in a small, forestedchannel using a remotely piloted aircraft.byCarina HelmB.Sc., The University of British Columbia, 2017A THESIS SUBMITTED IN PARTIAL FULFILMENT OFTHE REQUIREMENTS FOR THE DEGREE OFMaster of ScienceinTHE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES(Geography)THE UNIVERSITY OF BRITISH COLUMBIA(Vancouver)August 2019©Carina Helm, 2019The following individuals certify that they have read, and recommend to the Faculty ofGraduate and Postdoctoral Studies for acceptance, a thesis/dissertation entitled:A rapid and objective characterization of channel morphology in a small, forested channelusing a remotely piloted aircraft.submitted by: Carina J. Z. Helm in partial fulfillment of the requirementsfor the degree of Master of Sciencein GeographyExamining Committee:Marwan Hassan, GeographySupervisorJohn Richardson, Forest and Conservation SciencesSupervisory Committee MemberJennifer Williams, GeographyAdditional ExamineriiAbstractThe use of remotely piloted aircrafts (RPAs) in fluvial geomorphology has improved theability to characterize streams at greater resolutions and spatial extents than was previouslyattainable using traditional survey techniques. However, their use has been generally limitedto streams under ideal conditions that differ from the small, forested mountain channelscommon in the Pacific Northwest. These channels have remained difficult to characterizeusing modern techniques due to their dense canopies and rough terrain. A rapid and ob-jective method of characterizing channel morphology across the river basin using a RPA ispresented in this dissertation to help overcome this challenge. First, the accuracy of RPAsfor extracting bed elevations, bathymetry and grain size along 3 km of Carnation Creek, asmall, forested stream on Vancouver Island, was investigated through a sub-canopy survey.Relevant cross-sectional channel variables were then extracted to objectively characterizechannel morphologies across the river basin using a principal component analysis-clustering(PCA-clustering) technique. Then the Shannon’s diversity index was used to characterizethe local diversity across the channel, and investigate the scale needed to study the system,to ensure its heterogeneity was characterized. The results demonstrate that RPAs providea rapid alternative to characterizing these systems, through the construction of a 2–cm res-olution digital elevation model spanning 3 km of channel, with a root-mean-square-errorof 0.093 m for exposed bed check points and 0.1 m for submerged bed check points. ThePCA-clustering analysis provided an objective means of classifying channel morphology witha correct classification rate of 85%. Altogether, the results provide a precedent for using aRPA to characterize the morphology and diversity of small, forested channels at a scale ofecological relevance to the life histories of Pacific salmonids.iiiLay SummaryThe use of remotely piloted aircrafts (also known as drones) in surveying stream systemshas become well established in the last decade. However, to date these applications havebeen limited to streams with wide channels and minimal overhanging forest vegetation. Theobjective of this dissertation is to demonstrate the utility of remotely piloted aircrafts forsub-canopy surveying of a small, forested mountain channel at the basin-scale. The resultsprovide a rapid and objective method for characterizing patterns in channel morphology andhabitat diversity at a scale that is relevant to the life cycles of stream fishes.ivPrefaceThis dissertation is the result of a field survey conducted at Carnation Creek designed inconjunction with my supervisor Marwan Hassan. The field data used for this research wascollected with help from Dave Reid, Kyle Wlodarczyk and Ryan Matheson. Reference dataon the study area was provided by the British Columbia provincial Ministry of Forests,Lands and Natural Resource Operations. A paper with Marwan Hassan as a co-author willbe submitted for publication, based on material in Chapters 2 and 3. I completed all analysisand processing of the data.vContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiLay Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xList of Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiiAcknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Classification of channel form . . . . . . . . . . . . . . . . . . . . . . . . . . 21.3 Application of remotely piloted aircrafts and classification schemes to smallstreams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 Sub-canopy, remotely piloted aircraft survey of channel features in asmall, forested mountain channel . . . . . . . . . . . . . . . . . . . . . . . . 62.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82.2.1 Study area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82.2.2 Remotely piloted aircraft survey . . . . . . . . . . . . . . . . . . . . . 112.2.3 Photogrammetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132.2.4 Dense point cloud cleaning . . . . . . . . . . . . . . . . . . . . . . . . 142.2.5 Water depth corrections . . . . . . . . . . . . . . . . . . . . . . . . . 14vi2.2.6 Sediment texture analysis . . . . . . . . . . . . . . . . . . . . . . . . 152.2.7 Large wood extraction . . . . . . . . . . . . . . . . . . . . . . . . . . 162.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172.3.1 Accuracy assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . 172.3.2 Survey coverage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313 Rapid and objective classification of channel morphology and diversityat the basin-scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343.2.1 Selection of channel variables . . . . . . . . . . . . . . . . . . . . . . 343.2.2 Extraction of channel variables . . . . . . . . . . . . . . . . . . . . . 363.2.3 Principal component analysis . . . . . . . . . . . . . . . . . . . . . . 383.2.4 K-means clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403.2.5 Channel diversity classification . . . . . . . . . . . . . . . . . . . . . 413.2.6 Scale effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423.3.1 Clustering analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443.3.2 Clustering accuracy assessment . . . . . . . . . . . . . . . . . . . . . 463.3.3 Scale effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 564 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59viiList of Tables2.1 Computer processing time required for each photoset and GCP combinationin PhotoScan. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182.2 Photoset and GCP and CP combinations. . . . . . . . . . . . . . . . . . . . 182.3 Correlation matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212.4 Percentages of the study section covered with the RPA . . . . . . . . . . . . 263.1 Means of channel variables for each cluster. . . . . . . . . . . . . . . . . . . . 453.2 Accuracy assessment of morphology classification . . . . . . . . . . . . . . . 473.3 Comparison of average values for variables of each morphology to those foundin previously published studies. . . . . . . . . . . . . . . . . . . . . . . . . . 54viiiList of Figures2.1 Carnation Creek watershed. . . . . . . . . . . . . . . . . . . . . . . . . . . . 92.2 Environmental characteristics of the study area . . . . . . . . . . . . . . . . 102.3 RPA flight path and camera orientations. . . . . . . . . . . . . . . . . . . . . 122.4 Predictive grain size relationships . . . . . . . . . . . . . . . . . . . . . . . . 162.5 Physical characteristics of the accuracy assessment site . . . . . . . . . . . . 172.6 GCP distribution for accuracy assessments . . . . . . . . . . . . . . . . . . . 192.7 Boxplots showing the distribution of vertical error for the GCPs and CPs ofeach photoset and GCP distribution. . . . . . . . . . . . . . . . . . . . . . . 202.8 Bed coverage with oblique imagery . . . . . . . . . . . . . . . . . . . . . . . 212.9 Density plots for vertical error of the modelled elevations . . . . . . . . . . . 222.10 Distribution of GCPs for bathymetry assessment. . . . . . . . . . . . . . . . 232.11 Submerged bed elevation corrections for site a . . . . . . . . . . . . . . . . . 242.12 Submerged bed elevation corrections for site b . . . . . . . . . . . . . . . . . 242.13 RPA Coverage for SAs 2–5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252.14 RPA Coverage for SAs 6–9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.15 Coverage of study section five. . . . . . . . . . . . . . . . . . . . . . . . . . . 272.16 Coverage of a deep pool in study section eight. . . . . . . . . . . . . . . . . . 283.1 Diagram of a typical pool-riffle transition . . . . . . . . . . . . . . . . . . . . 353.2 Feature extraction set up. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373.3 Scree plot of the variance explained by the principal components. . . . . . . 393.4 Eigenvalues representing variance explained by the PCs. . . . . . . . . . . . 403.5 K-means clustering plot demonstrating how the within groups sum of squaresdecreases with the incorporation of more clusters. . . . . . . . . . . . . . . . 413.6 Variable contributions to Principal Components. . . . . . . . . . . . . . . . . 433.7 Correlation circle plot. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443.8 Biplot of each observation along PC1 and PC2. . . . . . . . . . . . . . . . . 453.9 Isolated biplot of each observation along PC1 and PC2 for the six clusters. . 463.10 Cumulative sum of channel morphologies . . . . . . . . . . . . . . . . . . . . 47ix3.11 Distribution of channel morphologies. . . . . . . . . . . . . . . . . . . . . . . 483.12 Longitudinal profile along the thalweg of Carnation Creek. . . . . . . . . . . 483.13 Riffle-pool transition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493.14 Influence of grain size and slope on the classification . . . . . . . . . . . . . . 513.15 Shannon’s diversity across the channel based on a 45-metre window . . . . . 523.16 Standard deviation of channel diversity. . . . . . . . . . . . . . . . . . . . . . 53xList of SymbolsSymbol Definition UnitsA cross-sectional area L2cm centimetre Ld water depth LD grain diameter LDx percentile of grain diameter LDEM digital elevation model -g acceleration due to gravity L T−2km kilometre Ll longitudinal distance along channel LH Shannon’s diversity index -LiDAR Light Detection And Ranging (Remote Sensing) -n Manning’s roughness coefficient T L−13n number of samples -m metre Lp proportion of channel type -P wetted perimeter LPC1, PC2, and PC3 principal component axis 1, 2, and 3 -PCA principal component analysis -Q discharge L3 T−1Qs sediment flux m T−1Rh hydraulic radius Lr2 coefficient of determination -SA study section -Sl local slope -Sws local slope of water surface -Sr reach slope -xit time TRPA remotely piloted aircraft -v velocity L T−1W wetted channel width Lz bed elevation Lρ density of water m L−3τ shear stress m L−1 T−2τ ∗ dimensionless shear stress -2D two-dimensional -3D three-dimensional -xiiAcknowledgementsI would like to extend my sincere gratitude to my supervisor Marwan Hassan. Without hiscontinual support, understanding and guidance I could not have completed this research.The wonderful people I have met, and projects I have been part of, have allowed me to learnand experience more both academically and personally than I could have anticipated, and Iam grateful for his role in this.I would also like to thank my lab-group and office mates for their encouragement andadvice along the way, especially to Dave Reid for his support in teaching the necessary fieldskills and always sparing time to share his advice on the project and anything graduate-school related. To Steve Bird of Fluvial Systems Research, and my committee member JohnRichardson, I am grateful for the discussions we’ve had and insightful comments shared,which have helped to shape my research. Thank you to Brett Eaton as well for helping inspiremy interest in fluvial geomorphology during my undergraduate studies and for continuing toprovide advice on the RPA-related aspects of this research.This project would not have been possible without the support of the British Columbiaprovincial Ministry of Forests Lands and Natural Resource Operations. Thank you to PeterTschaplinksi and Robin Pike for their interest in this research, for sharing data and for helpingfacilitate the field work. To Steve Voller and Andrew Westerhof, many thanks for making ourtime at Carnation Creek so memorable and helping to ensure the smooth operation of thecamp. I was very fortunate to be part of a great field crew for this research, and thank KyleWlodarczyk, Ryan Matheson, Kevin Pierce, Dave Reid, Dave Adams, and Jack Carrigan fortheir hard work in the field and positive attitudes.Thank you to my parents, Charles and Linda Helm, for instilling my interest in thenatural sciences from a young age and their unwavering love and support in my pursuits.Thank you to my brother Daniel, for always being available to take stressed phone calls andoffer wise words of advice. Last but not least, thank you to my partner, Jack Carrigan forhis patience, belief in me, and for helping me stay grounded throughout this process.xiiiChapter 1Introduction1.1 BackgroundGlobally, wild salmon stocks are in decline. Their stressors are numerous, but their declineis often attributed in part to degradation of their freshwater habitat (Nelson et al., 2015).To remedy this, millions of dollars are annually put into restoration programs to improvefreshwater conditions (Bisson et al., 2009). Despite this large investment, the declines andlow rates of recovery in several stocks are still persistent, and may point to a broader misun-derstanding of salmonids and their life requirements (Fausch et al., 2002; Bisson et al., 2009).This may be because of an emphasis on designing channels to contain what is considered tobe ideal habitat, rather than restoring natural processes of the stream system itself. Pacificsalmon have diverse life histories, so it is logical to assume that they are adapted and well-suited to the spatial and temporal diversity in habitat provided by the natural processes ofstream systems (Bisson et al., 2009). This is seen today by restoration frameworks that callfor the reestablishment of natural processes in streams, rather than altering these systemsto contain what is considered to be “good” habitat (Beechie et al., 2010).The neglect for restoring natural processes in previous conservation frameworks likelystems from prior conceptual theories that the ideal ecosystem is static and in equilibrium,which leads to the notion that disturbance should be discouraged (Wallington et al., 2005).However, there has since been a shift in ecological thinking that celebrates the variabilityprovided by the natural processes of stream systems and their capacity to provide an arrayof aquatic habitats. The 1990s and early 2000s saw the rise of many notable studies thatencouraged the characterization of ecological processes in stream systems continuously andat a scale relevant to species’ life histories (Fausch et al., 2002; Schlosser , 1991). The“Riverscape Approach”to viewing aquatic habitat posited by Fausch et al. (2002), emphasizesheterogeneity in streams as a critical factor influencing stream fish distribution and structure,1and has been observed in other studies (e.g. White et al., 2008; Kim and Lapointe, 2011).The logic for this approach is evident when considering that fish complete their life cyclesin streams at intermediate spatial scales between 1 and 100 kilometres in length (Fauschet al., 2002). It therefore encourages characterization of streams at a scale relevant to theseprocesses. However, this has been a scale that has traditionally been difficult to characterize.The present understanding of streams has been well informed by studies conducted atboth small scales (from the microhabitat and channel unit to reach extent) and large scales(basin extent). It is the intermediate scale that has been lacking and that requires moreinvestigation (Fausch et al., 2002). With advances in remote sensing technologies, it isbecoming easier to survey this intermediate scale at high resolutions to address the gapbetween conceptual theories such as the “Riverscape Approach” and empirical observation.However, with the ability to collect high resolution and continuous data comes the challengeof translating it into a meaningful form that can aid stream scientists in understanding andmanaging these systems.1.2 Classification of channel formClassification schemes for characterizing channel morphology have held an important placein fluvial geomorphology to help understand and analyze the natural world. While thechannel characteristics that they stem from, such as topography, grain size, and channelwidth and depth, are continuous variables, at the morphological unit scale they do appear tocorrespond with predictable channel controls. This can be observed in a recent study notingthe influence of characteristic width gradients on pool-riffle formation (Chartrand et al.,2018). Discernable morphologies, such as pools and riffles, can also provide functionallydifferent habitat for aquatic species (e.g. Bisson et al., 1988; Flitcroft et al., 2012; Namanet al., 2017). Therefore, classification of these channel types should provide a useful toolfor investigating relationships between salmonids and their physical environment. Severalclassification schemes have been proposed that aid this goal (Frissell et al., 1986; Hawkinset al., 1993; Rosgen, 1994), each of which may be tailored to the environment they weredeveloped in, thus limiting their broad-scale application (Nestler et al., 2016). Furthermore,the use of these early classification schemes often requires qualitative and subjective decisionsfrom the observer for which channel type to prescribe to a channel unit. Because quantitativerelationships between channel characteristics and channel form exist, there should be anopportunity for an improved objective classification scheme to be developed.Early geomorphologists qualitatively noted the relationship between channel characteris-tics and form through the Lane Balance (Equation 1.1) (Lane, 1955). Equation 1.1 reflects2a balance between the slope (S), discharge (Q), sediment flux (Qs) and grain size (D) in achannel under equilibrium conditions, as can be observed in nature.D αQSQs(1.1)Fluvial geomorphologists have since developed relationships that explain the processesleading to a channel’s form. Chief among them is the shear stress equation (Equation 1.2),which describes the force exerted by water on the bed, where water density (p) and gravity(g) are constants and depth (d) and slope (S) describe the channel’s shape.τ = ρgdS (1.2)Equation 1.3 elucidates how the channel slope, sediment characteristics and scale caninform a channel’s sediment transport regime (described by the shields number - the non-dimensionalized shear stress) (Church, 2006). It can then be inferred that along a typicalchannel, the transport regime will change along with the slope and roughness of the channel.As demonstrated by Tamminga (2016), it is important to consider these relations to providea better understanding of how a stream may adjust to channel conditions and how that willinfluence its morphology.S = 1.65 τ ∗Dd(1.3)With the channel characteristics noted above, there may be an opportunity to conduct anobjective characterization of channel morphologies that builds on a quantitative understand-ing of these systems. As described above, there is a general theme for the prior equations toinclude the following channel properties: hydraulic geometry, slope, and grain size charac-teristics. Therefore, there is a theoretical basis for inclusion of these variables in a statisticalclassification of channel types that could be conducted on and therefore best suited to aparticular region of interest. Development of such an objective and quantitative scheme islikely advantageous as it would protect against human pre-conceived notions of channel type,and may highlight patterns that would otherwise be missed in a visual classification.1.3 Application of remotely piloted aircrafts and clas-sification schemes to small streamsThe use of remotely piloted aircrafts (RPAs) (e.g. Tamminga, 2016; Dietrich, 2017; Woodgetand Austrums , 2017) and efforts to objectively classify channel morphology at large spatial3scales (Lindenschmidt and Long , 2013; Hugue et al., 2016) have both become emergingresearch themes in fluvial geomorphology. However, the use of RPAs for studying riverscapeshas been limited in the types of environmental conditions it has been applied to. Theenvironments in which they have been tested in can be divided into two broad categories:ˆ Segments of larger channels (e.g. MacVicar et al., 2009; Tamminga, 2016).ˆ Small streams lacking a dense riparian canopy (e.g. Woodget and Austrums , 2017)It appears that application of this tool to small, forested mountain channels is lacking.Yet the ecological importance of these streams cannot be understated, and is seen by thehigh numbers of salmonid species such as cutthroat trout and coho salmon that reside inthese systems (Rosenfeld et al., 2000). It has been postulated that several unique charac-teristics of small streams may have led to this preference, including that the influence oflarge wood on providing complex habitats is most effective in small, gravel-cobble beddedstreams (Rosenfeld et al., 2000). This suggests that despite their small size, these streamsare morphologically unique and can provide important habitat for salmonids. They alsocomprise a significant portion of the channel network and are often degraded by industrial,urban and agricultural activities (Rosenfeld et al., 2000). Acknowledgement of this can beseen by the efforts in the 1970s to construct a long-running study at Carnation Creek toassess the influence of logging treatments on ecological and geomorphological processes (seeTschaplinski and Pike, 2017). Results of the study show that forestry can have long-termimpacts on habitat, through changing the temperature regime of the channel to changingsedimentological processes (Tschaplinski and Pike, 2017). In conclusion, not only are thedynamics of small stream systems unique, but these systems can also be severely influencedby watershed management decisions, thus highlighting the need to provide observations thatcan inform protection of these systems.Despite a need to better understand small, forested stream systems, it appears that effortsto survey them have not kept pace with remote sensing techniques seen in other streamsystems, likely because dense canopies often make viewing them from aerial platforms achallenge. However, with the greater maneuverability and availability of RPAs, there maybe an opportunity to bring modern techniques to the surveying of small, forested channels.Given the importance of small, forested streams in providing aquatic habitat to fish species,improving field techniques and classification schemes in this environment may be beneficial.With the advent of RPA technology there may lie an opportunity to more easily acquiredata at the basin-scale to link patterns in ecological and geomorphological processes in thesesystems. To further this goal, the objective of this research is to provide a means of rapidly4and objectively surveying and classifying small, forested stream systems to better informtheir management. This objective was investigated through answering the following researchquestions:ˆ How can RPAs be used to characterize channel characteristics of a small, forestedchannel at the basin-scale using a sub-canopy survey?ˆ How can continuous measurements of channel characteristics be used to rapidly andobjectively characterize patterns in channel morphology?ˆ What metrics can be used to characterize patterns in channel diversity, and what isthe necessary scale to study these systems for watershed management?These research goals are addressed over two chapters of this thesis. The first chapterprovides an overview of the new technique employed to survey a small, forested mountainchannel and of how relevant channel features were extracted. The second chapter investigateshow these features can be used to objectively and near-continuously characterize patternsin channel morphology, and provides some of the basin-scale conclusions that can be drawnfrom the data.5Chapter 2Sub-canopy, remotely piloted aircraftsurvey of channel features in a small,forested mountain channel2.1 IntroductionThe use of RPAs in fluvial geomorphology has become well established in the last decadewith applications ranging from inventory of channel features to hydrodynamic flow mod-elling of streams (e.g. Tamminga, 2016; Woodget and Austrums , 2017). To date, the useof RPAs in such studies has several features in common. They are generally conducted onsmaller segments of the larger stream network and are limited to channels with ideal surveyconditions, where the streambed can be clearly viewed from a RPA flying above the forestcanopy. Such conditions are not commonly found in small, forested mountain channels thatare common in the Pacific Northwest. The temperate rainforest that these channels reside inoften results in the forest canopy blocking observations of the channel bed and banks fromRPA imagery collected above the forest canopy. While these channels compose a significantportion of the channel network in the Pacific Northwest and provide important habitat formany salmonids, it has remained a challenge to characterize them using techniques thatlarger streams have benefited from.The difficult terrain imposed by streams makes ground surveys of channel morphology ademanding and time-consuming task, resulting in such surveys often being conducted in shortsegments of a channel. Similarly, traditional methods of estimating sediment texture involvea significant field component and are not well suited to surveying large areas frequently.Because of the human resources required, these methods are at odds with the argument that6it is better to attain coarser data on aquatic habitat at a scale relevant to aquatic processesthan detailed data at the wrong scale (Fausch et al., 2002). While imagery from airplaneand satellite platforms has provided an excellent method for landscape classification, itsapplicability has been limited to environments such as those that lack riparian vegetation (inarid or cleared agricultural environments) or wide streams with minimal riparian vegetationin dense forested environments. Furthermore, these platforms are often limited in theirability to map sediment texture due to their lower spatial resolution. Similarly, LiDAR hasproven successful in habitat studies such as large wood (LW) identification (see Abalharthet al., 2015), but the method is often costly for commercial use. Furthermore, unless greenband LiDAR is used in conjunction with infrared LiDAR, the bathymetry cannot be mapped,which makes this tool ineffective for characterizing the submerged bed (Muste et al., 2012).In contrast, RPAs are becoming increasingly affordable and flight paths can be user-definedto meet the unique needs of a study, but their use has also been limited to studies underideal environmental conditions.Challenges in employing sub-canopy surveys may stem from the fact that image acqui-sition for RPA-surveys often involves using vertical images (images taken where the cameralens axis is perpendicular to the ground) with parallel flight paths to characterize an area, ashas been traditionally employed in airplane surveys for the last several decades (James andRobson, 2014). However, the techniques used for traditional photogrammetry differ fromthose employed in structure-from-motion (SFM) based photogrammetry, which may allowfor development of flight paths better suited to a sub-canopy survey. For traditional pho-togrammetry, metric cameras with known internal camera geometries and lens distortions areused in combination with two overlapping photos of known position and distance from oneanother, to establish distances from the camera to features in the imagery (Dietrich, 2014).With SFM-based photogrammetry, common features are identified in multiple images toreconstruct the geometry of a site from which arbitrary camera positions and orientationsare generated automatically. This approach allows non-metric cameras to be used and thefreedom to collect images from a range of unknown positions and orientations.Recently, there has been greater emphasis on testing flight patterns that deviate fromthose used in traditional photogrammetry. Perhaps the most notable finding is the positiveeffect incorporation of oblique and highly convergent imagery can have on overcoming errorsintroduced from the non-metric cameras frequently used in RPA-surveys (Wackrow andChandler , 2011; James and Robson, 2014; Harwin et al., 2015). Non-metric cameras are lightand less costly than their metric counterparts, but are prone to greater distortions as theyare manufactured to a lower standard (Harwin et al., 2015). While this can be overcome bya self-calibration of the camera in the SFM software, the incorporation of oblique-convergent7imagery has been shown to minimize systematic errors in model outputs that often plagueprojects where only imagery with a vertical orientation is employed (Harwin et al., 2015).Through embracing these findings, there may be an opportunity to overcome the surveychallenges imposed by small, forested stream systems. Oblique imagery is likely advantageousin streams where riparian vegetation may prevent the RPA from flying directly over the bank.In this situation it may be possible to characterize these obstructed portions of the channelthrough oblique and convergent imagery originating from a clear flight path below the canopy.However, few studies to date provide guidelines for RPA-surveys conducted below the canopyto characterize small, forested stream systems. This work seeks to provide a framework forcharacterizing channel features in these systems and a first assessment of the accuracy andcoverage that can be acquired when employing RPAs in this new environment. To build thisframework, this chapter aims to meet three objectives:ˆ Investigate the influence of low elevation oblique-convergent imagery on the accuracyof model outputs.ˆ Discuss techniques to filter overhanging vegetation that plagues models in these systemsand tools for extracting channel features such as bathymetry, sediment texture andlarge wood.ˆ Evaluate the method against traditional total station survey data collected in thechannel.2.2 Methods2.2.1 Study areaCarnation Creek is a small gravel bed stream located on the southwest coast of VancouverIsland (Figure 2.1). The watershed it resides in has been the site of a long-running, fish-forestry interactions study, initiated in the 1970s, to assess the impact of different loggingtreatments on the watershed. The stream is ∼8 km long and has a drainage area of ∼11km2 (Tschaplinski and Pike, 2017).8Figure 2.1: The Carnation Creek watershed, located on the south-west coast of VancouverIsland. The start and end points of the RPA-survey are indicated on the map as thumbnails,as well as the midpoints of the eight geomorphology study sections (SAs 2–9). Initiallystudy section one was established in the estuary, but has since been abandoned and was notincluded in the survey. Coordinate system is NAD83 UTM Zone 10.The watershed bears evidence of its glacial history through a stepped longitudinal profilewith alternating segments of mild and steep gradients (Reid et al., 2019), as is characteristicof deglaciated systems (Brardinoni and Hassan, 2007). The first approximately 3 km ofthe channel are mildly sloped and dominated by riffle-pool morphologies, after which thechannel narrows into a canyon that sees the onset of step-pool morphologies and provides abarrier to fish passage (Reid et al., 2019). The average bankfull width of the lower channelis approximately 15 m and the sediment texture of the bed varies from small gravels nearthe mouth of the channel to coarser cobbles and boulders as the canyon is approached (Reidet al., 2019). Anadromous salmon that inhabit the lower 3 km of channel include coho salmon(Oncorhynchus kisutch), chum salmon (Oncorhynchus keta), rainbow trout (Oncorhynchusmykiss) and cutthroat trout (Oncorhynchus clarki) (Tschaplinski and Pike, 2017).Detailed channel morphology data have been collected through annual total station sur-veys in eight study sections (SAs 2–9), seven of which are located downstream of the canyon(SAs 2–8) (Figure 2.1). Initially a study section was established in the estuary (SA1) how-ever, the site has since been abandoned and was not covered for this research. The mostupstream study section is study section nine, which is located above the canyon. The lowerstudy sections are 300–500 m apart and 5–10 bankfull widths in length (Reid et al., 2019).The channel is located within the Coastal Western Hemlock Biogeoclimatic Zone of9British Columbia (CWH) (Hartman et al., 1982). Western hemlock is the most commontree species of the CWH, followed by red cedar and Douglas-fir (Pojar et al., 1991). Inlogged areas, red alder is common and widespread (Pojar et al., 1991). As illustrated inFigure 2.2, much of the channel is hidden below a dense forest canopy composed of bothconiferous and deciduous tree species. In order to characterize the majority of the channelbed it would therefore be necessary to fly the RPA below the forest canopy. In July of 2018,approximately 3.5 km of the channel was surveyed from the mouth of the channel to thebeginning of the canyon, as well as the lower half of SA9, over a period of ten days.Figure 2.2: (a) Google Earth satellite imagery of Carnation Creek (Map data©Google 2019,Maxar Technologies), showing the topography of the site, midpoints of study sections twothrough eight, and location of the channel in yellow. (b) Canopy coverage in a section ofthe channel upstream of a weir with a cable car that provides a relatively clear view of thechannel bed. (c) Nearly closed canopy coverage that partially obstructs view of the bed, as istypical throughout the channel. Note that (b) and (c) were taken from an aircraft after fall,which differ from the summer conditions present during the RPA-survey where the presenceof leaves led to more pronounced canopy coverage.102.2.2 Remotely piloted aircraft surveyThe RPA-survey involved low-level flights conducted in tandem with placement of surveyedGround Control Points (GCPs). The RPA-survey was conducted using the DJI Phantom4 Advanced, which contains a camera with a focal length of 8.8 mm (24 mm in 35 mmformat equivalent) and a field of view of 84◦. To remove the influence of the forest canopyobstructing view of the channel, the RPA was flown manually below the canopy, whichresulted in flying heights ranging from approximately 3–15 m above ground level. Channelsegments were chosen that split the channel into approximately straight sections that were40–60 m long and bounded by areas with rough terrain (e.g. channel-spanning log jams).These segments provided manageable blocks along the channel to navigate with the RPA,and would comprise individual photosets for image processing. The images were acquired inautomatic mode to mitigate the influence of changing lighting conditions along the channeland at a capture rate of a photo every 2 s while moving at approximately 1 m/s to attainmaximum overlap between photos. In total, the approximately 3 km of surveyed channelwas broken into approximately 80 segments with 300–1000 photos each.Each channel segment was initially flown with a traditional flight path consisting ofparallel flight lines with vertical imagery while striving to attain at least 60% side and forwardoverlap, as has been noted as convention in other studies (Ortega-Terol et al., 2014). Thereach was then flown with oblique and convergent imagery, with an emphasis on ensuringthat the channel banks were well covered. To do this the camera was tilted at a low-oblique angle (20–30◦ from vertical, see Figure 2.3a) and a flight path as shown in Figure2.3b was employed. As illustrated, the RPA was first flown parallel to the bank with thecamera oriented in an upstream direction, and then the path was repeated with the cameraorientated in a downstream direction to acquire oblique-convergent imagery. These stepswere completed at multiple distances from the bank to ensure sufficient overlap with thevertical photosets.11Figure 2.3: (a) Channel cross section showing the oblique angles of the RPA’s camera (indi-cated by the solid black line) for image acquisition of the banks. To characterize the channelbanks, the camera was tilted 20–30◦ from vertical. (b) Plan view of the flight path of theRPA with the parallel flight lines shown as dashed lines. The outlined circles show the lo-cations of a vertical image, and the arrows show the horizontal orientation of the cameratowards the channel banks for the oblique images described in (a)We aimed to spread at least ten GCPs across each reach to georeference the imagery,with the remainder serving as checkpoints to assess the accuracy of the model outputs. Themajority of the GCPs were distributed in a zig-zag fashion along dry exposed bars in theperiphery of the reaches, with a smaller number situated towards the centre. While theredoes not appear to be a universally accepted number or distribution of GCPs required foroptimal model outputs at the time of writing (see Sanz-Ablanedo et al., 2018), this set upwas a balance between findings in the literature that suggest using between 4–15 GCPs12with distributions either purely in the periphery, periphery with the addition of centralpoints, or uniformly across a site (Harwin et al., 2015; Agu¨era-Vega et al., 2016; Tonkin andMidgley , 2016; Sanz-Ablanedo et al., 2018). Each GCP consisted of an orange 10 cm x 10 cmceramic tile with a central X marking the surveyed location. The location of the GCPs weremeasured using a Leica TPS 1100 total station through a local survey that included shotsof well-established benchmarks in the study sections. These benchmarks were then used totransform the coordinates into the projected coordinate system NAD83 UTM Zone 10 byapplying an affine transformation through an in-house script and the package ‘vec2dtransf’in R (Carrillo, 2014).2.2.3 PhotogrammetryAgisoft PhotoScan general workflowThe software Agisoft PhotoScan Professional (version 1.4.3) was used to generate densepoint clouds of each reach. These are 3D point clouds of the reach, containing millions ofpoints that are built from the matched points in the photoset (Henry , 2018). The steps itemploys to do this include aligning the photos through matching features across images, andperforming a bundle block adjustment in an arbitrary coordinate system that leads to theformation of a sparse point cloud (Mayer and Kersten, 2018). The addition of GCPs forgeoreferencing the model allows for this adjustment to be further optimized, which refinesthe camera calibration. This can be an important step in mitigating the effect of systematicerrors in the model output derived from non-metric cameras. Following this a dense pointcloud of the optimized sparse cloud can be created. The incorporation of these steps in theworkflow is outlined in greater detail below.Application to channel survey dataFor this analysis, the Estimate Image Quality function in PhotoScan was first used to identifylow quality images, which were then visually assessed to identify blurry photos for removal.Due to erroneous location measurements by the RPA’s on-board GPS, the image coordinateswere cleared prior to alignment. The erroneous measurements mostly occurred early on inthe flights, suggesting that the GPS had not yet connected to a suitable number of satellitesfor accurate positioning prior to the beginning of some flights. The photos were then alignedusing medium accuracy, a key point limit of 40,000 and tie point limit of 2,000. A low densitypoint cloud was then generated to allow the GCPs to be easily identifiable. GCPs were thenselected on the point cloud and the appropriate coordinates were attributed to their positionin the photos. After updating the model, image residuals for the GCPs and tie points were13viewed and adjusted in the reference pane as suggested by James et al. (2017) to ensure thatthey would not be weighted too heavily in the optimization. Then the camera positions wereoptimized with the GCPs selected to refine the camera calibration, leading to the generationof a re-adjusted sparse point cloud. The final dense point cloud was then generated undermedium quality and with aggressive depth filtering. This workflow was repeated for eachreach to generate dense point clouds across the entire study area. Setting the accuracy of thephoto alignment and dense point cloud quality to medium was deemed a suitable balancebetween processing time and representative model outputs.2.2.4 Dense point cloud cleaningThe open source software Cloud Compare (version 2.1) was used to further clean the pointclouds. To remove the influence of vegetation obstructing bed points, the Cloth SimulationFilter (Zhang et al., 2016) was applied to the clouds. Initially developed for processing ofLiDAR point clouds, this tool has rarely been applied to RPA derived point clouds in pre-viously published studies, despite evidence suggesting its high performance relative to otherground filtering algorithms (Yilmaz et al., 2017). Incorporation of this tool may allow RPAsto overcome drawbacks noted by Tamminga et al. (2015), namely their explicit suitabilityto streams with wide active channels where trees minimally obscure view of the channels.To remove the influence of riparian vegetation, the tool inverts the point cloud and ‘drapes’a simulated cloth (Zhang et al., 2016) over the inverted surface to approximate the terrainof an area. The resolution of the cloth, and the maximum distance a point may be fromthe cloud to still be considered part of the ground surface, can be altered to best suit theenvironmental conditions. Setting the cloth resolution to a value of 0.1 m and maximumdistance between 0.5–1.0 m was found to adequately filter the bed points.2.2.5 Water depth correctionsPoint clouds generated from SFM suffer from the effect of refraction causing submerged bedelevations to be over-estimated. To correct for this and develop bathymetry maps of thechannel, a script developed by Dietrich (2017) was used. A requirement of this script isclear water allowing the channel bed to be viewed. During the study period, few opaque orturbulent wetted channel areas were present due to low flow conditions, though a polarizingfilter was also fitted to the drone to mitigate the influence of surface reflections on the watersurface.The method requires a water surface mesh to be delineated across the channel. Bydetermining the distance from the mesh to the SFM bed elevations below, the corrected14water depth for a location on the bed can be calculated as a function of the multiple viewingangles used to observe the location and through the application of a correction factor. Priorto running the script, the clouds were sub-sampled to a spacing of 2 cm while retaining theminimum height in each cell to further clean the clouds and ensure anomalous points fromoverhanging vegetation were removed in Cloud Compare. For more details on the applicationof this method, and how the input files were generated, see Dietrich (2017). Following thecorrection, the uncorrected point clouds of the entire channel and corrected point cloudsfor the wetted area were then transformed into 2 cm resolution Digital Elevation Models(DEMs). These were merged together so that overlapping areas only retained the correctedbed elevations. The corrected water depths were also extracted from the submerged pointclouds to create bathymetry rasters across the channel.2.2.6 Sediment texture analysisGrain size estimates of the exposed bed were acquired through establishing a relationshipbetween the roughness of the point cloud and the D50 and D84 of 22 field training sitesacross the channel (Woodget and Austrums , 2017). Each field site was ∼1 m x 1 m andwas photographed by hovering the RPA ∼2 m above ground level. Using an in-house photo-sieving GUI in Matlab the grain size distributions of each training site was determined. Firstthe photos of each site were scaled based on the known diamter of the GCPs in the images.Then a 10x5 grid was overlaid on the photos, which allowed the b-axis of the 50 particles thatfell below the grid nodes to be measured for calculation of each site’s grain size distribution.The roughness of each point in the point clouds was determined by using the ‘roughnesstool’ in Cloud Compare. This tool computes the roughness of each point relative to thedistance from the point to an ordinary least squares plane, fitted through the point andits user specified neighbourhood. Prior to using this tool, areas of the cloud that did notcontain bed material (e.g. log jams or vegetated banks) were removed. Then the toolwas run with a specified neighborhood of 40 cm. As found by Woodget and Austrums(2017), this neighborhood size appeared to provide a good balance between ensuring thatthe largest clasts in the channel would be encompassed by the window and ensuring thatthe neighborhood was likely to contain homogeneous material. Following the roughnesscalculation, the training sites were extracted from the roughness clouds and exported as 2cm rasters from which the mean roughness value for each raster was determined.A linear model was then fit between both the D50 and D84 (Figure 2.4) of the trainingsites against their mean roughness value. The equation of this line was then used to estimategrain sizes across the entire channel. This was achieved in R by running a moving window15of 1 m x 1 m across the roughness rasters. The mean roughness value for each window wasdetermined and its centre cell replaced with the estimated grain size based on the predictiverelationship (Equation 2.1), where D50p is the predicted grain size for the centre cell of thewindow and Rw is the average roughness value of the cells in the window.y = 0.44 + 445.05 ⋅ x, r2 = 0.922468100.005 0.010 0.015 0.020Roughness of point cloud (mean)D50 (cm) of B−axisy = − 1.57 + 1196.02 ⋅ x, r2 = 0.7201020300.005 0.010 0.015 0.020Roughness of point cloud (mean)D84 (cm) of B−axisFigure 2.4: Predictive grain size relationships between a given grain size (Dx) and the averageroughness value of the training sites.D50p = 0.44 + 445.05Rw (2.1)2.2.7 Large wood extractionLarge wood is an important variable to consider in aquatic habitat delineations as it canprovide both channel complexity and cover for aquatic species. Large wood was manuallydigitized across the channel and the area of each piece calculated using the DEMs andorthomosaics in ArcMap (version 10.6.1). Pieces of wood (larger than approximately 10 cmin diameter and 1 m in length), that appeared to be exerting a control on channel form weredigitized individually, whereas log jams were digitized as polygons.162.3 Results2.3.1 Accuracy assessmentRemotely piloted aircraft surveyA detailed assessment of the accuracy and reproducibility of the sub-canopy RPA resultswas undertaken to determine the suitability of the technique. To do this, a small reachwas chosen early on in the photogrammetric processes that included the use of vertical andintensive oblique imagery to characterize the banks. The reach was chosen as it was relativelysmall for repeat processing (approximately 30 m in length and 10 m in width) and becausethe presence of overhanging vegetation (Figure 2.5) made it a challenging reach to surveythat was representative of difficult portions of the channel. As shown in Table 2.1, a greaternumber of photos included in the photoset was associated with a much greater time requiredto generate the dense point cloud.Figure 2.5: (a) Downstream view to a log jam at the beginning of the study reach. (b)Upstream view showing the presence of deciduous branches and overhanging vegetation thatwould require the RPA flight be low altitude and include oblique imagery. Note that theimages were taken during the fall total station survey, and therefore the water levels arehigher and leaf cover is lower than it was during the summer RPA-survey. Photos courtesyof Iain Reid.17Table 2.1: Computer processing time required for each photoset and GCP combination inPhotoScan.PhotomatchingPhotoalignmentDepth mapgenerationDense cloudgenerationTotal timeNumber ofphotosNumber ofpoints in cloudp1t1 13min 15s 5min 46s 35min 47s 3h 6min ∼4h 592 26,015,291p2t1 4min 3s 2min 27s 8min 19s 30min 31s ∼45min 311 21,252,408p3t1 3min 42s 3min 29s 8min 24s 25min 52s ∼40min 293 21,063,688p4t1 5min 41s 1min 3s 18min 3s 9min 5s ∼35min 234 17,616,174p1t2 13min 15s 5min 46s 36min 58s 3h 46min ∼4h 592 26,092,604p2t2 4min 03s 2min 27s 8min 15s 29min 43s ∼45min 311 21,243,270p3t2 3min 42s 3min 29s 8min 1s 24min 57s ∼40min 293 21,195,514p4t2 5min 41s 1min 03s 17min 53s 9min 02s ∼35min 234 17,531,918p1t3 13min 15s 5min 46s 36min 45s 3hr 5min ∼4h 592 29,941,363p2t3 4min 03s 2min 27s 9min 07s 29min 37s ∼45min 311 21,311,393p3t3 3min 42s 3min 29s 8min 01s 23min 29s ∼40min 293 20,824,947p4t3 5min 41s 1min 03s 17min 45s 9min 03s ∼35min 234 17,608,597Four combinations of photos were used, along with three combinations of GCPs as out-lined in Table 2.2 and Figure 2.6, giving rise to twelve combinations of GCPs and photosetsthat were run through PhotoScan to generate DEMs of the site. Using only odd or evenlynumbered photographs would allow for the creation of DEMs using two entirely uniquephotosets, that allow for an assessment of the reproducibility of the model outputs.Table 2.2: Photoset and ground control point (GCP) and check point (CP) combinations.Modifier Descriptionp1 all photosp2 odd photosp3 even photosp4 vertical photos onlyt1 11 GCPs and 11 CPst2 5 edge GCPS and 17 CPst3 5 edge GCPs, 3 centre GCPs and 14 CPs18Figure 2.6: (a) Ground control point (CP) and check point (CH) distribution for modifiert1.(b) Ground control point (CP) and check point distribution (CH) for modifier t2.(c)Ground control point (CP) and check point (CH) distribution for modifier t3. Coordinatesystem is NAD83 UTM Zone 10.19Following this, the difference between the estimated elevations from each DEM and theactual elevations for the GCPs and CPs measured using the total station was determined. Asseen in Figure 2.7, the majority of the errors for all models was within 2 cm, suggesting thatthe choice of flight path and GCP configuration had a small influence on the error. This wasfurther confirmed by plotting the values from each DEM against one another to calculatetheir correlation coefficients (Table 2.3), which showed strong correlation between all DEMsregardless of the combination used. The difference in the proportion of channel that couldbe captured with the incorporation of oblique imagery, in contrast to the photosets whereonly vertical images were used, can be visually compared in Figure 2.8.Control points Check pointsp1t1p1t2p1t3p2t1p2t2p2t3p3t1p3t2p3t3p4t1p4t2p4t3p1t1p1t2p1t3p2t1p2t2p2t3p3t1p3t2p3t3p4t1p4t2p4t3−0.020.000.02Photoset & GCP combinationVertical error (m)Figure 2.7: Boxplots showing the distribution of vertical error for the GCPs and CPs ofeach accuracy assessment. The boxes represent the interquartile range (IQR) and the solidblack lines represent the median error for each photoset and GCP combination. The verticalwhiskers represent the extent of the smallest and largest errors within 1.5 times the IQRbelow the 25th (Q1) or above the 75th (Q3) percentile, respectively. The dashed horizontalline represents a vertical error of 0. The dots are outliers which are either less than Q1 -1.5*IQR or greater than Q3 + 1.5*IQR.20Table 2.3: Correlation matrix describing the similarity of the rasters (through their correla-tion coefficients) generated for all twelve unique GCP and photoset combinations.p1t2 p1t3 p2t1 p2t2 p2t3 p3t1 p3t2 p3t3 p4t1 p4t2 p4t3p1t1 0.992 0.992 0.987 0.987 0.987 0.986 0.987 0.987 0.986 0.956 0.958p1t2 0.997 0.991 0.993 0.992 0.99 0.992 0.991 0.991 0.959 0.962p1t3 0.99 0.993 0.992 0.989 0.991 0.99 0.991 0.958 0.961p2t1 0.99 0.994 0.988 0.989 0.989 0.987 0.956 0.958p2t2 0.991 0.986 0.989 0.988 0.987 0.954 0.956p2t3 0.988 0.99 0.989 0.987 0.957 0.959p3t1 0.992 0.994 0.986 0.958 0.96p3t2 0.994 0.987 0.958 0.96p3t3 0.988 0.958 0.96p4t1 0.963 0.966p4t2 0.983Figure 2.8: Difference in the proportion of bed that can be covered between only a verticalimage photoset (a) and the incorporation of oblique images (b). The DEM with both verticaland oblique imagery (b) has more well defined channel banks as seen by the channel roughnessprovided by bank vegetation towards the margins of the channel.21The vertical error throughout the entire channel was estimated by comparing the eleva-tions of check points not used for georeferencing the models to the elevations estimated bythe DEMs. In total, 1724 checkpoints were spread across the channel, 1203 of which werelocated on the exposed bed and 521 of which were submerged below the water surface. Themetrics used to characterize the vertical error in the models were the root-mean-square-error(RMSE) and the mean error (ME). The RMSE provides a measure of the spread of thesquared residuals (Equation 2.2) whereas the ME provides a measure of any potential posi-tive or negative bias to the data (Equation 2.3). These values were chosen as they providecomparable metrics to other accuracy assessments of RPA-surveys completed in the litera-ture (e.g. Tamminga, 2016). The RMSE of the exposed bed points was found to be 0.093m and 0.025 m for the ME (n = 1203), whereas for the submerged bed points the RMSEand ME was 0.1 m and 0.025 m (n = 521), respectively. The overall spread of this error isillustrated in Figure 2.9.RMSE =√∑ni (zmod − zobs)2n(2.2)ME =∑ni (zmod − zobs)n(2.3)05101520−0.4 −0.2 0.0 0.2 0.4Vertical error (m)Density DatasetSubmergedExposedFigure 2.9: Density plot showing the distribution of vertical errors between the modelledand total station measured elevations. The dashed line represents the mean error for thedatasets, and the solid line a vertical error of 0 m.22Bathymetry analysisTo assess the accuracy of the corrected submerged bed elevations, two locations were chosenin the channel where dense measurements of the bed were collected using the total station(Figure 2.10). The values from the uncorrected and corrected DEMs were compared to thesetrue values as seen in Figures 2.11 and 2.12. Prior to the correction, these values deviatedfrom the 1:1 line due to the influence of refraction, with the deviation greater at larger waterdepths. Following the correction, the observations plot along the 1:1 line, indicating that thecorrections were successful.(a) (b)Figure 2.10: Distribution of GCPs for the accuracy assessment of the bathymetry correctionat sites (a) and (b).The bed texture is displayed through a hillshade layer, with smoothertextures corresponding with finer material, and very rough textures towards the margins ofthe channel corresponding with vegetated areas. The extent of the water surface is indicatedin blue, with darker blues corresponding with deeper sections of the channel.23Uncorrected Corrected10.5 11.0 11.5 10.5 11.0 11.510.511.011.5Estimated elevation (m)True elevation (m)1:1best−fitFigure 2.11: Comparison of the submerged bed elevations to the total station measuredelevation before and after the bathymetry correction for the site in Figure 2.10a. The 1:1line is represented by the solid black line.Uncorrected Corrected18.5 19.0 19.5 20.0 18.5 19.0 19.5 20.018.519.019.520.0Estimated elevation (m)True elevation (m)1:1best−fitFigure 2.12: Comparison of the submerged bed elevations to the total station measuredelevation before and after the bathymetry correction for the site in Figure 2.10b. The 1:1line is represented by the solid black line.2.3.2 Survey coverageThe area of channel covered by the RPA and total station in the eight geomorphology studysections were compared to evaluate the total coverage obtainable from the RPA-survey. Thearea of the channel covered compared to the study section boundaries is illustrated in Figures2.13 and 2.14, and in Table 2.4. When including side channels in channel boundary, which24were generally difficult to access with the RPA due to dense vegetation, it was possibleto capture approximately 80% of the bounding area, whereas when only considering themain channel as a boundary, the RPA-survey captured approximately 87% of the boundingarea. As time didn’t permit all of SA9 to be surveyed with the RPA, the comparison isbased on just the lower half of the study section. Figure 2.15 illustrates a situation wheredense vegetation and large wood prevented the middle portion of the study section frombeing captured. Figure 2.16 illustrates how oblique imagery allowed a deep pool covered byoverhanging vegetation to be characterized.Figure 2.13: RPA coverage overlapped by the study section boundaries indicated in blackfor SAs 2–5. The texture of the bed surface is displayed through a hillshade layer and thewater surface extent in blue, with deeper areas of the channel represented by a darker blue.25Figure 2.14: RPA coverage overlapped by the study section boundaries indicated in blackfor SAs 6–9. The texture of the bed surface is displayed through a hillshade layer and thewater surface extent in blue, with deeper areas of the channel represented by a darker blue.Table 2.4: Percentages of the study section covered with the RPA relative to the totalstation surveys. The comparisons are based on whether the reference boundary includedside channels (with side) or just the main channel (no side).Study section RPA: with side (%) RPA:no side (%)SA 2 90 90SA 3 83 83SA 4 54 74SA 5 81 84SA 6 91 91SA 7 67 79SA 8 94 94SA 9 89 99Average 81 8726Figure 2.15: Coverage of study section five in an area with low canopy (a) and a missedportion due to dense vegetation and a small log jam (b).The channel’s sediment texture ischaracterized by a hillshade layer and the water extent in blue, with deeper areas corre-sponding with darker blues. Note that the photo was taken in the fall when the water levelwas higher than that during the RPA-survey. Photo courtesy of Iain Reid.27Figure 2.16: Coverage of a deep pool in study section eight under dense riparian vegetation.The channels sediment texture is characterized by a hillshade layer and the waters extent inblue, with deeper areas corresponding with darker blues. Note that the photo was taken inthe fall when the water level was higher than it was during the RPA-survey. Photo courtesyof Iain Reid.2.4 DiscussionThe results of this study provide a precedent for using RPAs to characterize channel mor-phology in small, forested streams using sub-canopy imagery. In twelve field days, threekilometers of channel were surveyed with an estimated coverage rate of 80%. The DEMsand orthophotos created from these images were extremely high resolution at 2 cm / pixelwith a RMSE for the exposed elevations of 0.093 m and 0.1 m for the submerged bed el-evations. This magnitude of error is comparable to values observed in other studies, suchas Tamminga et al. (2015), who reported a RMSE of 0.088 m for exposed portions of thebed and 0.11 m for submerged portions of the bed, and Flener et al. (2013), who createda DTM of a stream channel using RPA imagery that yielded a RMSE of 0.088 m as well.In comparison to other airborne remote sensing techniques such as LiDAR, these results are28comparable to those found by Legleiter (2012), who reported a vertical RMSE of 0.21 m.However, the spacing of LiDAR datasets is often on the decimeter scale (Notebaert et al.,2009; Legleiter , 2012; Abalharth et al., 2015), which may not be suitable for resolving thesediment texture of the bed in the same manner as has been demonstrated in RPA-surveys(e.g. Woodget et al., 2017). Furthermore, in comparison to LiDAR, RPA imagery offers anaffordable and often more accessible tool for characterizing these systems, especially whentrying to characterize the submerged bed, which requires the use of expensive green bandLiDAR (Muste et al., 2012).From this high resolution imagery, we demonstrated the utility of methodologies devel-oped by Woodget et al. (2017) and Dietrich (2017) for accurately extracting sediment textureand bathymetry across the channel in a small, forested stream system. The RMSE valuesfor the submerged bed elevations are comparable to the RMSE values of the field studiesconducted by Dietrich (2017), who reported RMSE values ranging from 0.06–0.077 m fordifferent surveys of his study site. Using the roughness of the point cloud to derive estimatesof grain size across the channel, a strong predictive relationship was acquired for the D50 withan r2 value of 0.92. This is comparable to the r2 values reported in other studies relatingimage texture or roughness to predicted grain size (see Carbonneau et al., 2004; Tammingaet al., 2015; Woodget et al., 2017).The nature of small, forested streams, with their dense canopies, has long left the im-pression that these characteristics preclude them from the benefits that larger streams haveexperienced with RPA-surveys. However, through the incorporation of oblique-convergentimagery these results demonstrate that sub-canopy flying of RPAs can provide reasonablecoverage of these systems at large spatial extents, as illustrated in Figure 2.8. This applica-tion has several advantages over techniques traditionally used to characterize small, forestedstreams. Pole-mounted photogrammetry, where a camera is attached to a pole suspendedseveral meters above the bed to collect imagery of a site (e.g. Bird et al., 2010) is perhapsthe most similar approach to surveying such streams to date. Using this method, Bird et al.(2010), surveyed approximately 200m2 of the bed using 44 GCPs and 14 images that took10 minutes to capture in a traditional photogrammetric workflow. This allowed the devel-opment of a DEM with a ground resolution of 0.03 m and errors within a range of -0.2 to0.2 m. In comparison to this technique, RPAs provide a greater capacity to collect a largernumber of images, which is required for SFM photogrammetry, in a shorter period of time.For example, the channel segment used in the accuracy assessment (Figure 2.5), was approx-imately 275 m2, and included the use of 592 photos which took approximately 20 minutes tocapture.In contrast to a total station topographic survey, the sub-canopy RPA-survey allows the29collection of a greater density of bed points that can be post-processed to provide continuousmeasurements of grain size, water depths, large wood and bed elevations across the channelat resolutions that would be challenging to achieve using a total station survey. It is alsoworthwhile noting that while the total station survey allows for total coverage of the bed,this is through much lower resolution data. The point clouds for the study sections from thetotal station tend to vary between 500–1000 points, whereas the dense point clouds from theRPA imagery contain millions of points per a reach.As with each of the previously listed techniques, there are still drawbacks to sub-canopyRPA-surveys. First, due to fallen trees obstructing flight paths and areas with dense low-lying riparian vegetation completely obstructing view of the bed, there is still about 20% ofthe channel that was not covered with this technique. Depending on the specific researchobjective, this missing portion may be negligible given the greater longitudinal distance thatcan be covered with the RPA. However, if assessment of the entire channel is required, it maybe necessary to consider a combined approach, such as inclusion of additional total stationmeasurements along with a spatial interpolation method to characterize portions of thechannel not observed with the RPA. For example, in delineating aquatic habitat, undercutbanks which are situated along the channel edges may be important to identify as they canprovide critical cover for salmonids (Nelson et al., 2015). These portions of the channel couldbe roughly characterized with a modified total station survey that included coarse dimensionmeasurements of the undercuts. Furthermore, although LiDAR is more expensive, it is bettersuited to penetrating through to the bed in densely vegetated portions of the channel. Whilethe cloth simulation filter did help to mitigate this limitation for the RPA imagery, it stillrequires that there be viewing angles present that provide perspectives of partially obstructedportions of the bed. In contrast to more traditional aerial platforms, the RPA provided theopportunity to fly below the canopy and capture portions of the bed that would typicallybe hidden from view of airplane or satellite platforms. However, a disadvantage of this closerange imagery is the greater number of photos required to characterize a site and thereforethe greater time required for processing the imagery.There are certain environmental conditions to which the methods discussed in this chap-ter may also not be suited. Notably, the techniques for extracting the bathymetry may notbe suitable for streams with highly turbid water that prevent observation of the submergedbed. Furthermore, areas that were particularly difficult to survey included sections of thechannel with low-lying deciduous branches that spanned the channel, that while not fullyobstructing the view of the bed, prevented the passing of the RPA. Therefore, these tech-niques may be most suited to small channels in relatively mature forests that have a moreopen understory. Finally, while the field time for collecting the data was markedly reduced30compared to a total station survey, this came at the cost of greater time required for pro-cessing of the imagery. However, with improvements in processing and greater availabilityof cloud computing resources, this is an area that may see improvement in the future.Overall, this research helps to expand the toolkit available to geomorphologists for char-acterizing small, forested streams. Prior to this work, the use of RPAs was restricted to widerchannels or those without dense riparian vegetation. The methods described provided highresolution and continuous measurements of bathymetry, grain size and bed elevations thatare typically difficult to acquire in small, forested streams. If the corresponding salmoniddata were available at the same scale, this spatially continuous dataset would help to bridgethe gap between the conceptual theories of the Riverscape Approach proposed by Fauschet al. (2002) and the distribution of salmonids in these systems. For example, through usingthese data to delineate patterns in aquatic habitat in multiple watersheds with the corre-sponding fish assemblage and abundance data, it may be possible to assess how patternsin aquatic habitat diversity and connectivity impact fish populations. The demonstratedability to acquire these data provides a step towards helping to fully realize the influence ofbasin-scale patterns in aquatic habitat on stream fishes.2.5 ConclusionThis study provided a first attempt at testing the applicability of a sub-canopy RPA-surveyfor characterizing channel morphology in a small, forested mountain channel. Through theincorporation of oblique-convergent imagery, it was possible to survey 3 km of CarnationCreek at a coverage rate of approximately 80%. Building on prior work analyzing RPA-derived point clouds, information on channel features such as grain size and bathymetry wereextracted at resolutions that would be difficult to acquire using traditional survey methodslike a total station topographic survey. A drawback to this survey is the significant com-putation power required to process the data. However, with recent technological advancesin cloud processing and photogrammetry software, this is an area that is likely to continueto improve and help make the use of RPA-surveys accessible to the general public. Overall,the methods were very successful in demonstrating the use of RPAs for characterizing smallstreams with clear water and a relatively open sub-canopy. In contrast to other remotesensing technologies, the RPA-survey described provides a less costly and readily deploy-able alternative. In particular, this research provides precedent and will hopefully initiatediscussion on the role that RPAs can have in characterizing sub-canopy stream systems.31Chapter 3Rapid and objective classification ofchannel morphology and diversity atthe basin-scale3.1 IntroductionAdvances in the understanding of the relationship between the life processes of some aquaticspecies and their use of aquatic habitat demand a new approach to aquatic habitat mappingthat can be conducted at the basin-scale. This need is well noted by several papers in the1990s that encouraged characterizing ecological processes in stream systems continuously ata scale relevant to species’ life histories, deemed intermediate at 1–100 km in length (e.g.Schlosser , 1991; Fausch et al., 2002). When considering the process-based channel morphol-ogy classification scheme proposed by Montgomery and Buffington (1997), it follows that atthese scales streams may display unique characteristics, depending on the arrangement ofdifferent morphologies and their driving processes. However, this presents a mismatch withthe scale generally used to study aquatic species. Traditional survey techniques are time-consuming and tend to be conducted in study reaches 30–50 channel widths in length. Thisscale may not be well-suited to characterizing heterogeneity in stream systems, and differentapproaches to broad-scale mapping of aquatic habitat are needed.In response to this need, there has been a recent proliferation of studies providing newmethodologies to characterize aquatic habitat at these large scales. For example, Linden-schmidt and Long (2013) developed a GIS and principal component analysis (PCA) approachto extract channel variables (sinuosity, fractal dimension, width and slope) from coarse DEMsof large streams to identify areas with unique geomorphic characteristics. Several studies32have since employed this approach to relate patterns in channel morphologies to patterns inaquatic species assemblages (e.g. Meissner et al., 2016; Liu et al., 2017). However, to datethese variables have only been extracted at large scales, with sample sites located as far asevery 50 m along the longitudinal profile. While this may help to pull out broad basin-scalepatterns, this method is likely to miss heterogeneities across smaller streams that may havean important effect on aquatic habitat. Furthermore, the model was developed for the LittleSaskatchewan River, Manitoba, an environment that is different from the small, forestedmountain channels common in the Pacific Northwest.Similar to the work by Lindenschmidt and Long (2013), Hugue et al. (2016), developeda technique to classify patterns in fast water and slow water channel morphologies across astream using a PCA. To do this, they used satellite imagery of a 17 km reach from whichthey derived bathymetry data with 0.5 m resolution in the panchromatic and 2 m in themultispectral. Application of a PCA to the wetted variables derived from a 2D hydrodynamicmodel allowed them to identify channel types and characterize diversity across the channelbased on patterns in velocities and water depths. The meso-scale variables of the analysisprovided a good framework suited to identifying local heterogeneities across the channel.An area that has not yet benefited from rapid characterization of channel morphologyusing remotely sensed data, as demonstrated by Lindenschmidt and Long (2013) and Hugueet al. (2016), is small, forested mountain channels. This research seeks to provide a frameworkfor characterizing basin-scale patterns in aquatic habitat of small, forested stream systemsthat does not require a hydraulic analysis and is rather based on easily extractable metricsfrom RPA imagery. To build this framework, the chapter aims to meet three objectives:ˆ Develop a rapid and objective means of characterizing channel morphology near-continuouslyacross the channel (at 1 m intervals) using a PCA-clustering technique on easily ex-tractable RPA data.ˆ Build upon metrics for characterizing diversity to investigate patterns in channel mor-phology diversity.ˆ Investigate the necessary scale required to study the system to ensure that its hetero-geneity is captured.A better understanding of the patterns in channel morphology and diversity is of impor-tance as it is likely to facilitate improved conservation monitoring of these systems.333.2 Methods3.2.1 Selection of channel variablesTo classify the channel along the longitudinal profile, it was necessary to define a standardizedlocation along which observations would be extracted across the channel. The thalweg waschosen for this purpose as it is a feature that could be easily identified across the channel andwould likely highlight patterns key to discriminating channel types, such as differences inslope and water depth. To characterize patterns in channel morphology, five variables werechosen for the classification: the hydraulic radius (Rh), local bed (Sl) and water surface slope(Sws), reach bed slope (Sr), as well as the average D50 for cross sections along the thalweg.These channel characteristics reflect larger basin-scale controls on channel morphology, suchas geology, climate and land-use that influence watershed conditions (e.g. streamflow andsediment supply) which ultimately give rise to patterns in grain size, channel geometry, bedslope and bed forms across the riverscape (Buffington and Woodsmith, 2003). The interplayof these channel characteristics gives rise to unique channel morphologies, and it thereforefollows that their extraction should aid in the classification of channel types.Incorporation of slopeBed slope was included in the classification, as it helps inform the distribution of chan-nel morphologies across the basin. Channel morphologies tend to progress from pool-riffle,plane-bed, and step-pool to cascade morphologies with increasing channel slope (Montgomeryand Buffington, 1997). Of this sequence of morphologies, it is the pool-riffle unit that hascharacteristics that provide both important and functionally different habitats for salmonids(Hawkins et al., 1993). Pools, by nature of their deep and slow moving water, can providecover and refuge in low flow conditions. In contrast, riffles, which are characterized by fastmoving and shallow water, can be an important food source for salmonids via invertebratedrift (Leung et al., 2009; Naman et al., 2018). It is generally accepted that glides (low ve-locity, depth and slope) and runs (moderate velocity, depth and slope) represent transitionalmorphologies between these two end members in channel morphology within the pool-riffleunit (Wyrick et al., 2014). It is in these transitional environments, between the tail endof pools and beginning of riffles, that fish tend to spawn, where differences in bed topogra-phy lead to hyporheic exchange (Tonina and Buffington, 2009), that promote the necessarywell-oxygenated conditions for spawning habitat (Pfeiffer and Finnegan, 2017). Collectively,the unique characteristics of each of these channel types within the pool-riffle unit providesalmonids with an array of mesohabitats suited to their life histories. Therefore, the reach34Figure 3.1: Diagram of a typical pool-riffle transition, based on Garcia et al. (2012). Notethe fining of bed material at the entry of the pool, and coarsening of bed material as theriffle is approached.scale slope was considered to help aid the discrimination between pool-riffle and plane-bedmorphologies closer to the canyon, and the local bed and water surface slope to discriminatebetween pool, riffle, glide and run morphologies within the pool-riffle unit (Figure 3.1).Incorporation of hydraulic geometryThe hydraulic radius (Rh) was extracted from the cross sections to provide a measure of thechannel’s hydraulic geometry. The hydraulic radius was calculated according to Equation3.1, where A is wetted area and P is the wetted perimeter for the cross section. Essentially,the hydraulic radius indicates how much of the area of water in a cross section is in contactwith the bed, with a larger hydraulic radius corresponding with portions of the channel thathave less friction imposed on the water and therefore the potential for greater flow velocities.Rh =AP(3.1)Incorporation of channel roughnessTo provide a measure of the grain roughness across the channel, the average D50 of the dryexposed bars in a 0.5 m buffer around each cross section was extracted. According to theLane Balance (3.2) (Lane, 1955), in an equilibrium channel, grain size reflects a balancebetween channel conditions, such as slope (S), discharge (Q) and sediment flux (Qs). Theinterplay of these governing conditions determines channel morphology (Church M., 2002).Indeed, there is a general coarsening in bed material from glides and pools with finer materialto riffles and runs (Garcia et al., 2012).35D αQSQs(3.2)3.2.2 Extraction of channel variablesThalweg extractionTo extract the thalweg, the River Bathymetry Toolkit (RBT), an ArcMap add-in, was appliedto the RPA-derived DEMs generated of the bed across Carnation Creek (McKean et al.,2009). Inputs for the thalweg tool of the RBT include a detrended DEM, bankfull polygonand channel centreline. The tool expects a DEM that approximates the valley and channelbed, as opposed to roughness elements such as large wood, which the RPA imagery doesnot provide. To mitigate this, a pseudo-valley was established by fitting a plane through thepoint clouds of the channel bed and adding 1.5 m to the planes to approximate the channelvalley. Channel obstructions such as large wood were also clipped from the point cloudsin Cloud Compare and their bed elevations interpolated using inverse distance weighting tocreate clean and complete DEMs of the channel bed. The cleaned DEMs were then mergedwith the valley planes and the ‘detrend’ tool in the RBT applied to them.To provide estimates of the channel centreline, the ‘collapse dual lines to centreline’ toolwas applied to polylines of the right and left bank in ArcMap. With the detrended DEM,bankfull polygon and centreline as inputs, the thalweg was extracted using the ‘thalweg’ toolof the RBT. The tool runs a least-cost-path analysis based on the centreline and channelelevations, with deeper sections of the channel receiving a higher weighting to determine thethalweg.Feature extraction along thalwegTo characterize patterns in channel morphology along the thalweg, a function was developedto sample points along the thalweg every metre at which the hydraulic radius (Rh), localbed (Sl) and water surface slope (Sws), reach bed slope (Sr), as well as the average D50for a neighbourhood were extracted from the RPA derived rasters of the bed elevations,bathymetry and D50 as described below and illustrated in Figure 3.2. Cross sections wherethe channel banks were not discernable (e.g. due to channel obstructions or dense low-lyingvegetation) were excluded from the analysis.36Figure 3.2: Feature extraction set up. The Rh, Sl, Sws, and Sr were extracted every metre.The average grain size was extracted every meter based on the average value of the D50raster in a 0.5 m buffer around each cross-section line.Slope extractionThe local slopes of the bed and water surface were calculated for each sampling location usinga function in R that fits a linear model through observations in a 15-metre window aroundeach sample site. This window size was chosen based on field observations at Carnation Creekthat suggested it would be sufficient for capturing smaller scale patterns in slope within thepool-riffle unit. This was repeated for the reach-scale bed slope using a 45-metre windowbased on previous knowledge that this is equal to the average pool-riffle spacing at CarnationCreek (S. Bird, personal communications, April 2018).Hydraulic radius extractionIn wide and shallow channels, the hydraulic radius can be well approximated by the waterdepth. However, many of the pools along Carnation Creek occur in deep and narrow sectionsof the channel, thus violating the wide channel approximation assumption. Therefore, thehydraulic radius for all cross sections across the channel was calculated using Equation 3.1.To calculate the length of the wetted perimeter, the wetted cross section lines were firstconverted to 3D polylines in ArcMap using the ‘Interpolate Shape’ tool from the 3D AnalystToolbox. This attributes depths to the cross section lines every 10 cm from the bathymetryrasters. The wetted perimeter of each cross section was determined by using the ‘Add Surface37Information’ tool in ArcMap from the 3D Analyst Toolbox. For each wetted perimeter line,the area below a plot of distance along cross section against water depth was then calculatedusing the ‘trapz’ function in the Practical Numerical Math Function (‘pracma’) package inR (Borchers , 2018). The hydraulic radius was then calculated as the wetted area at eachcross section divided by its wetted surface length.D50 extractionTo provide a measure of the grain roughness across the channel, the average D50 of the dryexposed bars in a 0.5 m buffer around each cross section was extracted. Together, thesevariables summarize the channel form (Rh and S) and roughness (n, approximated by D50)of each cross section, akin to the variables used for the Manning Equation (1891) (Equation3.3), which provides an estimate of the channel’s velocity. Therefore, there is mathematicalimpetus to using these variables to discriminate between slow-moving residual pools andfaster riffle morphologies that will provide unique aquatic habitats.v =1nR23hS15 (3.3)3.2.3 Principal component analysisUsing the extracted channel variables, an attempt was made to rapidly and objectivelycharacterize channel morphologies using a PCA and k-means clustering algorithm using thepackage ‘stats’ in R (R Core Team, 2018). The general purpose of a PCA is to reduce thenumber of dimensions in a dataset that contains interrelated variables while describing themaximum amount of variation present (Jolliffe, 2002). This is achieved by a transformationthat projects the data along principal components, which are axes that best explain thevariation of the data, with the first component explaining the most variation, followed bythe remaining components (Jolliffe, 2002). Throughout the transformation, the relativepositions of the observations to one another remain the same. This allows componentswhich do not explain much of the variation to be removed, thereby allowing the data tobe visualized using fewer dimensions. Because the dataset was multi-dimensional, with fivevariables over 2362 sampling sites, a PCA was an appropriate analysis tool to help simplifyand pull out patterns in the data, which is beneficial for k-means clustering.Eigenvector selectionThe PCA was first run using all five dimensions, from which a decision was made on howmany dimensions were necessary to explain the variation in the data. To make this decision,38the cumulative percentage of variation explained by the principal components, as well asthe variation explained by each eigenvector was examined. Eigenvectors are vectors thatexplain the directions along which the data varies, whereas eigenvalues explain the amount ofvariation explained by each of the principal components. A commonly used rule to determinethe number of components to use is to choose a cut-off value between 70% and 90% for thecumulative variance explained by the components (Jolliffe, 2002). The scree plot in Figure3.3 shows that the first two and first three components explain 64.6% and 79.0% of thevariation in the dataset, respectively. This suggests that three components would be asuitable number to retain. Considering the eigenvalues of the principal components (Figure3.4) shows that the first two components have eigenvalues greater than or close to one. Aneigenvalue of less than one represents components which explain less variance than one ofthe original variables (Jolliffe, 2002). However, to allow for sampling variation, it has beensuggested that an eigenvalue cut-off value of 0.7 may be appropriate (Jolliffe, 2002). As thethird component has an eigenvalue of 0.73, an argument can be further made for retainingjust the first three principal components.45.1%19.3%14.6% 12.4%8.7%020401 2 3 4 5DimensionsPercentage of explained variancesScree plotFigure 3.3: Scree plot of the variance explained by the principal components.390.51.01.52.01 2 3 4 5Dimensions (m)EigenvalueFigure 3.4: Eigenvalues representing variance explained by the principal components (di-mensions).3.2.4 K-means clusteringFollowing the PCA, a k-means clustering algorithm was run to identify clusters in the dataalong the first three components. A k-means clustering algorithm is an unsupervised classifi-cation that assigns observations from ‘n’ dimensions to clusters that allow the within-clustersum of squares to be minimized (Hartigan and Wong , 1979). A key decision in a clusteringanalysis is determining the number of clusters to be used to classify the data. Initially whenadding more clusters, the variation within each group will steeply decrease, but eventuallythere will be a point where the addition of more clusters has a marginal influence on decreas-ing the within-cluster variance. This point, which would be located at an elbow or plateauin a plot of clusters against within-cluster variance is often deemed an appropriate point tochoose as a cut-off (Flynt and Dean, 2016). To identify this location, the k-means clusteringalgorithm was run with the number of clusters ranging from one to 20 (a value greater thanthe number of morphologies to be expected at Carnation Creek) and a plot of within-clustervariance against number of clusters examined. As seen in Figure 3.5, at about 5–7 clustersthe curve becomes much less steep, which matches well with the number of morphologiesone may expect at Carnation Creek, being pool, riffle, run, glide and plane-bed morphologiesnear the canyon. Therefore, the k-means clustering algorithm was run with the number ofclusters set to six.405 10 15 2020006000Within groups sum of squaresNumber of ClustersFigure 3.5: K-means clustering plot demonstrating how the within groups sum of squaresdecreases with the incorporation of more clusters.3.2.5 Channel diversity classificationFollowing clustering of the cross-sectional variables, the mean values for each of the channelvariables of each cluster were examined and a channel type was attributed to each cluster.This provided a continuous classification of channel types across Carnation Creek from themouth to the beginning of the canyon at 1 m intervals. From these data, a measure of chan-nel diversity was acquired throughout the stream. To characterize the diversity of channelmorphology across the stream, a moving analysis using the Shannon diversity index (Shan-non and Weaver , 1964) was conducted. The Shannon diversity index provides a measureof the abundance and evenness in an area (Lloyd and Ghelardi , 1964). Whereas it is oftencalculated with regard to species types in ecology, in the context of this research channeltypes are instead considered, similar to the work of Harris et al. (2009). To calculate it, theproportion of each channel type (pi) in an area is multiplied by the natural logarithm of theproportion. This value for all the channel types present in the system is summed up yieldingthe Shannon diversity index (Equation 3.4).H = −∑pilnpi (3.4)413.2.6 Scale effectsFinally, a relevant question with regards to managing these systems is determining the scalethat needs to be surveyed to ensure that their heterogeneity is likely captured. To answerthis, the standard deviation of diversity was calculated for window sizes ranging from 15–600m in length. It would be expected that eventually the standard deviation would approach afinite value with increasing sample size, providing an estimate of how much of the channelit is necessary to survey to ensure that its heterogeneity is captured.3.3 ResultsInterpretation of the PCAThe first three components from the PCA explained approximately 80% of the variation inthe data, with components one, two and three reflecting 45.11%, 19.3% and 14.6% of thevariation respectively. The contributions of each variable to each component are illustratedin Figure 3.6, with larger contributions indicating that the variable has a greater influence onthe component. The first component is most largely represented by Sr, D50 and Sws (Figure3.6a), the second by the Rh (Figure 3.6b), and the third by the Sl and D50 (Figure 3.6c).42(a)01020D 50Slope rSlope wsSlope l R hContributions (%)Contribution of variables to Dim−1(b)020406080R hSlope lSlope rSlope ws D 50Contributions (%)Contribution of variables to Dim−2(c)0204060Slope l D 50Slope r R hSlope wsContributions (%)Contribution of variables to Dim−3Figure 3.6: Variable contributions to Principal Components. The red-dashed line representsthe mean contribution of the variables at 20%.The correlation circle, Figure 3.7, provides a representation of the quality and correlationof the variables with one another along the first two components. Variables that are of highquality along a dimension are represented by their high cos2 value and the greater distancethey plot from the circle’s centre. Variables that are closely related tend to plot close to oneanother, while those that are inversely related tend to plot at 180◦ from one another, andthose with no relation tend to plot perpendicular to one another. As seen in Figure 3.7, D50,Sr and Sws are of relatively high quality and plot in a slightly opposite direction to the Rh.Sl is of relatively low quality, but plots in the same direction as D50, Sr and Sws, suggestingthat the variables are weakly correlated.43SlopewsD50SlopelSloperRh−1.0−0.50.00.51.0−1.0 −0.5 0.0 0.5 1.0Dim1 (45.1%)Dim2 (19.3%)0.60.70.80.9cos2Variables − PCAFigure 3.7: Correlation circle plot. The colour ramp describes the quality of the representa-tion along the component.3.3.1 Clustering analysisAfter running the k-means clustering algorithm using six groupings on the first three compo-nents, similar patterns appear in the biplot (Figure 3.8 and Figure 3.9), compared to thoseobserved in the correlation circle (Figure 3.7). For each cluster, the mean of each variablewas determined and the likely morphology attributed to it based on these values (Table 3.1).Moving from left to right along the first dimension there is a transition from morphologieswith shallow bed and water surface slopes with finer material, to steeper morphologies withcoarser material (Figure 3.8). This appears to represent a transition from pool to rifflemorphologies along the first component. Within the riffle channel type, the classificationalso captures a distinction between riffle morphologies with slightly coarser bed material,which received the channel type “Riffle-coarse” (RiffleC) to characterize this change. Movingdown the y-axis, shallower environments are encountered with decreasing local bed slopes,as indicated by the transition from pool to glide morphologies (Figure 3.8).44Table 3.1: Means of channel variables for each cluster. l is the longitudinal distance, d isthe thalweg depth, Rh is the hydraulic radius, Sl is the local slope, Sws is the water surfaceslope, Sr is the reach slope, D50 is the median grain size and W is the channel width.Cluster l (m) d (m) Rh (m/m) Sl (m/m) Sws (m/m) Sr (m/m) D50 (cm) W (m)RiffleC 2980 0.16 0.12 0.018 0.018 0.024 6.74 4.13Plane-bed 3160 0.20 0.14 0.054 0.047 0.042 8.21 3.47Riffle 1650 0.13 0.090 0.027 0.016 0.012 4.10 3.65Glide 1470 0.28 0.16 -0.020 0.003 0.003 3.68 4.99Run 1435 0.61 0.35 0.044 0.005 0.016 3.92 4.94Pool 1420 1.04 0.60 -0.031 -0.004 0.000 3.66 5.99−2024−5.0 −2.5 0.0 2.5 5.0 7.5PC1 (45.11 % explained var.)PC2 (19.3 % explained var.)MorphologyGlidePoolRifflePlane−bedRiffleCRunFigure 3.8: Biplot of each observation along PC1 and PC2. The groupings from the k-meansclustering analysis are colour-coded and their centroid outlined.45RiffleC RunRiffle Plane − bedGlide Pool−5.0 −2.5 0.0 2.5 5.0 −5.0 −2.5 0.0 2.5 5.00.02.55.00.02.55.00.02.55.0PC1 (45.11 % explained var.)PC2 (19.3 % explained var.)Figure 3.9: Isolated biplot of each observation along PC1 and PC2 for the six clusters. Thegroupings from the k-means clustering analysis are colour-coded and their centroid outlined.3.3.2 Clustering accuracy assessmentTo assess the accuracy of the clustering algorithm, 100 locations were randomly selectedalong the thalweg and visually assigned to either glide, pool, run, riffle, riffleC , or plane-bedmorphologies according to Figure 3.1. These values were then compared to the morphologiespredicted by the PCA. As seen in Table 3.2, there was 85% agreement between the PCAclassified morphologies and visually classified morphologies. The riffle morphologies had thelowest correct classification rate at 71.8% and 77.0% for the riffle and riffleC categories,respectively. Glide and plane-bed morphologies had the highest classification rate at 97.4%and 100%, respectively.46Table 3.2: Accuracy assessment of morphology classificationMorphology % CorrectGlide 97.4Pool 80.0Run 84.6Riffle 71.9RiffleC 77.7Plane-bed 100 .0All 85.0The trends of the clustering analysis generally match what is observed in the field. Theincrease in the plane-bed morphology observed around 3000 m in Figure 3.10 correspondsto what is observed at Carnation Creek (Figure 3.11). When approaching the canyon thereis a sudden and marked change in morphology, with the slope becoming steeper (Figure3.12) and bed material much coarser, which the plane-bed cluster appears to capture. Thischannel type is characterized by shallower water depths, very steep reach scale slopes (Sr =0.042 m) and nearly double the average grain size (D50 = 8.21 cm) found in lower reachesof the channel.RiffleC RunRiffle Plane − bedGlide Pool0 1000 2000 3000 0 1000 2000 3000020040060080002004006008000200400600800Distance (m)Cumulative sum of channel typeFigure 3.10: Cumulative sum of channel morphologies along the stream’s longitudinal profile.47Figure 3.11: Distribution of channel morphologies at Carnation Creek. At ∼3000 m up-stream, there is a marked change in channel morphologies from the typical riffle-pool mor-phologies to much steeper and shallower channel morphologies.01020300 1000 2000 3000Distance along thalweg (m)Elevation (m)Figure 3.12: Longitudinal profile along the thalweg of Carnation Creek.In a heterogeneous area further downstream in the channel, the classification matches thetypical progression expected of channel morphologies (Figure 3.13) in a pool-riffle transitionarea. The exit of the pool is classified as a glide, with negative channel slopes. As the slopegets a little steeper we see shallow riffle type morphologies that meld into a deeper run atthe entry of the pool (Figure 3.13).48Figure 3.13: The transition of channel morphology from riffles to pools in a heterogeneoussection of channel.49The channel morphology classification is also sensitive to changes in bed slope and grainsize (Figure 3.14). This is evidenced in study section seven, where there is an initial coarsen-ing in bed material and steepening in bed slope where a tributary enters the channel. Ratherthan receiving a riffle classification (typical Sr = 0.012 m/m and D50 = 3.65 cm), this uniquesection of the channel received the riffleC morphology classification which is characterizedby a slightly steeper slope and larger grain sizes (typical Sr = 0.024 m/m and D50 = 6.74cm). Figure 3.14 demonstrates how the PCA-clustering technique can identify local areaswith unique morphological characteristics that may be missed in a traditional view of thesystem.50Figure 3.14: The influence of grain size and slope is indicated by the classification of riffleCmorphology in study section seven, where a tributary enters the channel.513.3.3 Scale effectsBecause fish require an array of physical habitats to complete their life histories, it wouldlikely be useful from a management perspective to have a metric that characterizes patternsin channel diversity across the river basin. Such a metric was provided by calculating theShannon’s diversity index of channel morphologies at one-metre intervals in a 45-metre win-dow. As demonstrated in Figure 3.15, diversity fluctuates greatly at the 45-m scale, however,there appears to be a trend towards decreased channel diversity as the canyon is approached.0.00.51.01.50 1000 2000 3000Distance (m)Channel DiversityFigure 3.15: Shannon’s diversity across the channel based on a 45-metre window. Gaps inthe data correspond with locations where there were no survey observations in a window.Finally, we were interested in assessing how much of the channel it was necessary to surveyto ensure that its heterogeneity was captured through analyzing the change in standarddeviation of channel diversity for different window sizes. This assessment was limited to thelower three kilometers of the channel, prior to the onset of the plane-bed morphologies, whichare not used by anadromous salmon. The results of calculating the standard deviation ofthe diversity metric for channel types from windows ranging from 15-600 m in length can beobserved in Figure 3.16. After a window size of around 200 m in length, there are marginaldecreases in the standard deviation of channel diversity with increasing window size. Thissuggests that at around this distance, it is likely that the diversity of channel types will becaptured.520.10.20.30 200 400 600Window size (m)Standard DeviationFigure 3.16: Standard deviation of channel diversity.3.4 DiscussionThis study shows how RPA imagery can be used to objectively and rapidly characterize pat-terns in channel morphology at the basin-scale of a small, forested mountain channel. Usingremotely sensed data and a PCA-clustering analysis, three kilometers of channel were char-acterized along the longitudinal profile at 1 m intervals. This is a resolution and extent tra-ditionally difficult to capture in small, forested channels, yet is of critical importance for thelife cycles of some aquatic species (Fausch et al., 2002). In contrast to prior PCA-clusteringanalysis (Lindenschmidt and Long , 2013; Hugue et al., 2016), this research demonstrateshow easily extractable RPA-centric variables can be used without running computationally-expensive flow models to classify the channel. Following a k-means clustering algorithm,six unique channel morphologies were identified, with an overall correct classification rate of85%.A strength of the PCA is the ability to investigate patterns in the data that influenced theobserved classification. As illustrated in Figures 3.6 and 3.7, the general relations betweenthe variables comprising the three components used in the analysis match our geomorphicunderstanding of stream systems. The first component is well represented by Sws, S45 andD50, the second by Rh, and the third by Sl. Therefore, the first component is character-ized by variables known to share clear trends along the longitudinal profile (Buffington andWoodsmith, 2003), whereas the second and third describe more local conditions, relating tothe local channel geometry and slope respectively. As shown in Figure 3.7, Sws, S45 andD50 are of relatively high quality and plot in a slightly opposite direction to Rh, indicatingthat these two sets of variables are inversely related. Indeed, one would expect pools, areasthat tend to have smaller water surface and reach slopes to have a greater hydraulic radius.As the D50 plots near Sws, S45, it is well correlated with reach scale patterns in slope, in-53dicating that steeper portions of the channel are likely linked with coarser grain sizes, ashas been noted in the literature (Buffington and Woodsmith, 2003). Local channel slope wasthe lowest quality variable, likely due to the array of channel slopes associated with similarchannel geometries in transitional areas. It was most well represented on PC3, which hadan eigenvalue of less than one, meaning that it explains less variation than a single variableof the original untransformed dataset (Jolliffe, 2002). There could therefore be an argumentmade for excluding PC3 from the analysis. However, it is the component which had thestrongest and highest quality representation of local channel slope (Figure 3.6c), which islikely an important discriminator of meso-scale morphologies such as glides, which tend tohave local negative slopes. Therefore, the third component was included in the analysis, tobe conservative in ensuring local channel slope was well represented.Examination of the classification from the PCA-clustering analysis revealed that therewas good agreement between the characteristics of the morphologies with those found inpreviously published studies. As shown in Table 3.3, the mean values of the variables foreach assigned morphology are close to reference values found for the slope, depth and grainsize characteristics of similar channels. The accuracy of the PCA-clustering technique wasalso tested against a visual classification of random locations along the channel. The greatestdisagreement in the channel morphologies came from the riffle and riffleC morphologies atclassification success rates of 71.9%, and 77.7% respectively. Determining the morphologyof transition zones, such as between runs and riffles appeared to be a source of confusionbetween the classification schemes. In contrast, plane-bed and glide morphologies were eas-ily distinguishable through their unique slopes and grain sizes (Table 3.1), thus leading totheir correct classifications. As the visual classification is subjective, it is not surprisingthe two classification schemes do not have 100% agreement. Ultimately, the PCA-clusteringtechnique presented provides an objective alternative with a statistical basis.Table 3.3: Comparison of average values for variables of each morphology to those foundin previously published studies. XChurch refers to values referenced from Church (1992),XHogan to values referenced from Anonymous (1996), and XBuff. to values referenced fromBuffington and Woodsmith (2003), and XHelm to values reported in this research.SChurch SHogan SBuff. SHelm D/dChurch D/dHogan D/dHelmMorphology (m/m) (m/m) (m/m) (m/m) (m) (m) (m)Riffle 0.02 0.005-0.015 0.001-0.02 0.012 <1.0 0.1-0.3 0.328RiffleC - 0.015-0.03 - 0.024 - 0.3-0.6 0.411Plane-bed 0.02-0.04 0.03-0.05 0.01-0.04 0.042 ∼1 0.6-1.0 0.419Glide - - - 0.003 - - 0.134Run - - - 0.016 - - 0.06454A useful contribution of this research for watershed management was formed by themetrics that were developed to characterize the channel’s heterogeneity. The application ofShannon’s diversity index, a frequently used metric in ecology for investigating the structureof ecosystems, was a new application in the assessment of diversity of channel morphologyof a small, forested stream. As demonstrated in Figure 3.15, there was a decrease in di-versity with distance upstream, which matches observations in the field. Certainly as thecanyon is approached at around 3000 m, there is a decrease in diversity as the channel tran-sitions to plane-bed morphologies. However, overall channel diversity fluctuates greatly atthe 45-metre reach scale, further elucidating the importance of considering the RiverscapeApproach for studying these systems that may not vary consistently and predictably. Inreality though, many long-term research experiments still rely on using discrete reaches tounderstand watershed processes. Therefore, to provide a useful tool for managers, an esti-mate was made on the required scale to study Carnation Creek to ensure its heterogeneitywas characterized. This was predicted to be approximately 200 m, and intriguingly is closeto the sediment storage wavelength identified by (Reid et al., 2019) for the system. Thisnumber can provide insights for setting up studies like the one observed in Carnation Creek.Perhaps the greatest strength of the PCA-clustering technique is its implicit suitability tothe study area of interest. Each stream is unique, and therefore it is possible that imposingunit boundaries discovered in one stream will not yield realistic groupings in a new system.Rather, the PCA-clustering technique finds natural groupings in the dataset that may bebetter suited to the area of interest. The RPA-survey as a means of data collection alsoprovides an improvement over traditional methods. Extraction of channel widths typically isa task that would require at least two people, and care must be taken in the field to properlycollect the data at systematic intervals and transcribe the measurements (Bisson et al.,2007). In contrast, extracting data from the RPA-derived rasters leaves a raw dataset thatmeasurements can always be checked against, as well as the ability to adjust window sizes forcollecting the data as the needs of a study change. However, there were characteristics of theanalysis that limited its broad-scale applicability. First, it relied on using wetted variables,in contrast to features like the bankfull width or depth to characterize the geometry of thechannel. This was done because some areas with low-lying overhanging vegetation made itimpossible to accurately determine the extent of the bankfull width. When considering theneeds of salmonids in the summer, the low flow conditions observed in July are what willbe of concern and what will determine the connectivity and distribution of certain channeltypes across the riverscape. However, for more generalizable results, it could be worthwhileto consider variables like the bankfull width, or the bankfull depth, which are less tied tothe wetted conditions observed at the time of the survey.55Despite the limitations of the PCA, and classification schemes in general, they remain avaluable tool in aiding the understanding of stream systems. These methods provide a toolto facilitate the rapid, objective and unsupervised classification of channel morphologies atthe basin-scale, a scale identified of great importance for the life histories of some aquaticspecies, but notoriously difficult to characterize. Future studies conducted at this scale,along with observation of aquatic species across multiple watersheds, may lead to greaterunderstanding of the influence that basin-scale patterns in aquatic habitat may have on someaquatic species, thereby helping to fully realize the utility of the Riverscape Approach.3.5 ConclusionThe methods illustrated in this chapter provide a rapid and objective technique for char-acterizing patterns in aquatic habitat at the basin-scale of a small, forested channel. Thisis an environment of historic importance for aquatic species like coho salmon (Tschaplinskiand Pike, 2017), yet has not benefited from the large scale and high resolution analysis ofaquatic habitat facilitated through remote sensing that larger streams have experienced. Useof RPA-derived rasters of bed morphology, bathymetry, and grain size in combination with aPCA-clustering analysis of the channel morphology at 1-metre intervals provided characteri-zation of this channel at an extent that would be difficult to attain using traditional methods.This allowed for the analysis of the channel’s local diversity at the basin-scale, and inves-tigation of the necessary scale required to capture its heterogeneity. For Carnation Creek,this was observed to be approximately 200 m. The implications of this research are twofold.First, prior to setting up monitoring experiments, care needs to be taken to ensure that thecumulative sum of the study sections at least equals this scale length. Second, investigationof the local diversity across the channel would help to ensure that study sections are welldistributed between areas of high and low diversity. Future research could involve a pairedcatchment analysis to assess how diversity indexes vary between catchments with differentcontrolling factors (e.g. land-use), to assess if diversity may be a proxy for habitat suitability.Ultimately, these results help to provide a means of bringing novel techniques used in largersystems to the challenging conditions provided by small, forested river systems. This willhopefully help to provide an opportunity for greater riverscape analysis in these systems.56Chapter 4ConclusionThe methodologies presented in this thesis advance the ability of stream scientists to objec-tively characterize small, forested mountain channels at the basin-scale using high resolutionRPA imagery. Small, forested streams have long been notoriously challenging to characterizeusing both traditional field and remote sensing techniques, by nature of their dense canopiesand challenging topography. Nonetheless, their importance cannot be understated, as seenby the strong preference of salmonids like coho salmon and cutthroat trout for these systems(Rosenfeld et al., 2000). The overall aim of this research was to provide tools to aid in theunderstanding of these systems through developing a rapid RPA-survey for extracting rele-vant channel metrics, so as to objectively characterize patterns in channel morphology at thebasin-scale. To do this, new techniques were required for the sub-canopy characterization ofthe system. Chapters 2 and 3 detail these new techniques and test their application to thelarger stream network.The research objective for Chapter 2 was to test the utility of a close range, sub-canopyRPA-survey for characterizing a small, forested mountain channel. The closed canopy andriparian vegetation of the channel precluded the ability to characterize the channel banksusing the vertical imagery that has dominated flight plans for stream surveys to date. Thischallenge was overcome through the inclusion of oblique-convergent imagery that facilitatedcharacterization of hidden portions of the channel banks. An in-depth accuracy assessmentwas conducted to investigate the difference in vertical error of checkpoints in DEMs createdthrough different flight path configurations. The assessment revealed that the inclusion ofoblique imagery provided greater coverage of the channel bed, as well as comparable verticalerrors to a flight plan with only vertical imagery. The vertical errors were within the cm scale,which is suited to the reach-scale geomorphic analysis these data would generally be used for.Techniques developed in previous studies were also successfully applied to the imagery forextracting bathymetry and grain sizes across the channel, yielding similar error estimates to57those reported in other studies. Application of this new data collection technique can providegeomorphologists with the necessary tools to characterize patterns in channel characteristicssuch as grain size, water depths and topography across the riverscape.Building upon the methodology and channel data extracted in Chapter 2, the thirdchapter provides a means of rapidly and objectively characterizing patterns in channel mor-phology. Using a PCA and variables grounded in a theoretical understanding of channelform, pool-riffle, transitional and plane-bed morphologies were classified continuously acrossthe stream network. The PCA-clustering technique is less subjective than a traditionalobservation-based characterization of channel morphology across the stream network, andavoids the potential for the human perception of the channel to influence the classification ofchannel form. The scale investigated was also greater than what would be easily attainableusing traditional methods that yield a similar dataset, such as a total station topographicsurvey. This is of importance, as the scale investigated with the RPA is a better match to thelife cycles completed by some aquatic species in the system (Fausch et al., 2002). Throughassessing the patterns in channel diversity across the riverscape, the necessary scale requiredto capture the stream’s heterogeneity was also investigated.Altogether, this dissertation demonstrates how the capabilities of RPA technology canbe used to improve the understanding of small forested mountain channels. The primarycontributions to fluvial geomorphology and ecology are:ˆ A new methodology using sub-canopy and oblique-convergent imagery to characterizethe topography, grain size and bathymetry of small, forested mountain channels.ˆ Demonstration of the applicability of this technique to surveying channels at a resolu-tion and extent relevant to the life cycles of some aquatic species.ˆ A new framework for the rapid and objective characterization of channel morphologyat the basin-scale using easily extractable channel features, and application of thesedata to investigating patterns in channel diversity across the riverscape.From a management perspective, these contributions can be applied to streams to providea holistic view of channel morphology across the riverscape that will allow for the assessmentof the links between channel morphology and aquatic species at the riverscape scale. Indeed,given the varying geology, climate and land-use conditions along streams, it is expectedthat individual streams will display unique characteristics. This leads to the timely researchquestion postulated by Lapointe (2012), “How does the unique structure of a riverscape affectsalmonid community composition, population size, and stability?”. Through incorporation ofthe techniques illustrated in this chapter, geomorphologists are now in a position to address58the physical habitat component of this question. With advances in the ability to study fishat this same scale, the ability to fully realize the utility and potential of the RiverscapeApproach is being approached.59BibliographyAbalharth, M., M. A. Hassan, B. Klinkenberg, V. Leung, and R. McCleary (2015), Us-ing LiDAR to characterize logjams in lowland rivers, Geomorphology, 246, 531–541, doi:10.1016/j.geomorph.2015.06.036.Agu¨era-Vega, F., F. Carvajal-Ramı´rez, and P. Mart´ınez-Carricondo (2016), Accuracy ofdigital surface models and orthophotos derived from unmanned aerial vehicle photogram-metry, Journal of Surveying Engineering, 143 (2), 04016,025, doi:10.1061/(asce)su.1943-5428.0000206.Anonymous (1996), Channel assessment procedure: Field guidebook, 95 pp., Ministry ofForest and British Columbia Environment, Victoria, British Columbia.Beechie, T. J., D. A. Sear, J. D. Olden, G. R. Pess, J. M. Buffington, H. Moir, P. Roni,and M. M. Pollock (2010), Process-based Principles for Restoring River Ecosystems, Bio-Science, 60 (3), 209–222, doi:10.1525/bio.2010.60.3.7.Bird, S., D. Hogan, and J. Schwab (2010), Photogrammetric monitoring of small streamsunder a riparian forest canopy, Earth Surface Processes and Landforms, 35 (8), 952–970,doi:10.1002/esp.2001.Bisson, P. A., K. Sullivan, and J. L. Nielsen (1988), Channel hydraulics, habitat use,and body form of juvenile coho salmon, steelhead, and cutthroat trout in streams,Transactions of the American Fisheries Society, 117 (3), 262–273, doi:10.1577/1548-8659(1988)117<0262:CHHUAB>2.3.CO;2.Bisson, P. A., J. M. Buffington, and D. R. Montgomery (2007), Valley segments, streamreaches, and channel units, Methods in Stream Ecology, pp. 23–49, doi:10.1016/B978-012332908-0.50004-8.Bisson, P. A., J. B. Dunham, and G. H. Reeves (2009), Freshwater ecosystems and resilienceof Pacific Salmon: Habitat management based on natural variability, Ecology and Society,14 (1), doi:10.5751/ES-02784-140145.60Borchers, H. W. (2018), pracma: Practical Numerical Math Functions, r package version2.2.2.Brardinoni, F., and M. A. Hassan (2007), Glacially induced organization of channel-reachmorphology in mountain streams, Journal of Geophysical Research: Earth Surface, 112 (3),1–18, doi:10.1029/2006JF000741.Buffington, J., and R. Woodsmith (2003), Fluvial processes in Puget Sound rivers and thePacific Northwest, Restoration of Puget Sound Rivers, (January 2014), 46 – 78.Carbonneau, P. E., S. N. Lane, and N. E. Bergeron (2004), Catchment-scale mapping ofsurface grain size in gravel bed rivers using airborne digital imagery, Water ResourcesResearch, 40 (7), 1–11, doi:10.1029/2003WR002759.Carrillo, G. (2014), vec2dtransf: 2D cartesian coordinate transformation, r package version1.1.Chartrand, S. M., A. M. Jellinek, M. A. Hassan, and C. Ferrer-Boix (2018), Morphodynamicsof a width-variable gravel bed stream: New insights on pool-riffle formation from physicalexperiments, Journal of Geophysical Research: Earth Surface, 123 (11), 2735–2766, doi:10.1029/2017JF004533.Church, M. (1992), Channel morphology and typology, in The Rivers Handbook, edited byP. Carlow and G. Petts, pp. 126–143, Blackwell, Oxford, UK.Church, M. (2006), Bed material transport and the morphology of alluvial riverchannels, Annual Review of Earth and Planetary Sciences, 34 (1), 325–354, doi:10.1146/annurev.earth.33.092203.122721.Church M. (2002), Geomorphic thresholds in riverine landscapes, Freshwater Biology, 47,541–557.Dietrich, J. T. (2014), Application of structure-from-motion photogrammetry to fluvial geo-morphology, PhD Thesis, (December), 109.Dietrich, J. T. (2017), Bathymetric structure-from-motion: extracting shallow streambathymetry from multi-view stereo photogrammetry, Earth Surface Processes and Land-forms, 42 (2), 355–364, doi:10.1002/esp.4060.Fausch, K. D., C. E. Torgersen, C. V. Baxter, and H. W. Li (2002), Landscapes to riverscapes: Bridging the gap between research and conservation of stream fishes, Bioscience, 52 (6),483–498, doi:10.1641/0006-3568(2002)052[0483:LTRBTG]2.0.CO;2.61Flener, C., M. Vaaja, A. Jaakkola, A. Krooks, H. Kaartinen, A. Kukko, E. Kasvi, H. Hyyppa¨,J. Hyyppa¨, and P. Alho (2013), Seamless mapping of river channels at high resolu-tion using mobile liDAR and UAV-photography, Remote Sensing, 5 (12), 6382–6407, doi:10.3390/rs5126382.Flitcroft, R. L., K. M. Burnett, G. H. Reeves, and L. M. Ganio (2012), Do network rela-tionships matter? Comparing network and instream habitat variables to explain densitiesof juvenile coho salmon (Oncorhynchus kisutch) in mid-coastal Oregon, USA, AquaticConservation: Marine and Freshwater Ecosystems, 22 (3), 288–302, doi:10.1002/aqc.2228.Flynt, A., and N. Dean (2016), A survey of popular R Packages for cluster analysis , Journalof Educational and Behavioral Statistics, 41 (2), 205–225, doi:10.3102/1076998616631743.Frissell, C. A., W. J. Liss, C. E. Warren, and M. D. Hurley (1986), A hierarchical frameworkfor stream habitat classification: Viewing streams in a watershed context, EnvironmentalManagement, 10 (2), 199–214, doi:10.1007/BF01867358.Garcia, X. F., I. Schnauder, and M. T. Pusch (2012), Complex hydromorphology of meanderscan support benthic invertebrate diversity in rivers, Hydrobiologia, 685 (1), 49–68, doi:10.1007/s10750-011-0905-z.Harris, C., M. Thoms, and M. Scown (2009), The ecohydrology of stream networks, Inter-national Association of Hydrological Sciences, 328, 127–136.Hartigan, J. A., and M. A. Wong (1979), Algorithm AS 136: A k-means clustering algorithm,Applied Statistics, 28 (1), 100, doi:10.2307/2346830.Hartman, G. F., B. C. Andersen, and J. C. Scrivener (1982), Seaward movement of cohosalmon ( Oncorhynchus kisutch ) fry in Carnation Creek, an unstable coastal stream inBritish Colombia, Canadian Journal of Fisheries and Aquatic Sciences, 39 (4), 588–597,doi:10.1139/f82-083.Harwin, S., A. Lucieer, and J. Osborn (2015), The impact of the calibration method onthe accuracy of point clouds derived using unmanned aerial vehicle multi-view stereopsis,Remote Sensing, 7 (9), 11,933–11,953, doi:10.3390/rs70911933.Hawkins, C. P., J. L. Kershner, P. A. Bisson, M. D. Bryant, L. M. Decker, S. V. Gregory,D. A. McCullough, C. K. Overton, G. H. Reeves, R. J. Steedman, and M. K. Young(1993), A hierarchical approach to classifying stream habitat features, Fisheries, 18 (6),3–12, doi:10.1577/1548-8446(1993)018<0003:AHATCS>2.0.CO;2.62Henry, N. (2018), Geogames and geoplay: game-based approaches to the analysis of geo-information, in Geogames and geoplay: game-based approaches to the analysis of geo-information, edited by O. Ahlqvis and C. Schlieder, pp. 177–194, Springer InternationalPublishing, doi:10.1007/978-3-319-22774-09.Hugue, F., M. Lapointe, B. C. Eaton, and A. Lepoutre (2016), Satellite-based remotesensing of running water habitats at large riverscape scales : Tools to analyze habi-tat heterogeneity for river ecosystem management, Geomorphology, 253, 353–369, doi:10.1016/j.geomorph.2015.10.025.James, M. R., and S. Robson (2014), Mitigating systematic error in topographic models de-rived from UAV and ground-based image networks, Earth Surface Processes and Landforms,39 (10), 1413–1420, doi:10.1002/esp.3609.James, M. R., S. Robson, S. D’Oleire-Oltmanns, and U. Niethammer (2017), Optimising UAVtopographic surveys processed with structure-from-motion: Ground control quality, quantityand bundle adjustment, Geomorphology, 280, 51–66, doi:10.1016/j.geomorph.2016.11.021.Jolliffe, I. (2002), Principal Component Analysis, Springer Series in Statistics, Springer-Verlag,New York, doi:10.1007/b98835.Kim, M., and M. Lapointe (2011), Regional variability in Atlantic salmon (Salmo salar) river-scapes: A simple landscape ecology model explaining the large variability in size of salmonruns across Gaspe´ watersheds, Canada, Ecology of Freshwater Fish, 20 (1), 144–156, doi:10.1111/j.1600-0633.2010.00471.x.Lane, E. (1955), The importance of fluvial geomorphology in river hydraulic engineering, inProceedings, pp. 1–17, American Society of Civil Engineers.Lapointe, M. (2012), River geomorphology and salmonid habitat: Some examples illustratingtheir complex association, from redd to riverscape scales, Gravel-Bed Rivers: Processes,Tools, Environments, pp. 191–215, doi:10.1002/9781119952497.ch17.Legleiter, C. J. (2012), Remote measurement of river morphology via fusion of LiDAR to-pography and spectrally based bathymetry, Earth Surface Processes and Landforms, 37 (5),499–518, doi:10.1002/esp.2262.Leung, E. S., J. S. Rosenfeld, and J. R. Bernhardt (2009), Habitat effects on invertebrate driftin a small trout stream: Implications for prey availability to drift-feeding fish, Hydrobiologia,623 (1), 113–125, doi:10.1007/s10750-008-9652-1.63Lindenschmidt, K.-E., and J. Long (2013), A GIS approach to define the hydro-geomorphological regime for instream flow requirements using geomorphic response units(GRU), River Systems, 20 (3-4), 261–275, doi:10.1127/1868-5749/2012/0053.Liu, N., A. Harwood, J. Kells, and K.-E. Lindenschmidt (2017), A geospatial model for charac-terizing the fish resources of the Similkameen River, British Columbia, Canada BYU Schol-arsArchive Citation, Open Water Journal, 4 (1).Lloyd, M., and R. . J. . Ghelardi (1964), A Table for Calculating the ‘ Equitability ’ Componentof Species Diversity, Journal of Animal Ecology, 33 (2), 217–225.MacVicar, B., H. Piegay, A. Henderson, F. Comiti, C. Oberlin, and E. Pecorari (2009), Quan-tifying the temporal dynamics of wood in large rivers: field trials of wood surveying, dating,tracking, and monitoring techniques, Earth Surface Processes and Landforms, 34 (15), 2031–2046, doi:10.1002/esp.1888.Mayer, C., and T. P. Kersten (2018), A Comprehensive Workflow to Process UAV Images forthe Efficient Production of Accurate Geo-information A Comprehensive Workflow to Pro-cess UAV Images for the Efficient Production of Accurate Geo-information, IX ConferenciaNacional de Cartografia e Geodesia, (November), 1–8.McKean, J., D. Nagel, D. Tonina, P. Bailey, C. W. Wright, C. Bohn, and A. Nayegandhi(2009), Remote sensing of channels and riparian zones with a narrow-beam aquatic-terrestrialLIDAR, Remote Sensing, 1 (4), 1065–1096, doi:10.3390/rs1041065.Meissner, A. G. N., M. K. Carr, I. D. Phillips, and K.-e. Lindenschmidt (2016), Using a geospa-tial model to relate fluvial geomorphology to macroinvertebrate habitat in a geomorphictypologies, (March), 1–17, doi:10.3390/w8020042.Montgomery, D. R., and J. M. Buffington (1997), Channel-reach morphology in moun-tain drainage basins, Bulletin of the Geological Society of America, 109 (5), 596–611, doi:10.1130/0016-7606(1997)109<0596:CRMIMD>2.3.CO.Muste, M., D. Kim, and V. Merwade (2012), Modern digital instruments and techniques forhydrodynamic and morphologic characterization of river channels, in Gravel-Bed Rivers, pp.315–341, John Wiley & Sons, Ltd, Chichester, UK, doi:10.1002/9781119952497.ch24.Naman, S. M., J. S. Rosenfeld, L. C. Third, and J. S. Richardson (2017), Habitat-specificproduction of aquatic and terrestrial invertebrate drift in small forest streams: implicationsfor drift-feeding fish, Canadian Journal of Fisheries and Aquatic Sciences, 74 (8), 1208–1217,doi:10.1139/cjfas-2016-0406.64Naman, S. M., J. S. Rosenfeld, P. M. Kiffney, and J. S. Richardson (2018), The energeticconsequences of habitat structure for forest stream salmonids, Journal of Animal Ecology,87 (5), 1383–1394, doi:10.1111/1365-2656.12845.Nelson, M. C., M. D. Hocking, J. N. Harding, J. M. Harding, and J. D. Reynolds (2015),Quantifying the effects of stream habitat on populations of breeding Pacific salmon, CanadianJournal of Fisheries and Aquatic Sciences, 72 (10), 1469–1476, doi:10.1139/cjfas-2014-0253.Nestler, J. M., C. Baigu´n, and I. Maddock (2016), Achieving the aquatic ecosystem perspective:integrating interdisciplinary approaches to describe instream ecohydraulic processes, RiverScience: Research and Management for the 21st Century, (i), 84–102.Notebaert, B., G. Verstraeten, G. Govers, and J. Poesen (2009), Qualitative and quantita-tive applications of LiDAR imagery in fluvial geomorphology, Earth Surface Processes andLandforms, 34 (2), 217–231, doi:10.1002/esp.1705.Ortega-Terol, D., M. A. Moreno, D. Herna´ndez-Lo´pez, and P. Rodr´ıguez-Gonza´lvez (2014),Survey and classification of large woody debris (LWD) in streams using generated low-costgeomatic products, Remote Sensing, 6 (12), 11,770–11,790, doi:10.3390/rs61211770.Pfeiffer, A. M., and N. J. Finnegan (2017), Basin-scale methods for predicting salmonid spawn-ing habitat via grain size and riffle spacing, tested in a California coastal drainage, EarthSurface Processes and Landforms, 42 (6), 941–955, doi:10.1002/esp.4053.Pojar, J., K. Klinka, and D. A. Demarchi (1991), Coastal western hemlock zone, Special ReportSeries 6: Ecosystems of British Columbia, pp. 95–111.R Core Team (2018), R: A Language and Environment for Statistical Computing, R Foundationfor Statistical Computing, Vienna, Austria.Reid, D. A., M. A. Hassan, S. Bird, and D. Hogan (2019), Spatial and temporal patterns ofsediment storage over 45 years in Carnation Creek , BC , a previously glaciated mountaincatchment, doi:10.1002/esp.4595.Rosenfeld, J., M. Porter, and E. Parkinson (2000), Habitat factors affecting the abundanceand distribution of juvenile cutthroat trout (Oncorhynchus clarki) and coho salmon (On-corhynchus kisutch), Canadian Journal of Fisheries and Aquatic Sciences, 57 (February),766–774, doi:10.1139/f00-010.Rosgen, D. L. (1994), A classification of natural rivers, Catena, 22, 169–199.65Sanz-Ablanedo, E., J. H. Chandler, J. R. Rodr´ıguez-Pe´rez, and C. Ordo´n˜ez (2018), Accu-racy of Unmanned Aerial Vehicle (UAV) and SfM photogrammetry survey as a functionof the number and location of ground control points used, Remote Sensing, 10 (10), doi:10.3390/rs10101606.Schlosser, I. J. (1991), Stream fish ecology: A landscape perspective, BioScience, 41 (10), 704–712, doi:10.2307/1311765.Shannon, C. E., and W. Weaver (1964), The mathematical theory of communication, 21–33pp., Urbana: University of Illinois Press.Tamminga, A. (2016), UAV-based remote sensing of fluvial hydrogeomorphology and aquatichabitat dynamics, (September).Tamminga, A., C. Hugenholtz, B. Eaton, and M. Lapointe (2015), Hyperspatial remote sensingof channel reach morphology and hydraulic fish habitat using an unmanned aerial vehicle(UAV): A first assessment in the context of river research and management, River Researchand Applications, 31 (3), 379–391, doi:10.1002/rra.2743.Tonina, D., and J. M. Buffington (2009), Hyporheic exchange in mountain rivers I : Mechanicsand environmental effects, 3, 1063–1086.Tonkin, T. N., and N. G. Midgley (2016), Ground-control networks for image based sur-face reconstruction: An investigation of optimum survey designs using UAV derivedimagery and structure-from-motion photogrammetry, Remote Sensing, 8 (9), 16–19, doi:10.3390/rs8090786.Tschaplinski, P. J., and R. G. Pike (2017), Carnation Creek watershed experiment—long-termresponses of coho salmon populations to historic forest practices, Ecohydrology, 10 (2), doi:10.1002/eco.1812.Wackrow, R., and J. H. Chandler (2011), Minimising systematic error surfaces in digital ele-vation models using oblique convergent imagery, Photogrammetric Record, 26 (133), 16–31,doi:10.1111/j.1477-9730.2011.00623.x.Wallington, T. J., R. J. Hobbs, and S. A. Moore (2005), Implications of current ecologicalthinking for biodiversity conservation: A review of the salient issues, Ecology and Society,10 (1), doi:10.5751/ES-01256-100115.White, S. M., F. J. Rahel, F. J. R. Ahel, S. M. White, and F. J. Rahel (2008), Complementationof habitats for Bonneville cutthroat trout in watersheds influenced by beavers, livestock, and66drought, Transactions of the American Fisheries Society, 137 (3), 881–894, doi:10.1577/T06-207.1.Woodget, A. S., and R. Austrums (2017), Subaerial gravel size measurement using topographicdata derived from a UAV-SfM approach, Earth Surface Processes and Landforms, 42 (9),1434–1443, doi:10.1002/esp.4139.Woodget, A. S., R. Austrums, I. P. Maddock, and E. Habit (2017), Drones and digital pho-togrammetry: from classifications to continuums for monitoring river habitat and hydromor-phology, Wiley Interdisciplinary Reviews: Water, 4 (4), e1222, doi:10.1002/wat2.1222.Wyrick, J. R., A. E. Senter, and G. B. Pasternack (2014), Revealing the natural complexity offluvial morphology through 2D hydrodynamic delineation of river landforms, Geomorphology,210, 14–22, doi:10.1016/j.geomorph.2013.12.013.Yilmaz, C. S., V. Yilmaz, and O. Gungor (2017), Ground filtering of a uav-based point cloudwith the cloth simulation filtering algorithm.Zhang, W., J. Qi, P. Wan, H. Wang, D. Xie, X. Wang, and G. Yan (2016), An easy-to-useairborne LiDAR data filtering method based on cloth simulation, Remote Sensing, 8 (6),1–22, doi:10.3390/rs8060501.67