Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Characterizing the link between fire history, productivity, and forest structure across Canada’s northern… Bolton, Douglas Kane 2016

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

Item Metadata

Download

Media
24-ubc_2016_november_bolton_douglas.pdf [ 6.06MB ]
Metadata
JSON: 24-1.0315350.json
JSON-LD: 24-1.0315350-ld.json
RDF/XML (Pretty): 24-1.0315350-rdf.xml
RDF/JSON: 24-1.0315350-rdf.json
Turtle: 24-1.0315350-turtle.txt
N-Triples: 24-1.0315350-rdf-ntriples.txt
Original Record: 24-1.0315350-source.json
Full Text
24-1.0315350-fulltext.txt
Citation
24-1.0315350.ris

Full Text

Characterizing the link between fire history, productivity, and forest structure across Canada’s northern boreal using multi-source remote sensing by Douglas Kane Bolton B.A., Boston University, 2009 M.A., Boston University, 2011  A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF  THE REQUIREMENTS FOR THE DEGREE OF  DOCTOR OF PHILOSOPHY in The Faculty of Graduate and Postgraduate Studies (Forestry)  THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver)  September 2016 © Douglas Kane Bolton, 2016ii  Abstract Forest structure is an important indicator of ecosystem function and carbon storage in above-ground biomass, yet observations of forest structure are scarce across Canada’s unmanaged bo-real. To reduce uncertainties in global carbon budgets, an improved understanding of spatial and temporal variability in forest structure is required across unmanaged boreal forests. The objective of this dissertation is to investigate how fire history and forest productivity together shape the structure of Canada’s boreal forests, and to develop methods to assess these relationships over large forested areas.  Transects of airborne light detection and ranging (lidar) data, totaling 25,000 km in length, were collected across northern Canada in 2010, providing a unique opportunity to study spatial varia-bility in forest structure. To elucidate on the relationships between fire, productivity, and struc-ture, lidar measures of forest structure were combined with optical satellite indicators of disturb-ance history and forest productivity. Specifically, a 25-year chronosequence of forest regenera-tion following fire was developed, and the relationship between forest structure and productivity was assessed as a function of time since fire. In addition, the relationship between structure and productivity was assessed in stands with no recorded disturbances.    Satellite-derived estimates of forest productivity were an important predictor of early stand de-velopment following fire, as lidar-derived estimates of canopy cover varied strongly along re-gional gradients of productivity after 15 years following fire (r = 0.63 – 0.72, p < 0.01). Addi-tionally, pre-disturbance conditions were a strong indicator of stand development following fire, as patches classified as dense forest (> 50% canopy cover) prior to burning displayed faster iii  growth and recovery compared to patches classified as open forest (20 – 50% canopy cover). Further, this research highlights the importance of monitoring multiple aspects of forest recov-ery, as lidar-derived estimates of canopy cover and stand height showed contrasting relationships to productivity in recently burned stands (1985 – 2009) as well as in stands with no recent dis-turbance. The results of this dissertation demonstrate the value of the airborne lidar transects for describing stand-level variability in forest structure over large areas, and demonstrate the need for lidar to validate wall-to-wall indicators of disturbance, productivity, and structure.    iv  Preface My supervisory committee and I developed the objectives of this dissertation through a series of discussions. The majority of the research presented in this dissertation has been published in peer-reviewed journals, with each publication listed below.  For each manuscript, I developed the specific approaches to data processing and analysis, interpreted the results, and wrote and pre-pared the manuscripts for publication. Co-authors provided critical feedback and ideas for each manuscript and in certain cases provided important data inputs.  Chapter 3: Bolton, D.K., Coops, N.C., and Wulder, M.A., 2013. Investigating the agreement between global canopy height maps and airborne lidar derived height estimates over Canada. Canadian Journal of Remote Sensing. 39(s1), S126 – S138.  Chapter 4: Bolton, D.K., Coops, N.C., and Wulder, M.A., 2013. Measuring forest structure along productivity gradients in the Canadian boreal with small-footprint lidar. Environmental Monitoring and Assessment. 185(8), 6617 – 6634.  Chapter 5: Bolton, D.K., Coops, N.C., and Wulder, M.A., 2015. Characterizing residual structure and forest recovery following high-severity fire in the western boreal of Canada using Landsat time-series and airborne lidar data. Remote Sensing of Environment. 163, 48 – 60.  Chapter 6: Bolton, D.K., Coops, N.C.,Hermosilla, T., Wulder, M.A., and White, J.C. Accepted with minor revisions in the Journal of Biogeography. Assessing variability in post-fire structure along gradients of productivity in the Canadian boreal using multiresolution remote sensing.       v  Table of Contents  Abstract ........................................................................................................................................... ii Preface............................................................................................................................................ iv Table of Contents ............................................................................................................................ v List of Tables ................................................................................................................................. ix List of Figures ................................................................................................................................. x List of Abbreviations ................................................................................................................... xiii Acknowledgements ....................................................................................................................... xv Dedication ................................................................................................................................... xvii  Introduction .................................................................................................................................. 1 1 Background and motivation .................................................................................................. 1 1.1 Fire and forest productivity as drivers of boreal forest structure ................................... 3 1.1.1 Remote sensing to assess fire, forest productivity, and forest structure ............................... 6 1.2 Objectives and research questions....................................................................................... 10 1.3 Dissertation overview .......................................................................................................... 11 1.4 Study area and data sources ....................................................................................................... 13 2 The Canadian boreal............................................................................................................ 13 2.1 Remotely sensed data sources ............................................................................................. 15 2.2 Measuring forest structure with airborne lidar data ...................................................... 16 2.2.1 Detecting disturbances with Landsat data .................................................................... 19 2.2.2 Assessing productivity with MODIS data .................................................................... 20 2.2.3 How well do wall-to-wall canopy height maps capture spatial variability in forest structure 3when compared to airborne lidar data? ......................................................................................... 23  Introduction ......................................................................................................................... 23 3.1 Materials and methods ........................................................................................................ 25 3.2 Data sources .................................................................................................................. 25 3.2.1 Comparison of airborne lidar and GLAS-derived canopy height estimates ................. 30 3.2.2 Additional comparison against the Lefsky (2010) height product ............................... 32 3.2.3vi   Results ................................................................................................................................. 34 3.3 Discussion ........................................................................................................................... 42 3.4 Questions answered ............................................................................................................. 46 3.5 How does forest structure vary along gradients of productivity in the absence of recent 4disturbance? .................................................................................................................................. 48  Introduction ......................................................................................................................... 48 4.1 Materials and methods ........................................................................................................ 51 4.2 Data sources .................................................................................................................. 51 4.2.1 Selection of mature unmanaged forest cells ................................................................. 55 4.2.2 Investigating the relationship between lidar-derived structure and MODIS GPP ........ 56 4.2.3 Results ................................................................................................................................. 59 4.3 Canopy cover ................................................................................................................ 59 4.3.1 Stand height .................................................................................................................. 64 4.3.2 Structural complexity ................................................................................................... 65 4.3.3 Discussion ........................................................................................................................... 66 4.4 Canopy cover ................................................................................................................ 66 4.4.1 Stand height .................................................................................................................. 69 4.4.2 Structural complexity ................................................................................................... 70 4.4.3 Considerations for interpreting results ......................................................................... 71 4.4.4 Questions answered ............................................................................................................. 72 4.5 How does forest structure vary as a function of time since fire for early successional stands? 74 5 Introduction ......................................................................................................................... 74 5.1 Material and methods .......................................................................................................... 77 5.2 Study area ..................................................................................................................... 77 5.2.1 Data sources .................................................................................................................. 78 5.2.2 Fire detection ................................................................................................................ 80 5.2.3 Forest classification ...................................................................................................... 85 5.2.4 Assessing structural response to fire ............................................................................ 89 5.2.5 Results ................................................................................................................................. 92 5.3vii   Detection of fires and classification of forests ............................................................. 92 5.3.1 Assessment of structure ................................................................................................ 96 5.3.2 Discussion ......................................................................................................................... 101 5.4 Residual structures dominated canopies in the first decade following fire ................ 102 5.4.1 Pre-disturbance imagery provided expectations for post-fire stand regeneration ...... 103 5.4.2 Growing space remained in stands at the end of the chronosequence ........................ 105 5.4.3 Regenerating stands were more homogenous than unburned stands ......................... 106 5.4.4 Considerations for interpreting results ....................................................................... 109 5.4.5 Questions answered ........................................................................................................... 110 5.5 How does early stand development vary along gradients of productivity? ............................. 113 6 Introduction ....................................................................................................................... 113 6.1 Materials and methods ...................................................................................................... 115 6.2 Data sources ................................................................................................................ 115 6.2.1 Selection of lidar cells ................................................................................................ 118 6.2.2 Assessment of post-fire structure ............................................................................... 119 6.2.3 Results ............................................................................................................................... 122 6.3 Discussion ......................................................................................................................... 124 6.4 Variability in canopy cover increases as time since fire increases ............................. 125 6.4.1 Stand height and canopy cover tell alternate stories of recovery ............................... 126 6.4.2 Considerations for interpreting results ....................................................................... 128 6.4.3 Questions answered ........................................................................................................... 130 6.5 Conclusions .............................................................................................................................. 132 7 Significance of research .................................................................................................... 132 7.1 Research questions addressed ........................................................................................... 133 7.2 Lessons learned for monitoring forest structure over large areas ..................................... 139 7.3 Pre-disturbance spectral information can provide expectations for regeneration ...... 139 7.3.1 Optical measurements of recovery may not tell the whole story ............................... 139 7.3.2 Do not mistake residual structure for recovery following fire ................................... 140 7.3.3 Limitations ........................................................................................................................ 141 7.4viii   Limitations of transect data ........................................................................................ 141 7.4.1 Unmeasured factors that influence forest structure .................................................... 142 7.4.2 Lack indicators of carbon stored in soils, deadwood, litter, and wetlands ................. 143 7.4.3 Directions for future research ............................................................................................ 144 7.5 Planning future large-area lidar transect collection .................................................... 144 7.5.1 Continued need for wall-to-wall estimates of structure .............................................. 146 7.5.2 Incorporation of Landsat MSS data to extend chronosequence of structure .............. 147 7.5.3References ................................................................................................................................... 149    ix  List of Tables Table 3.1 Derivation of wall-to-wall canopy height from GLAS waveforms for the Lefsky (2010) and Simard et al. (2011) products. ................................................................................................ 28 Table 3.2 Bias and RMSE statistics for the Lefsky (2010) product when compared against the mean 95th height percentile and the 90th percentile of Lorey’s height from airborne lidar ......... 35 Table 4.1 The correlation coefficients, slopes and modified t-test results for the relationship between percent cover above 2 m (X) and MODIS GPP, MAT and TAP (Y). A distance interval of 10 km was used to calculate the effective sample size. Slopes are only displayed for the statistically significant relationships (p< 0.05*, p < 0.01**, p < 0.001***). ................................ 58 Table 4.2 The correlation coefficients, slopes and modified t-test results for the relationship between the 95th height percentile (X) and MODIS GPP (Y). A distance interval of 10 km was used to calculate the effective sample size. Slopes are only displayed for the statistically significant relationships (p< 0.05*, p < 0.01**, p < 0.001***). ................................................... 63 Table 5.1 Number of images acquired by path/row between June-Sept 1984-2013 .................... 79 Table 5.2 Frequency of each canopy cover class along the lidar transects. Training patches were randomly sampled to match the frequency of each class across the landscape. ........................... 87 Table 5.3 Landsat data inputs to the classification tree ................................................................ 88 Table 5.4 Cross-validated accuracy assessment for classifying canopy cover using a single Landsat composite centered on August 15th, 2010. Minimum observations per node were set to 100, and the maximum number of tree splits set to 15. Correctly classified pixels are underlined........................................................................................................................................................ 95 Table 5.5 The number of dense and open patches sampled with airborne lidar in each YSF group........................................................................................................................................................ 95    x  List of Figures Figure 1.1 Flow chart of core research chapters ........................................................................... 12 Figure 2.1 Boreal ecozones clipped to the Brandt (2009) definition of the Canadian boreal zone....................................................................................................................................................... 15 Figure 2.2 Boreal-focused lidar transects overlaid on the Brandt (2009) definition of the Canadian boreal zone .................................................................................................................... 17 Figure 2.3 Average annual estimates of Gross Primary Productivity (GPP) from 2001 to 2010 from MODIS at 1-km spatial resolution ....................................................................................... 22 Figure 3.1 GLAS-derived canopy height maps over the forested ecozones of Canada. (a) Lefsky (2010) and (b) Simard et al., (2011) ............................................................................................. 28 Figure 3.2 Heat maps between Lefsky (2010) and airborne lidar canopy heights at 925 m spatial resolution. The relative RMSE is displayed in parentheses. ......................................................... 34 Figure 3.3Heat maps between Simard et al. (2011) and airborne lidar canopy heights at 925 m spatial resolution. The relative RMSE is displayed in parentheses. ............................................. 36 Figure 3.4 Ecodistrict-level RMSE for (a) Lefsky (2010) and (b) Simard et al. (2011). Ecodistrict-level bias for (c) Lefsky (2010) and (d) Simard et al. (2011). ................................... 38 Figure 3.5 Relationship between ecodistrict-level RMSE and terrain roughness for (a) Lefsky (2010) and (b) Simard et al. (2011). Ecodistricts are labeled by ecozone: Taiga Plains (TP), Taiga Shield East (TSE), Taiga Shield West (TSW), Boreal Shield East (BSE), Boreal Shield West (BSW), Atlantic Maritime (AM), Boreal Plains (BP), Boreal Cordillera (BC), and Hudson Plains (HP). .............................................................................................................................................. 39 Figure 3.6 Example segments of the airborne lidar dataset in (a) the Boreal Shield East and (b) the Boreal Cordillera. Airborne lidar height estimates are displayed as a distribution of the 95th height percentiles from the 25 m plots within each 925 m pixel. A single outlier for the Lefsky (2010) product at approximately 20 km in (a) is not displayed (height estimate of 58 m). Flight paths are displayed from south to north. ....................................................................................... 41 Figure 4.1 a) Path of 34 small-footprint lidar transects flown by CFS in 2010 b) Average annual MODIS GPP from 2001-2010 c) Percent of each 1-km MODIS cell classified as forest by the EOSD d) Presence or absence of fire, roads or anthropogenic change within each 1-km MODIS cell e) Selected mature unmanaged MODIS cells shaded by lidar-derived canopy cover f) Number of MODIS cells selected for analysis within each boreal ecozone ................................. 52 xi  Figure 4.2 Relationship between percent cover above 2 m and MODIS GPP for a) Boreal Shield East (scatterplot), shaded by dominant forest type  b) Boreal Shield East (boxplot) c) Boreal Shield West d) Boreal Plains e) Boreal Cordillera f) Taiga Shield East  g) Taiga Plains  h) Hudson Plains. The number above each bin corresponds to the number of samples within the bin....................................................................................................................................................... 60 Figure 4.3 Relationship between percent cover above 2 m and Minimum annual temperature for a) Boreal Shield East (scatterplot), shaded by dominant forest type  b) Boreal Shield East (boxplot) c) Boreal Shield West d) Boreal Plains e) Boreal Cordillera  f) Taiga Shield East  g) Taiga Plains  h) Hudson Plains. The number above each bin corresponds to the number of samples within the bin .................................................................................................................. 62 Figure 4.4 Relationships between 95th height percentile and MODIS GPP for a) Boreal Shield East (scatterplot), shaded by dominant forest type b) Boreal Shield East (boxplot) c) Boreal Shield West d) Boreal Plains e) Boreal Cordillera f) Taiga Shield East g) Taiga Plains h) Hudson Plains. The number above each bin corresponds to the number of samples within the bin ......... 63 Figure 4.5 a) Relationship between the CV of return height and MODIS GPP in the Boreal Shield East, shaded by a) dominant forest type b) the 95th height percentile (coniferous dominated stands only) ................................................................................................................. 65 Figure 4.6 Schematic representations of the observed relationships in the Boreal Shield East between MODIS GPP and a) percent cover above 2 m b) 95th height percentile c) CV of return height............................................................................................................................................. 68 Figure 5.1 The airborne lidar transects and selected Landsat scenes overlaid on the Brandt (2009) definition of the boreal. The Boreal Shield West ecozone is the focus of this research chapter. . 78 Figure 5.2 Flow chart of processing steps to derive lidar metrics for open and dense forest patches that burned between 1985 and 2010. These steps are described in detail in Sections 5.2.3 – 5.2.5............................................................................................................................................ 82 Figure 5.3 Burned patches shaded by the year of detection across 13 Landsat scenes ................ 92 Figure 5.4 Cross-validated error using a target date of August 15th for classifying canopy cover as a function of a) minimum observations per node and b) maximum number of tree splits. No tree pruning was used for panel a, while minimum observations per leaf was set to 100 for panel b..................................................................................................................................................... 94 Figure 5.5 Distribution of sampled dense and open patches across latitude and longitude. When bars exceed the dashed grey lines, more than 50% of the patches within the group fall within the corresponding latitude or longitude bin. ....................................................................................... 96 Figure 5.6 Median lidar metrics for previously dense and open forest patches, with error bars displaying the interquartile range for a) Cover above 2m, b) the 75th height percentile, c) xii  Skewness of return heights above 2m, d) Kurtosis of return heights above 2m, and e) Rumple derived from the Canopy Height Model (CHM). Rumple was derived from canopy pixels (CHM height > 2m) only. Asterisks represent statistical differences between dense and open patches in each group (* p < 0.05, ** p < 0.01, *** p < 0.001) .................................................................... 98 Figure 5.7 Median percentage of Canopy Height Model (CHM) area in each height class, with error bars displaying the interquartile range, for a) previously dense and b) previously open patches. Only clumps > = 5 pixels (20m2) were included in the area calculation. ..................... 101 Figure 6.1Areas detected as burned (1985–2011) across the Canadian boreal following the Composites 2 Change (C2C) approach. Panels S1, S2, S3 are examples of the intersect between detected fires and lidar transects ................................................................................................. 117 Figure 6.2 Scatterplots between lidar metrics and 2010 gross primary productivity (GPP) estimates for patches in five years since fire (YSF) groups. For comparison, patches that were undisturbed between 1985 and 2010 are displayed in the top panel, as well as in the background of subsequent panels. Summary statistics are provided for the lidar metrics. The significance of correlation coefficients were calculated using a distance interval of 20 km in the modified t-test (p< 0.05*, p < 0.01**, p < 0.001***). ........................................................................................ 121 Figure 6.3 a) Relationship between lidar-derived estimates of canopy cover (cover above 2m) and stand height (75th percentile) for burned patches, with points shaded according to years since fire (YSF). For comparison, patches that were unburned between 1985 and 2010 are displayed in the background. b) Schematic interpretation of structural development following boreal fire, as assessed using lidar structural metrics. Dashed lines represented expected height gains after 25 YSF. ............................................................................................................................................ 124   xiii  List of Abbreviations APAR – Absorbed Photosynthetically Active Radiation AVHRR – Advanced Very High Resolution Radiometer C2C  – Composite 2 Change  C-CLEAR – Canadian Consortium for Lidar Environmental Applications Research CDED  –  Canadian Digital Elevation Dataset  CFS – Canadian Forest Service CHM  – Canopy Height Model  CNFDB – Canadian National Fire Database CV – Coefficient of Variation dNBR  – Differenced Normalized Burn Ratio EOSD – Earth Observation for Sustainable Development of Forests ETM+ – Enhanced Thematic Mapper Plus FPAR – Fraction of Absorbed Photosynthetically Active Radiation GLAS – Geoscience Laser Altimeter System GPP – Gross Primary Productivity IQR  – Interquartile Range  LEDAPS – Landsat Ecosystem Disturbance Adaptive Processing System Lidar  –  Light Detection and Ranging LUE – Light Use Efficiency MAT – Minimum Annual Temperature MODIS – Moderate Resolution Imaging Spectroradiometer MSS – Multispectral Scanner xiv  NASA – National Aeronautics and Space Administration   NBR – Normalized Burn Ratio NDVI – Normalized Difference Vegetation Index NDWI – Normalized Difference Wetness Index NPP – Net Primary Productivity NTEMS – National Terrestrial Ecosystem Monitoring System  OLI – Operational Land Imager PAR – Photosynthetically Active Radiation RMSE – Root Mean Squared Error SRTM – Shuttle Radar Topography Mission TAP – Total Annual Precipitation TCT  – Tasseled Cap Transformation  TM – Thematic Mapper USGS – United States Geological Survey VPD – Vapor Pressure Deficit YSF – Years Since Fire   xv  Acknowledgements I owe so much of my success in this degree to my supervisor, Dr. Nicholas Coops, whose devo-tion to his students is second to none. He has provided me with more resources and support than any graduate student could ask for, which has allowed me to thrive both scientifically and per-sonally during my degree. I would like to thank Dr. Michael Wulder of the Canadian Forest Ser-vice for both his professional and personal guidance on my research and career path. His com-mitment to my success has inspired me to challenge myself and continue to push the envelope. I would like to thank my committee members Dr. Sarah Gergel and Dr. Allan Carroll for encour-aging me to think critically about the direction of my research and for spending hours discussing, reading drafts, and providing critical feedback. Further, this research has benefited tremendously from discussions and collaboration with Joanne White of the Canadian Forest Service. I would also like to thank my former advisors, Dr. Mark Friedl and Dr. Curtis Woodcock of Boston Uni-versity, for inspiring my interest in remote sensing and encouraging me to pursue a PhD.  I am grateful to the many funders who have supported my work over the course of my degree. Substantial financial support was provided by the Canadian Forest Service (CFS), along with funding through the National Terrestrial Ecosystem Monitoring System (NTEMS)  program, which is jointly funded by CFS and the Canadian Space Agency (CSA) Government Related Ini-tiatives Program (GRIP). I would also like to express gratitude for the scholarships I received through the University of British Columbia, including the Theodore E. Arnold Fellowship, the David W. Strangway Fellowship, the Asa Johal Graduate Fellowship in Forestry, the Mary and David Macaree Fellowship, the Canfor Corporation Fellowship in Forest Ecosystem Manage-xvi  ment, the Elwyn Gregg Memorial Fellowship, the Faculty of Forestry Strategic Recruitment Scholarship, and the ESRI Canada GIS Scholarship.  There are many researchers who played a critical role in the collection and processing of data sources used in this dissertation. I would like to thank Dr. Chris Hopkinson for his leadership role in collecting transects of light detection and ranging (lidar) data, Christopher Bater for his analysis efforts and insights into the processing of these lidar transects, and Trevor Milne of Gaiamatics for assisting with the development of customized code for processing these data. I would like to thank Dr. Txomin Hermosilla for his work in processing Canada-wide forest dis-turbance products used in the later stages of this dissertation. Most importantly, without the shared vision of Dr. Michael Wulder and Dr. Nicholas Coops for the collection and processing of these critical data sources, none of the research in this dissertation would have been possible.  I owe so much of what I have accomplished in this degree to the friends, family, and colleagues who have surrounded me. I would like to thank the members of the Integrated Remote Sensing Studio for the lunch walks, donuts, camping trips, and for making Vancouver my home. I am in-credibly grateful to my parents for the endless opportunities and support they have given me, my brother for inspiring my interest in science, and my sister for reminding me that there is more to life than work.  Thank you to my closest friends, Paddy Kelly, Kevin Overshiner, and Brendon O’leary, for shaping the way I view the world, and my uncle, Ralph Kane, for his continued in-terest in my education, from elementary school through the end of this degree.  Finally, I would like to thank my editor-in-chief, Natalie Bolton, for always challenging me to live outside of my comfort zone. I am a better person and researcher because of it.   xvii   Dedication   To my mom, For teaching me to make life fun and for introduc-ing me to coffee.     And my dad, For following me into ice cold lakes and supporting my craziest ideas. You always have my back.     1  Chapter 1 Introduction  Introduction 1 Background and motivation 1.1Forests play a critical role in the global carbon cycle by regulating exchanges of carbon between the atmosphere and the terrestrial biosphere (Pregitzer and Euskirchen, 2004; Houghton et al., 2009). Between 1990 – 2007, forests were a net carbon sink, sequestering an estimated 1.1 ± 0.8 petagrams of carbon each year from the atmosphere (Pan et al., 2011). While carbon sequestra-tion is a vital ecosystem service provided by forests, the future of this terrestrial carbon sink re-mains uncertain in the face of a changing climate. Specifically, changes in forest productivity (Boisvenue and Running, 2006), disturbance regimes (Flannigan et al., 2005), and continued land conversion are expected to alter rates of carbon sequestration and carbon storage in the world’s forests (Le Quéré et al., 2009). In order to understand potential feedbacks between for-ests and global climate, a better understanding of current carbon storage in the world’s forests, and how carbon storage varies both spatially and temporally, is first required (Houghton et al., 2009; Goetz and Dubayah, 2011) Boreal forests, which represent 30% of the world’s forested area (Gauthier et al., 2015), account-ed for an estimated 22% of the terrestrial carbon sink between 1990 – 2007 (Pan et al., 2011). Despite playing an important role in the global carbon cycle, spatial and temporal variability in aboveground biomass remains poorly quantified in many boreal regions due to a scarcity of field measurements (Gillis et al., 2005; Kurz et al., 2013). Across the Canadian boreal, for example, 2  roughly 60% of forests are unmanaged (Venier et al., 2014), and therefore not subjected to rou-tine forest inventory (Gillis et al., 2005). Without sufficient field measurements, Canada’s un-managed boreal remains a source of uncertainty in both national (Kurz et al., 2013) and global (Pan et al., 2011) efforts to characterize forest carbon budgets. In order to reduce uncertainties around the amount of carbon stored in aboveground biomass, additional measurements of three-dimensional forest structure are needed across Canada’s unmanaged boreal, in addition to an im-proved characterization of how structure varies through time.   Across unmanaged boreal forests, temporal variability in forest structure is driven primarily by natural disturbance and recovery dynamics (Kasischke et al., 1995; Kurz et al., 2013) . Stand-replacing disturbances, principally fire, result in large fluxes of carbon from forests to the atmos-phere through the combustion of biomass and the decay of dead plant material (Kasischke et al., 1995; Amiro et al., 2001). In the years following a stand-replacing disturbance, carbon is typical-ly re-sequestered as pioneer trees establish and grow (Johnstone et al., 2004; Kurz et al., 2013). The rate at which carbon is sequestered, referred to here as forest productivity, will dictate how quickly the biomass lost to disturbance is recovered. A quantification of the relationships be-tween fire history, forest productivity, and forest structure, and methods to monitor these rela-tionships over large forested areas, can provide an improved understanding of both spatial and temporal variability in boreal forest structure. In the proceeding section, I review the importance of fire and forest productivity for shaping structure across Canada’s boreal forests. 3   Fire and forest productivity as drivers of boreal forest structure  1.1.1Across Canada, an average of 1.8 million ha burned each year between 1959 – 1997 (Stocks et al., 2002), with direct carbon emissions estimated at 27 teragrams of carbon per year (Amiro et al., 2001). These fires, which are typically large, stand-replacing disturbance events, occurred primarily in Canada’s boreal (95% of the total area burned, Stocks et al., 2002).  While fires in the southern boreal are often suppressed for public safety and resource protection, fires in the unmanaged boreal are typically left to burn naturally, resulting in greater fire size and area burned in unmanaged boreal forests (Stocks et al., 2002). Fire frequency generally increases from east to west across the Canadian boreal, as conditions in the west are drier and the probabil-ity of lightning strikes is higher, with fire frequency varying from several decades to several cen-turies (Ryan, 2002; Brassard and Chen, 2006). Varying fire return intervals across Canada influ-ence the age and structure of forest stands. Kneeshaw and Gauthier (2003) demonstrated that western boreal stands have a lower proportion of old growth forests and stands > 200 years old than eastern boreal stands, due to the shorter fire return interval in the west. Where return inter-vals are very short, the initial cohort of trees can dominate stands until the next fire, while longer return intervals allow for the breakup of initial cohort and the formation of gaps, leading to more structurally complex forest stands (Brassard and Chen, 2006).  Following fire, a four stage model of forest succession defined by Oliver and Larson (1990)  is typically used to describe forest regeneration (Chen and Popadiouk, 2002) . In the first stage, termed stand initiation, the open space created by fire is colonized by pioneer trees (Oliver and Larson, 1990). Broadleaf species, such as trembling aspen (Populus tremuloides)  or white birch (Betula papyrifera), are capable of sprouting from roots or stumps (e.g., vegetative reproduc-4  tion), while coniferous species establish primarily from seed sources (Chen and Popadiouk, 2002; Johnstone et al., 2004). Once the available growing space is re-occupied, and the canopy begins to close, trees begin to compete for resources (e.g., nutrients and sunlight) in the stem ex-clusion stage of succession. This stage is often accompanied by a decrease in stem density, as some trees outcompete others for resources (Chen and Popadiouk, 2002). Over time, canopy gaps are created in forest stands due to intermediate disturbances (e.g., insects or disease) or age-related mortality, which allows shade-tolerant trees to emerge into the overstory (Chen and Popadiouk, 2002; Brassard and Chen, 2006). This stage, called understory regeneration, often creates an un-evened age structure, and stands are typically more structurally complex than in earlier successional stages (Brassard and Chen, 2006). Finally, stands reach the gap dynamics stage when individual tree mortality and regeneration processes dominate, and most of the trees that initially colonized the site have died.   While time since fire is an important determinant of forest structure, aboveground biomass does not increase linearly through these four stages of succession (Pare and Bergeron, 1995; Harper et al., 2002; Lecomte et al., 2006b). In the immediate years following fire, forests typically remain a source of carbon as the loss of biomass through decay outweighs the biomass accumulated by newly established trees (Kasischke et al., 1995; Pregitzer and Euskirchen, 2004; Kurz et al., 2013).  Sites colonized by broadleaf species may transition to a carbon sink faster than sites col-onized by coniferous species, as broadleaf species are able to rapidly establish from roots or stumps and have faster growth rates, while coniferous species often take several years to estab-lish from seed (Johnstone et al., 2004; Mack et al., 2008; Kurz et al., 2013).  While rapid coloni-zation by broadleaf species is common in boreal mixedwood forests (Bergeron et al., 2004), 5  stands often transition to coniferous dominance once the initial cohort of broadleaf trees begin to die (Chen and Popadiouk, 2002). This transition may be accompanied by a decrease in above-ground biomass where trembling aspen is in high abundance, as trembling aspen can reach heights unmatched by other boreal forest species (Pare and Bergeron, 1995) Forest productivity, which varies at multiple scales across the boreal, influences forest regenera-tion and structure in the years between stand-replacing fires (Johnstone et al., 2004; Harper et al., 2005; Boisvenue and Running, 2006; Mack et al., 2008). Regionally, forest productivity var-ies due to climate (Churkina and Running, 1998). Temperature is the main limiting factor to productivity in the Canadian boreal, with rates of photosynthesis and decomposition decreasing from the southern to northern boreal in response to decreasing temperature and growing season length (Churkina and Running, 1998). Due to cold winters and short growing seasons, productiv-ity is low relative to temperate and tropical ecosystems (Bonan and Shugart, 1989; Boisvenue and Running, 2006). While temperature is a dominant limiting factor to growth in the boreal, wa-ter stress has also been shown to be a limiting factor to growth in certain boreal regions (Peng et al., 2011; Mansuy et al., 2012; Bond-Lamberty et al., 2014). While climate drives regional varia-tion in forest productivity, productivity also varies from stand to stand due to a multitude of fac-tors, including site conditions, fire severity, and species composition (Arseneault, 2001; Johnstone et al., 2004; Harper et al., 2005; Johnstone and Chapin, 2006; Lecomte et al., 2006a,b).  Under a changing climate, fires are expected to become more frequent and more severe across most boreal forests (Flannigan et al., 2005), resulting in increased emissions of carbon from for-6  ests to the atmosphere (Amiro et al., 2009). Conversely, forest productivity is expected to in-crease where water and nutrients are not limiting (Boisvenue and Running, 2006), allowing for-ests to recover biomass at faster rates following disturbance events. Developing methods to char-acterize the relationships between fire, forest productivity, and forest structure is therefore of particular importance for assessing potential feedbacks between boreal forests and global cli-mate.   Remote sensing to assess fire, forest productivity, and forest structure 1.2While localized studies have provided strong characterizations of the impacts of fire and forest productivity on forest structure (e.g., Boucher et al., 2006; Larson et al., 2008; Mack et al., 2008), few studies have attempted to quantify these relationships over large forested regions. Further, due to the remoteness and inaccessibility of most unmanaged boreal forests (Andrew et al., 2012), it is difficult and expensive to quantify these relationships through field campaigns alone (Gillis et al., 2005; Kurz et al., 2013). Alternatively, remote sensing technologies are capa-ble of detecting forest disturbances (e.g., Huang et al., 2010; Kennedy et al., 2010), monitoring forest productivity (e.g., Hicke et al., 2003; Running et al., 2004), and measuring forest structure (Hansen et al., 2014; Kane et al., 2014). Optical imagery from the Landsat series of satellites has been used extensively to detect and de-scribe forest disturbances at regional scales for decades (e.g., Vogelmann and Rock, 1988; Cohen et al., 2002; Schroeder et al., 2011). The change in the normalized burn ratio (NBR) be-tween Landsat images, for example, has been used to detect fires and estimate burn severity (López García and Caselles, 1991; Hall et al., 2008; Soverel et al., 2010). The opening of the 7  Landsat archive in 2008, along with advances in cloud screening (Zhu et al., 2012) and atmos-pheric correction (Masek et al., 2006), has led to a drastic increase in the volume of Landsat data used in disturbance detection studies, both spatially and temporally (Hansen and Loveland, 2012; Wulder et al., 2012a). By developing approaches to analyze dense time-series of Landsat images (Huang et al., 2010b; Kennedy et al., 2010; Zhu et al., 2012), it is now possible to reconstruct the history of forest disturbances over the Landsat data record. For example, Goodwin and Col-lett (2014) processed thousands of images across 20 Landsat scenes to map fire history in Queensland, Australia from 1986-2013.  Optical imagery, in combination with ancillary spatial layers, is also capable of providing esti-mates of forest productivity. In particular, optical imagery from the MODerate Resolution Imag-ing Spectrometer (MODIS) has been used to derive 1-km estimates of Gross Primary Productivi-ty (GPP) globally for the past 15 years (Running et al., 2004; Zhao and Running, 2010). Optical imagery is a critical data input for deriving estimates of absorbed photosynthetically active radia-tion (APAR). When estimates of APAR are combined with estimates of vegetation light use effi-ciency (LUE), estimates of GPP can be derived (Monteith, 1972; Running et al., 2004).  While optically-derived indicators of forest disturbance and productivity can provide valuable insights into the drivers of forest structure, optical imagery cannot directly inform on three-dimensional forest structure (Goetz and Dubayah, 2011). Therefore, without linking to actual measurements of forest structure, the information that these data sources provide on forest struc-ture remains unclear.  8  Alternatively, light detection and ranging (lidar), an active remote sensing technology, can measure three-dimensional forest structure over larger spatial scales and at higher sampling fre-quencies than with conventional field methods (Dubayah and Drake, 2000; Lefsky et al., 2002; Lim et al., 2003). Lidar systems measure the distance to objects by emitting pulses of near-infrared laser energy and recording the timing and intensity of pulse returns (Wehr and Lohr, 1999). When millions of lidar pulses are emitted over forest canopies (e.g., >1 pulse/m2), typical-ly from airborne platforms, discrete return lidar systems produce a cloud of points describing the structure of forest stands (Wehr and Lohr, 1999; Lim et al., 2003). Most structural information in a lidar point cloud can be summarized into three basic attributes: stand height, canopy cover, and stand structural complexity (Lefsky et al., 2005a; Kane et al., 2010b). Stand height and canopy cover, which are important indicators of aboveground biomass, can be estimated directly from a point cloud (Wulder et al., 2008a), while structural complexity can be inferred by the variation in point height (Zimble et al., 2003) or the variation in maximum height across the canopy surface (Kane et al., 2010a). In addition, a number of studies have demonstrated the value of lidar data for characterizing forest successional stage (e.g., Falkowski et al., 2009; Kane et al., 2011; van Ewijk et al., 2011). For example, Falkowski et al. (2009) distinguished six stand development stages with over 95% accuracy in northern Idaho, using a range of lidar metrics that described vegetation height and canopy cover.  Recent attempts to monitor the impacts of disturbance on structure have relied on a fusion of Landsat time-series and lidar data, which together provide the means to detect disturbances and quantify their impact on forest structure (Wulder et al., 2009; Kane et al., 2013, 2014; Pflugmacher et al., 2014). For example, Kane et al. (2014) differentiated the structure of forests 9  following varying levels of burn severity in Yosemite National Park using Landsat data to de-termine burn severity and airborne lidar to assess structural response. In addition to assessing the direct impact of disturbance on structure, the regeneration of vegetation following disturbance can be tracked through the fusion of Landsat time-series and lidar data. Lefsky et al. (2005b) used Landsat to determine stand age and airborne lidar to assess biomass accumulation and forest productivity in western Oregon. Similarly, Goetz et al. (2010) used lidar data to assess vegeta-tion regrowth following fire in Alaskan boreal forests.  While capable of providing detailed information on forest structure, lidar is not ideally suited to provide wall-to-wall information on forest structure at continental to global extents, due to the cost and time associated with data collection. Alternatively, two recent research programs de-rived wall-to-wall estimates of canopy height, a key indicator of aboveground biomass, by ex-trapolating lidar data through empirical relationships with spatially contiguous variables (Lefsky, 2010; Simard et al., 2011). Both programs relied on a global sample of lidar data from the Geo-science Laser Altimeter System (GLAS), a spaceborne lidar instrument which collected data from 2003 to 2007 (Zwally et al., 2002). As GLAS laser pulses are separated by 172 m along the flight path, and up to 14.5 km between flight paths, wall-to-wall estimates of structure cannot be derived from GLAS data alone, and therefore ancillary spatial variables are required (Zwally et al., 2002). Lefsky (2010) extrapolated GLAS-derived heights for the world’s forested areas using optical satellite imagery, while Simard et al. (2011) used climatic, topographic and other globally available ancillary variables. While the method of extrapolation is the most notable difference between the two products, differences also extend to the selection, processing, and derivation of height estimates from the lidar data, leading to large discrepancies in canopy height estimates 10  between the two products in many regions of the world (Simard et al., 2011). As GLAS-derived height products can complement field-based inventories and provide valuable information on carbon storage in forests, the accuracy of these products and the errors they will introduce to car-bon models must be further assessed and better understood.  In the summer of 2010, airborne lidar transects totaling 25,000 km in length were collected across the Canadian boreal (Wulder et al., 2012b). This dataset provides an unprecedented op-portunity to study spatial variation in forest structure across the boreal, and a chance to investi-gate the link between optically derived indicators of disturbance and productivity and lidar-derived estimates of forest structure. In addition, these transects provide a unique opportunity to investigate discrepancies in globally-derived estimates of canopy height across wide-swaths of Canada’s boreal.   Objectives and research questions  1.3The primary objective of this dissertation is to improve our understanding of the link between fire history, forest productivity, and forest structure across the Canadian boreal, and to develop methods to assess these relationships over large forested areas using remotely sensed data. As wall-to-wall canopy height maps and airborne lidar transects represent two potential sources of information for describing spatial variability in forest structure, the first objective is to better un-derstanding the strengths and weaknesses of each data source for assessing forest structure.  To meet these objectives, this dissertation addresses the following core questions: 1) How well do wall-to-wall canopy height maps capture spatial variability in forest structure when compared to airborne lidar data?  11  2) How does forest structure vary along gradients of productivity in the absence of re-cent disturbance? 3) How does forest structure vary as a function of time since fire for early successional stands? 4) How does early stand development vary along gradients of productivity? By addressing these research questions, I aim to demonstrate new techniques for monitoring for-est structure over large areas and improve our understanding of the information provided by ex-isting remote sensing indicators of disturbance, productivity, and structure.   Dissertation overview 1.4The remainder of this dissertation is split into six chapters (Figure 1.1). In Chapter 2, I provide an overview of Canadian boreal forests and a description of the primary remote sensing data sources used in this dissertation. In Chapter 3, I assess existing wall-to-wall predictions of canopy height to better understand how well variability in forest structure is currently described across the Canadian boreal. In Chapter 4, I quantify the link between lidar-derived estimates of structure and satellite-derived estimates of productivity in forest stands with no record of disturbance. In Chapter 5, I assess the impacts of fire on boreal forest structure and observe early stand suc-cession by linking Landsat-derived disturbance information with lidar-derived structural attrib-utes.  12  In Chapter 6, I bring together both disturbance and productivity as drivers of boreal structure, and investigate how forest regeneration following fire varies along regional gradients of produc-tivity.  In Chapter 7, I draw conclusions and discuss insights gained in this dissertation and directions for future research.  Figure 1.1 Flow chart of core research chapters   13  Chapter 2 Study area and data sources  Study area and data sources 2  The Canadian boreal 2.1The Canadian boreal spans over 550 million ha, and consists of a mosaic of forests, wetlands, and lakes (Brandt et al., 2013). Over 270 million ha of the Canadian boreal is treed, which repre-sents 8% of the world’s forested area, and is dominated by cold-tolerant coniferous species, such as black spruce (Picea mariana), white spruce (Picea glauca), and balsam fir (Abies balsamea). Broadleaf species, such as trembling aspen (Populus tremuloides) and white birch (Betula pa-pyrifera), are more abundant in the southern boreal (Ecological Stratification Working Group, 1995; Farrar, 1995; Brandt et al., 2013). More than 70% of the Canadian boreal is dominated by coniferous forest types, followed by mixedwood (14%), and broadleaf dominated stands (8%, NFI, 2013). Throughout this dissertation, the extent of the Canadian boreal zone is determined using the Brandt (2009) definition of the boreal in combination with boreal ecozones defined by the Eco-logical Stratification Working Group (1995, Figure 2.1).  The Ecological Stratification Working Group divided the Canadian boreal into seven ecozones, representing broad scale divisions of geomorphology, soil, vegetation and climate characteristics. These ecozones are further divided into ecoregions and ecodistricts, with ecodistricts representing the finest scale division of these characteristics. As the Boreal Shield and Taiga Shield span a wide range of climatic and ecosys-14  tem conditions from east to west, these ecozones are divided into east and west compartments in this dissertation (Stocks et al., 2002; Stinson et al., 2011)  Human population densities are low throughout most of the Canadian boreal, and the majority of the northern boreal is ‘de facto’ protected due to inaccessibility (Andrew et al., 2012), resulting in forested ecosystems that are dominated by natural disturbance and recovery processes. In ad-dition to fire, these northern forests are altered by insects, disease, and windthrow, which can lead to mortality and gap formation in late successional stands (Chen and Popadiouk, 2002; Brassard and Chen, 2006). In addition to forming canopy gaps, severe insect outbreaks can sig-nificantly alter boreal forest structure and lead to widespread mortality, such as the outbreak of eastern spruce budworm (Choristoneura fumiferana) that impacted over 50 million ha of forests in the eastern boreal in the 1970s (Bergeron et al., 1995). While the northern boreal is de facto protected, southern regions of the boreal are actively man-aged for timber (Brandt et al., 2013). Between 1990– 2008, an average of approximately 350,000 ha were harvested each year across the Canadian boreal (Stinson et al., 2011; Gauthier et al., 2014). Clear-cutting harvesting, which aims to emulate stand-replacing fire,  is the most common form of harvesting, where contiguous groups of trees are removed and carbon is transferred from aboveground biomass into the forestry sector (Kurz et al., 2009). Therefore, in addition to fire, harvesting is an important form of stand-replacing disturbance in managed boreal forests.    15    Remotely sensed data sources 2.2This dissertation utilizes three primary sources of remotely sensed data to measure vegetation structure, detect and identify forest disturbances, and measure landscape productivity. First, I ob-tain fine scale measurements of forest structure from airborne lidar data. Second, I detect and identify forest disturbances using moderate-resolution Landsat data. Third, I assess landscape productivity using coarse-resolution MODIS data. In addition to these core datasets, a number of  Figure 2.1 Boreal ecozones clipped to the Brandt (2009) definition of the Canadian boreal zone     16  ancillary data sources are used throughout the dissertation. These ancillary datasets are intro-duced in the relevant research chapters.   Measuring forest structure with airborne lidar data 2.2.1The Canadian Forest Service (CFS), in coordination with the Applied Geomatics Research Group and the Canadian Consortium for Lidar Environmental Applications Research (C-CLEAR), acquired 34 transects of small-footprint airborne lidar data across Canada’s boreal for-ests in 2010 with an Optech ALTM 3100 discrete return sensor (Wulder et al., 2012b). The 34 transects totaled nearly 25,000 km in length with a minimum swath width of 400 m, spanning from Newfoundland in the east to the Yukon in the west (Figure 2.2).  The data were collected with a fixed scan angle of 15° and a pulse repetition frequency of 70 kHz between the altitudes of 450-1900 m, resulting in a nominal pulse density of approximately 2.8 returns/m2  (Wulder et al. 2012). The average transect length, largely determined by the location of suitable airports, was approximately 700 km (Hopkinson et al., 2011). Preprocessing of the lidar dataset, including the classification of points into ground and non-ground returns, was completed using customized software tools designed to deal with large transect files (Hopkinson et al., 2011). The data were divided into 25- by 25-m plots and a suite of lidar metrics that describe the distri-bution and density of lidar returns was calculated for each plot in FUSION (available at: http://forsys.cfr.washington.edu/fusion/fusionlatest.html), a free software package developed by the US Forest Service for lidar data processing. From the over 18 billion lidar points collected during the national transects campaign, lidar metrics were generated for more than 17 million 25- by 25-m plots (Wulder et al., 2012b). 17    In this dissertation, I utilize a number of lidar metrics to assess canopy cover, stand height, and stand structural complexity.  Vegetation cover within any vertical position of a canopy can be estimated by calculating the ratio of lidar pulses intercepted by a canopy layer to the total number of returns that entered the layer with well-established accuracy (Wulder et al., 2008a). For example, Andersen (2009) used the percentage of first returns above 2 m to assess canopy closure in boreal Alaska while Solberg  Figure 2.2 Boreal-focused lidar transects overlaid on the Brandt (2009) definition of the Canadian boreal zone    18  et al. (2006) used the percentage of returns above 1 m to assess insect defoliation in Norway. Here, canopy cover was calculated as the ratio of first returns above 2 m to the total number of first returns, which conforms closely to most field definitions of canopy cover (Jennings et al., 1999; USDA Forest Service, 2003) Stand height can be assessed using height percentiles, which describe the cumulative height dis-tribution of lidar returns and correlate strongly to plot-level inventory attributes such as mean tree height, dominant tree height, and stand volume (Wulder et al., 2008a). The 95th height per-centile was used in this dissertation instead of the maximum return height or 99th height percen-tile as these latter metrics can provide unrepresentative estimates of stand height in the presence of physical (e.g., birds, power lines) or atmospheric anomalies (Magnussen and Boudewyn, 1998; Kane et al., 2010b). As the 95th height percentile will represent the top of the canopy, this metric will be strongly influenced by the presence of tall residual structures (e.g., snags or sur-viving trees) in the immediate years after fire. Therefore, in Chapters 5 and 6, where I assess ear-ly stand development following fire, I use the 75th height percentile in place of the 95th as an in-dicator of stand height, as the 75th height percentile will be less impacted by residual structures and is more likely to inform on vegetation regrowth. Both the 95th and 75th height percentiles were calculated using only first returns above 2 m. While structural complexity cannot be measured directly with lidar, and is difficult to define in the field, I assess a few potential indicators of complexity throughout this dissertation. In Chapter 4, I use the coefficient of variation (CV) of return height as an indicator of stand structural com-plexity, as variability in return height can inform on the variability of structural elements within 19  the canopy (Zimble et al., 2003). In Chapter 5, with a focus on post-fire structure, I assess two additional indicators of complexity: skewness and kurtosis. Skewness describes the symmetry of the vertical distribution of lidar returns. Following fire, a positively skewed distribution of lidar returns could signify the presence of several tall, residual trees among a stand dominated by short, regenerating vegetation. Kurtosis describes the peakedness of the distribution. If a stand consists of dense, regenerating vegetation at a uniform height, the distribution of lidar returns would display a strong peak at the height of this dense vegetation (i.e., high kurtosis). Alterna-tively, if a stand consists of a range of tree heights or the foliage is dispersed over a wide vertical range, the peak of lidar returns would be less well defined (i.e., lower kurtosis). Similar to the calculation of the 75th height percentile, only returns above 2 m were used to calculate CV, skewness, and kurtosis. Structural complexity can also be assessed by detecting individual trees within a lidar point cloud and characterizing the variability in tree height and crown size of the detected trees. However, given the relatively low point density of this lidar dataset (2.8 re-turns/m2), this dataset is not ideally suited for individual tree analyses. Therefore, plot-based as-sessments of complexity, as described above, are more suitable and reliable.    Detecting disturbances with Landsat data 2.2.2The Landsat series of satellites, a joint effort of the National Aeronautics and Space Administra-tion (NASA) and the U.S. Geological Survey (USGS), has collected multispectral imagery of the earth continuously from 1972 until the present. The Multispectral Scanner (MSS), the primary sensor aboard Landsats 1, 2, and 3 (1972-1983), collected data at 80-m spatial resolution for four spectral bands with a revisit time of 18-days. Alternatively, the Thematic Mapper (TM) aboard Landsats 4 and 5 (1982-2012) and the Enhanced Thematic Mapper Plus (ETM+) aboard Landsat 20  7 (1999-Present) collected data at 30-m spatial resolution for seven spectral bands with a revisit time of 16-days. The multispectral data collected by these Landsat sensors, in addition to the re-cently launched Operational Land Imager (OLI) aboard Landsat 8 (2013-Present), provide over four decades of continuous earth observation which can be used to monitor and detect changes to Earth’s landscapes (Loveland and Dwyer, 2012). Recent advances in both atmospheric correc-tion (Masek et al., 2006) and cloud masking (Zhu and Woodcock, 2012) have allowed for the development and distribution of analysis ready Landsat products, which can be accessed at earthexplorer.usgs.gov. All analysis ready imagery has been processed to surface reflectance through the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS, Masek et al., 2006), and is delivered with Fmask, a cloud and cloud shadow mask (Zhu and Woodcock, 2012). As analysis ready products do not currently exist for MSS data, this dissertation focuses on the TM and ETM+ Landsat era.  As the lidar data collected by the CFS in 2010 only provides a “snapshot” of forest structure, Landsat time-series data from 1984-2010 is used in this dissertation to assess disturbance history. By sampling Landsat data along the lidar transects, I will detect and identify the timing and spa-tial extent of fires to construct a 25-year disturbance history for the sampled forests.   Assessing productivity with MODIS data 2.2.3The MODIS Gross Primary Productivity (GPP) algorithm provides 8-day estimates of GPP glob-ally at 1-km spatial resolution. Derived following the principles of Monteith (1972), GPP is de-termined for each 1-km cell as a function of the absorbed photosynthetically active radiation (APAR) and the light-use efficiency (LUE) of vegetation: 21  GPP = εmax * 0.45 * SWrad * FPAR * fVPD * fTmin     Eq. 2.1 where εmax is the maximum LUE; SWrad is the incident short-wave solar radiation, multiplied by 0.45 to derive photosynthetically active radiation (PAR); FPAR is the fraction of incident PAR that is absorbed by vegetation; and fVPD and fTmin are reductions in LUE from high vapor pres-sure deficits (VPD) that lead to water stress in plants and low temperatures that limit plant func-tion (Zhao and Running, 2010). The algorithm defines εmax by vegetation type according to the MODIS Land Cover Type prod-uct (MOD12Q1, Friedl et al., 2010). Daily meteorological data is used to calculate minimum dai-ly temperature (Tmin), VPD, and SWrad (Zhao and Running, 2010). FPAR is determined using the 1-km MODIS FPAR product (MOD15A2, Myneni et al., 2011), which is computed from at-mospherically corrected MODIS surface reflectances.  The MODIS GPP algorithm has been implemented in NASA’s MOD17 product to provide 8-day and annual estimates of GPP from 2000-2015 (Running et al., 2004). Heinsch et al. (2006) showed that annual MODIS GPP (MOD17A3) had a relatively strong correlation to annual flux tower estimates of GPP across North America (r = 0.859 ± 0.173), but overestimated the tower estimates at most sites (relative error = 24%). A re-processed version of MOD17A3, which ad-dresses cloud and aerosol contamination issues (Zhao and Running, 2010), will be used in this dissertation (available at: http://www.ntsg.umt.edu/project/mod17). For Chapter 4, the annual GPP products were compiled into a ten-year average (2001-2010), serving as a long-term esti-mate of landscape productivity (Figure 2.3). To ensure that GPP estimates were representative of post-fire conditions in Chapter 6, estimates of annual GPP from 2010 were used.    22      Figure 2.3 Average annual estimates of Gross Primary Productivity (GPP) from 2001 to 2010 from MODIS at 1-km spatial resolution  23  Chapter 3 How well do wall-to-wall canopy height maps capture spatial varia-bility in forest structure when compared to airborne lidar data?   How well do wall-to-wall canopy height maps capture spatial variability in forest structure when compared to airborne lidar data? 3 Introduction 3.1Airborne lidar transects and GLAS-derived canopy height products represent two potential sources of information for describing spatial variability in forest structure across the Canadian boreal. While the airborne lidar transects collected over the Canadian boreal in 2010 offer stand-level information on forest structure, these transects only represent a sample of structure. Alter-natively, GLAS-derived height products are spatially exhaustive, but the ability of these products to describe spatial variability in forest structure remains relatively unexplored. The goal of this chapter is to determine which data source is more valuable for assessing the relationships be-tween forest structure, disturbance history, and forest productivity across the Canadian boreal. Through a comparison of height estimates derived from airborne lidar transects and GLAS-derived height products developed by Lefsky (2010) and Simard et al. (2011), I address the fol-lowing questions:  Do GLAS-derived height products correlate strongly to estimates of height from airborne lidar? As GLAS laser footprints are separated by 172 m along track and up to 14.5 km across tracks (Zwally et al., 2002), extrapolation with ancillary variables is critical for deriving wall-to-wall maps of canopy height. However, it remains unclear how well these extrapolated height estimates capture variability in structure where GLAS data are not present. If GLAS-derived 24  height products correlate strongly to airborne lidar height estimates, then these products would be a valuable source of information for describing variability in structural conditions across the boreal. Alternatively, if the correlation to airborne lidar height estimates is relatively weak, dif-ferences in forest structure due to varying disturbance histories and forest productivity may not be captured by these products, and height estimates may instead represent average landscape conditions.  Does the relationship between GLAS-derived height products and airborne lidar height esti-mates vary depending on terrain roughness? Research has demonstrated that GLAS waveforms broaden over sloped terrain, leading to biases in canopy height predictions over areas of high terrain relief (Duncanson et al., 2010a; Xing et al., 2010; Pang et al., 2011). Therefore, it is hypothesized that the differences between global canopy height products and airborne lidar height estimates will be greatest in areas with rough terrain. As the Lefsky (2010) and Simard et al. (2011) products use different approaches to ad-dress the terrain issue (details in Section 3.2.1.1), the results of this chapter will also help deter-mine which approach is more effective.  In addition to determining which data source is most suitable to address the primary objectives of this dissertation, the findings of this chapter are intended to inform interpretations of global can-opy height products and provide insights for the generation of future large area height products with lidar, models, and remotely sensed data. While Simard et al. (2011) demonstrated the large discrepancies that existed between their product and the one produced by Lefsky (2010), this 25  chapter provides the most extensive comparison and assessment to date of the ability of these global products to capture spatial variability in forest structure over boreal forests.   Materials and methods 3.2 Data sources 3.2.1 GLAS-derived canopy height products  3.2.1.1Canopy height estimates from Lefsky (2010) and Simard et al. (2011) are displayed in Figure 3.1 over the forested ecozones of Canada. Data collection periods and analysis approaches differed between the two products (Table 3.1). Lefsky (2010) used cloud-free GLAS waveforms acquired between September 2003 and November 2007 (L2A to L3I, with the exception of L2C), while Simard et al. (2011) used cloud-free GLAS waveforms acquired from May and June of 2005 (L3C). Lefsky (2010) developed least squares regression models to estimate Lorey’s height for each GLAS waveform using the waveform extent and the height of the 10th and 90th percentile of waveform energy. Models were developed independently for broadleaf and needleleaf forests, using data from 484 plots across seven study areas in the United States and Brazilian Amazon (results were averaged to develop a mixed forest model). Forest type (coniferous, deciduous, or mixed) was determined for each GLAS waveform using the MODIS land cover product (MOD12Q1). While Lefsky (2010) developed empirical models to derive estimates of canopy height, Simard et al. (2011) used the RH100 metric calculated from the GLA14 land product as a measure of canopy height for each GLAS waveform. The RH100 is a measure of the distance between the signal beginning and the ground peak for each GLAS waveform (Harding and Carabajal, 2005; Sun et al., 2008; Simard et al., 2011). To prevent biases in height estimates in 26  areas with rough terrain, Lefsky (2010) applied a terrain correction that utilized information within the GLAS waveforms to adjust the waveform extent (approach explained in detail in Lefsky et al., 2007). Alternatively, Simard et al. (2011) produced a slope map using 90-m Shut-tle Radar Topography Mission (SRTM) data between 60°S and 60°N to estimate and correct the potential bias in canopy height introduced by slope. Most notably, Simard et al. (2011) removed all waveforms from the analysis that were located in areas of high slope (> 5 degrees) or where the slope correction was >25% of the measured RH100. As SRTM data is not available above 60°N, GLAS waveforms above 60°N were not used. Additionally, Simard et al. (2011) only used GLAS waveforms that fell within forested areas according to the Globcover landcover map, pro-duced by the European Space Agency and international partners (Arino et al., 2007).      After GLAS waveforms were selected and canopy height estimates derived for each waveform, Lefsky (2010) and Simard et al. (2011) extrapolated these height estimates to produce wall-to-wall canopy height maps. Lefsky (2010) used image segmentation to derive 4.4 million forest “patches” from monthly composites of 500 m MODIS spectral data. Forest patches ranged from 1-900 pixels with an average of 100 pixels (25 km2). Each forest patch was assigned the 90th per-centile of all GLAS-derived heights that intersected the patch. As only 23.9% of forest patches contained GLAS-derived height estimates, MODIS spectral data and the MODIS land cover product were used to extrapolate GLAS-derived heights to the remaining forest patches using the Cubist statistical model. Simard et al. (2011) used the Random Forest regression tree method to extrapolate RH100 values based on seven globally available variables: mean precipitation, pre-cipitation seasonality, mean temperature, temperature seasonality, elevation, MODIS tree cover and protection status. While forest patch size varied in the Lefsky (2010) product, the Simard et 27  al. (2011) height estimates were derived at a constant resolution of 30 arc seconds (approximate-ly 1 km).  To allow for a direct comparison between map products, the products were reprojected using nearest neighbor interpolation into Lamberts Conformal Conic projection. A 925-m grid was chosen, which closely corresponds to the 30 arc-second resolution of the Simard et al. (2011) product over the study area. While the Lefsky (2010) product is delivered at a finer resolution (500 m) than the Simard et al. (2011) product (30 arc-seconds), very few single pixel forest patches exist over the study area (i.e., only 180 of 8.5 million pixels had a height estimate that differed from all eight neighboring pixels), and therefore a minimal amount of information is lost by scaling the Lefsky (2010) product to 925-m.  28    Figure 3.1 GLAS-derived canopy height maps over the forested ecozones of Canada. (a) Lefsky (2010) and (b) Simard et al., (2011)  Table 3.1 Derivation of wall-to-wall canopy height from GLAS waveforms for the Lefsky (2010) and Simard et al. (2011) products. Product GLAS data Derived height metric Terrain correction Extrapolation approach Resolution Lefsky (2010) Sept 2003 – Nov 2007 Lorey's height Used information within the GLAS waveforms to correct for terrain  (see Lefsky et al. 2007 for details) Segmented MODIS spectral data to derive forest patches. Assigned patches with the 90th percentile of Lorey's height from intersect-ing GLAS footprints. Used MODIS spectral data and land cover information to extrapolate to all patches. 500m (Patches range from 1-900 pixels) Simard et al. (2011) May 2005 - June 2005 RH100 Used SRTM data to correct RH100 val-ues. Removed foot-prints on steep terrain (> 5 degrees or where slope correction was >25% of RH100 val-ue) Extrapolated RH100 at a constant resolution using seven globally available variables 30 arc seconds (approx. 1 km)  29   Airborne lidar data 3.2.1.2Airborne lidar transects collected in 2010 across the Canadian boreal were used to compare against the global canopy height products. The 95th height percentile was chosen as a measure of canopy height to compare against the GLAS-derived height products, representing a direct measure of vertical structure. Details on this lidar dataset can be found in Section 2.2.1.   Additional datasets 3.2.1.3To focus the analysis on forested areas, land cover information was obtained from the Earth Ob-servation for Sustainable Development of Forests (EOSD; Wulder et al., 2008b) which is a 25-m spatial resolution land cover classification derived from Landsat ETM+ images (circa 2000) for the forested ecozones of Canada. The EOSD consists of 23 land cover classes, including conifer-ous, broadleaf and mixed forest classes at three densities (i.e., dense, open, sparse), for a total of 9 forest classes. These 9 forest classes were used to calculate the percentage of each 925-m pixel that was forested.  To investigate the effect of slope on GLAS-derived height predictions, the Canadian Digital Ele-vation Dataset (CDED) was obtained. The CDED is produced by Natural Resources Canada at scales of 1:50,000 and 1:250,000, derived primarily from data from the National Topographic Data Base (NTDB). The 1:50,000 product, which varies from a spatial resolution of 0.75-3 arc seconds, was obtained for this analysis (http://www.geobase.ca/geobase/en/data/cded/index.html).  The standard deviation of elevation within each ecodistrict according to the CDED was calculated as a measure of ecodistrict-level 30  terrain roughness. The SRTM data was not used for this purpose as the lidar transects extend north of 60° N, where SRTM data is unavailable.   Comparison of airborne lidar and GLAS-derived canopy height estimates 3.2.2To facilitate a direct comparison between the airborne and GLAS-derived canopy height esti-mates, the 95th height percentiles derived from airborne lidar were averaged from the 25-m plot-level to the 925-m pixel-level. Non-forested lidar plots (according to the EOSD) were removed prior to calculating the 925-m pixel averages. Plots with < 10% vegetation cover, determined by the percentage of small-footprint airborne lidar first returns above 2 m, were also removed. In addition, plots with a 95th height percentile above 50 m were removed, as these values were as-sumed to be erroneous – too tall – in the boreal study area (based upon ecological understanding and interrogation of national plot databases). Of the 9.4 million forested lidar plots, only 591 had a 95th height percentile above 50 m.  Finally, a ‘spatial uniqueness’ test was performed on the lidar plots to insure that no areas were double counted in locations where flight lines crossed when calculating the 925-m pixel averages.  To determine which Lefsky (2010) and Simard et al. (2011) 925-m pixels were used in the com-parison, several conditions needed to be met. First, the pixel needed to contain both a Lefsky (2010) and Simard et al. (2011) height estimate above 2 m. Second, the pixel needed to contain at least 137 25-m lidar plots, which equals 10% pixel coverage. Finally, the pixel needed to be at least 75% forested according to the EOSD, to ensure the analysis was restricted to forested areas. This resulted in a total of 8184 pixels suitable for comparison across the three data sources.  31  The bias and root mean square error (RMSE) between each GLAS-derived canopy height map and the mean 95th height percentile from airborne lidar were calculated within each ecozone as follows  𝐵𝑖𝑎𝑠𝑧 =  1𝑛∑ (𝐺𝐿𝐴𝑆𝑖𝑛𝑖=1 −  𝐴𝑖𝑟𝑏𝑜𝑟𝑛𝑒𝑖)        Eq. 3.1  𝑅𝑀𝑆𝐸𝑧 =  √1𝑛∑ (𝐺𝐿𝐴𝑆𝑖  −  𝐴𝑖𝑟𝑏𝑜𝑟𝑛𝑒𝑖) 2𝑛𝑖=1       Eq. 3.2 where Biasz and RMSEz are the bias and RMSE of the zth ecozone, GLASi and Airbornei are the GLAS and airborne lidar derived canopy height estimates for the ith pixel, and n is the number of pixels. Conventionally, bias and RMSE are used to compare model predicted values against measured values to assess the accuracy of a model. Here, these metrics are not used to assess model accuracy, but simply to assess the agreement between two predictions (i.e., GLAS predic-tions against airborne lidar derived predictions of canopy height). The RMSE will reveal the av-erage difference between GLAS and airborne lidar height estimates within each ecozone, while the bias will reveal if the GLAS-derived heights tend to be higher or lower than the airborne es-timates. In addition to calculating the bias and RMSE, correlation coefficients (r) were calculated within each ecozone to assess the association between GLAS and airborne lidar derived height estimates. Bias, RMSE and correlation coefficients were calculated using the R statistical pack-age (R Core Team, 2015). Bias and RMSE were also calculated for each ecodistrict and compared against terrain roughness (i.e., the standard deviation of elevation within each ecodistrict according to the CDED) to assess 32  the impact of slope on the relationship between GLAS and airborne lidar derived height esti-mates. Only forested lidar plots meeting the previous described criteria were included in the cal-culation. Only ecodistricts containing > 50 pixels were considered.   Additional comparison against the Lefsky (2010) height product 3.2.3The 95th height percentile was selected for this analysis as it provided a direct measure of vertical structure within each 25-m plot. However, direct measures of vertical structure may not compare favorably to the Lorey’s height estimates in the Lefsky (2010) product. Furthermore, for this analysis, the 95th height percentiles were averaged across each 925-m pixel to capture average stand-level conditions, while Lefsky (2010) assigned each forest patch with the 90th percentile of Lorey’s height. To assess if these differences would impact the results, Lorey’s height estimates computed by Wulder et al. (2012b) for each lidar plot were used to provide an additional com-parison against the Lefsky (2010) dataset. Estimates of Lorey’s height were derived for each li-dar plot using empirically derived relationships between lidar metrics and fields measured attrib-utes collected in Quebec, Ontario, and the Northwest Territories (see Wulder et al. 2012b for de-tails). Based on these empirical models, Wulder et al.  (2012b) estimated Lorey’s height (R2 = 0.83, RMSE = 1.34 m) for each plot using the 95th height percentile (HP95) from airborne lidar: Lorey’s = exp(0.7341  + (0.7215 *ln(HP95))) *1.0037      Eq. 3.3 The 90th percentile of Lorey’s height from the 25-m plots was assigned to each 925-m cell to best match the definition of canopy height used by Lefsky (2010). It should be noted, however, that as most forest patches in the Lefsky (2010) product are larger than a single pixel, this additional height estimate from airborne lidar does not directly match the Lefsky (2010) definition (90th 33  percentile of Lorey’s height within each forest patch). As information on the Lefsky (2010) patch boundaries was not available, the calculation of the 90th percentile of Lorey’s height within each 925-m pixel was the best alternative. Only forested lidar plots meeting the previous described criteria were included in the calculation of the 90th percentile of Lorey’s height within each 925-m pixel. 34    Results 3.3Figure 3.2 compares the Lefsky (2010) canopy height estimates against the mean 95th height per-centile from airborne lidar data within each sampled ecozone. The Taiga Plains, Boreal Cordille-ra, Boreal Shield East, and Boreal Shield West have the largest sample sizes and therefore pro- Figure 3.2 Heat maps between Lefsky (2010) and airborne lidar canopy heights at 925 m spatial resolution. The relative RMSE is displayed in parentheses. 35  vide the most insight into the agreement between the canopy height estimates. Correlation be-tween canopy height estimates was low in most ecozones (r = -0.02 − 0.43). The bias was posi-tive (i.e., Lefsky 2010 predicted taller canopies than the airborne lidar) in all ecozones except the Hudson Plains and Boreal Plains, with the largest biases occurring in the Boreal Cordillera (8.09 m) and the Taiga Shield East (6.06 m). The RMSE ranged from 4.03 m in the Hudson Plains to 10.53 m in the Boreal Cordillera and the relative RMSE (i.e., RMSE expressed as a percentage of the mean) averaged 62% across all ecozones. Lefsky (2010) predicted notably taller canopies for many pixels in the Boreal Shield East and the Boreal Cordillera than the airborne lidar and produced a number of canopy height estimates beyond the expected range in these ecozones (i.e., > 40 m). The Lefsky (2010) product compared more favorably against the 90th percentile of Lo-rey’s height from airborne lidar (Table 3.2), with lower bias and RMSE values in most ecozones. The average relative RMSE decreased to 46%, however the RMSE was above 5 m in all ecozones. Table 3.2 Bias and RMSE statistics for the Lefsky (2010) product when compared against the mean 95th height percentile and the 90th percentile of Lorey’s height from airborne lidar  Mean 95th height percentile  90th percentile of Lorey's height Ecozone Bias (m) RMSE (m)*  Bias (m) RMSE (m)* Boreal Shield E 2.50 8.28 (65)  -0.26 7.99 (51) Boreal Shield W 2.90 6.62 (61)  -0.17 5.91 (43) Boreal Plains -0.07 7.29 (47)  -2.26 7.09 (40) Taiga Shield E 6.06 8.57 (85)  3.57 6.99 (56) Taiga Shield W 3.10 6.71 (80)  0.10 6.00 (53) Taiga Plains 1.06 6.68 (50)  -2.20 6.47 (39) Boreal Cordillera 8.09 10.53 (74)  5.67 8.59 (52) Hudson Plains -0.40 4.03 (39)  -3.53 5.40 (40) Atlantic Maritime 1.60 7.53 (53)  -0.92 7.42 (44)      * Relative RMSE in parentheses                   36    Figure 3.3 compares the Simard et al. (2011) canopy height estimates against the mean 95th height percentile from airborne lidar. In all ecozones, the RMSE was lower for the Simard et al. (2011) product than the Lefsky (2010) product, with an average relative RMSE of 33%. The RMSE was lowest in the Hudson Plains (2.15 m) and highest in the Taiga Plains (5.85 m), Taiga  Figure 3.3Heat maps between Simard et al. (2011) and airborne lidar canopy heights at 925 m spatial resolution. The relative RMSE is displayed in parentheses.   37  Shield West (5.31 m) and Boreal Cordillera (4.83 m).  The bias was largest (and positive) in the Taiga Plains (3.26 m), Taiga Shield West (3.89 m), and the Boreal Cordillera (2.94 m).  The cor-relation was stronger between the Simard et al. (2011) product and airborne lidar than the Lefsky (2010) product in all but the Taiga Shield West (r = 0.18 – 0.61). Where the 95th height percen-tile was low in the Taiga Plains, Simard et al. (2011) tended to estimate significantly taller cano-pies (bias = 5.4 m when 95th height percentile <15 m).  Figure 3.4 displays the ecodistrict-level RMSE and bias between the GLAS-derived canopy height maps and the airborne lidar 95th height percentile. The RMSE is higher in most ecodis-tricts for the Lefsky (2010) product than Simard et al. (2011) product, with an RMSE > 12 m in some ecodistricts. The bias is generally positive for both GLAS products, indicating that the GLAS products predict taller canopies than are predicted with airborne lidar. The RMSE for the Lefsky (2010) product appeared to increase from the southern to the northern Boreal Shield East. The five ecodistricts with an RMSE > 6 m for the Simard et al. (2011) product occurred above 60° N. The bias was positive in these ecodistricts.     38      Figure 3.4 Ecodistrict-level RMSE for (a) Lefsky (2010) and (b) Simard et al. (2011). Ecodistrict-level bias for (c) Lefsky (2010) and (d) Simard et al. (2011).  39   Figure 3.5 displays the relationship between ecodistrict-level RMSE and terrain roughness (i.e., the standard deviation of elevation within each ecodistrict according to the CDED) for each GLAS-derived product. The majority of ecodistricts with high terrain roughness were in the Bo-real Cordillera (i.e., 7 of the 11 highest standard deviations of elevation were in the Boreal Cor-dillera). Ecodistricts with high terrain roughness also tended to have high RMSEs for the Lefsky (2010) product, with the RMSE ranging from 8.1 – 13.8 m in ecodistricts with a standard devia- Figure 3.5 Relationship between ecodistrict-level RMSE and terrain roughness for (a) Lefsky (2010) and (b) Simard et al. (2011). Ecodistricts are labeled by ecozone: Taiga Plains (TP), Taiga Shield East (TSE), Taiga Shield West (TSW), Boreal Shield East (BSE), Boreal Shield West (BSW), Atlantic Maritime (AM), Boreal Plains (BP), Boreal Cordillera (BC), and Hudson Plains (HP).  40  tion of elevation > 140 m. With the exception of the Taiga Plains ecodistricts, the RMSE tended to increase slightly for the Simard et al. (2011) product as terrain roughness increased. However, most RMSEs were lower for the Simard et al. (2011) product than the Lefsky (2010) product in ecodistricts with high terrain roughness.  In Figure 3.6, two 110 km segments of the airborne lidar dataset are displayed for the Boreal Shield East (panel a) and the Boreal Cordillera (panel b). The Lefsky (2010) and Simard et al. (2011) height estimates are shown for each 925-m pixel sampled along these segments, along with the distribution of stand-level height estimates derived from airborne lidar (the 95th height percentile) within each coarse resolution pixel. Specifically, the area between the 25th and 75th percentile and the area between the 10th and 90th percentile of the 25-m canopy height estimates falling within each 925-m pixel are displayed. Recall that a minimum of 137 25-m plots (10% pixel coverage) was required within a 925-m pixel for it to be included in the analysis. These segments illustrate the variability in stand height that is captured by airborne lidar both within and between coarse resolution pixels. The GLAS products display similar trends as the 95th height percentile along the Boreal Shield East segment (e.g., peak at 20 km and gradual increase from 30-110 km).  The patch structure of the Lefsky (2010) dataset is evident in Figure 3.6 (i.e., continuous regions of a single height estimate), with many patches containing a range of stand conditions (i.e., high variability in the 95th height percentile). Neither GLAS product captures the extreme features of the 95th height percentile in the Boreal Cordillera segment (e.g., valley at 45 km and peak at 65 km). The Simard et al. (2011) product appears to serve more as an average height estimate over this 110 km segment. Additionally for the Boreal Cordillera segment, the 41  canopy height estimates for several Lefsky (2010) forest patches are > 10 m outside of the area between the 10th and 90th percentile of height estimates from airborne lidar.      Figure 3.6 Example segments of the airborne lidar dataset in (a) the Boreal Shield East and (b) the Boreal Cor-dillera. Airborne lidar height estimates are displayed as a distribution of the 95th height percentiles from the 25 m plots within each 925 m pixel. A single outlier for the Lefsky (2010) product at approximately 20 km in (a) is not displayed (height estimate of 58 m). Flight paths are displayed from south to north.  42   Discussion 3.4GLAS provides a valuable source of information on forest structure at the global extent. The der-ivation of a wall-to-wall map of canopy height from GLAS data involves four critical steps: 1) the selection of suitable GLAS waveforms, 2) the correction of biases introduced by terrain slope, 3) the derivation of canopy height estimates and 4) the extrapolation of height estimates using spatially contiguous variables. Differences in these four steps can result in large discrepan-cies between height products, as demonstrated through the comparison of Lefsky (2010) and Simard et al. (2011) height products over the Canadian boreal. The large discrepancies between canopy height estimates over Canada make selecting the appropriate map for a specific applica-tion a non-trivial process, as these discrepancies could translate into potentially large differences in aboveground biomass estimates for carbon modeling activities. It is important to note that while comparisons between these GLAS-derived height products and height estimates from air-borne lidar data were made, the definition of canopy height and the scale of the predictions vary between products, and as a result the height estimates are not directly comparable. Therefore, the results of this analysis are not intended to suggest one product over the other, but to acknowledge that large differences exist between GLAS and airborne derived height estimates that could translate into potentially large differences in carbon storage estimates depending on the defini-tion and scale of the canopy height product that is used.    The nature of the comparison can partially explain why the airborne lidar derived heights related more closely to the Simard et al. (2011) height estimates than the Lefsky (2010) estimates. While the comparison was conducted at a spatial resolution of 925m (< 1 km2), the average forest patch 43  size reported by Lefsky (2010) was 25 km2. This resulted in multiple comparisons to each Lefsky (2010) patch, which likely amplified the differences between the products and resulted in large differences where canopy height variation within a single patch was high. In addition, this chap-ter compares direct measurements of vertical structure from airborne lidar (i.e., the 95th height percentile) against empirically derived estimates of canopy height from the Lefsky (2010) prod-uct (i.e., Lorey’s height). The method of scaling also differs, as each 925-m pixel is assigned the average 95th height percentile while Lefsky (2010) assigned each forest patch with the 90th per-centile of Lorey’s height. These differences likely contribute to the discrepancies observed be-tween the Lefsky (2010) product and airborne lidar height estimates. The airborne lidar derived heights compared more favorably to the Simard et al. (2011) product as the 95th height percentile and the RH100 metric are both direct measures of vertical structure and because the comparison was performed at the native resolution of the Simard et al. (2011) product. The RH100 metric used by Simard et al. (2011) represents the tallest portion of the canopy within a 65-m GLAS footprint, while the airborne estimate is the mean 95th height percentile from 25-m lidar plots. This difference in definition likely contributed to the slight positive bias in most ecozones for the Simard et al. (2011) product when compared to the airborne lidar metric. While the nature of the comparison can partially explain the stronger relationship between the Simard et al. (2011) and airborne lidar height estimates, the reasons likely extend beyond this. The Lefsky (2010) height estimates appeared to be more affected by slope, as the RMSEs were larger for ecodistricts with high terrain roughness compared to ecodistricts with less rough ter-rain. It is expected that sloped terrain was responsible for most of the large outliers (i.e., > 40 m forest patches) observed in the Lefsky (2010) product. The removal of GLAS waveforms col-44  lected on > 5 degree slopes is likely why similar outliers were less obvious in the Simard et al. (2011) product and why the RMSE was lower in ecodistricts with high terrain roughness. It should be noted that the Boreal Cordillera and Taiga Plains ecodistricts with high terrain rough-ness were at or north of 60° latitude, meaning that many or all of the GLAS waveforms in these ecodistricts were not used to derived the Simard et al. (2011) height estimates regardless of slope. The lack of GLAS data north of 60° latitude could explain why the RMSE was highest for the Simard et al. (2011) in the northern Taiga Plains and Boreal Cordillera ecodistricts. Regional and national DEMs, such as the CDED for Canada or the ASTER GDEM (available at: https://asterweb.jpl.nasa.gov/gdem.asp), should be incorporated in future development of GLAS height products to improve the estimation of canopy height north of 60° latitude.   In this research, the aim was not to assess the accuracy of these canopy height products. Instead, the aim was to understand how they compared to average estimates of stand height derived from airborne lidar. Despite this, an opportunity existed to produce an additional metric that corre-sponded closely to the definition of canopy height used by Lefsky (2010). The results suggest that the height metric and method of scaling can influence the correspondence between airborne and GLAS-derived height products. However, the RMSE across all transects remained higher for the Lefsky (2010) product than the Simard et al. (2011) product when the mean 95th height per-centile was replaced by the 90th percentile of Lorey’s height within each pixel. While the metric more accurately matches the definition of canopy height used by Lefsky (2010), these results should still be interpreted with caution as the 90th percentile was calculated within 925-m pixels while Lefsky (2010) took the 90th percentile within much larger forest patches. Additionally, er-45  ror in the calculation of Lorey’s height for each 25-m lidar plot must also be considered (see Wulder et al. 2012b). The segments of the airborne lidar dataset displayed in Figure 3.6 emphasize the contrasting na-ture of airborne lidar and GLAS-derived height estimates. The airborne lidar dataset is rich in stand-level information, yet lacks the spatial coverage provided by the wall-to-wall GLAS-derived products. In general, a tendency towards the mean is observed for the larger pixels of the global GLAS products. For example, large contiguous areas of small trees could be common within single 25-m pixels, but would be unusual to occur in areas as large as the GLAS-derived pixels. A single coarse resolution pixel may contain a range of forest conditions, yet global prod-ucts describe these varying conditions with a single value. Depending on the application, a single value may or may not be appropriate for representing the pixel. For example, an understanding of the average forest condition over a coarse resolution pixel may lead to an acceptable above-ground biomass estimate. However, if drivers of change in carbon stocks occur at the stand-level, characterizing changes in carbon stocks may be inappropriate with a product that averages condi-tions over a large area (Houghton et al., 2009).  In this analysis, GLAS-derived height products were only assessed in one biogeographical region (Canadian boreal). The comparison could lead to different results in other biomes, and therefore caution should be taken when interpreting the results of this analysis for studies outside of Cana-da. As field validation of large area products can be time consuming and costly, wide ranging transect-based airborne lidar campaigns, such as the one used in this analysis, are an important 46  source of data for validating and interpreting the results of GLAS-derived products over varying topographic and ecosystem conditions.    Questions answered 3.5This chapter proposed and addressed two research questions. The main findings for each ques-tion are briefly reviewed here and expanded upon in the dissertation conclusions (Section 7.2).  Do GLAS-derived height products correlate strongly to estimates of height from airborne lidar? In most sampled ecozones, the Simard et al. (2011) product was more strongly correlated to air-borne lidar height estimates (r = 0.18 - 0.61) than the Lefsky (2010) product (r = -0.02 – 0.43). While these GLAS-derived height products represent a critical step in the right direction for characterizing variability in forest structure globally, these products are not ideally suited for de-scribe structural differences due to varying disturbance histories or productivity within individual boreal landscapes, as these products tended to represent average landscape conditions. Therefore, the remainder of this dissertation focuses on structural metrics derived from the airborne lidar transects, as these data represent a critical source of information on stand-level variability in for-est structure over the large and remote northern boreal of Canada.  Does the relationship between GLAS-derived height products and airborne lidar height esti-mates vary depending on terrain roughness? Disagreement between airborne lidar derived heights and Lefsky (2010) predictions were gener-ally highest where terrain roughness was highest, suggesting that despite terrain correction, the Lefsky (2010) product was still adversely affected by slope. The lower RMSEs for Simard et al. 47  (2011) in ecodistricts with high terrain slope suggest that the removal of GLAS waveforms over high slopes is a superior approach.   48  Chapter 4 How does forest structure vary along gradients of productivity in the absence of recent disturbance?  How does forest structure vary along gradients of productivity in the absence of recent disturbance? 4 Introduction 4.1In the years between stand-replacing disturbances, forest productivity plays an important role in shaping the structure of forest stands. While plot-based studies have provided a strong under-standing of the relationships between forest productivity and forest structure (e.g., Boucher et al., 2006; Larson et al., 2008), the scarcity of field measurements across the unmanaged boreal have prevented these relationships from being explored over a range of boreal ecosystems (Gillis et al., 2005). Further, while optically-derived estimates of forest productivity exist across the Cana-dian boreal (Running et al., 2004), we lack a strong understanding of how these productivity es-timates are realized structurally. An improved understanding of the information that satellite-derived estimates of productivity provide for describing forest structure can aid in the characteri-zation of structural variability across the Canadian boreal.   In this chapter, estimates of productivity from MODIS are related to airborne lidar-derived indi-cators of canopy cover, stand height, and stand structural complexity within six boreal ecozones.  To reduce the impact of recent disturbance and management on the observed structure, infor-mation on land cover, fire history, anthropogenic change, and roads is used to restrict the study to mature unmanaged forest stands. Once stratified, the following research questions are ad-dressed.    49  Do satellite-derived estimates of productivity relate more closely to lidar-derived estimates of canopy cover or stand height? I expect canopy cover to be more strongly related to satellite-derived estimates of productivity for two reasons. First, canopy cover is an important indicator of absorbed photosynthetically ac-tive radiation (APAR), a critical input to the MODIS GPP algorithm (Running et al., 2004), which will likely lead to strong correlations between canopy cover and MODIS GPP. Second, while forest productivity will dictate vertical growth rates, the realized height of a forest stand will largely be determined by stand age and disturbance history, leading to weaker relationships between productivity and stand height. Available information on fire history is used to remove recently burned stands, but this information is not inclusive of all fire events and lacks infor-mation on other disturbances that influence vertical structure, such as disease or insect infesta-tions (Chen and Popadiouk, 2002; Brassard and Chen, 2006). Therefore, while I expect both canopy cover and stand height to be positively related to forest productivity across the boreal, the relationship will likely be strongest between canopy cover and productivity.   Does the relationship between canopy cover and productivity vary between ecozones? While canopy cover will relate strongly to satellite-derived estimates of productivity, I expect the slope of this relationship to vary substantially between ecozones. Canopy cover is an important indicator of APAR, but APAR is only one component that influences productivity (Running et al., 2004). Specifically, productivity can vary between two stands with similar APAR due to dif-ferences in the light use efficiency (LUE) of vegetation (Monteith, 1972). As LUE will vary spa-tially due to differences in temperature and VPD limitations, the slope of the relationship be-tween canopy cover and productivity will also likely vary between ecozones. 50  Where do the most structurally complex forest stands occur? Stand age and successional stage are critically important for predicting the structural complexity of a forest stand (Oliver and Larson, 1990; Chen and Popadiouk, 2002). Specifically, structural complexity is typically low through the stand initiation and stem exclusion stages of succession, as stands are even-aged. In the understory regeneration stage of succession, when the initial co-hort of trees begins to die, stands typically become more structurally complex as stands shift to uneven-aged, multi-strata canopies (Oliver and Larson, 1990; Chen and Popadiouk, 2002). While successional stage is an important determinant of structural complexity, plot-based studies have demonstrated that high productivity stands can reach an un-even aged structure faster than low productivity stands (Boucher et al., 2006; Larson et al., 2008).  In addition, insufficient resources at low productivity sites can restrict maximum tree dimensions, limiting tree size diversity and structural complexity in low productivity sites (Boucher et al., 2006). Therefore, while a strong relationship between structural complexity and forest productivity is not expected, I do expect that the most structurally complex stands will be found on high productivity sites.  By addressing these research questions, this chapter aims to better understand how forest struc-ture varies regionally along gradients of forest productivity across the Canadian boreal, and aims to assess how satellite-derived estimates of forest productivity are realized structurally across boreal landscapes.  51   Materials and methods 4.2 Data sources 4.2.1 Lidar data 4.2.1.1Indicators of canopy cover, stand height, and stand structural complexity were derived along 25,000 km of lidar transects collected in 2010 (Figure 4.1a). Percent cover above 2 m was used as an indicator of canopy cover, the 95th height percentile was used as an indicator of stand height, and the coefficient of variation (CV) was used an indicator of structural complexity. Li-dar metrics were derived at a spatial resolution of 25 m. See Section 2.2.1 for details on the lidar data collection, processing, and selection of lidar metrics.   MODIS GPP 4.2.1.2The MODIS GPP product provides annual estimates of GPP for the entire planet at 1-km spatial resolution from 2001 to the present (Running et al., 2004). As inter-annual variability and tem-poral trends exist within these data (Zhao and Running, 2010), GPP estimates from a single year are likely unrepresentative of long-term forest productivity. Therefore, the annual GPP products were compiled into a ten-year average (2001-2010), serving as a long-term estimate of produc-tivity to relate to the lidar-derived structural metrics (Figure 4.1b). All processing in this analysis was then performed on the 1-km MODIS sinusoidal grid. See Section 2.2.3 for details on the MODIS GPP product.  52    Figure 4.1 a) Path of 34 small-footprint lidar transects flown by CFS in 2010 b) Average annual MODIS GPP from 2001-2010 c) Percent of each 1-km MODIS cell classified as forest by the EOSD d) Presence or absence of fire, roads or anthropo-genic change within each 1-km MODIS cell e) Selected mature unmanaged MODIS cells shaded by lidar-derived canopy cover f) Number of MODIS cells selected for analysis within each boreal ecozone    53   Climate data 4.2.1.3In addition to the primary comparison of structural metrics against GPP estimates, lidar metrics were also compared against climate variables to provide an understanding of climate variability within each ecozone. Minimum annual temperature (MAT) and total annual precipitation (TAP) data for North America were obtained from the Pacific Climate Impacts Consortium (PCIC, http://pacificclimate.org). These climate datasets were derived at 32-km spatial resolution from 1979-2010 by the National Centers for Environmental Prediction (NCEP) North American Re-gional Reanalysis (NARR) project (Mesinger et al. 2006). A natural neighbor interpolation ap-proach was used to produce annual maps of MAT and TAP on the 1-km MODIS sinusoidal grid. The annual maps were averaged together to derive average MAT and TAP for 1979-2010.  Additional datasets 4.2.1.4Land cover was obtained from the Earth Observation for Sustainable Development of Forests (EOSD) program led by the CFS (Wulder et al., 2008b). The EOSD is a 25-m spatial resolution land cover classification of the forested ecozones of Canada derived from Landsat-7 Enhanced Thematic Mapper Plus (ETM+) images (circa 2000) and consists of 23 land cover classes, in-cluding 9 forest classes (coniferous, mixedwood and broadleaf / dense, open and sparse). These 9 forest classes were used to estimate the forested percentage of each 1-km MODIS cell (Figure 4.1c). In addition, the 500 m MODIS Land Cover Type product (MOD12Q1, Friedl et al., 2010) was obtained from 2001-2010 to compare against the EOSD classification.  All classes matching the EOSD definition of a forest (i.e., > 10% tree covered) according to the University of Mary-54  land classification scheme (Friedl et al., 2010) were selected and used to calculate the forested percentage of each 1-km cell in each year. Fire, road and anthropogenic disturbance layers were used to identify 1-km MODIS cells that potentially contained recent disturbances (Figure 4.1d). The 2010 Canadian National Fire Data-base (CNFDB, available at: http://cwfis.cfs.nrcan.gc.ca/ha/nfdb) is a collection of fire polygons recorded by provincial and territorial fire management agencies and Parks Canada. While fire records in the CNFDB date back to 1917 in British Columbia, the oldest recorded fire to intersect a CFS lidar transect was in 1941. The methods for recording fires have changed with time and vary by agency, ranging from sketches of fire boundaries to the interpretation of aerial photog-raphy and the classification of satellite imagery.  The 2010 Road Network File is a compilation of all Canadian roads recorded in Statistics Cana-da’s National Geographic Database (available at: https://www12.statcan.gc.ca/census-recensement/2011/geo/RNF-FRR/index-eng.cfm). In this analysis, the Road Network File acts as an indicator of forest management: if a 1-km MODIS cell contains a road, then the forests within that cell are potentially managed. Logging roads that provide access to managed forests from ex-isting roads may be absent from the Road Network File. Therefore, a 1-km cell was flagged as containing a road if one existed in a neighboring cell (3 by 3 cell window).  Lastly, Global Forest Watch Canada analyzed Landsat data (30-m spatial resolution) to produce anthropogenic change maps for areas in Nova Scotia (Cheng and Lee, 2008), Saskatchewan and Manitoba (Stanojevic et al., 2006a), Ontario (Cheng and Lee, 2008), Québec (Stanojevic et al., 2006b) and British Columbia (Lee and Gysbers, 2008). The major anthropogenic changes identi-55  fied and mapped in these studies include development, clear-cutting, road construction, agricul-tural clearing, reservoir construction and petroleum and natural gas exploration (Stanojevic et al., 2006b). The areas mapped by Global Forest Watch Canada do not cover the entire boreal, plac-ing more importance on the Road Network File to identify potentially managed and anthropo-genic disturbed forests.   Selection of mature unmanaged forest cells 4.2.2Indicators of canopy cover, canopy height and structural complexity were derived for each 1-km MODIS cell by averaging together the plot level (25- by 25-m) lidar metrics within each cell. Only lidar plots classified as forest by the EOSD and meeting the structural definition of a forest according to the 2005 Global Forest Resources Assessment (height [95th percentile] > 5 m, cano-py cover [percent cover above 2 m] > 10%) were used to calculate the 1-km cell averages (FAO, 2006). A “spatial uniqueness” test was performed on the lidar plots to insure that no area was double counted in the MODIS cell averages where flight lines crossed. Lidar plots with a 95th height percentile above 50 m were assumed to be erroneous and were therefore removed prior to the calculation of the MODIS cell averages. In total, only 591 of 9.4 million forested lidar plots had a 95th height percentile above 50 m.  MODIS cells containing less than 100 forested lidar plots were removed from the analysis, in addition to cells where less than 75% of the lidar plots were forested. MODIS cells that were less than 75% forested according to the EOSD (Figure 4.1c) were also removed, as the GPP estimate could become unrepresentative of the forested portion of the cell with the presence of additional 56  land covers. To remove the effects of disturbance and management on forest structure, cells that contained a fire, anthropogenic change, or a road were also removed (Figure 4.1d). Given that vegetation type is a critical input to the MODIS GPP algorithm, misclassifications in the MODIS Land Cover Type product could result in less reliable GPP estimates (Zhao et al., 2005). Therefore, cells that were less than 75% forested in any year (2001-2010) according to the MODIS Land Cover Type product were also removed, as discrepancies between EOSD and MODIS land cover could signify incorrect vegetation inputs to the GPP calculation. Averaging the 25- by 25-m plot metrics up to 1-km and applying this set of rules allowed for a direct comparison between lidar structural metrics and MODIS GPP for mature unmanaged stands. Figure 4.1e shows the distribution of the 5675 MODIS cells that met this set of rules (shaded by percent cover above 2 m), while Figure 4.1f  shows the number of cells that fall with-in each boreal ecozone. The Boreal Shield East is of particular interest in this study because of its large sample size (1809) and large latitudinal gradient in GPP (Figure 4.1b). The calculation of MODIS cell averages and the stratification of mature unmanaged stands was performed in R (R Core Team, 2015).  Investigating the relationship between lidar-derived structure and MODIS 4.2.3GPP The relationship between lidar-derived structural metrics and satellite-derived GPP was assessed using Pearson’s correlation coefficient and the modified t-test proposed by Clifford et al. (1989) and altered by Dutilleul (1993).  In the presence of positive spatial autocorrelation, a standard t-test is unfit for assessing the significance of a correlation coefficient as each sample does not 57  constitute a full degree of freedom (Clifford et al., 1989). The modified t-test adjusts the degrees of freedom by calculating an “effective sample size” that is inversely proportional to the extent of spatial autocorrelation in each variable (full details can be found in Dutilleul 1993). To assess the extent of spatial autocorrelation, the distances between all pairs of points are divided into k distance strata and spatial autocorrelation is assessed for both variables in each strata. The speci-fication of k impacts the calculation of the effective sample size as shorter distance intervals (i.e., larger value of k) will result in a higher calculated spatial autocorrelation (Fortin, 1999) and a lower effective sample size. When relating wildfire and forest regeneration in Canadian boreal forests, Fortin and Payette (2002) found that the effective sample size increased as k decreased (i.e., larger distance interval), but decreasing k did not affect the rejection of the null hypothesis. To assess the effect of k in this analysis, four distance intervals were tested in each ecozone: 5, 10, 20, and 40 km. The modified t-test was calculated for each ecozone using the Dutilleul (1993) modification in Pattern Analysis, Spatial Statistics and Geographic Exiegesis (PAS-SaGE), a freely available spatial analysis software package (Rosenberg and Anderson, 2011). In addition to testing the relationship between lidar-derived structure and MODIS GPP, the rela-tionship between structure and the climate variables (i.e., MAT and TAP) was also assessed. Fi-nally, linear regressions were developed in R to assess the slope of the relationships. Results for the Taiga Shield West are not communicated due to the small sample size in this ecozone (38 MODIS cells remained following stratification).  58    Table 4.1 The correlation coefficients, slopes and modified t-test results for the relationship between percent cover above 2 m (X) and MODIS GPP, MAT and TAP (Y). A distance interval of 10 km was used to calculate the effective sample size. Slopes are only displayed for the statistically significant relationships (p< 0.05*, p < 0.01**, p < 0.001***). Ecozone Variable Sample Size k Effective Sample Size r Slope*100 Boreal Shield E. GPP 1809 257 12.82 0.74** 0.71  MAT   11.77 0.68* 6.66  TAP   11.35 -0.51  Boreal Shield W. GPP 842 121 57.92 0.27* 0.18  MAT   43.01 0.22   TAP   65.16 -0.04  Boreal Plains GPP 145 112 50.62 0.44** 0.22  MAT   78.52 0   TAP   51.56 -0.11  Boreal Cordillera GPP 1488 84 54.38 0.58*** 0.36  MAT   27.55 0.58** 3.07  TAP   81.49 0.19  Taiga Shield E. GPP 465 112 15.14 0.57* 0.51  MAT   15.34 0.41   TAP   11.94 0.49  Taiga Plains GPP 701 144 17.23 0.70** 0.46  MAT   12.72 0.46   TAP   16.11 0.37  Hudson Plains GPP 136 84 9.98 0.47   MAT   13.74 0.29   TAP   22.08 -0.25     59   Results 4.3 Canopy cover 4.3.1Table 4.1 presents the correlation coefficients (r), slopes and modified t-test results for the rela-tionship between percent cover above 2 m and MODIS GPP, as well as both climate variables, using a distance interval of 10 km for the calculation of effective sample size. The number of strata (k) needed for a distance interval of 10 km varied from 84 in the Hudson Plains and Boreal Cordillera to 257 in the Boreal Shield East. The effective sample sizes were significantly smaller than the original sample sizes in all ecozones. While the Boreal Shield East has the most MODIS cells (1809), the effective sample sizes in the Boreal Shield East are among the smallest, with values between 11 and 13. The Boreal Shield West, Boreal Plains and Boreal Cordillera had the largest effective sample sizes, with each ecozone averaging > 50. The effective sample size in-creased as the distance interval increased from 5 to 40 km (results not shown), however this had no effect on the rejection of the null hypothesis (α = 0.05). The level of significance did vary (i.e., from p < 0.05 to 0.01 or p < 0.01 to 0.001) in several cases when the distance interval was changed.  60  Figure 4.2 displays the relationship between percent cover above 2 m and MODIS GPP for each sampled ecozone as a series of boxplots. To investigate the differences between forest types, Figure 4.2a displays the relationship in the Boreal Shield East as a scatterplot, with points shaded by the dominant (> 50%) forest type within the cell according to the EOSD land cover classifica-tion. A statistically significant (α = 0.05) correlation between lidar-derived canopy cover and MODIS GPP was found in all but the Hudson Plains, with the strongest link occurring in the Bo-real Shield East (r = 0.74, p < 0.01, Figure 4.2b). The correlation was weakest in the Boreal  Figure 4.2 Relationship between percent cover above 2 m and MODIS GPP for a) Boreal Shield East (scatterplot), shaded by dominant forest type  b) Boreal Shield East (boxplot) c) Boreal Shield West d) Boreal Plains e) Boreal Cordillera f) Taiga Shield East  g) Taiga Plains  h) Hudson Plains. The number above each bin corresponds to the number of samples within the bin    61  Shield West (r = 0.27, p < 0.05, Figure 4.2c) and the Boreal Plains (r = 0.44, p < 0.01, Figure 4.2d), both of which have a narrow sampled range in GPP. The link was strong between lidar-derived canopy cover and MODIS GPP in the Boreal Cordillera (r = 0.58, p < 0.001, Figure 4.2e) and the Taiga Plains (r = 0.70, p < 0.01, Figure 4.2g), but the slope was shallower than in the Boreal Shield East. The sampled range in GPP was larger in the Boreal Shield East than in other ecozones, with a mean GPP value of less than 0.6 kgC m-2yr-1 for the 20 – 30 % cover group, increasing to over 1.0 kgC m-2yr-1 for the 80 – 90 % cover group. Markedly more stands had a canopy cover > 90% in the Boreal Shield East than in other ecozones and these stands had the highest mean GPP of all sampled cover groups (≈1.1 kgC m-2yr-1). Figure 4.2a reveals a distinct separation between coniferous, mixedwood and broadleaf dominat-ed stands in the Boreal Shield East. Broadleaf dominated stands had the highest canopy cover (generally > 80%) and the highest GPP (generally 1.0 - 1.3 kgC m-2yr-1). Mixedwood stands had high GPP (generally 0.9 - 1.2 kgC m-2yr-1) however a wider range in canopy cover as most stands are concentrated between 50 – 95 % cover. Coniferous stands had the largest sampled ranges in both canopy cover and GPP, with the majority of stands having a cover between 20 - 85 % and GPP values between 0.3 - 1.0 kgC m-2yr-1. A positive trend between lidar-derived can-opy cover and MODIS GPP is clearly apparent within coniferous stands, while no trend is appar-ent within broadleaf or mixedwood stands.  62  Figure 4.3 provides insight to the drivers of GPP by displaying the relationship between canopy cover and Minimum Annual Temperature (MAT). The relationship between lidar-derived cano-py cover and MAT was only statistically significant in the Boreal Shield East (r = 0.68, p < 0.05, Figure 4.3b), where the sampled range of MAT was highest, and the Boreal Cordillera (r = 0.58, p < 0.01, Figure 4.3e). Similarly to GPP, the 90 – 100 % cover group in the Boreal Shield East had a higher mean MAT than any other sampled cover group across all ecozones. The relation-ship between lidar-derived canopy cover and Total Annual Precipitation (TAP) was not signifi-cant in any ecozone.   Figure 4.3 Relationship between percent cover above 2 m and Minimum annual temperature for a) Boreal Shield East (scatter-plot), shaded by dominant forest type  b) Boreal Shield East (boxplot) c) Boreal Shield West d) Boreal Plains e) Boreal Cordil-lera  f) Taiga Shield East  g) Taiga Plains  h) Hudson Plains. The number above each bin corresponds to the number of samples within the bin    63    Figure 4.4 Relationships between 95th height percentile and MODIS GPP for a) Boreal Shield East (scatterplot), shaded by dominant forest type b) Boreal Shield East (boxplot) c) Boreal Shield West d) Boreal Plains e) Boreal Cordillera f) Taiga Shield East g) Taiga Plains h) Hudson Plains. The number above each bin corresponds to the number of samples within the bin    Table 4.2 The correlation coefficients, slopes and modified t-test results for the relationship between the 95th height percentile (X) and MODIS GPP (Y). A distance interval of 10 km was used to calculate the effective sample size. Slopes are only dis-played for the statistically significant relationships (p< 0.05*, p < 0.01**, p < 0.001***). Ecozone Sample Size k Effective Sample Size r Slope*100 Boreal Shield E. 1809 257 19.35 0.49* 2.99 Boreal Shield W. 842 121 26.45 0.47* 1.45 Boreal Plains 145 112 40.47 0.12  Boreal Cordillera 1488 84 100.19 0.33*** 0.89 Taiga Shield E. 465 112 24.03 0.45* 3.34 Taiga Plains 701 144 16.72 0.59* 1.51 Hudson Plains 136 84 11.10 0.47     64   Stand height 4.3.2Table 4.2 presents the correlation coefficients, slopes and modified t-test results for the relation-ship between the 95th height percentile and MODIS GPP using a distance interval of 10 km, while Figure 4.4 displays the relationship as a series of boxplots. The effective sample sizes were relatively similar to Table 4.1, with the exception of a large increase in the Boreal Cordillera (54.38 to 100.19) and a large decrease in the Boreal Shield West (57.92 to 26.45). Similarly to lidar-derived cover, changing the distance interval had no effect on the rejection of the null hy-pothesis (α = 0.05). The level of significance did change from p < 0.01 to 0.001 when the dis-tance was increased from 5 km to 10 km in the Boreal Cordillera. Correlation coefficients were significant in the Boreal Shield East (r = 0.49, p < 0.05, Figure 4.4b) and West (r = 0.47, p < 0.05, Figure 4.4c), the Boreal Cordillera (r = 0.33, p < 0.001, Figure 4.4e), Taiga Plains (r = 0.59, p < 0.05, Figure 4.4g) and Taiga Shield East (r = 0.45, p < 0.05, Figure 4.4f). With the exception of the Boreal Shield West and Hudson Plains, the correlation coefficients between lidar-derived canopy height and MODIS GPP were lower in each ecozone than for lidar-derived canopy cover. The majority of stands were concentrated into relatively few height bins compared to canopy cover, with nearly 75 % of the stands in the Boreal Shield East (Figure 4.4b) between 9-15 m. The Taiga Plains (Figure 4.4g) contained the tallest stands, while few tall stands were sampled in the Taiga Shield East (Figure 4.4f) or Hudson Plains (Figure 4.4h). Approximately 4 % of stands in the Boreal Shield East had a 95th height percentile above 18 m. Most of these regionally tall stands in the Boreal Shield East are dominated by broadleaf and mixedwood forest types, with coniferous stands reaching a maximum lidar-derived height near 18 m (Figure 4.4a). Compared to the link between lidar-derived canopy cover and MODIS GPP, the link between the 95th 65  height percentile and MODIS GPP is not as linear, which is apparent by comparing the Boreal Shield East scatterplots (Figure 4.2a vs. Figure 4.4a). The most notable trend in Figure 4.4a is that the maximum sampled height derived from lidar increases as GPP increases.     Structural complexity 4.3.3The relationship between the CV of return heights and MODIS GPP is more complex than per-cent cover above 2 m or the 95th height percentile. Figure 4.5a displays the relationship between the CV and MODIS GPP in the Boreal Shield East, with points shaded according to dominant forest type. At low levels of MODIS GPP, the range of sampled CV values was narrow and cen-tered near 0.4. As GPP increases, the range of sampled CV values became wider but remained  Figure 4.5 a) Relationship between the CV of return height and MODIS GPP in the Boreal Shield East, shaded by a) dominant forest type b) the 95th height percentile (coniferous dominated stands only)    66  centered near 0.4. Broadleaf dominated stands generally had the lowest CV, while mixedwood and coniferous stands had a larger range in CV than broadleaf stands.  Figure 4.5b displays the relationship between the CV of return heights and MODIS GPP for co-niferous cells shaded by the 95th height percentile. Short stands tended to have lower CV values than taller stands with similar GPP and the CV of short stands decreased slightly as GPP in-creased. Taller stands had a wider range of CV values than short stands, and the maximum sam-pled CV for tall stands increased as GPP increased.   Discussion 4.4 Canopy cover 4.4.1Strong links between lidar-derived canopy cover and satellite-derived GPP in the boreal are an-ticipated for two reasons: 1) more productive sites can support a higher density of trees with more dense canopies; and, 2) canopy cover relates to the amount of foliage, which is a key driver of productivity (Schulze et al., 2002). The strength of the relationship between lidar-derived can-opy cover and MODIS GPP across ecozones was related to the sampled range of Minimum An-nual Temperature (MAT), as temperature is the main climatic driver of productivity in Canadian boreal forests (Churkina and Running, 1998; Boisvenue and Running, 2006). The largest gradi-ent in MAT occurred in the Boreal Shield East, with cold temperatures limiting the productivity and observed stand density in northern coniferous forests compared to southern broadleaf forests. The observed differences in lidar-derived canopy cover between forest types in the Boreal Shield East is likely caused by this strong latitudinal gradient, as forest type transitions along with tem-perature from broadleaf dominated stands in the south to coniferous dominated stands in the 67  north. To investigate the differences in structure across forest types, forest stands under similar site conditions would need to be isolated to remove this latitudinal effect.  While temperature is a main limiting factor of productivity in the Canadian boreal, productivity is fundamentally restricted by the amount of foliage that is absorbing solar radiation (Schulze et al., 2002). The MODIS GPP algorithm accounts for the amount of absorbed solar radiation with the MODIS FPAR product, explaining why MODIS GPP correlates more closely to lidar-derived canopy cover within most ecozones than temperature data alone. In addition, as the MODIS FPAR product is essentially measuring foliage amounts, FPAR relates directly to canopy cover.  Several reasons exist for why canopy cover and GPP would not be even more strongly related. GPP can vary between stands with similar canopy cover values if differences exist in LUE, re-ceived solar radiation or the fraction of the canopy that is composed of foliage (Figure 4.6a). Productivity can also vary as a function of stand age and successional stage as younger stands are often more productive than older stands (Ryan et al., 1997).    The slope between canopy cover and GPP likely varied between ecozones because of spatial var-iability in LUE. Specifically, while dense forest canopies were found across multiple ecozones (e.g., canopy cover > 80%), the productivity of dense forest canopies was highest in the Boreal Shield East, because these forests had warmer temperatures, and therefore higher LUE.  There-fore, while canopy cover is an important indicator of APAR, the relationship between canopy cover and productivity will vary depending on climatic conditions and LUE.   68  The lack of statistically significant relationships between lidar-derived canopy cover and Total Annual Precipitation (TAP) are in agreement with past studies showing temperature, not precipi-tation, to be the primary factor limiting growth in the boreal (Churkina and Running, 1998). If precipitation does play a role in determining canopy cover in the boreal, it would be obscured by the strong latitudinal effects in this analysis.  Figure 4.6 Schematic representations of the observed relationships in the Boreal Shield East between MODIS GPP and a) percent cover above 2 m b) 95th height percentile c) CV of return height    69   Stand height 4.4.2The correlations between the 95th height percentile and MODIS GPP highlight the importance of successional stage and stand age in shaping structure in Canadian boreal forests. Stand-replacing disturbances are the main determinant of age, and therefore height, in the boreal (Kurz and Apps, 1999; Bond-Lamberty et al., 2007; Amiro et al., 2009), explaining why lidar-derived stand height was not linked as strongly to MODIS GPP as lidar-derived canopy cover in all but the Bo-real Shield West and Hudson Plains. Productivity affects stand height by determining the rate of growth between disturbances and restricting growth in the prolonged absence of disturbance in low productivity sites (Boucher et al., 2006). The relationship between the 95th height percentile and MODIS GPP in the Boreal Shield East likely provides insights to the effects of both produc-tivity and age on forest structure (Figure 4.6b). Relatively young stands were likely sampled across a wide range of MODIS GPP, explaining why the minimum sampled stand height re-mained relatively constant as GPP increased. Alternatively, the maximum sampled 95th height percentile increased along with MODIS GPP, as growth is faster and less restricted by resources at high levels of GPP. Therefore, stands can become taller in the prolonged absence of disturb-ance on more productive sites. It should be noted that stand height will not continuously increase with time since stand-replacing disturbance, as the transition from deciduous to coniferous dom-inance and non-stand-replacing disturbances can reduce stand height (Pare and Bergeron, 1995; Brassard et al., 2008) and stands will not grow indefinitely.  The finding that most coniferous stands reach a maximum 95th height percentile around 18 m in the Boreal Shield East corresponds well to other studies of forest structure in the Boreal Shield (Pare and Bergeron, 1995; Brassard et al., 2008). Higher productivity in the southern portion of 70  the ecozone and tall broadleaf species such as trembling aspen allow mixedwood and broadleaf stands to grow taller than sampled coniferous stands (Pare and Bergeron, 1995; Brassard et al., 2008).  Structural complexity 4.4.3Successional stage and age also play an important role in determining the structural complexity of forests. The range of sampled CV values became wider in the Boreal Shield East as GPP in-creased because of several competing factors (Figure 4.6c). First, fewer young stands were likely sampled in cells with low GPP compared to cells with high GPP, as growth rates are likely slow-er where GPP is low, requiring more time for stands to reach five meters in height (i.e., the min-imum height considered in this analysis). The inclusion of younger stands at higher levels of GPP could explain why the CV of return height in short stands decreases as GPP increases. Can-opy gaps, uneven-aged structure and less dense vegetation in mature, low productivity stands will generally result in more complex forest structures than in young, highly productive stands. Alternatively, maximum tree dimensions are less restricted on highly productive sites and stands can reach an uneven-aged structure faster (Boucher et al., 2006; Larson et al., 2008). Therefore, mature forest stands will be more structurally complex on high productivity sites than low productivity sites. As a result, the differences between the structural complexities of young and mature stands appear to become greater as GPP increases.  The spherical shape of broadleaf crowns and the greater height of the sampled broadleaf stands in the Boreal Shield East results in generally low CV values for broadleaf dominated stands. The 71  presence of multi-aged and multi-species canopies in mixedwood stands is the expected cause of higher CV values for many mixedwood stands compared to broadleaf stands.   Considerations for interpreting results 4.4.4In this analysis, a ten-year average of MODIS GPP, which acts as a long-term indicator of forest productivity, was compared to a single snapshot in time of forest structure from airborne lidar data. As most sampled stands are older than ten years and have varying disturbance histories, productivity over the most recent ten years would only reflect part of the observed stand struc-ture. To better quantify the relationship between productivity and forest structure, disturbance history, successional stage and stand age must be accounted for. To do so, I attempted to restrict this study to mature unmanaged forests. However, the presence of short stands in highly produc-tive forests suggests that all young stands were not removed from the analysis, as natural disturb-ance is a fundamental component of the ecosystem, and is difficult to monitor in its entirety. It is critical to note that the strong correlations observed between MODIS GPP and canopy cover do not necessary imply causation, as satellite-derived estimates of productivity and lidar-derived estimates of canopy cover are inherently linked. Specifically, the MODIS GPP product relies on the MODIS FPAR product to estimate the fraction of radiation that is absorbed by forest cano-pies, and the MODIS FPAR product is sensitive to variations in canopy cover (i.e., denser forest canopies can absorb more radiation). While causation should not be assumed when interpreting the results of this chapter, the results do provide important insights into the information provided by satellite-derived estimates of forest productivity and how these products can be used to de-scribe structural variability.    72   Questions answered  4.5This chapter proposed and addressed three research questions. The main findings for each ques-tion are briefly reviewed here and expanded upon in the dissertation conclusions (Section 7.2).   Do satellite-derived estimates of productivity relate more closely to lidar-derived estimates of canopy cover or stand height? Satellite-derived estimates of forest productivity were more strongly related to canopy cover (r = 0.27 – 0.74) than stand height (r = 0.12 – 0.59) in five of the seven sampled ecozones. Stand height varied widely within narrow ranges of forest productivity, as stand height is largely a function of stand age and disturbance history. By incorporating Landsat time-series information to control for stage age and disturbance history, the relationship between satellite-derived productivity and stand height could be better assessed.   Does the relationship between canopy cover and productivity vary between ecozones? The slope of the relationship between canopy cover and productivity varied widely between ecozones, likely due to spatial variability in LUE across the boreal. The slope of the relationship was steepest in the Boreal Shield East, which also had the widest range in temperature, an im-portant driver of LUE.  Do the most structurally complex stands occur where productivity is highest?  The most structural complexity stands, as measured with the Coefficient of Variation (CV) of lidar return heights, occurred in high productivity, mixedwood forests. Data from the Boreal 73  Shield East were used to address this question, given the wide sampled range of productivity in that ecozone.     74  Chapter 5 How does forest structure vary as a function of time since fire for early successional stands?  How does forest structure vary as a function of time since fire for early successional stands? 5 Introduction 5.1Post-fire regrowth is an important component of carbon dynamics in Canada’s boreal forests, yet observations of structural development following fire are lacking across this remote and expan-sive region. When combined with ancillary information on disturbance history, airborne lidar transects can provide direct measures of post-fire structure over large forested areas (Goetz et al., 2010; Kane et al., 2014). While Chapter 4 investigated spatial variability in forest structure along gradients of forest productivity, this chapter aims to assess variability in structure as a function of time since fire.  Previously, using the airborne lidar transects collected over the Canadian boreal, Magnussen and Wulder (2012) developed a relationship between canopy height and time since fire for 163 fires that occurred from 1942 to 2007, which were recorded in the Canadian National Fire Database (CNFDB), a compilation of historical fire data from fire management agencies across Canada.  As most fire perimeters in the CNFDB are generalized and can contain a mosaic of burned and unburned forest patches, Magnussen and Wulder (2012) required a statistical approach based on maximum expected growth to separate burned, regenerating vegetation from unburned vegeta-tion within the fire perimeters. Using Landsat time-series data to delineate burned areas in place of the CNFDB would facilitate a more precise assessment of post-fire structure.  75  In this chapter, high-severity burned patches are detected from 1985-2010 using Landsat time-series data across 40 million ha of Canada’s Boreal Shield West ecozone, and the structural re-sponse to these fires is assessed using airborne lidar transects. By sampling patches that burned over a 25-year period, a 25-year chronosequence of structural development is constructed to ad-dress the following questions:  Do lidar metrics capture residual forest structures (e.g., snags, surviving trees) as well as tree regeneration during the first 25 years since fire (YSF)?  High-intensity crown fires, which dominate the Canadian boreal, leave little to no live tree cover in the immediate years following fire (Heinselman, 1981; Viereck, 1983). Before regenerating trees begin to form an overstory, lidar metrics will capture and relate to residual structures such as snags (i.e., standing dead wood) or trees that survive the fire. As tree growth is restricted to short growing seasons (Bonan and Shugart, 1989) and tree establishment can take a number of years following fire in the boreal (Johnstone et al., 2004), an overstory is not expected to begin forming until at least ten years after fire (e.g., Gralewicz et al., 2012).  Given the short chronosequence of structure observed in this chapter (25 years following fire), I expect most stands to remain in the stand initiation stage of succession through the entire chron-osequence (Chen and Popadiouk, 2002; Harper et al., 2005). In black spruce forests in Ontario, Harper et al. (2005) found that stand initiation lasted 34-39 years, which suggests that growing space will remain in most stands at the end of the chronosequence. However, on sites that are rapidly colonized by broadleaf species following fire, it is possible that stem exclusion will be reached within 25 years (Lieffers et al., 2003; Johnstone et al., 2004).  76  How does structure at the end of the chronosequence compare to structure in patches with no record of burning? In particular, I am interested in observing if stands reach crown closure by the end of the chrono-sequence (21 – 25 YSF), and how estimates of stand height, an indicator of aboveground bio-mass, compare to stands with no record of burning. Harper et al. (2002) found that canopy cover was highest in eastern boreal stands from 50-100 YSF, while stand height peaked from 75-150 YSF, suggesting that structure at the end of the chronosequence will remain significantly differ-ent than in patches with no recorded burns. Additionally, I expect stand structure to be less com-plex at the end of the chronosequence compared to stands with no record of burning, as canopy breakup and gap filling (Chen and Popadiouk, 2002) lead to more complex canopies in older bo-real stands (Brassard et al., 2008).  Does canopy cover prior to fire relate to stand development post-fire?  Site-level conditions and pre-fire stand composition have been shown to strongly influence stand development following fire in boreal forests (e.g., Epting and Verbyla, 2005; Harper et al., 2005; Johnstone and Chapin, 2006a; Greene et al., 2007; Chen et al., 2009). As structural measure-ments prior to fire are not available, burned patches in this chapter are classified into previously open (20 – 50% canopy cover) and previously dense (> 50% canopy cover) forest using pre-fire Landsat imagery. I expect newly established trees to grow faster in patches classified as dense forest prior to fire, as these sites are likely more productive. However, it is unclear the degree to which structural differences between previously dense and open patches will be detectable with lidar during the first 25 YSF.   77  By addressing these questions, this chapter aims to 1) develop improved techniques for assessing structural response to fire over large and remote forested areas, and 2) provide an improved char-acterization of structural variability following fire in boreal forests that lack sufficient field measurements of post-fire structure.   Material and methods 5.2 Study area 5.2.1The Boreal Shield is the largest Canadian ecozone, spanning from Newfoundland in the east to Saskatchewan in the west (Ecological Stratification Working Group, 1995). As the Boreal Shield spans a wide range of climatic and ecosystem conditions from east to west, the ecozone is often divided into east and west compartments for analysis (Stocks et al., 2002; Stinson et al., 2011).The percent annual area burned between 1959-1997 was more than five times higher in the Boreal Shield West (0.76%) compared to the Boreal Shield East (0.15%, Stocks et al. 2002), mainly due to drier conditions  in the west and a higher probability of lightning strikes (Brassard and Chen, 2006). Due to this large difference in percent annual area burned, the Boreal Shield West was selected as the focus of this study.  With a relatively short fire cycle in the Boreal Shield West (~130 years based on the above esti-mate of percent annual area burned), the initial cohort of trees that establishes following fire can dominate a stand until the next stand-replacing disturbance (Johnstone et al., 2010). Coniferous species establish most sites following fire, with broadleaf establishment restricted to sites with thin organic layers (Greene et al., 2007).  78   Data sources 5.2.2 Airborne lidar data 5.2.2.1Forest structure was assessed along approximately 4,500 km of lidar transects that fell within the Boreal Shield West ecozone of Canada (Figure 5.1). Details on the lidar transects, which were collected in the summer of 2010, are provided in Section 2.2.1.    Figure 5.1 The airborne lidar transects and selected Landsat scenes overlaid on the Brandt (2009) definition of the boreal. The Boreal Shield West ecozone is the focus of this research chapter.   79   Landsat data 5.2.2.2Thirteen Landsat scenes that intersect the airborne lidar transects within the Boreal Shield West were selected for the analysis (Figure 5.1). Landsat scene selection was guided by the CNFDB, with selected scenes containing 6-26 fires in the CNFDB which intersected the airborne lidar transects between 1985 and 2010. All available imagery from the Landsat Surface Reflectance Climate Data Record (CDR) collected between June-September of 1984 to 2013 and containing < 60% cloud cover were downloaded from earthexplorer.usgs.gov. The CDR consists of Landsat Thematic Mapper (TM) and Enhanced Thematic Mapper (ETM+) data processed to surface re-flectance through the Landsat Ecosystem Disturbance Adaptive Processing System ( LEDAPS, Masek et al., 2006). Each image in the CDR is delivered with Fmask, a cloud and cloud shadow mask produced at Boston University (Zhu and Woodcock, 2012). The number of downloaded images per scene ranged from 146 to 186 (Table 5.1). Approximately half the area of the Boreal Shield West was sampled using these 13 Landsat scenes (~ 40 million ha).  Table 5.1 Number of images acquired by path/row between June-Sept 1984-2013 Path Row L5 TM L7 ETM+ Total  25 23 113 49 162  27 23 109 59 168  27 24 117 55 172  29 22 114 63 177  31 21 104 59 163  31 22 117 54 171  31 23 130 56 186  33 21 111 50 161  35 21 102 56 158  36 20 92 54 146  37 21 117 65 182  38 20 117 49 166  40 19 121 52 173     80   Land cover 5.2.2.3To remove wetlands and water from the analysis, information on land cover was obtained from the Earth Observation for Sustainable Development of Forests (EOSD) dataset (available at: http://tree.pfc.forestry.ca/). The circa 2000 EOSD land cover dataset is a Landsat-based classifi-cation of the forested ecosystems of Canada led by the Canadian Forest Service, with federal, provincial, and university partners. The EOSD product is of a 25 m spatial resolution, following a classification scheme derived from the National Forest Inventory, as well as the information needs of provincial and terrestrial forest management agencies (Wulder et al., 2008b).  Fire detection 5.2.3The detection of fires was performed on a scene by scene basis, using all Landsat imagery within 45 days of August 1st from 1984-2013 (Figure 5.). A 90-day window was selected to maximize the number of Landsat images while avoiding images before vegetation greenup or after vegeta-tion senescence, and reducing images with snow. As fire results in large spectral changes in Landsat time-series analysis, slight variability between images due to phenological differences will have minimal impact on fire detection. The approach relies on detecting changes in NBR, which is calculated using Landsat band 4 (near-infrared) and band 7 (shortwave-infrared).  NBR = (B4 - B7) / (B4 + B7)        Eq. 5.1  The change in NBR from pre- to post-fire, known as the differenced NBR (dNBR), has been shown to correlate to burn severity (Hall et al., 2008; Soverel et al., 2010). dNBR = NBRPre –NBRPost        Eq. 5.2 81  Healthy vegetation is highly reflective in the near-infrared and absorptive in the mid-infrared, resulting in a drastic change in NBR when vegetation is removed by fire (i.e., decrease in near-infrared reflectance, increase in mid-infrared reflectance).  Hall et al. (2008) found a strong rela-tionship (R2 = 0.88) between dNBR and the composite burn index for 161 field plots across four large boreal fires in Saskatchewan, the Northwest Territories, and the Yukon. Based on this rela-tionship, Hall et al. (2008) developed dNBR thresholds for low, moderate, and high severity burns, which were used to detect burned patches in this analysis. Prior to detecting fires with dNBR, a number of pre-processing steps were required.   82    Generation of Landsat stacks and removal of noise 5.2.3.1Landsat data for each scene were stacked using a nearest neighbor resampling. Pixels detected as cloud or cloud shadow by Fmask were masked in each image. Additionally, as pixels in close  Figure 5.2 Flow chart of processing steps to derive lidar metrics for open and dense forest patches that burned between 1985 and 2010. These steps are described in detail in Sections 5.2.3 – 5.2.5.      83  proximity to clouds can be contaminated or contain undetected cloud edges, any pixel within 500 m of a detected cloud or cloud shadow was removed from the analysis. A water mask was creat-ed for each Landsat scene by identifying pixels consistently classified as water by Fmask (> 75% of images). Requiring that a pixel was consistently classified as water ensured that a misclassi-fied pixel in one image is not removed unnecessarily. The EOSD  LC 2000 dataset was used to mask any remaining water pixels as well as wetland areas from the Landsat stacks.  While most noise associated with clouds and cloud shadows was removed by the above proce-dures, undetected clouds, cloud shadows, haze, or smoke can lead to the false detection of change in the time-series. While the spectral signal of disturbances likely persists for multiple images, noise will typically result in a “spike” in the time-series. Kennedy et al. (2010) devel-oped a method to detect and remove spikes in Landsat time-series data by testing the similarity of a spectral index before and after the spike in the time-series. Essentially, if the pre- and post-spike values are similar to each other, relative to the magnitude of the spike, then the spike is not considered real change and is removed based on the following equation:        Eq. 5.3 Where Despike is a user-defined threshold which determines how aggressively spikes are re-moved. Despike can range from 0 to 1, with smaller values resulting in more aggressive despik-ing. To identify time-series spikes in this analysis, the Kennedy et al. (2010) approach was ap-plied to both NBR and Tasseled Cap Transformation (TCT) brightness component (Crist, 1985). DespikeIndIndIndIndInd12PrePost SpikePrePost84  Brightness was included in addition to NBR as it is well suited to capture spikes associated with smoke (which is a common source of noise when detecting fires), as smoke is reflective across all Landsat bands, with greater reflectance found with the shorter, visible, bands. The approach was applied iteratively through the time-series for each pixel, until all spikes were removed using a despike value of 0.75, which was proposed by Kennedy et al. (2010) as an aggressive value for removing time-series spikes.    Detection of burned patches 5.2.3.2Following the application of masks and removal of time-series spikes, dNBR was calculated at each image time-step. To limit the false detection of fires (i.e., errors of commission) and to fo-cus on high-severity burns, only pixels meeting the high-severity threshold defined by Hall et al. (2008) were categorized as burned (dNBR > 0.514). Pixels detected as burned within the same year were clustered into burned patches (i.e., pixels neighboring on one of eight sides). If noise (e.g., clouds) occludes a portion of a burned area, a burned patch may be detected as several patches across two or more years. To address this issue, each missing pixel was temporality filled with the next valid pixel in the time-series. If a filled pixel was detected as burned and joined or created a burned patch > 10 ha, then the pixel was labeled as burned in the year to which it was moved.  To ensure that entire burned patches were not falsely moved to a previous year, it was required that > 5% of the burn patch were observed in the first year to justify moving pixels.  85  To remove small-area disturbances and image noise, burned patches < 10 ha were removed from the analysis.  Additionally, to limit confusion between fire and anthropogenic disturbances (e.g., harvesting), any patch within 10 km of a road was required to be > 50 ha for inclusion in the analysis. While potentially removing some fires, this approach ensures all identified patches are likely fire in origin. The location of roads was determined using the 2010 Road Network File, which is a compilation of all Canadian roads recorded in Statistics Canada’s National Geograph-ic Database (available at: https://www12.statcan.gc.ca/census-recensement/2011/geo/RNF-FRR/index-eng.cfm). By detecting large, discrete changes in dNBR, confusion between fire and spruce budworm outbreaks should be minimal, as spectral changes associated with repeat defoli-ation are typically more gradual through time (Meigs et al., 2011).  In areas of Landsat scene overlap, the same fire events were detected across multiple scenes. Overlapping burned patches that occurred within +/- one year were joined into the same burned patch. The year of the fire was determined as the earliest year of detection between the overlap-ping scenes.  Forest classification 5.2.4To restrict the analysis to forested areas and differentiate the structural response based on pre-fire conditions, information on land cover prior to fire was required. A classification tree was devel-oped which stratified pixels into three canopy cover classes (0-20%, 20-50%, and > 50%) based on pre-fire Landsat data. The classification tree was developed using circa 2010 Landsat compo-sites from each Landsat scene, and trained on canopy cover estimates derived from the 2010 air-borne lidar transects.  86   Development of Landsat image composites 5.2.4.1A “best available pixel” approach (e.g., Griffiths et al., 2012; White et al., 2014) was employed to generate a circa 2010 image composite for each Landsat scene. Only images within +/- 15 days of a specified target date were used to develop the composites, as phenological stage can significantly influence the spectral characteristics of vegetation (Song and Woodcock, 2003).  Multiple target dates were tested for creating image composites (August 1st and August 15th) to determine which target date produced the highest classification accuracy. To begin, time-series spikes and pixels within 500 m of clouds or shadows were removed for each image following the procedures in Section 5.2.3.1. Starting in 2010, each pixel in the composite was filled using the closest available image to the target date. If a pixel could not be filled using 2010 data, it was filled with the closest image to the target date in 2009, and so on, back to 2005.  A pixel was marked as missing if a change in NBR > 0.514 (i.e., the threshold used for fire detection) oc-curred between the date of the selected image and the target year of the classification (2010).   Selection of training data 5.2.4.2Wulder et al. (2012b) calculated a range of lidar metrics on a 25 m grid along the lidar transects, including an estimate of canopy cover (percentage of first returns above 2 m). These canopy cover estimates were assigned to each overlapping Landsat pixel. Neighboring pixels that be-longed to the same canopy cover class (0-20%, 20-50%, and > 50%) were then grouped into training clusters. Pixels located at transitions between classes were not included in the training clusters to minimize the impact of any geolocation error and presence of transitional, possibly 87  mixed, classes. Without a water class, water pixels not captured by the applied masks would con-sistently be classified as > 50% canopy cover (i.e., dense, dark vegetation). Therefore, clusters of water were included as training data, which were extracted from the water mask derived in Sec-tion 5.2.3.1. As wetland areas can be structural similar to forested areas but quite different spec-trally, pixels classified as wetland in the EOSD were masked from the training data. To give equal weight to each training cluster, 25 pixels were randomly sampled from each (clus-ters with less than 25 non-edge pixels were removed). Additionally, to provide a training set that was representative of the sampled landscapes, training clusters were randomly sampled propor-tionally to the frequency of each class along the lidar transects (Table 5.2).    Development of classification tree 5.2.4.3Landsat bands 3, 4, 5, and 7 and a series of spectral indices were used to develop the classifica-tion tree (Table 5.3). Indices were selected which relate to overall reflectivity (TCT brightness), photosynthetic activity (NDVI), and moisture content (NDMI and NBR). The classification tree Table 5.2 Frequency of each canopy cover class along the lidar transects. Training patches were randomly sampled to match the frequency of each class across the landscape.  Class frequency Training clusters Class Pixels Percent Patches Pixels > 50% 451,752 22.7 340 8,500 20 - 50% 504,943 25.4 382 9,550 < 20% 545,905 27.5 412 10,300 Water 483,587 24.3 364 9,100    88  was developed in MATLAB using a ten-fold cross validation approach, which consists of ran-domly dividing the training clusters into ten groups of equal size and iteratively training the clas-sification on nine groups of clusters while testing on the held-out group (see Friedl and Brodley, 1997 for decision tree approaches to classifying remotely sensed imagery). Two classification tree parameters were iteratively varied to test their influence on classification accuracy: the min-imum number of observations (pixels) per node and the maximum number of tree splits. Based on the cross-validated results, the best parameters were selected, and the classification was de-veloped using all data.   Application of classification tree 5.2.4.4To apply the classification, pre-fire image composites were derived for each burned patch fol-lowing the procedure above (Section 5.2.4.1). The target year of each composite was set to the year in which the fire was detected, and filled only with images collected prior to the detection date. The classification tree was then applied to each pre-fire composite. To restrict the analysis to areas that were forested prior to burning, focused was placed on the 20-50% and >50% canopy cover classes, which are refer to as ‘previously open’ and ‘previously dense’ forest, respectively. Table 5.3 Landsat data inputs to the classification tree Inputs Citation Landsat bands 3,4, 5, and 7 - Tasseled Cap Brightness  Crist (1985) Normalized Difference Vegetation Index (NDVI)  Tucker (1979) Normalized Difference Moisture Index (NDMI)  Gao (1996), Wilson and Sader (2002) Normalized Burn Ratio (NBR) López García and Caselles (1991)    89   Identification of unburned forest stands 5.2.4.5To compare the structure of burned patches against forest patches with no recorded burns, the classification was applied across the 2010 Landsat composites. Any pixel detected as burned (dNBR > 0.514) in the time-series (1985-2010) was masked, as well as any pixel that intersected a fire in the CNFDB prior to 1985. The earliest recorded fire in the CNFDB for the Boreal Shield West occurred in 1948; however, the quality and completeness of fire records in the CNFDB var-ies through time and between contributing agencies. Pixels with no recorded burns were clus-tered to create unburned patches of open (20-50% canopy cover) and dense (> 50% canopy cov-er) forest. By applying the classification to the 2010 composites, the selected unburned patches should be spectrally similar to the burned forest patches prior to fire.  Assessing structural response to fire 5.2.5Airborne lidar data were extracted and summarized for each burned forest patch using FUSION, a software package produced by the US Forest Service for processing and visualizing lidar data (available at: http://forsys.cfr.washington.edu/fusion/fusionlatest.html). While other chapters in this dissertation utilize the lidar metrics calculated by Wulder et al. (2012b), metrics were recal-culated in this chapter for two reasons. First, because each burned patch was split into subpatches of dense and open forest, the patches in this chapter were often small (median patch size = 12.4 ha). Therefore, spatial mismatches between the 30-m Landsat pixels and the 25-m lidar cells cal-culated by Wulder et al. (2012b) may have a larger impact on these small patches than in other chapters. Second, the smaller study area of this chapter allowed for additional metrics to be cal-culated and explored which were not included in the Wulder et al. (2012b) dataset.  90  The analysis focused on several lidar metrics which describe canopy cover, stand height, and the distribution shape of lidar returns, which were calculated directly from the point clouds for each patch. Canopy cover was calculated as the percentage of first returns intercepted above 2 m to the total number of first returns, while stand height was assessed as the 75th height percentile of first returns. The 95th or 99th percentiles were avoided as these metrics will be less sensitive to regenerating vegetation if residual structures are present in a stand (i.e., inform only on the tallest objects). Only returns above 2 m were used in the calculation of the 75th height percentile to re-move the impact of returns from low vegetation and the ground.  Skewness and kurtosis, metrics describing the shape of a distribution, were also calculated for each lidar point cloud in FUSION. Skewness describes the symmetry of a distribution. Following fire, a positively skewed distribution of lidar returns could signify the presence of several tall, residual trees among a stand dominated by short, regenerating vegetation. Kurtosis describes the peakedness of a distribution. If a stand consists of dense, regenerating vegetation at a uniform height, the distribution of lidar returns would display a strong peak at the height of this dense vegetation (i.e., high kurtosis). Alternatively, if a stand consists of a range of tree heights or the foliage is dispersed over a wide vertical range, the peak of lidar returns would be less well de-fined (i.e., lower kurtosis). Similar to the calculation of the 75th height percentile, only returns above 2 m were used to calculate skewness and kurtosis.  In addition to point cloud metrics, a canopy height model (CHM) was developed for each burned patch on a 2 m grid in FUSION, representing the tallest vegetation across the landscape. To re-91  duce the influence of edge effects, the CHM was derived using a 10 m buffer around each burned patch and subsequently clipped to the patch boundary.  Rumple, a measure of canopy surface roughness and an indicator of stand structural complexity (Kane et al., 2010b), was also derived in FUSION for each burned patch. Rumple is calculated as the ratio of CHM surface area to the ground area. The surface area of each 2 m pixel in the CHM is calculated by fitting triangles between the center point of the pixel and the center points of all neighboring pixels (Kane et al., 2010b). Only canopy pixels (CHM > 2m) were used to calculate the average rumple across each patch.  Additionally, each CHM cell was stratified by height (0-2, 2-5, > 5m) to represent various layers of the canopy. Clumps within each class were derived by clustering neighboring pixels. Small clumps consisting of less than 5 pixels (20 m2) were removed from the analysis. The remaining clumps were used to calculate the percentage of area in each height class for each patch.  To provide a structural comparison between burned and unburned patches, lidar metrics were also calculated for all unburned forest patches.  Only burned and unburned forest patches for which > 5 ha were sampled with the lidar transects were analyzed. Additionally, areas that burned more than once between 1985-2010 were masked before the calculation of these lidar metrics, as the structural response of vegetation could be more complicated in these cases.   92    Results 5.3 Detection of fires and classification of forests 5.3.1Over 30,000 burned patches were detected across the Landsat scenes between 1985 and 2010 (Figure 5.3), totaling ~ 4 million ha or 17.2% of the sampled land area (after water and wetland areas were removed). Approximately 44,000 ha, or 0.2% of the area, was detected as burned more than once between 1985-2010.   Figure 5.3 Burned patches shaded by the year of detection across 13 Landsat scenes   93  The highest cross-validated accuracy for detecting canopy cover classes was achieved using a target date of August 15th  (87.2%)  compared to August 1st (85.8% ).  Cross-validated error (i.e., the percentage of pixels misclassified in the cross-validation) was lowest when minimum obser-vations per node was set between 100 and 200 (Figure 5.4a), and increased sharply when the tree was pruned to fewer than ten branches (Figure 5.4b). For the remainder of results, the minimum observations per leaf was set to 100 and the maximum number of branches was set to 15, result-ing in an overall cross-validated accuracy of 86.7% (Table 5.4). Of the ~ 4 million ha detected as burned, 30.9% was classified as dense forest (> 50 % canopy cover) prior to burning, 52.9% as open forest (20 – 50 % canopy cover) prior to burning, 15.6% as < 20% canopy cover, and <1% as water. Table 5.5 displays the number of patches sampled with airborne lidar data for an area > 5 ha, with burned patches aggregated into five year since fire (YSF) groups. More previously open patches (372 burned, 957 unburned) were sampled along the lidar transects than previously dense patches (234 burned, 629 unburned). The spatial distribution of open and dense patches was sim-ilar (Figure 5.5), suggesting that dense and open patches were typically sampled from the same landscapes and fire events. The number of burned patches sampled with lidar allowed the struc-tural response to fire to be assessed over 13,583 ha of forests.   94    Figure 5.4 Cross-validated error using a target date of August 15th for classifying canopy cover as a function of a) minimum observations per node and b) maximum number of tree splits. No tree pruning was used for panel a, while minimum observa-tions per leaf was set to 100 for panel b   95   Table 5.4 Cross-validated accuracy assessment for classifying canopy cover using a single Landsat composite centered on Au-gust 15th, 2010. Minimum observations per node were set to 100, and the maximum number of tree splits set to 15. Correctly classified pixels are underlined.  Reference Data (2010 Airborne lidar data) > 50%*  20 - 50%  < 20%  Water Total Classified Data > 50%  6939 1012 130 20 8101 20 - 50%  1410 7220 942 1 9573 < 20%  122 1311 9228 0 10661 Water 29 7 0 9079 9115 Total 8500 9550 10300 9100 37450       Producer’s 81.6 75.6 89.6 99.8  User’s 85.7 75.4 86.6 99.6  Overall accuracy 86.7        *Percentages represent cover > 2m   Table 5.5 The number of dense and open patches sampled with airborne lidar in each YSF group. Class Unburned 1-5  6-10  11-15  16-20  21-25  YSF Dense forest prior to fire (>50% cover)  629 27 14 88 53 52 Open forest prior to fire (20-50% cover)  957 57 41 126 84 64    96    Assessment of structure 5.3.2Mean canopy cover, as estimated with the percentage of first returns above 2m, was lowest 6 – 10 YSF for previously dense forest patches (8.4%) and 6 – 15 YSF for previously open forest patches (5.0 – 5.6%), followed by increasing trends to 21 – 25 YSF (Figure 5.6a). Of the burned groups, mean canopy cover was highest 21 – 25 YSF for both previously dense (41.9%) and pre-viously open (18.6%) patches, but remained significantly lower (p < 0.001) than the unburned groups (63.3% for dense and 38.6% for open). Canopy cover was statistically higher (p < 0.001)  Figure 5.5 Distribution of sampled dense and open patches across latitude and longitude. When bars exceed the dashed grey lines, more than 50% of the patches within the group fall within the corresponding latitude or longitude bin.   97  for previously dense patches compared to previously open patches in the 11 – 15 YSF, 21 – 25 YSF, and unburned groups.  Stand height, as estimated with the 75th height percentile, displayed contrasting trends to canopy cover (Figure 5.6b). For previously dense patches, the 75th height percentile decreased from 1 – 5 to 11 – 15 YSF before increasing gradually to the 21 – 25 YSF group. The 75th height percentile did not show signs of increasing in previously open patches near the end of the chronosequence. The difference between unburned and 1 – 5 YSF groups was less pronounced than for canopy cover, with no statistical difference observed between the unburned and 1-5 YSF group for pre-viously open forests. The 75th height percentile was significantly higher (p < 0.001) for previous-ly dense patches compared to previously open patches in both the unburned and 21 – 25 YSF groups. At 21 – 25 YSF, the mean 75th height percentile was approximately half as tall (4.9 m for previously dense, 4.2 m for previously open) as unburned patches (9.8 m for dense, 7.7 m for open). The interquartile range (IQR) of the 75th height percentile was lower for the 16 – 25 YSF groups (IQR = 0.9 – 1.1 m) compared with the unburned forest patches (IQR = 2.7 and 2.4 m). While variability was relatively low for the 75th height percentile from 16 – 25 YSF, variability in canopy cover was relatively high in these groups (IQR = 14.4 – 20.6 %).  The coefficient of variation (CV), which allows variability to be compared across metrics (i.e., standard deviation divided by the mean), was lower for the 75th height percentile from 15 – 25 YSF (CV = 0.12 – 0.20) compared to canopy cover (CV = 0.36 – 1.02).   98    Figure 5.6 Median lidar metrics for previously dense and open forest patches, with error bars displaying the interquartile range for a) Cover above 2m, b) the 75th height percentile, c) Skewness of return heights above 2m, d) Kurtosis of return heights above 2m, and e) Rumple derived from the Canopy Height Model (CHM). Rumple was derived from canopy pixels (CHM height > 2m) only. Asterisks represent statistical differences between dense and open patches in each group (* p < 0.05, ** p < 0.01, *** p < 0.001)   99  Skewness was significantly higher (p < 0.001) at 11 – 20 YSF compared to 1 – 10 YSF and un-burned groups for both previously dense and open patches (Figure 5.6c). Skewness increased sharply between 6 – 10 and 11 – 15 YSF for previously dense patches, followed by a decreasing trend to 21 – 25 YSF. For previously open patches, the increase in skewness was more gradual from 6 – 10 to 16 – 20 YSF.  Skewness was significantly higher (p < 0.001) for previously dense patches compared to previously open at 11 – 15 YSF, and significantly higher (p < 0.001) for previously open patches at 21 – 25 YSF and in the unburned groups. For both previously dense and open patches, skewness remained significantly higher (p < 0.001) at 21 – 25 YSF compared to the unburned groups. Skewness was more variable in the 11 – 25 YSF groups (IQR = 0.88 – 1.52) compared to the 1 – 10 YSF and unburned groups (IQR = 0.33 – 0.48).  The kurtosis, or the strength of the distribution peak of first returns, was significantly higher (p < 0.001) 11 – 25 YSF compared to the 1 – 10 YSF and unburned groups for both previously dense and open patches (Figure 5.6d). Kurtosis also increased sharply from 6 – 10 to 11 – 15 YSF for previously dense patches, followed by a gradual decrease, while kurtosis increased more gradual-ly for previously open patches from 6 – 10 to 16 – 20 YSF. Kurtosis was significantly higher (p < 0.001) for previously dense patches at 11 – 15 YSF, while significantly (p < 0.001) higher for open patches in the unburned groups. Kurtosis was also more variable in the 11 – 25 YSF groups (IQR = 5.81 – 12.17) compared to the 1 – 10 YSF and unburned groups (IQR = 0.59 – 1.37). Figure 5.6e displays the rumple of canopy pixels (i.e., pixels > 2m in height) for each YSF group derived from the Canopy Height Model (CHM). The rumple was significantly lower (p < 0.01) for the 11 – 25 YSF groups than 1 – 10 YSF and unburned groups for both previously dense and 100  open patches. The mean rumple was highest 1 – 5 YSF for previously dense (3.0 m) and 1 – 10 YSF for previously open (2.6 - 2.8 m).  Figure 5.7 displays the area of the CHM within each height class, expressed as a percentage. The percentage of area below 2 m (i.e., open areas) was highest 6 – 10 YSF for previously dense (mean = 83.7%) and 5-15 YSF for previously open (mean = 88.9 - 89.9%), followed by a de-creasing trend to 21 – 25 YSF in each case. The area below 2 m remained significantly higher (p < 0.001) in previously dense and open patches at 21 – 25 YSF compared to the unburned groups. The average area between 2 – 5 m increased from 6 – 10 to 21 – 25 YSF for previously dense (8.8% to 44.1%) and from 11 – 15 to 21 – 25 YSF for previously open (8.7% to 31.2%). For pre-viously dense patches, the area between 2 – 5 m was more than three times higher at 21 – 25 YSF (mean = 44.1%) compared to unburned (mean = 12.5%). The area above 5 m was relatively high in the first five YSF (mean = 38.3% for dense, 27.1% for open), but decreased sharply by 6 – 10 YSF (mean = 7.4% for dense, 6.7% for open). The average area above 5 m increased for previously dense patches from 16 – 20 to 21 – 25 YSF (7.7% to 23.4%), but remained signifi-cantly lower (p < 0.001) than unburned patches (75.9%). No significant increase was observed in the area above 5 m for previously open patches at the end of the chronosequence.   101    Discussion 5.4The integration of historical disturbance detection from Landsat and structural measurements from lidar provides powerful means for characterizing the response of forests to disturbance over large areas (Kane et al., 2013, 2014). As lidar structural measurements were limited to a single snapshot in time (summer 2010), pre- and post-burn structure could not be directly compared. Therefore, an approach was demonstrated in this chapter that relies on pre-fire spectral infor-mation from Landsat to differentiate varying levels of canopy cover prior to disturbance. By comparing structural attributes from burned stands against those from spectrally similar un-burned stands, the impacts of fire on structure could be assessed.  Figure 5.7 Median percentage of Canopy Height Model (CHM) area in each height class, with error bars displaying the interquartile range, for a) previously dense and b) previously open patches. Only clumps > = 5 pixels (20m2) were included in the area calculation.   102  Statistical differences in structural attributes between unburned and 21 – 25 YSF patches confirm the significant and lasting impact of high-severity fire on structure in the Boreal Shield West. In addition, the results demonstrate the large proportion of the landscape impacted by these burns (17.2% of the sampled area over 25 years, or 0.68%/year). Stocks et al. (2002) found a similar burn rate for the Boreal Shield West (0.76%/year) between 1959-1997, however several key dif-ferences exist between the estimates: 1) the estimates from this analysis are for 13 Landsat scenes and not the entire ecozone, 2) Only burn patches that met the high-severity threshold de-fined by Hall et al. (2008) were considered in this analysis, and 3) Stocks et al. (2002) only con-sider burns > 200 ha.  Forest structure immediately following stand-replacing fire is characterized by standing dead wood (i.e., snags) and little to no live tree cover. Over time, the open space created by the dis-turbance is filled with new vegetation and standing dead wood begins to fall (Chen and Popadiouk, 2002; Harper et al., 2005; Angers et al., 2011). This transition from a canopy of re-sidual structures (i.e., snags or surviving trees) to one dominated by young, even-aged trees was captured clearly by lidar metrics during the 25-year chronosequence.   Residual structures dominated canopies in the first decade following fire 5.4.1In the first decade following fire, canopies remained relatively open (i.e., low canopy cover), as most regenerating trees remained below 2 m in height. Slow boreal growth rates and delayed tree establishment would prevent most trees from growing more than 2 m in height during this short time span (Sirois and Payette, 1989; Johnstone et al., 2004; Harper et al., 2005; Bartels et al., 2016). At boreal sites in Alaska and the Yukon, Johnstone et al. (2004) found that most trees 103  took 3 – 7 years to establish following fire. Therefore, instead of detecting new tree growth, lidar metrics were sensitive to residual canopy structures in the first ten YSF. Specifically, the 75th height percentile provided evidence of tall residual trees, as regenerating trees could not achieve the observed heights in less than ten years. The loss of residual structure through time was cap-tured by the CHM metrics, as the percentage of area above 5 m decreased sharply from 1 – 5 to 6 – 10 YSF (Bond-Lamberty and Gower, 2008; Angers et al., 2011). This loss of residual structure as snags begin to fall, in addition to new tree growth above 2 m in height, explains the decreasing trend in the 75th height percentile between 1 – 5 YSF to 11 – 15 YSF.   Pre-disturbance imagery provided expectations for post-fire stand 5.4.2regeneration From 11 – 25 YSF, estimates of canopy cover and stand height captured the emergence and growth of new trees into the canopy. While previously dense and open patches were structurally similar in the first ten YSF, estimates of canopy cover and stand height were statistically higher in previously dense patches by 21 – 25 YSF, suggesting faster growth and recovery of trees compared to previously open patches. Differences in growth rates between previously dense and open patches were further confirmed by CHM metrics, as considerably more canopy area was above 5 m in height in previously dense patches at 21 – 25 YSF compared to previously open patches. While temperature is the main limiting factor to growth in most boreal forests (Churkina and Running, 1998), differences in site-conditions, species composition, and fire severity can lead to local variability in site productivity and stand regeneration following fire (Arseneault, 2001; Johnstone et al., 2004; Harper et al., 2005; Johnstone and Chapin, 2006; Lecomte et al., 2006a,b). The classification of burned patches into previously dense or open forest appears to 104  have captured some of the underlying differences in site productivity and stand development across the landscape. Specifically, the factors that limited growth prior to disturbance, which de-termined if a patch was classified as dense or open forest prior to burning, were also limiting the growth of new trees following fire. Patches classified as previously dense forest had faster growth  and recovery likely due to more suitable soil conditions for seed germination (Certini, 2005; Johnstone and Chapin, 2006; Lecomte et al., 2006a,b), higher nutrient availability (Bhatti et al., 2002; Payette and Delwaide, 2003), or more favorable positions in the landscape ( e.g., more sunlight, Bonan and Shugart, 1989). Further, the species that colonize a site can play a sig-nificant role in site productivity (Johnstone et al., 2004; Mack et al., 2008). For example, canopy closure would be reached faster in stands dominated by trembling aspen compared to black spruce, due to differences in species growth rates (Chen and Popadiouk, 2002; Johnstone et al., 2004; Mack et al., 2008). As tree establishment post-fire is strongly linked to pre-fire species composition (Turner et al., 1997; Epting and Verbyla, 2005; Chen et al., 2009; Johnstone et al., 2010), structural differences between stands due to species composition could re-develop during the early stages of succession.  While pre-fire species composition and soil conditions are important indicators of post-fire re-generation, the severity of a fire can also influence the successional pathway of a stand, as severe fires can alter soil conditions and destroy seed sources (Certini, 2005; Johnstone and Chapin, 2006; Lecomte et al., 2006a; Johnstone et al., 2010). Johnstone et al. (2010) found that stands previously dominated by black spruce in Alaska can shift to broadleaf dominance following se-vere fire, as these fires can destroy the serotinous cones of black spruce  and consume the surface organic layer on which black spruce can outcompete broadleaf species. Similarly, Lecomte et 105  al.(2006b) found that stands produced 50% more tree biomass in the first 100 years following high severity fire compared to low severity fire in the eastern Canadian boreal, as soil conditions following severe fire are more suitable for the growth of most boreal tree species (i.e., removal of organic layer). Therefore, while pre-disturbance spectral information from Landsat was an important indicator of post-fire stand development in this study, productivity may shift depend-ing on fire severity due to changes in soil conditions and the removal of biological legacies (e.g., seed sources, roots).  Growing space remained in stands at the end of the chronosequence 5.4.3When the open space created by a disturbance is filled, stand initiation is complete, and estab-lished trees begin to compete for available sunlight and nutrients (Oliver and Larson, 1990; Chen and Popadiouk, 2002). The finding that canopy cover remained significantly lower 21 – 25 YSF compared to stands with no record of burning suggests that available growing space remained at the end of the chronosequence and stem exclusion had not yet been reached. Higher percentages of open areas (CHM < 2m) at 21 – 25 YSF compared to patches with no recorded burns provides further evidence that available growing space remained. Harper et al. (2005) found that stand initiation lasted 34-39 years in black spruce boreal forests in Ontario, while Lieffers et al. (2003) suggests 5 – 15 years for aspen or pine dominated boreal stands and up to 40 years for boreal stands dominated by spruce. Using lidar metrics to assess successional stage in Pacific Northwest forests, Kane et al. (2011) found that canopies in the stem exclusion stage were homogenous and contained relatively few gaps.  Most patches did not reach this structural definition of stem ex-clusion by the end of the chronosequence, as CHMs remained relatively open. However, it should be noted that site limitations on establishment and growth may prevent stands from reach-106  ing canopy closure for many decades, if at all, in low productivity boreal stands (Harper et al., 2005). In these cases, competition for other resources besides sunlight (e.g., nutrients and water) may lead to stem exclusion before canopy closure is reached (Carleton and Wannamakerf, 1987). The availability of growing space is therefore only one indicator of the transition from the stand initiation to stem exclusion stage of succession.   Regenerating stands were more homogenous than unburned stands  5.4.4Stands typically maintain an even-aged structure for many years following fire, until shade-tolerant trees establish in the understory and trees in the initial cohort begin to die (Chen and Popadiouk, 2002; Taylor and Chen, 2011).  The even-aged nature of regenerating trees was first observed at 11 – 15 YSF, when lidar metrics began to describe the young, emerging canopy. At this time, kurtosis increased sharply in previously dense patches, suggesting a strong peak in li-dar returns from vegetation at a uniform height. Skewness also increased sharply for previously dense patches at 11 – 15 YSF, as the combination of short, dense vegetation and tall, sparse veg-etation (i.e., residual vegetation) led to positively skewed distributions of lidar returns. Slower growth rates and less dense vegetation explain why skewness and kurtosis increased later, and not as sharply, in previously open patches. As time since fire increases and trees grow taller, fo-liage becomes more dispersed vertically, explaining why kurtosis decreased in previously dense patches at 21 – 25 YSF.  Skewness also decreased in previously dense patches by the end of the chronosequence, as a smaller proportion of lidar returns will be from residual structures as the density of new trees increases, and snags continue to fall. The length of time that boreal snags remain standing varies widely in the literature (e.g., Angers et al., 2011; Bond-Lamberty and Gower, 2008; Boulanger and Sirois, 2006). For example, Angers et al. (2011) reported a half-life 107  (i.e., length of time for half of snags to fall after mortality) of 4.4 years while Boulanger and Si-rois (2006) reported 16.2 years for black spruce in Quebec following fire. This variability in snag persistence, in addition to variability in tree survival, may contribute to the high variability (i.e., high interquartile range) observed within groups for skewness and kurtosis between 11 – 25 YSF. Low skewness for unburned stands suggests that vegetation elements are more normally or evenly distributed in the canopy than recently burned patches, while low kurtosis suggests that vegetation elements are spread over a wide vertical range, either due to a range of tree heights or a broad canopy layer. Ewijk et al. (2011) noted a similar difference in the vertical distribution of lidar returns between early and late successional stands in central Ontario.  Tree size diversity is low in early successional stands, as trees are even-aged and maximum tree size is limited by a short growth period (Boucher et al., 2006; Bradford and Kastendick, 2010).  Through time, tree size diversity increases, as differences in growth rates become more pro-nounced (Boucher et al., 2006), and tree mortality and gap forming disturbances lead to un-even aged structures (Chen and Popadiouk, 2002; Brassard et al., 2008). Rumple, a metric that assess-es the roughness of the canopy surface, provided evidence of this contrast in stand complexity between early successional stands and stands with no record of burning.  Specifically, low rum-ple values from 11 – 25 YSF indicate homogenous structure across the young, regenerating stands, while higher rumple values indicate more complex canopy surfaces in patches with no recorded burns. From 1 – 10 YSF, rumple captured the large contrast in height between tall, re-sidual vegetation and surrounding open areas. In the presence of canopy gaps, the height of a stand can influence the rumple (i.e., larger distance from canopy to ground), which could con-tribute to the differences in rumple observed between early successional stands and stands with 108  no record of burning. In the Pacific Northwest, Kane et al. (2010, 2011) also found early succes-sional stands to be less complex than late successional stands using the rumple metric.   If the majority of trees in early successional stands occupy a narrow range in size, variability in stand height between stands would also be low. This was the case in this analysis, as within-group variability was low for the 75th height percentile in the 16 – 20 YSF and 21 – 25 YSF groups (i.e., low interquartile range). On the other hand, variability in canopy cover was relative-ly high in the 16 – 20 YSF and 21 – 25 YSF groups, as differences in tree establishment (i.e., number and density of trees) and the timing of tree emergence into the canopy can quickly lead to structural differences between canopies (Chen and Popadiouk, 2002; Greene et al., 2004; Johnstone et al., 2004). A range of factors can influence tree establishment following fire, such as the availability of seed sources, the suitability of a site for growth, and the severity of the fire (Greene et al., 2004; Johnstone et al., 2004, 2010; Harper et al., 2005; Mack et al., 2008).  These results suggest that cover metrics capture more variability between stands than height metrics in early stand development in boreal forests, due to the low diversity in tree size but potentially large diversity in tree number and density.    Variability in stand height was inherently higher in the unburned groups, as these groups contain a wide range of stand ages. While age was unknown for these patches, stands were likely older than 25 years and younger than approximately 130 years (i.e., the approximate fire cycle length for the Boreal Shield West, derived from the results of Stocks et al. 2002). Stochastic processes, such as disturbance and tree mortality, also contribute to higher structural variability between these older stands (Chen and Popadiouk, 2002; Boucher et al., 2006; Taylor and Chen, 2011).   109   Considerations for interpreting results 5.4.5Post-fire structural development depends largely on fire severity, as fire severity will influence rates of tree mortality, seed availability, and soil conditions, among other factors (Greene et al., 2004; Lecomte et al., 2006b; Johnstone et al., 2010; Angers et al., 2011). As the Canadian boreal is dominated by large, stand-replacing crown fires (Heinselman, 1981; Viereck, 1983; de Groot et al., 2013), the aim of this study was to  characterize the response of structure to severe fires only, which was accomplished by applying a dNBR threshold for high-severity fire (Hall et al., 2008). While Hall et al. (2008) found a strong relationship between dNBR and fire severity, a number of factors can confound this relationship, such as the length of time between pre- and post-fire imagery, pre-fire vegetation characteristics, and variation in vegetation and soil mois-ture (Epting et al., 2005; Wulder et al., 2009). Additionally, the assessment of fire severity itself is a highly subjective process, as fire severity is an interpretation of a fire’s impact on the envi-ronment, not a direct measure (Lentile et al., 2006; Hall et al., 2008).  Therefore, while a single dNBR threshold was applied to detect high-severity fires, a range of fire severities may be in-cluded in this analysis, contributing to the variability in structural responses observed. The re-sults suggest that most sampled patches did endure high-severity, stand replacing fires, and the high percentage of the study area impacted by these fires (17.2% in 25 years) demonstrates their significance across the landscape. However, this focus on high severity fire and the potential for confusion between severity classes should be considered when interpreting the results of this study.  The result of this study are for a single Canadian ecozone (the Boreal Shield West), and therefore, care must be taken when extrapolating these results to other boreal environments. Spe-110  cies composition, climate, and fire characteristics vary regionally, which leads to variability in post-fire structure and forest regeneration (Bergeron et al., 2004; Lecomte et al., 2006b; Burton et al., 2008; Bartels et al., 2016).  For example, faster growth and regeneration is expected in bo-real mixedwood stands that are rapidly colonized by broadleaf species (Bergeron et al., 2004; Johnstone et al., 2004; Mack et al., 2008), and regional differences in temperature and growing season length will impact growth rates (Churkina and Running, 1998; Kurz et al., 2013; Bartels et al., 2016). Further, the number of samples within each YSF group (Table 5.5) should be con-sidered when extrapolating results. Most notably, only 14 previously dense patches were sam-pled from 6 – 10 YSF. Despite this small sample size, it should be noted that these 14 patches still represented a total area of 536 ha. Sampling this large of an area with field measurements alone would be difficult, given the limited accessibility to these remote forested areas. However, more emphasis should be placed on other YSF groups with larger samples sizes, as these groups likely capture a wider range of structural variability.   Questions answered 5.5This chapter proposed and addressed three research questions. The main findings for each ques-tion are briefly reviewed here and expanded upon in the dissertation conclusions (Section 7.2).  Do lidar metrics capture residual forest structures (e.g., snags, surviving trees) as well as tree regeneration during the first 25 YSF? During the 25-year chronosequence, lidar metrics did capture the transition from open canopies of residual structures to canopies dominated by young, even-aged trees. Height estimates were higher than expected in the first decade following fire, as the 75th height percentile described re-111  sidual structures in the canopy. From 11 – 25 YSF, estimates of canopy cover captured the re-generation of young forest stands, while estimates of canopy roughness captured the even-aged structure of these regenerating stands.   How does structure at the end of the chronosequence compare to structure in patches with no record of burning? All lidar metrics remained significantly different at 21 – 25 YSF compared to stands with no rec-ord of burning. The lasting impact of high-severity fire on structure was confirmed by estimates of stand height, which were approximately half as tall for patches at 21 – 25 YSF (4.9 m for pre-viously dense, 4.2  m for previously open) compared to patches with no recorded burns (9.8 m for previously dense, 7.7 m for previously open). Further, many patches likely remained in the stand initiation stage of succession at the end of the chronosequence, as canopy cover was signif-icantly lower (p < 0.001) for patches at 21 – 25 YSF (mean = 41.9% for previously dense, 18.6% for previously open) compared to patches with no recorded burns (mean = 63.3% for previously dense, 38.6% for previously open).  Does canopy cover prior to fire relate to stand development post-fire?  Canopy cover prior to fire, as determined using a classification of pre-fire Landsat imagery, was an important indicator of post-fire stand development. Specifically, patches classified as dense forest ( > 50% canopy cover) prior to burning displayed faster growth and recovery than  patches classified as open forest (20 – 50% canopy cover) prior to burning. The classification of pre-fire Landsat imagery likely captured differences in site productivity and species composition across 112  the landscape. As site productivity and species composition prior to fire are important indicators of post-fire regeneration (Turner et al., 1997; Epting and Verbyla, 2005; Chen et al., 2009; Johnstone et al., 2010),  pre-fire Landsat imagery is a valuable source of information for estab-lishing expectations for growth and recovery.    113  Chapter 6 How does early stand development vary along gradients of produc-tivity?  How does early stand development vary along gradients of productivity? 6 Introduction 6.1While Chapter 5 provided a strong characterization of early stand development for one boreal ecozone, forest regeneration and productivity following fire will vary regionally due to differ-ences in climate, species composition, and fire characteristics, among other factors (Churkina and Running, 1998; Johnstone et al., 2004; Lecomte et al., 2006b; Kurz et al., 2013). In order to consider the combined influence of both productivity and disturbance on forest structure, varia-bility in forest regeneration along regional gradients of productivity must be characterized.    Recently, Hermosilla et al. (2016) utilized Landsat imagery to produce annual records of forest disturbance across all forested ecosystems of Canada from 1985 to 2011 at a spatial resolution of 30 m. This dataset provides an unprecedented look at forest change across Canada, and allows the location and spatial extent of recent boreal fires to be accurately assessed. In this chapter, fire disturbances mapped by Hermosilla et al. (2016) across the entire Canadian boreal from 1985-2009 are combined with structural metrics along the entire 25,000 km of lidar transects to assess variability in post-fire structure. Post-fire structural metrics are related to satellite-derived esti-mates of GPP from MODIS to address the following research questions: How does variability in stand structure change as a function of time since fire?  114  As high-intensity crown fires dominate the Canadian boreal, stands typically consist of little to no live tree cover in the first years following fire (Mack et al., 2008; de Groot et al., 2013). For-est regeneration and carbon uptake following fire will vary both locally and regionally due to a multitude of factors, including regeneration method (i.e., vegetative regeneration or from seed), species composition, site conditions, fire severity, and climate (Johnstone et al., 2004; Harper et al., 2005; Burton et al., 2008; Kurz et al., 2013; Bartels et al., 2016). Here, I am interested in de-termining how quickly differences in early stand succession are realized structurally following fire across the boreal. As boreal trees take several years to establish from seed (Johnstone et al., 2004), and growth rates are typically slow (Bonan and Shugart, 1989; Harper et al., 2005), I ex-pect variability in structure  between stands to be relatively low in the first decade following fire, and increase in the second decade as differences in newly formed canopies become pronounced. Earlier regeneration may be possible on sites dominated by broadleaf species, such as trembling aspen or white birch, as these species are capable of sprouting from roots or stumps (i.e., vegeta-tive regeneration) and have faster growth rates than coniferous species (Chen and Popadiouk, 2002; Johnstone et al., 2004; Bartels et al., 2016). While I expect variance in structure to in-crease during the length of this study (25 years following fire), I would not expect variance to increase indefinitely (Pare and Bergeron, 1995; Harper et al., 2002; Lecomte et al., 2006b). For example, high productivity sites colonized by trembling aspen may decrease in height once the initial cohort of trees die and is replaced by late successional conifers (Pare and Bergeron, 1995), reducing the difference in height compared to less productive stands.    Does this variability in stand structure correlate to coarse-resolution satellite-derived estimates of productivity?  115  Coarse-resolution (1-km) estimates of GPP from MODIS are a valuable source of information for describing regional and landscape-level variability in forest productivity (Running et al., 2004). Here, I am interested in understanding if these estimates of productivity can inform on the variability observed in early stand structure. In the immediate years following fire, prior to the formation of new canopies, productivity will not likely be an important factor in describing structural variability, as variability will be more a function of fire severity and pre-disturbance structure (Boulanger and Sirois, 2006; Angers et al., 2011). As time since fire increases, differ-ences in structure that are driven by regional variability in productivity will become more pro-nounced, and the correlation between structural attributes and productivity should increase. However, local variations in fire severity, site conditions, and species composition may leave a large portion of variation in structure unexplained by coarse-resolution productivity estimates.  By addressing these questions, this chapter provides an improved understanding of the influence of both time since fire and forest productivity on boreal forest structure and recovery, and a char-acterization of how forest recovery varies spatially across the Canadian boreal.  Materials and methods 6.2 Data sources 6.2.1 Lidar data 6.2.1.1Estimates of canopy cover and stand height, two key attributes that are indicative of carbon stor-age in aboveground biomass, were assessed along 25,000 km of lidar data collected in 2010. Per-centage of first returns above 2 m was used as an indicator of canopy cover, while stand height was assessed as the 75th height percentile of first returns. Similar to Chapter 5, the 75th height 116  percentile was chosen over the 95th or the 99th as the 75th height percentile will be less impacted by residual structures (e.g., snags or surviving trees) than these latter metrics. Lidar metrics were derived at a spatial resolution of 25 m. See Section 2.2.1 for details on the lidar data collection, processing, and selection of lidar metrics.  Landsat time-series data 6.2.1.2To identify burned patches, Landsat-derived disturbance information produced by Hermosilla et al. (2016) following the Composite 2 Change (C2C) approach was used (Figure 6.1). Specifical-ly, best-available pixel (BAP) image composites were first produced from Landsat imagery by selecting the best observations for each pixel within a specific date range (August 1+/-30 days) based on the scoring functions defined by White et al. (2014), which rank the presence and dis-tance to clouds and their shadows, the atmospheric quality, and the acquisition sensor. Next, these image composites were further refined by removing noisy observations (e.g., haze and smoke) and infilling data gaps using spectral trend analysis of pixel time series (Hermosilla et al., 2015a). This step results in the production of seamless annual surface reflectance composites for all of Canada from 1984 to 2012, as well as the detection and characterization of forest change events. Following the object-based image analysis approach introduced in Hermosilla et al. (2015b), the changes detected were attributed to a change type (i.e., fire, harvesting, road, or non-stand-replacing), based on their spectral, temporal, and geometrical characteristics using a Random Forests classifier, with an overall accuracy of 92%. Fire detection had the highest pro-ducer’s (98%) and user’s (93%) accuracy.    117    Figure 6.1Areas detected as burned (1985–2011) across the Canadian boreal following the Composites 2 Change (C2C) ap-proach. Panels S1, S2, S3 are examples of the intersect between detected fires and lidar transects   118   MODIS data 6.2.1.3The MODIS GPP product provides annual estimates of GPP for the entire planet at 1-km spatial resolution from 2001 to the present (Running et al., 2004). To ensure that GPP estimates were representative of post-fire conditions, and coincided with the collection of lidar data, estimates of annual GPP from 2010 were used. As vegetation type is an important input to the GPP algorithm and can significantly influence GPP (Running et al., 2004), only cells classified as forest (i.e., > 10% tree cover) in the 2010 MODIS Land Cover Type product were included in the analysis. See Section 2.2.3 for details on the MODIS GPP product.  Selection of lidar cells 6.2.2Lidar cells classified as water, wetlands, agriculture, and developed areas were removed using information on landcover from the Earth Observation for Sustainable Development of Forests (EOSD) dataset (available at: http://tree.pfc.forestry.ca/). This land cover dataset is a circa 2000 Landsat-derived classification of Canada’s forested ecosystems, produced by the Canadian For-est Service, along with federal, provincial, and university partners (Wulder et al., 2008b). To augment the EOSD in the removal of wetlands, a high spatial resolution map of global inundated areas (15 arc-seconds) was obtained, which was produced by downscaling the Global Inundation Extent from Multi-Satellites (GIEMS) dataset using topographic and hydrographic variables (Fluet-Chouinard et al., 2015). Any areas mapped as inundated in this dataset, called GIEMS-15, were considered wetlands. While this approach removes landcover categories that are not of in-terest, low productivity sites that are unable to support tree growth will likely remain within burn patches. To remove areas with high anthropogenic impact, lidar cells within 1-km of a road were 119  also removed, according to the 2010 Canadian Road Network File (available at: https://www12.statcan.gc.ca/census-recensement/2011/geo/RNF-FRR/index-eng.cfm).  Assessment of post-fire structure 6.2.3Lidar-derived estimates of canopy cover and stand height, and estimates of GPP from MODIS, were averaged across each burned patch. The contribution of each 1-km GPP pixel to the patch average was proportional to the area of lidar data that the pixel contained. Only burned patches containing > 5 ha of suitable lidar data were analyzed (i.e., lidar cells meeting the criteria out-lined above). Only lidar cells with canopy cover > 0% were used to calculate patch averages for the 75th height percentile. Burned patches were split into five groups based on years since fire (1–5, 6–10, 11–15, 16–20, 21–25 years since fire), and the relationship between lidar-derived structural metrics and satel-lite-derived GPP was assessed within each group using Pearson’s correlation coefficient and the modified t-test developed by Clifford et al. (1989) and altered by Dutilleul (1993).  When data is spatially autocorrelated, standard t-tests are not valid, as each sample does not represent a full degree of freedom (Clifford et al., 1989). In the modified t-test, the degrees of freedom are re-duced through the calculation of an “effective sample size”, which is inversely proportion to the amount of spatial autocorrelation in each variable (see Dutilleul 1993). To calculate the effective sample size, the distances between all pairs of points are divided into k distance strata and spatial autocorrelation is assessed for both variables of interest. The selection of k impacts the calcula-tion of the effective sample size, as larger values of k (i.e., shorter distance interval) will result in a higher calculated spatial autocorrelation and, therefore, a lower effective sample size (Fortin 120  and Payette, 2002). Fortin and Payette (2002) varied k between 5 and 15 in a study of boreal fire events (distance interval = 20–60 km), and found that while the effective sample size did change, varying k did not affect the rejection of their null hypothesis. To assess the sensitivity of the re-sults to the selection of k, three distance intervals were tested: 10, 20, and 40 km. The modified t-test was calculated using the Dutilleul (1993) modification in Pattern Analysis, Spatial Statistics and Geographic Exiegesis (PASSaGE; Rosenberg and Anderson 2011).  In total, structure was assessed for 417 patches that burned from 1985 to 2009. These patches covered a total area of 36,674 ha, with a median patch size of 37.7 ha. Patches that burned in 2010 were not included as the fires may have burned after the lidar flight (June 2010). To provide a comparison to post-burn structure, lidar metrics were also calculated for areas that were not disturbed for 1985–2010 according to the C2C Landsat record.  Lidar metrics were av-eraged across each 1-km MODIS pixel using lidar cells that were not disturbed between 1985 and 2010 and that met the criteria in Section 6.2.2. MODIS pixels were analyzed if they con-tained > 5 ha of suitable lidar cells. In total, structure was assessed for 15,642 undisturbed patch-es, with a median patch size of 14.6 ha. As the modified t-test is computationally intensive, 2000 undisturbed patches were randomly sampled to test the significance of the relationship between lidar metrics and productivity.    121    Figure 6.2 Scatterplots between lidar metrics and 2010 gross primary productivity (GPP) estimates for patches in five years since fire (YSF) groups. For comparison, patches that were undisturbed between 1985 and 2010 are displayed in the top panel, as well as in the background of subsequent panels. Summary statistics are provided for the lidar metrics. The significance of correlation coefficients were calculated using a distance interval of 20 km in the modified t-test (p< 0.05*, p < 0.01**, p < 0.001***).   122   Results 6.3Scatterplots between lidar metrics and productivity are displayed in Figure 6.2, with burned patches separated into five years since fire (YSF) groups. For comparison, the relationships for undisturbed patches (i.e., no disturbance detected for 1985–2010) are shown in the background of each panel. Varying the distance interval from 10 km to 40 km in the modified t-test did not change the significance of any correlation coefficients (α = 0.05), however, the level of signifi-cance did change from p < 0.05 to p < 0.01 in several cases.  Canopy cover (cover above 2m) was moderately related to GPP in patches with no record of dis-turbance between 1985 and 2010 (r = 0.56, p < 0.001). However, from 1–10 YSF, canopy cover was relatively low across all sampled productivities (mean = 13.3–16.0%), resulting in low vari-ance between patches (sd = 8.4–8.9%) and weak correlations to GPP (r = 0.18–0.34). While can-opy cover remained low in most patches at 11–15 YSF (mean = 11.0%, sd = 7.8%), canopy cov-er was moderately related to GPP at this time (r = 0.48, p < 0.05). By 16–20 YSF, a marked dif-ference in canopy cover existed between patches with GPP < 0.7 KgCm2yr-1 (mean = 9.1%) and patches with GPP > 0.7 KgCm2yr-1 (mean = 29.5%), leading to an increase in variance (sd = 16.0 %) and a strong relationship to GPP (r = 0.63, p < 0.01). Canopy cover exceeded 40% in 8 of 90 patches at 16–20 YSF, while only one patch out of 254 exceeded 40% between 1–15 YSF. By 21–25 YSF, mean canopy cover increased to 21.7% and the correlation to GPP was strongest (r = 0.72, p < 0.01). At 21–25 YSF, canopy cover remained below 10% in more than half of low productivity patches (GPP < 0.6 KgCm2yr-1), compared to only 5% of patches with GPP > 0.6 123  KgCm2yr-1. The variability between patches at 16–20 YSF (sd = 16.0%) and 21–25 YSF (sd = 17.1%) was nearly has high as in undisturbed patches (sd = 21.9%).  The results of stand height (75th height percentile) displayed a number of key differences to can-opy cover (Figure 6.2).  First, the correlation to GPP was weaker for stand height than canopy cover for undisturbed patches (r = 0.40, p < 0.001) and in all burned groups from 6–25 YSF (r = 0.01–0.25). While variance between patches increased for canopy cover from 11–15 YSF (sd = 7.8 %) to 21–25 YSF (sd = 17.1 %), variance decreased for stand height from 6–10 YSF (sd = 2.2 m) to 21–25 YSF (sd = 0.9 m). Mean stand height also decreased from 6–10 YSF (mean = 7.2 m) to 21–25 YSF (mean = 4.9 m), at which time stand height was low across all sampled productivities. While the relationship between canopy cover and GPP was stronger at 21–25 YSF than in undisturbed patches, the correlation between stand height and GPP remained weak and insignificant at 21–25 YSF (r = 0.21). Figure 6.3a displays the relationship between canopy cover and stand height for burned patches as well as undisturbed patches. Canopy cover and stand height were strongly correlated in patch-es with no record of disturbance (r = 0.79, p< 0.001), but were weakly related in stands that burned between 1985–2009 (r = 0.21, p < 0.05). Patches at 1–10 YSF tended to be taller than undisturbed patches with similar canopy cover, and tall relative to burned patches from 11–25 YSF. For example, when canopy cover was low (< 40 %), stand heights > 7m were common be-tween 1–10 YSF (37% of patches), but rare in other burned groups (6% of patches) and in undis-turbed patches (7% of cells). In contrast, when canopy cover exceed 40% in burned patches, stands tended to be short relative to undisturbed patches with similar cover. The mean height for 124  undisturbed patches was nearly double that of burned patches when canopy cover exceeded 40% (mean = 10.2m and 5.4m, respectively).   Discussion 6.4The results of this chapter clearly demonstrate the influence of both time since fire and site productivity on forest structure across the Canadian boreal. While estimates of canopy cover and stand height were related to satellite-derived estimates of GPP in stands with no record of dis- Figure 6.3 a) Relationship between lidar-derived estimates of canopy cover (cover above 2m) and stand height (75th percentile) for burned patches, with points shaded according to years since fire (YSF). For comparison, patches that were unburned be-tween 1985 and 2010 are displayed in the background. b) Schematic interpretation of structural development following boreal fire, as assessed using lidar structural metrics. Dashed lines represented expected height gains after 25 YSF.   125  turbance (1985-2009), these same relationships did not exist in the first decade following fire, highlighting the stand replacing nature of the sampled fires (Bond-Lamberty et al., 2007; Groot et al., 2013) and the slow establishment and growth of boreal trees (Johnstone et al., 2004; Harper et al., 2005; Bartels et al., 2016). Specifically, low canopy cover estimates from 1–5 YSF imply that stands were relatively open regardless of site productivity, as most tree cover was re-moved by fire. Canopy cover was below 10 % in more than half of the patches at 11–15 YSF, as insufficient time had passed for new overstory canopies to form. At boreal sites in Alaska and the Yukon, Johnstone et al. (2004) found that trees typically took 3–7 years to establish after fire. This delayed establishment time, coupled with slow boreal growth rates, supports the finding that variability in canopy structure would not be observable along productivity gradients until at least the second decade after stand-replacing fire.   Variability in canopy cover increases as time since fire increases 6.4.1Following fire, the canopies of high productivity sites will begin to refill first, as favorable site conditions and longer growing seasons allow for faster growth (Bonan and Shugart, 1989; Johnstone et al., 2004; Mack et al., 2008). In addition to reaching canopy closure first, these canopies will also be the most densely vegetated, as competition for resources and poor growing conditions can limit the number of trees that establish and grow on lower productivity sites (Harper et al., 2005; Johnstone and Chapin, 2006; Lecomte et al., 2006b). Between 16–25 YSF, estimates of canopy cover clearly captured this variability in the timing of canopy refilling and the density of regenerating canopies.  In particular, variability in productivity had finally trans-lated into structural differences by 16 – 20 YSF, as sufficient regeneration had occurred above 2-m height in high productivity stands (GPP > 0.7 KgCm2yr-1), while the canopies of lower 126  productivity stands remained relatively open. Pioneer trees likely remained below 2-m in height by the end of the chronosequence in many of the lowest productivity patches (GPP < 0.6 KgCm2yr-1), or few trees had established these sites, as canopy cover estimates remained below 10% in more than half of these low productivity patches at 21 – 25 YSF. However, canopy cover may have been below 10% in some patches in the analysis because trees never occupied the patch.  Stand height and canopy cover tell alternate stories of recovery 6.4.2While estimates of canopy cover captured the opening of forest canopies by fire and the slow formation of new canopies, estimates of stand height told a different story during the first 25 YSF. In the first decade after fire, stand height estimates were typically higher than expected from young, regenerating vegetation, suggesting the presence of residual structures (i.e., snags or surviving trees) in the canopy. The height of residual structures was not related to productivity, as the characteristics of snags and surviving trees are primarily a function of fire severity, pre-disturbance structure, and stochastic processes that determine if trees survive and remain stand-ing (e.g., Angers et al., 2011). The transition from canopies dominated by snags and surviving trees to canopies dominated by regenerating trees was captured by the decrease in mean stand height from 6–10 YSF (7.2 m) to 21–25 YSF (4.9 m). While snags may remain standing past 10 YSF, they no longer contribute significantly to the calculation of lidar metrics once canopies are dominated by pioneer trees, as they represent a smaller proportion of the vegetation above 2 m in height. Rapid regeneration and vertical growth by broadleaf species may also contribute to the stand heights observed in the first 10 YSF. However, the absence of these tall stands later in the chronosequence, and the relatively low canopy cover estimates for these stands (Figure 6.3a), 127  suggests that these early height estimates are from residual structures. In studies that use lidar-derived height metrics to assess post-fire regeneration, the presence of residual structures in the canopy must be considered, or the rate of regeneration could be overestimated. While the number of trees that establish a site can vary widely between early successional stands, height differences between stands will be minimal in the first 25 YSF, as short growing seasons limit the rate of growth, and therefore, the range of heights of pioneer trees (Boucher et al., 2006). This was confirmed by the low variability in stand height between patches from 16–25 YSF, as insufficient time had passed for differences in vertical growth to become pronounced along gradients of productivity. Alternatively, variability in canopy cover was nearly as high as in undisturbed stands by 16–25 YSF, as variability in tree establishment can be observed as soon as new canopies begin to form.  By bringing together the trends observed for canopy cover and stand height, a schematic model of forest regeneration following boreal fire  can be build (Figure 6.3b). In the first decade follow-ing fire, canopy cover will likely be low regardless of productivity, while stand height estimates will vary depending on fire severity, pre-disturbance structure, and the stochastic processes that influence the characteristics of residual structures. The results suggest that canopies will first fill in laterally prior to making significant gains in height, as the available growing space is re-occupied by pioneer trees. Once the growing space is filled, stands will continue to make vertical gains in height, and differences in height will likely become apparent across gradients of produc-tivity. Due to the short chronosequence in this analysis, and the slow boreal growth rates, these increases in height were not observed. However, the large differences that exist between the 128  height of stands at 21–25 YSF and stands with similar canopy cover, but no record of disturb-ance, suggest that significant vertical gains will be made in high productivity stands.   Considerations for interpreting results 6.4.3 The influence of averaging across burned patches 6.4.3.1Structural attributes were averaged across burned patches in this chapter, as the goal was to ob-serve how post-fire structure and regeneration vary regionally, not locally, along gradients of productivity. Differences in fire severity, species composition, and site conditions likely existed within many of the sampled burned patches, leading to variability in forest regeneration within each patch (Arseneault, 2001; Johnstone et al., 2004; Harper et al., 2005; Johnstone and Chapin, 2006; Lecomte et al., 2006a,b). While substantial regeneration above 2-m was not observed across entire patches to signal regeneration until 16 – 20 YSF, tree regeneration could occur ear-lier at the sub-patch level on sites rapidly colonized by broadleaf species, or on high productivity sites colonized by coniferous species (Johnstone et al., 2004; Bartels et al., 2016). Averaging across patches also has important implications on the assessment of stand height variability. Dif-ferences in height would be expected between rapidly established broadleaf trees and slow grow-ing conifers early after fire; however, averaging to the patch level appears to have masked these differences. The low variance in height that was observed from 16 – 25 YSF at the patch-level suggests that few burned patches were dominated entirely by rapidly growing broadleaf trees, as taller stands would be expected (Bartels et al., 2016). While understanding fine-scale variability in forest regeneration is critical for many applications, the results of this chapter provide im-portant insights into how average post-fire conditions vary regionally, as this information is im-129  portant for characterizing regional variations in carbon uptake following disturbance (Kurz et al., 2013).   Confusing residual structure and stand regeneration 6.4.3.2Determining the precise timing of residual structure loss and replacement by pioneer trees is dif-ficult with lidar metrics, as these processes are gradual and occur simultaneously. The length of time that snags remain standing in boreal stands also varies widely (Boulanger and Sirois, 2006; Angers et al., 2011). For black spruce stands in Quebec, for example, Boulanger and Sirois (2006) reported a half-life of 16.2 years (i.e., length of time for half of the snags to fall after mor-tality), while Angers et al. (2011) reported a half-life of only 4.4 years. By 11–15 YSF, when mean canopy cover was lowest, canopies may have been transitioning from residual structure dominance to dominance by pioneer trees. However, the amount of regeneration (i.e., increase in canopy cover) did not appear to outweigh the loss of residual structure (i.e., decrease in canopy cover), preventing clear evidence of regeneration from being observed for this group.   MODIS GPP is inherently related to vegetation cover 6.4.3.3A final consideration for interpreting the results of this analysis involves the use of satellite-derived estimates of GPP to assess differences in forest productivity. As vegetation greenness (i.e., as related to the calculation of MODIS FPAR) is an important input to the MODIS GPP al-gorithm, GPP is inherently related to vegetation cover, which is one reason why the relationship was consistently stronger between canopy cover and GPP than stand height and GPP. Therefore, is it important to note that the strong relationship between MODIS GPP and canopy cover does 130  not necessarily imply causation, as the observed canopy cover values will influence the satellite-derived productivity estimates. While a relationship between canopy cover and satellite GPP is therefore expected based on the inputs to the GPP algorithm, the results of this chapter provide important insights into how these relationships vary through time, and how these satellite-derives estimates of productivity are realized structurally. Additionally, as a single 1-km MODIS cell covers 100 ha, and the median burned patch size was 37.7  ha, these GPP estimates serve more as an indicator of landscape productivity, as opposed to a precise measure of productivity within each burned patch.   Questions answered 6.5This chapter proposed and addressed two research questions. The main findings for each ques-tion are briefly reviewed here and expanded upon in the dissertation conclusions (Section 7.2).  How does variability in stand structure change as a function of time since fire? Canopy cover became more variable between patches through time from 11–15 YSF (sd = 7.8 %) to 21–25 YSF (sd = 17.1 %). Alternatively, stand height became less variable through time from 6 – 10 YSF (sd = 2.2 m) to 21–25 YSF (sd = 0.9 m). Canopy cover estimates increase in variability through time as these estimates captured differences in tree establishment, while vari-ability in vertical growth is less pronounced due to slow boreal growth rates. The contrasting trends observed for canopy cover and stand height metrics point to the importance of monitoring multiple aspects of forest recovery.  131  Does this variability in stand structure correlate to coarse-resolution satellite-derived estimates of productivity? Satellite-derived estimates productivity captured differences in canopy structure across the bore-al, but only after 15 YSF, when canopy cover and MODIS GPP became strongly correlated (r = 0.63 – 0.71, p < 0.01). Conversely, estimates of stand height and MODIS GPP were weakly re-lated through the entire chronosequence (r = 0.01 – 0.25). As the variability in stand height re-mained low between patches at 21 – 25 YSF, a longer chronosequence would be needed to as-sess the ability of coarse-resolution estimates of productivity to predict differences in vertical growth.     132  Chapter 7 Conclusions  Conclusions 7 Significance of research 7.1Forest disturbance and productivity play an integral role in shaping forest structure and above-ground biomass in boreal forests. As field-based measurements of forest structure, and the fac-tors that drive structure, are scarce across unmanaged boreal forests, I employed novel remote sensing techniques to elucidate on spatial and temporal variations in forest structure over the Ca-nadian boreal. Specifically, this dissertation provides:  An assessment of existing wall-to-wall predictions of canopy height across the Canadian boreal (Chapter 3)  A quantification of how structural attributes vary along gradients of productivity in the boreal in the absence of disturbance based on lidar-derived observations (Chapter 4).  A characterization of the impacts of high-severity fire on forest structure and variability in early stand development (Chapter 5).  An investigation of how early stand development following fire varies along regional gradients of productivity (Chapter 6).  This dissertation presents a number of key innovations for assessing forest structure with remote-ly sensed data that can improve on-going and future monitoring efforts, in Canada and interna-tionally. To the best of my knowledge, I provide for the first time: 133   A method for describing variability in early stand development over large forested areas using a combination of lidar-derived structure metrics, Landsat-derived disturbance histo-ry, and MODIS-derived productivity information (Chapter 6).   A novel approach that uses Landsat imagery not only to detect fires, but also to inform on forest conditions prior to burning, which can help establish expectations for early stand development following fire (Chapter 5).   Research questions addressed 7.2This dissertation first assesses existing wall-to-wall global estimates of forest structure (Chapter 3) prior to addressing three primary research questions (Chapter 4 – 6). Each research question, along with the approach and key findings, is provided below:  Chapter 3: How well is variability in forest structure captured by existing wall-to-wall products across the boreal? Approach: I compared average estimates of canopy height from airborne lidar (95th height per-centile) against global canopy height products produced by Lefsky (2010) and Simard et al. (2011) at 1-km spatial resolution.  Main Findings: The agreement between the global height products and airborne lidar-derived height estimates varied strongly between the products (average ecozone RMSE = 3.9 and 7.4 m), demonstrating that differences in data selection, processing, and extrapolation can greatly influ-ence height predicts derived from lidar data. Differences between airborne lidar and Simard et al. (2011) height estimates were highest in ecodistricts above 60 ° N (RMSE > 6 m), suggesting that 134  high-latitude forests remain a source of uncertainty in global height products, and continued ef-fort is needed to capture spatial variability in these areas.  The global products tended to act as indicators of average landscape canopy height. Therefore, continued effort must be placed on deriving estimates of structure at the scale at which structure varies temporally. For example, quantifying the amount of carbon removed by a fire requires an estimate of pre-disturbance structure that is representative of the burned area, not the entire land-scape. If biomass in the burned portion of the landscape is higher than the landscape average, carbon emissions may be underestimated with a landscape indicator of structure (Houghton et al., 2009). While these global height products are a valuable source of information on forest structure, and represent an important step in the right direction, my results demonstrate that sig-nificant work remains to derive highly accurate stand-level estimates of forest structural attrib-utes over remote boreal forests. Airborne lidar remains an important data source for characteriz-ing variability in forest structure at the stand-scale across boreal forests, and was therefore used in place of the global products for the remainder of the dissertation.  Chapter 4: How does boreal forest structure vary along gradients of productivity in the ab-sence of recent disturbance? Approach: I related lidar-derived indicators of canopy cover (cover above 2 m), stand height (95th height percentile), and stand structural complexity (coefficient of variation of return height) to satellite-derived estimates of GPP within seven ecozones of the Canadian boreal. While my capacity to control for stand age was limited, I removed recently disturbed and managed forests 135  using information on fire history from a large fire database, anthropogenic change layers, and information on road locations.  Main findings: The strength of the relationship between canopy cover and productivity varied between ecozones (r = 0.27 – 0.74), but was strong within ecozones that had a wide range in productivity. The relationship was strong for two main reasons: first, high productivity sites can support denser canopies and reach canopy closure faster than low productivity sites. Even in the prolonged absence of disturbance, many low productivity sites will not reach canopy closure due to resource limitations, site conditions, or harsh climate (Harper et al., 2005). Second, canopy cover is an indicator of absorbed radiation (Schulze et al., 2002), which is a key driver of productivity (i.e., canopies that absorb more radiation can sequester more carbon). However, as light use efficiency will vary regionally due to differences in temperature and water stress, the slope of the relationship between canopy cover and productivity varied between ecozones.  The relationship between estimates of stand height and productivity were typically weaker (r = 0.12 – 0.59).  While productivity dictates height growth, the realized height of a stand will large-ly be a function of stand age and disturbance history. While I attempted to remove recently dis-turbed stands using information on fire and anthropogenic disturbances, stands likely varied from young to hundreds of years old in this chapter, resulting in a wide range of heights observed within narrow ranges of productivity. Further, intermediate disturbances, such as insects or dis-ease, can alter forest structure and increase variability in stand height within narrow ranges of productivity.  In the Boreal Shield East, which had the widest range of sampled productivity and temperature, a relationship was apparent between productivity and maximum sampled height, as 136  growth rates and resource availability limit maximum tree height in the prolonged absence of disturbance.  The most structurally complex stands, as measured by the coefficient of variation of lidar return heights, also occurred where productivity was highest, as maximum tree sizes are less restricted and therefore a wider range of heights is possible within a stand. Therefore, while satellite-derived estimates of productivity are strong indicators of canopy cover, these estimates act more as an indicator of potential vertical structure when detailed information on stand age is not available. Incorporating disturbance history from Landsat can provide estimates of stand age to fill this information gap.  Chapter 5: How does forest structure vary as a function of time since fire for early succes-sional stands? Approach: I used Landsat time-series data (1985-2010) to detect high-severity fires in the Boreal Shield West ecozone of Canada across 40 million ha, and assessed post-fire structure for > 600 burned patches using lidar data. I stratified burned areas into patches of dense (> 50% canopy cover) and open (20-50% canopy cover) forest based on a classification of pre-fire Landsat im-agery, and used these patches to establish a 25-year chronosequence of structural development for each class.   Main findings: During the 25-year chronosequence, lidar metrics captured the transition from open canopies of residual structures (e.g., snags and surviving trees) to canopies dominated by young, even-aged trees. While structural attributes were similar between dense and open patches during the first ten years since fire (YSF), canopy cover (cover above 2m) and stand height (75th height percentile) were significantly higher (p < 0.001) for dense patches by the end of the 137  chronosequence (21 – 25 YSF), suggesting that differences in site productivity or species com-position were driving patches towards pre-disturbance structure. The results suggest that growing space remained in stands at the end of the chronosequence, and therefore stem exclusion was not yet reached, as canopy cover was significantly lower (p < 0.001) for patches at 21 – 25 YSF (mean = 41.9% for dense, 18.6% for open) compared to patches with no recorded burns (mean = 63.3% for dense, 38.6% for open).  The lasting impact of high-severity fire on structure was fur-ther confirmed by estimates of stand height, which were approximately half as tall for patches at 21 – 25 YSF (4.9 m for dense, 4.2  m for open) compared to patches with no recorded burns (9.8 m for dense, 7.7 m for open). Lidar measures of canopy roughness (i.e., rumple) and the distribu-tion shape of lidar returns (i.e., skewness and kurtosis) provided evidence of young, even-aged structure once new overstory canopies had formed in the second decade after fire. The fusion of Landsat time-series and airborne lidar data provided powerful means for assessing changes in forest structure following disturbance over this large forested area. As field measurements of post-fire structure are lacking across the unmanaged forests of the Boreal Shield West, this chap-ter provided critical knowledge on the range of structural responses that occur following high severity fire. While this chapter provided a strong characterization of early stand development for one ecozone, forest regeneration will also vary regionally across the boreal due to gradients in forest productivity. Chapter 6: How does early stand development vary along gradients of productivity? Approach: I related structural measurements from lidar data to GPP estimates from MODIS along a 25-year chronosequence of forest regeneration following fire. Over 400 patches that 138  burned from 1985–2009 were analyzed along the entire transects, with fire information obtained from a national Landsat-derived record of forest change. As the goal of this chapter was to char-acterize regional variability in forest regeneration, not local variability, burned patches were not pre-stratified into open or dense forest as in Chapter 5.   Main findings: In the first 15 years since fire (YSF), estimates of percent canopy cover (> 2m) were typically low regardless of GPP (mean = 11.0–16.0%, sd = 7.8-8.9%) and correlations to GPP were relatively weak (r = 0.18–0.48), as insufficient time had passed for most overstory canopies to form. Canopy cover was more variable between stands by 16–20 YSF (sd = 16.0%) as forest canopies began to close in higher productivity stands (GPP > 0.7 KgCm2yr-1), leading to strong correlations between canopy cover and GPP at 16–20 YSF (r = 0.63, p < 0.01) and 21–25 YSF (r = 0.71, p<0.01). While canopy cover was highly variable between stands at 16–25 YSF (mean = 16.2 – 21.7 %, sd = 16.0–17.1 %), variability in stand height (75th height percentile) was low (mean = 4.9 – 5.0 m, sd = 0.9–1.1 m) and weakly related to GPP (r = 0.16–0.21). My results suggest that canopy cover estimates exhibit more variability along gradients of productivity in early successional stands (YSF < 25), as canopy cover estimates capture differences in the densi-ty of tree establishment, while variability in vertical growth is less pronounced due to slow bore-al growth rates. Further, these results demonstrate that satellite-derived estimates of productivity can inform on variability in forest regeneration; however, significant variation remains unex-plained by these coarse-resolution estimates.  139   Lessons learned for monitoring forest structure over large areas 7.3This dissertation provided a number of key insights for future efforts to monitor forest structure across boreal forests. Each of these insights is reviewed below.   Pre-disturbance spectral information can provide expectations for regenera-7.3.1tion Pre-disturbance stand composition and structure are strong indicators of forest regeneration fol-lowing fire (Epting and Verbyla, 2005; Harper et al., 2005; Boucher et al., 2006; Johnstone and Chapin, 2006; Greene et al., 2007; Chen et al., 2009). However, direct measurements of pre-disturbance stand composition and structure are not typically available, as was the case in this dissertation. To address this information gap,  I demonstrated in Chapter 5 that pre-fire Landsat imagery can inform on expectations for regeneration, as patches classified as dense forest (> 50 % canopy cover) prior to burning displayed faster growth than patches classified as open forest (20 – 50 % canopy cover). Therefore, when direct measurements are not available, the extensive Landsat archive should be utilized to elucidate on pre-fire conditions and expected growth, al-lowing for more spatially explicit predictions of stand development and carbon uptake following disturbance.  Optical measurements of recovery may not tell the whole story 7.3.2Optical measures of recovery following disturbance remain critically important, as these measures are spatially continuous. However, it has been well documented that optical measures of recovery are more sensitive to canopy infilling than vertical height gains (Wulder et al., 1996; Duncanson et al., 2010b; Goetz and Dubayah, 2011).  In Chapter 6, I demonstrated that stands were capable of refilling quickly following fire (high canopy cover within 25 YSF), but the 140  height of these stands remained low relative to stands with no record of disturbance. Therefore, while optical indices may return to pre-disturbance values once canopies reach crown closure, large differences in the vertical structure of stands, and subsequently carbon storage, will remain.  Following stand replacing disturbance, Pickell et al. (2016) demonstrated how rapidly some Landsat vegetation indices return to pre-disturbance conditions from selected locations across the Canadian boreal.  For instance, when applying the normalized difference vegetation index (NDVI), Pickell et al. (2016) found that 93.4% of disturbed pixels recovered within five years (i.e., pixels reached at least 80% of pre-disturbance NDVI). Similar results were found using the normalized burn ratio (77% of pixels recovered in 5 years).  At a coarser spatial-resolution, Hicke et al. (2003) found that Net Primary Productivity (NPP) returned to pre-disturbance values in approximately 9 years after fire in the boreal using 8-km estimates of NPP from the AVHRR. Using direct indicators of structure, I demonstrated in Chapter 5 that stands remained roughly half as tall at 21 – 25 YSF (4.9 m for dense, 4.2  m for open) compared to stands with no record of disturbance (9.8 m for dense, 7.7 m for open), and canopy cover also remained significantly lower (p < 0.001). Therefore, samples of structure derived from lidar remain a critical source of information for understanding optically-derived estimates of recovery and productivity, and the limitations of optical remote sensing data.  Do not mistake residual structure for recovery following fire 7.3.3In chapter 5, I demonstrated that stand height estimates are often higher than expected in the first decade following fire due to the presence of residual structures in the canopy. Additionally, can-opy cover estimates were higher in the first five YSF compared to 6 – 15 YSF, likely due to the 141  loss of residual structures through time as snags begin to fall. The presence of residual structures, and the impact these structures have on lidar metrics, is therefore critical to consider when as-sessing carbon loss in disturbance and carbon gained in regeneration. If lidar data is collected in the first five YSF and the presence of residual structures are not acknowledged, the impact of the fire on biomass loss may be underestimated, or the recovery of regenerating stands overestimat-ed. Residual structures can also impact attempts to model aboveground biomass empirically. The results of Chapter 6 demonstrated that many recently burned stands (1 – 10 YSF) were structur-ally unique, as these stands had low canopy cover estimates (< 40%) but relatively tall height estimates (> 7 m). As many empirical models to predict aboveground biomass from lidar rely heavily on height percentiles (e.g., Næsset, 2004; Wulder et al., 2012b), biomass estimates may be overestimated when height percentiles are dominated by residual structures. Therefore, these structurally unique stands must be included in model development to derive accurate above-ground biomass predictions in recently burned stands.   Limitations 7.4 Limitations of transect data 7.4.1The lidar transects used in this dissertation provide unprecedented insights into structural varia-bility across the Canadian boreal, and significant effort was involved in the planning and execu-tion of these transects by members of both CFS and C-CLEAR. However, as these transects were not designed to capture a representative sample of forest structure across the boreal, due to limi-tations imposed by suitable airports and flight times, these transects are not ideally-suited to de-142  scribe the full range of variability in boreal forest structure or to characterize the full range of structural drivers. For example, in Chapter 4, I reported that the maximum height that stands reached was restricted in low productivity sites (GPP < 0.6 kgC m-2yr-1) in the Boreal Shield East compared to higher productivity forests. However, only 5.7% of the sampled MODIS cells in the ecozone had a GPP value < 0.6 kgC m-2yr-1. It is possible that with increased sampling in low productivity forests, taller stands would have been sampled. Equal sampling across a range of productivities would therefore improve our understanding of structural variability within produc-tivity groups.   Further, the swath width of transect data can influence comparisons to coarse spatial resolution remote sensing data. Specifically, the swath width of the products generated from the lidar tran-sects (400m) were narrower than a single MODIS cell (1 km), preventing the structure across entire MODIS cells from being measured. In Chapter 4, the average MODIS cell contained 461 lidar plots, which accounts for approximately 29% of the area of a single MODIS cell. Therefore, the forest stands sampled with lidar may not accurately represent the productivity of an entire cell. This should not be a major issue, as any MODIS cell that was less than 75% forested was removed. It is assumed that variations in productivity are minimal within each 1-km cell, which may not always be the case as differences in nutrient and water availability as well as varying species and stages of succession can occur at finer scales.  Unmeasured factors that influence forest structure 7.4.2This dissertation demonstrated that satellite-derived estimates of productivity and disturbance provide valuable information on structural variability. However, unexplained variance in struc-143  ture remains due to a range of factors that are currently infeasible to measure wall-to-wall, such as stand age (if the stand was not disturbed during the Landsat record), soil conditions (Certini, 2005; Johnstone and Chapin, 2006; Lecomte et al., 2006a,b), nutrient availability (Bhatti et al., 2002; Payette and Delwaide, 2003), species composition (Johnstone et al., 2004; Mack et al., 2008), and the influence of additional disturbance types such as insect and disease (Chen and Popadiouk, 2002; Boulanger and Arseneault, 2004; Brassard and Chen, 2006). In particular, it is critically important to account for stand age in order to assess other drivers of forest structure, as differences in stand age will likely outweigh most other drivers, such as differences in site condi-tions or growing season length. In chapter 4, stand age was accounted for by removing areas of known disturbance, but a wide-range of age, and therefore stand height, remained within narrow ranges of productivity.  I was able to accurately control for stand age in Chapters 5 and 6 by limiting the analysis to stands that burned during the relevant portion of the Landsat TM/ETM+ record (1985 – 2009). However, due to the slow growth rates of boreal stands, variability in stand height remained low between patches even at the end of the chronosequence, preventing a detailed analysis of the varying drivers of vertical structure. It is expected that significant variability in vertical structure would be observable within 25 YSF in faster growing temperate and tropical ecosystems, but a longer chronosequence is needed in boreal environments to assess a range of structural drivers.   Lack indicators of carbon stored in soils, deadwood, litter, and wetlands 7.4.3This dissertation focused primarily on estimates of canopy cover and stand height, which are im-portant indicators of carbon storage in aboveground biomass. While boreal forests contain 32% 144  of the carbon stored in the terrestrial biosphere, only an estimated 20% of that is stored in bio-mass (above- and below-ground), with the rest stored in soils, deadwood, and litter (Pan et al., 2011). Therefore, characterizing temporal and spatial variability in forest structure represents only a fraction of the information needed to monitor boreal carbon budgets. However, as above-ground biomass is one of the most dynamic stores of carbon in the boreal due to frequent dis-turbances, monitoring variability in aboveground biomass remains critically important (Goetz and Dubayah, 2011; Kurz et al., 2013).  Further, forest structure was not assessed in wetland ecosystems in this dissertation, as the ability to assess vertical structure with lidar in wetlands is prone to errors, due to the weak pulse returns associated with water saturated surfaces (Hopkinson et al., 2005). As wetlands make up an esti-mated 14% of Canada’s land area (Price and Waddington, 2000), carbon storage in wetlands re-mains critical to characterize with alternative approaches (Vitt et al., 2000).   Directions for future research 7.5 Planning future large-area lidar transect collection 7.5.1Several spaceborne lidar missions are planned over the next few years. The Ice, Cloud and land Elevation Satellite-2 (ICESat-2) will be launched in 2017, and will collect a sample of lidar data globally. However, as these data are designed to measure snow and ice, data quality for charac-terizing vegetation will be not be optimal due to low laser power (Escobar and Brown, 2014), and these data have been shown in a simulation study to be particularly limited in low biomass boreal environments (Montesano et al., 2015). Alternatively, a spaceborne lidar designed to measure forests, called Global Ecosystem Dynamics Investigation (GEDI), will be mounted on 145  the International Space Station in 2018, and will collect a sample of lidar data for large areas of the world’s forests (Dubayah et al., 2014). However, as data collection will be limited to be-tween 52° S and 52° N latitude, most boreal forests will not be sampled by this mission (Patterson and Healey, 2015). Therefore, as in the past, structural information on northern boreal forests will be lacking. Without data from the GEDI mission, lidar transect collection will remain one of the best ways to provide a representative sample of structural variability across boreal en-vironments.   Based on the findings of this dissertation, future collection efforts should utilize ancillary spatial layers to design flights paths in order to ensure a range of forest conditions are captured. First, each boreal ecozone should be pre-stratified based on satellite-derived productivity estimates, Landsat-derived disturbance and recovery information, historical fire data, and topographic vari-ables (e.g., elevation, slope, aspect). Second, flight lines between suitable airports should be de-signed to adequately sample the variability in each of these spatial layers. As suitable airports and flight times are restrictive in unmanaged boreal forests, sampling the full range of variability will not be feasible in many cases, but efforts can be made to sample the greatest range of varia-bility possible. While transect data collection is limited by suitable airports and flight times, it is important to note that these factors influence the collection of field-based measurements as well. Therefore, field-based measurements are often more limited than transects in their ability to characterize the full range of forest conditions.  With the transects used in this dissertation, sampling was often too sparse within individual land-scapes to disentangle local and regional drivers of structure, as slight deviations in flight lines 146  could result in different age classes or forest types sampled within a landscape. If researchers want to characterize both local and regional drivers of forest structure, the transects should be designed to capture variability at multiple scales. For example, representative landscapes could be identified across the boreal using ancillary spatial layers, and each chosen landscape could be sampled densely with lidar transects. With this collection method, local variability in structure could be assessed within each collection, while regional drivers could be assessed between each collection. From an analysis such as this, the importance of individual drivers of structure could be compared and contrasted across a range of boreal landscapes.  Continued need for wall-to-wall estimates of structure 7.5.2As significant time and resources would be required to provide wall-to-wall estimates of forest structure with lidar over continental scales, with airborne or spaceborne data, we will continue to rely on extrapolating lidar data using spatially continuous variables. As demonstrated in Chapter 3, existing wall-to-wall estimates of canopy height can describe average conditions across land-scapes, but this may be unfit for monitoring changes in carbon storage that occur at finer scales (i.e., stand-level changes). Alternatively, the spatial-resolution of Landsat data (30 m) is well suited to capture stand-level variability in forest conditions over large study areas, and can be used to extrapolate lidar structural information (Pflugmacher et al., 2012; Zald et al., 2016). While this dissertation used pre-disturbance spectral information from Landsat to elucidate on post-disturbance structure, spectral indicators of recovery following disturbance are also a valua-ble source for informing on variability in post-fire regeneration. However, as Landsat lacks in-formation on three-dimensional forest structure, actual structural measures from lidar will remain crucial to understanding what information Landsat spectral data provides on structure and carbon 147  storage. Current and future research is focused on using a suite of Landsat-derived spectral met-rics to extrapolate lidar-derived structural information wall-to-wall across Canada’s forested ecozones (e.g., building upon Zald et al., 2016). This work will represent a significant advance-ment in our understanding of structural variability across Canadian boreal landscapes, and reduce uncertainties in carbon storage and sequestration estimates. However, many challenges will re-main in order to achieve highly accurate predictions of height and biomass, as Landsat spectral measures are known to saturate when canopies close (Wulder et al., 1996; Duncanson et al., 2010b; Avitabile et al., 2011, 2012). Therefore, biomass gains beyond canopy closure may re-main difficult to detect (Houghton et al., 2009), signifying the continued importance of lidar structural metrics and productivity estimates to predict growth beyond canopy closure.   Incorporation of Landsat MSS data to extend chronosequence of structure 7.5.3While the Landsat data record extends back to 1972, most analyses of forest change, including this dissertation, only use data from the Landsat TM record and on (1984 – present), as these da-ta have narrower band passes, higher spatial resolution  for detecting disturbances (Wulder et al., 2012a), and analysis ready products exist (e.g., surface reflectance and cloud screening). In chap-ter 6, I demonstrated that strong variability in canopy cover between boreal patches is not detect-able until the second decade after fire, while variability in stand height remained low through the end of the chronosequence (25 YSF). If the chronosequence is extended from 25- to 37-years by incorporating Landsat MSS data, additional insights into successional pathways of forest stands may be gained, and more stages of stand succession may be observable.  Additionally, subtle dif-ferences in growth due to factors such as nutrient availability or site position in the landscape (e.g., slope and aspect) may become more pronounced if stands have grown for a longer period 148  of time, allowing for a more detailed analysis of a range of structural drivers. Currently, re-searchers are working to provide analysis ready MSS data that can be used to bridge the gap be-tween MSS data and the TM/ETM+ era, allowing for the analysis of change through the entire Landsat record (e.g., Braaten et al., 2015). While the incorporation of Landsat MSS data will al-low time since stand-replacing disturbance to be determined from 1972 until the present, these data will remain limited in their ability to inform on spectral trends through time due to limita-tions imposed by data quality, spectral bands, and spatial resolution. However, by combining MSS data with lidar structural metrics, and extending the chronosequence back to 1972, more graduated and detailed characterizations of structural development following disturbance can be derived.   149  References Amiro, B.D., Cantin, A., Flannigan, M.D. and de Groot, W.J., 2009. Future emissions from Canadian boreal forest fires. Can. J. For. Res. 39, 383–395. Amiro, B.D., Stocks, B.J., Alexander, M.E., Flannigan, M.D. and Wotton, B.M., 2001. Fire, climate change, carbon and fuel management in the Canadian boreal forest. Int. J. Wildl. Fire. 10, 405–413. Andersen, H.E., 2009. Using airborne light detection and ranging (LIDAR) to characterize forest stand condition on the Kenai Peninsula of Alaska. West. J. Appl. For. 24 (2), 95–102. Andrew, M.E., Wulder, M.A. and Coops, N.C., 2012. Identification of de facto protected areas in boreal Canada. Biol. Conserv. 146, 97–107. Angers, V.A., Gauthier, S., Drapeau, P., Jayen, K. and Bergeron, Y., 2011. Tree mortality and snag dynamics in North American boreal tree species after a wildfire: a long-term study. Int. J. Wildl. Fire. 20, 751–763. Arino, O., Gross, D., Ranera, F., Leroy, M., Bicheron, P., Brockman, C., Defourny, C., Achard, F., Durieux, L., Bourg, L., Latham, J., Di Gregorio, A., Herold, M., Sambale, J., Plummer, S. and Weber, J.-L., 2007. GlobCover: ESA service for Global Land Cover from MERIS. 2007 IEEE International Geoscience and Remote Sensing Symposium. pp. 2412–2415. Arseneault, D., 2001. Impact of fire behavior on postfire forest development in a homogeneous boreal landscape. Can. J. For. Res. 31, 1367–1374. Avitabile, V., Baccini, A., Friedl, M.A. and Schmullius, C., 2012. Capabilities and limitations of Landsat and land cover data for aboveground woody biomass estimation of Uganda. Remote Sens. Environ. 117, 366–380. Avitabile, V., Herold, M., Henry, M. and Schmullius, C., 2011. Mapping biomass with remote sensing: a comparison of methods for the case study of Uganda. Carbon Balance Manag. 6 (7), doi:10.1186/1750-0680-6-7. Bartels, S.F., Chen, H.Y.H., Wulder, M.A. and White, J.C., 2016. Trends in post-disturbance recovery rates of Canada’s forests following wildfire and harvest. For. Ecol. Manage. 361, 194–207. Bergeron, Y., Gauthier, S., Flannigan, M. and Kafka, V., 2004. Fire Regimes At the Transition Between Mixedwood and Coniferous Boreal Forest in Northwestern Quebec. Ecology. 85 (7), 1916–1932. Bergeron, Y., Leduc, A., Morin, H. and Claude, J., 1995. Balsam fir mortality following the last spruce budworm outbreak in northwestern Quebec. Can. J. For. Res. 25, 1375 – 1384. Bhatti, J.S., Apps, M.J. and Jiang, H., 2002. Influence of nutrients , disturbances and site conditions on carbon stocks along a boreal forest transect in central Canada. Plant Soil. 242 (1), doi:10.1023/A:1019670619316. 150  Boisvenue, C. and Running, S.W., 2006. Impacts of climate change on natural forest productivity–evidence since the middle of the 20th century. Glob. Chang. Biol. 12 (5), 862–882. Bonan, G.B. and Shugart, H.H., 1989. Environmental factors and ecological processes in boreal forests. Annu. Rev. Ecol. Syst. 20, 1–28. Bond-Lamberty, B. and Gower, S.T., 2008. Decomposition and fragmentation of coarse woody debris: Re-visiting a boreal black spruce chronosequence. Ecosystems. 11, 831–840. Bond-Lamberty, B., Peckham, S.D., Ahl, D.E. and Gower, S.T., 2007. Fire as the dominant driver of central Canadian boreal forest carbon balance. Nature. 450 (7166), 89–92. Bond-Lamberty, B., Rocha, A. V., Calvin, K., Holmes, B., Wang, C. and Goulden, M.L., 2014. Disturbance legacies and climate jointly drive tree growth and mortality in an intensively studied boreal forest. Glob. Chang. Biol. 20 (1), 216–227. Boucher, D., Gauthier, S. and De Grandpré, L., 2006. Structural changes in coniferous stands along a chronosequence and a productivity gradient in the northeastern boreal forest of Québec. Ecoscience. 13 (2), 172–180. Boulanger, Y. and Arseneault, D., 2004. Spruce budworm outbreaks in eastern Quebec over the last 450 years. Can. J. For. Res. 34 (5), 1035–1043. Boulanger, Y. and Sirois, L., 2006. Postfire dynamics of black spruce coarse woody debris in northern boreal forest of Quebec. Can. J. For. Res. 36, 1770–1780. Braaten, J.D., Cohen, W.B. and Yang, Z., 2015. Automated cloud and cloud shadow identification in Landsat MSS imagery for temperate ecosystems. Remote Sens. Environ. 169, 128–138. Bradford, J.B. and Kastendick, D.N., 2010. Age-related patterns of forest complexity and carbon storage in pine and aspen-birch ecosystems of northern Minnesota, USA. Can. J. For. Res. 40, 401–409. Brandt, J.P., Flannigan, M.D., Maynard, D.G., Thompson, I.D. and Volney, W.J.A., 2013. An introduction to Canada’s boreal zone: ecosystem processes, health, sustainability, and environmental issues. Environ. Rev. 21 (4), 207–226. Brandt, J.P.P., 2009. The extent of the North American boreal zone. Environ. Rev. 17 (NA), 101–161. Brassard, B.W. and Chen, H.Y.H., 2006. Stand structural dynamics of North American boreal forests. CRC. Crit. Rev. Plant Sci. 25 (2), 115–137. Brassard, B.W., Chen, H.Y.H., Wang, J.R. and Duinker, P.N., 2008. Effects of time since stand-replacing fire and overstory composition on live-tree structural diversity in the boreal forest of central Canada. Can. J. For. Res. 38, 52–62. Burton, P.J., Parisien, M., Hicke, J.A., Hall, R.J. and Freeburn, J.T., 2008. Large fires as agents 151  of ecological diversity in the North American boreal forest. Int. J. Wildl. Fire. 17, 754–767. Carleton, T.J. and Wannamakerf, B.A., 1987. Mortality and self-thinning in postfire black spruce. 59, 621–628. Certini, G., 2005. Effects of fire on properties of forest soils: A review. Oecologia. 143 (1), doi:10.1007/s00442-004-1788-8. Chen, H.Y.H. and Popadiouk, R. V, 2002. Dynamics of North American boreal mixedwoods. Environ. Rev. 10, 137–166. Chen, H.Y.H., Vasiliauskas, S., Kayahara, G.J. and Ilisson, T., 2009. Wildfire promotes broadleaves and species mixture in boreal forest. For. Ecol. Manage. 257 (1), 343–350. Cheng, R. and Lee, P., 2008. Urban Sprawl and Other Major Land Use Conversions in Ontario’s Greenbelt From 1993 to 2007. Global Forest Watch Canada, Edmonton, Alberta. http://www.globalforestwatch.ca/publications/20080702A. Churkina, G. and Running, S.W., 1998. Contrasting climatic controls on the estimated productivity of global terrestrial biomes. Ecosystems. 1, 206–215. Clifford, P., Richardson, S. and Hémon, D., 1989. Assessing the significance of the correlation between two spatial processes. Biometrics. 45 (1), 123–134. Cohen, W.B., Spies, T.A., Alig, R.J., Oetter, D.R., Maiersperger, T.K. and Fiorella, M., 2002. Characterizing 23 years (1972-95) of stand replacement disturbance in western Oregon forests with Landsat imagery. Ecosystems. 5, 122–137. Crist, E.P., 1985. A TM Tasseled Cap Equivalent Transformation for Reflectance Factor Data. Remote Sens. Environ. 17, 301–306. Dubayah, R., Goetz, S., Blair, J.B., Luthcke, S., Healey, S., Hansen, M., Hofton, M., Hurtt, G., Kellner, J., Fatoyinbo, T., Swatantran, A. and Papathanassiou, K., 2014. The Global Ecosystem Dynamics Investigation (GEDI) Lidar. ForestSAT2014 open conference system. Dubayah, R.O. and Drake, J.B., 2000. Lidar remote sensing for forestry. J. For. 98 (6), 44–46. Duncanson, L.I., Niemann, K.O. and Wulder, M.A., 2010a. Estimating forest canopy height and terrain relief from GLAS waveform metrics. Remote Sens. Environ. 114, 138–154. Duncanson, L.I., Niemann, K.O. and Wulder, M.A., 2010b. Integration of GLAS and Landsat TM data for aboveground biomass estimation. Can. J. Remote Sens. 36 (2), 129–141. Dutilleul, P., Clifford, P., Richardson, S. and Hemon, D., 1993. Modifying the t test for assessing the correlation between two spatial processes. Biometrics. 49 (1), 305–314. Ecological Stratification Working Group, 1995. A National Ecological Framework for Canada. Centre for Land and Biological Resources Research, Research Branch, Agriculture and Agri-Food Canada. http://sis.agr.gc.ca/cansis/publications/ecostrat/cad_report.pdf. Epting, J. and Verbyla, D., 2005. Landscape-level interactions of prefire vegetation, burn 152  severity, and postfire vegetation over a 16-year period in interior Alaska. Can. J. For. Res. 35, 1367–1377. Epting, J., Verbyla, D. and Sorbel, B., 2005. Evaluation of remotely sensed indices for assessing burn severity in interior Alaska using Landsat TM and ETM+. Remote Sens. Environ. 96, 328–339. Escobar, V. and Brown, M., 2014. ICESat--2 Applications Vegetation Tutorial with Landsat 8 Report. Ann Arbor, Michigan. http://icesat.gsfc.nasa.gov/icesat2/applications/ICESat2_Vegetation_Report_FINAL.pdf. van Ewijk, K.Y., Treitz, P.M. and Scott, N.A., 2011. Characterizing Forest Succession in Central Ontario using Lidar-derived Indices. Photogramm. Eng. Remote Sensing. 77 (3), 261–269. Falkowski, M.J., Wulder, M.A., White, J.C. and Gillis, M.D., 2009. Supporting large-area, sample-based forest inventories with very high spatial resolution satellite imagery. Prog. Phys. Geogr. 33 (3), 403–423. FAO, 2006. Global Forest Resources Assessment 2005: Progress towards sustainable forest management. Rome. ftp://ftp.fao.org/docrep/fao/008/A0400E/A0400E00.pdf. Farrar, J.L., 1995. Trees in Canada. Fitzhenry and Whiteside Ltd., Natural Resources Canada, Canada Forest Service, Ottawa, Ontario. Flannigan, M.D., Logan, K.A., Amiro, B.D., Skinner, W.R. and Stocks, B.J., 2005. Future area burned in canada. Clim. Change. 72, doi:10.1007/s10584-005-5935-y. Fluet-Chouinard, E., Lehner, B., Rebelo, L.-M., Papa, F. and Hamilton, S.K., 2015. Development of a global inundation map at high spatial resolution from topographic downscaling of coarse-scale remote sensing data. Remote Sens. Environ. 158, 348–361. Fortin, M. and Payette, S., 2002. How to test the significance of the relation between spatially autocorrelated data at the landscape scale : A case study using fire and forest maps. Ecoscience. 9 (2), 213–218. Fortin, M.-J., 1999. Effects of sampling unit resolution on the estimation of spatial autocorrelation. Ecoscience. 6 (4), 636–641. Friedl, M.A. and Brodley, C.E., 1997. Decision tree classification of land cover from remotely sensed data. Remote Sens. Environ. 61, 399 – 409. Friedl, M.A., Sulla-Menashe, D., Tan, B., Schneider, A., Ramankutty, N., Sibley, A. and Huang, X., 2010. MODIS Collection 5 global land cover: Algorithm refinements and characterization of new datasets. Remote Sens. Environ. 114 (1), 168–182. Gao, B.-C., 1996. NDWI - A normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sens. Environ. 58, 257–266. Gauthier, S., Bernier, P., Burton, P.J., Edwards, J., Isaac, K., Isabel, N., Jayen, K., Le Goff, H. and Nelson, E.A., 2014. Climate change vulnerability and adaptation in the managed 153  Canadian boreal forest. Environ. Rev. 22, 256 – 285. Gauthier, S., Bernier, P., Kuuluvainen, T., Shvidenko, A.Z. and Schepaschenko, D.G., 2015. Boreal forest health and global change. Science. 349, 819–822. Gillis, M.D., Omule, A.Y. and Brierley, T., 2005. Monitoring Canada’s forests: the National Forest Inventory. For. Chron. 81 (2), 214–221. Goetz, S. and Dubayah, R., 2011. Advances in remote sensing technology and implications for measuring and monitoring forest carbon stocks and change. Carbon Manag. 2 (3), 231–244. Goetz, S.J., Sun, M., Baccini, A. and Beck, P.S.A., 2010. Synergistic use of spaceborne lidar and optical imagery for assessing forest disturbance: An Alaska case study. J. Geophys. Res. 115, doi:10.1029/2008JG000898. Goodwin, N.R. and Collett, L.J., 2014. Development of an automated method for mapping fire history captured in Landsat TM and ETM+ time series across Queensland, Australia. Remote Sens. Environ. 148, 206–221. Gralewicz, N.J., Nelson, T.A. and Wulder, M.A., 2012. Factors influencing national scale wildfire susceptibility in Canada. For. Ecol. Manage. 265, 20–29. Greene, D.F., Macdonald, S.E., Haeussler, S., Domenicano, S., Noël, J., Jayen, K., Charron, I., Gauthier, S., Hunt, S., Gielau, E.T., Bergeron, Y. and Swift, L., 2007. The reduction of organic-layer depth by wildfire in the North American boreal forest and its effect on tree recruitment by seed. Can. J. For. Res. 37, 1012–1023. Greene, D.F., Noël, J., Bergeron, Y., Rousseau, M. and Gauthier, S., 2004. Recruitment of Picea mariana , Pinus banksiana , and Populus tremuloides across a burn severity gradient following wildfire in the southern boreal forest of Quebec. Can. J. For. Res. 34, 1845–1857. Griffiths, P., van der Linden, S., Kuemmerle, T. and Hostert, P., 2013. A pixel-based Landsat vompositing algorithm for large area land cover mapping. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 6 (5), 2088–2101. de Groot, W.J., Cantin, A.S., Flannigan, M.D., Soja, A.J., Gowman, L.M. and Newbery, A., 2013. Forest Ecology and Management A comparison of Canadian and Russian boreal forest fire regimes. For. Ecol. Manage. 294, 23–34. Groot, W.J. De, Flannigan, M.D. and Cantin, A.S., 2013. Forest Ecology and Management Climate change impacts on future boreal fire regimes q. 294, 35–44. Hall, R.J., Freeburn, J.T., de Groot, W.J., Pritchard, J.M., Lynham, T.J. and Landry, R., 2008. Remote sensing of burn severity: experience from western Canada boreal fires. Int. J. Wildl. Fire. 17, 476–489. Hansen, A.J., Phillips, L.B., Dubayah, R., Goetz, S. and Hofton, M., 2014. Regional-scale application of lidar: Variation in forest canopy structure across the southeastern US. For. Ecol. Manage. 329, 214–226. 154  Hansen, M.C. and Loveland, T.R., 2012. A review of large area monitoring of land cover change using Landsat data. Remote Sens. Environ. 122, 66–74. Harding, D.J. and Carabajal, C.C., 2005. ICESat waveform measurements of within-footprint topographic relief and vegetation vertical structure. Geophys. Res. Lett. 32 (L21S10), doi:10.1029/2005GL023471. Harper, K.A., Bergeron, Y., Drapeau, P., Gauthier, S. and De Grandpré, L., 2005. Structural development following fire in black spruce boreal forest. For. Ecol. Manage. 206 (1-3), 293–306. Harper, K.A., Bergeron, Y., Gauthier, S. and Drapeau, P., 2002. Post-Fire Development of Canopy Structure and Composition in Black Spruce Forests of Abitibi , Québec : A Landscape Scale Study. Silva Fenn. 36 (1), 249 – 263. Heinsch, F.A., Running, S.W., Kimball, J.S., Nemani, R.R., Davis, K.J., Bolstad, P. V, Cook, B.D., Desai, A.R., Ricciuto, D.M., Law, B.E., Oechel, W.C., Wofsy, S.C., Dunn, A.L., Munger, J.W., Baldocchi, D.D., Hollinger, D.Y., Richardson, A.D., Stoy, P.C., Siqueira, M.B.S., Monson, R.K., Burns, S.P. and Flanagan, L.B., 2006. Evaluation of remote sensing based terrestrial productivity from MODIS using regional tower eddy flux network observations. IEEE Trans. Geosci. Remote Sens. 44 (7), 1908–1925. Heinselman, M.L., 1981. Fire and Succession in the Conifer Forests of Northern North America. Forest Succesion: Concepts and Application. pp. 374–405. Hermosilla, T., Wulder, M., White, J.C., Coops, N. and Hobart, G.W., 2015a. An integrated Landsat time series protocol for change detection and generation of annual gap-free surface reflectance composites. Remote Sens. Environ. 158, 220–234. Hermosilla, T., Wulder, M.A., White, J.C., Coops, N.C. and Hobart, G.W., 2015b. Regional detection, characterization, and attribution of annual forest change from 1984 to 2012 using Landsat-derived time-series metrics. Remote Sens. Environ. 170, 121–132. Hermosilla, T., Wulder, M.A., White, J.C., Coops, N.C., Hobart, G.W. and Campbell, L.B., 2016. Mass data processing of time series Landsat imagery: pixels to data products for forest monitoring. Int. J. Digit. Earth., doi:10.1080/17538947.2016.1187673. Hicke, J.A., Asner, G.P., Kasischke, E.S., French, N.H.F., Randerson, J.T., Collatz, G.J., Stocks, B.J., Tucker, C.J., Los, S.O. and Field, C.B., 2003. Postfire response of North American boreal forest net primary productivity analyzed with satellite observations. Glob. Chang. Biol. 9 (8), 1145–1157. Hopkinson, C., Chasmer, L.E., Sass, G., Creed, I.F., Sitar, M., Kalbfleisch, W. and Treitz, P., 2005. Vegetation class dependent errors in lidar ground elevation and canopy height estimates in a boreal wetland environment. Can. J. Remote Sens. 31 (2), 191–206. Hopkinson, C., Wulder, M.A., Coops, N.C., Milne, T., Fox, A. and Bater, C.W., 2011. Airborne lidar sampling of the Canadian boreal forest: Planning, execution, and initial processing. Proceedings of the SilviLaser 2011 Conference. Hobart, Australia. 155  Houghton, R.A., Hall, F. and Goetz, S.J., 2009. Importance of biomass in the global carbon cycle. J. Geophys. Res. 114, doi:10.1029/2009JG000935. Huang, C., Goward, S.N., Masek, J.G., Thomas, N., Zhu, Z. and Vogelmann, J.E., 2010a. An automated approach for reconstructing recent forest disturbance history using dense Landsat time series stacks. Remote Sens. Environ. 114, 183–198. Huang, C., Thomas, N., Goward, S.N., Masek, J.G., Zhu, Z. and Townshend, J.R.G., 2010b. Automated masking of cloud and cloud shadow for forest change analysis using Landsat images. Int. J. Remote Sens. 31 (20), 5449–5464. Jennings, S.B., Brown, N.D. and Sheil, D., 1999. Assessing forest canopies and understorey illumination: canopy closure, canopy cover and other measures. Forestry. 72 (1), 59–73. Johnstone, J.F., Chapin III, F.S., Foote, J., Kemmett, S., Price, K. and Viereck, L., 2004. Decadal observations of tree regeneration following fire in boreal forests. Can. J. For. Res. 34, 267–273. Johnstone, J.F. and Chapin, F.S., 2006. Effects of soil burn severity on post-fire tree recruitment in boreal forest. Ecosystems. 9, 14–31. Johnstone, J.F., Hollingsworth, T.N., Chapin III, F.S. and Mack, M.C., 2010. Changes in fire regime break the legacy lock on successional trajectories in Alaskan boreal forest. Glob. Chang. Biol. 16, 1281–1295. Kane, V.R., Bakker, J.D., McGaughey, R.J., Lutz, J.A., Gersonde, R.F. and Franklin, J.F., 2010a. Examining conifer canopy structural complexity across forest ages and elevations with LiDAR data. Can. J. For. Res. 40, 774–787. Kane, V.R., Gersonde, R.F., Lutz, J.A., McGaughey, R.J., Bakker, J.D. and Franklin, J.F., 2011. Patch dynamics and the development of structural and spatial heterogeneity in Pacific Northwest forests. Can. J. For. Res. 41, 2276–2291. Kane, V.R., Lutz, J.A., Roberts, S.L., Smith, D.F., McGaughey, R.J., Povak, N.A. and Brooks, M.L., 2013. Landscape-scale effects of fire severity on mixed-conifer and red fir forest structure in Yosemite National Park. For. Ecol. Manage. 287, 17–31. Kane, V.R., McGaughey, R.J., Bakker, J.D., Gersonde, R.F., Lutz, J.A. and Franklin, J.F., 2010b. Comparisons between field- and LiDAR-based measures of stand structural complexity. Can. J. For. Res. 40, 761–773. Kane, V.R., North, M.P., Lutz, J.A., Churchill, D.J., Roberts, S.L., Smith, D.F., McGaughey, R.J., Kane, J.T. and Brooks, M.L., 2014. Assessing fire effects on forest spatial structure using a fusion of Landsat and airborne LiDAR data in Yosemite National Park. Remote Sens. Environ. 151, 89–101. Kasischke, E.S., Christensen, N.L. and Stocks, B.J., 1995. Fire , global warming , and the carbon balance of boreal forests. Ecol. Appl. 5 (2), 437–451. Kennedy, R.E., Yang, Z. and Cohen, W.B., 2010. Detecting trends in forest disturbance and 156  recovery using yearly Landsat time series : 1 . LandTrendr — Temporal segmentation algorithms. Remote Sens. Environ. 114, 2897–2910. Kneeshaw, D. and Gauthier, S., 2003. Old growth in the boreal forest: A dynamic perspective at the stand and landscape level. Environ. Rev. 11, S99–S114. Kurz, W.A. and Apps, M.J., 1999. A 70-year retrospective analysis of carbon fluxes in the Canadian forest sector. Ecol. Appl. 9 (2), 526–547. Kurz, W.A., Dymond, C.C., White, T.M., Stinson, G., Shaw, C.H., Rampley, G.J., Smyth, C., Simpson, B.N., Neilson, E.T., Trofymow, J.A., J, M., J, A.M., Metsaranta, J. and Apps, M.J., 2009. CBM-CFS3: A model of carbon-dynamics in forestry and land-use change implementing IPCC standards. Ecol. Modell. 220, 480–504. Kurz, W.A., Shaw, C.H., Boisvenue, C., Stinson, G., Metsaranta, J., Leckie, D., Dyk, A., Smyth, C. and Neilson, E.T., 2013. Carbon in Canada ’ s boreal forest — A synthesis. Environ. Rev. 21, 260–292. Larson, A.J., Lutz, J.A., Gersonde, R.F., Franklin, J.F. and Hietpasi, F.F., 2008. Potential site productivity influences the rate of forest structural development. Ecol. Appl. 18 (4), 899–910. Lecomte, N., Simard, M. and Bergeron, Y., 2006a. Effects of fire severity and initial tree composition on stand structural development in the coniferous boreal forest of northwestern Québec, Canada. Ecoscience. 13 (2), 152–163. Lecomte, N., Simard, M., Fenton, N. and Bergeron, Y., 2006b. Fire severity and long-term ecosystem biomass dynamics in coniferous boreal forests of eastern Canada. Ecosystems. 9 (8), 1215–1230. Lee, P. and Gysbers, J.D., 2008. Recent Anthropogenic Changes within the Inland Temperate Rainforest of British Columbia: Interim Report. Global Forest Watch Canada, Edmonton, Alberta. Lefsky, M.A., 2010. A global forest canopy height map from the Moderate Resolution Imaging Spectroradiometer and the Geoscience Laser Altimeter System. Geophys. Res. Lett. 37 (L15401), doi:10.1029/2010GL043622. Lefsky, M.A., Cohen, W.B., Parker, G.G. and Harding, D.J., 2002. Lidar remote sensing for ecosystem studies. Bioscience. 52 (1), 19–30. Lefsky, M.A., Hudak, A.T., Cohen, W.B. and Acker, S.A., 2005a. Patterns of covariance between forest stand and canopy structure in the Pacific Northwest. Remote Sens. Environ. 95, 517–531. Lefsky, M.A., Keller, M., Pang, Y., de Camargo, P.B. and Hunter, M.O., 2007. Revised method for forest canopy height estimation from Geoscience Laser Altimeter System waveforms. J. Appl. Remote Sens. 1 (013537), doi:10.1117/1.2795724. Lefsky, M.A., Turner, D.P., Guzy, M. and Cohen, W.B., 2005b. Combining lidar estimates of 157  aboveground biomass and Landsat estimates of stand age for spatially extensive validation of modeled forest productivity. Remote Sens. Environ. 95, 549–558. Lentile, L.B., Holden, Z.A., Smith, A.M.S., Falkowski, M.J., Hudak, A.T., Morgan, P., Lewis, S.A., Gessler, P.E. and Benson, N.C., 2006. Remote sensing techniques to assess active fire characteristics and post-fire effects. Int. J. Wildl. Fire. 15, 319–345. Lieffers, V.J., Messier, C., Burton, P.J., Ruel, J.-C. and Grover, B.E., 2003. Nature-based silviculture of sustaining a variety of boreal forest values. Towards Sustainable Management of the Boreal Forest.(ed. by P.J. Burton,C. Messier,D.W. Smith,and W.L. Adamowicz). pp. 481–550. N.R.C. Press, Ottawa, Canada. Lim, K., Treitz, P., Wulder, M.A., St-Onge, B. and Flood, M., 2003. LiDAR remote sensing of forest structure. Prog. Phys. Geogr. 27 (1), 88–106. López García, M.J. and Caselles, V., 1991. Mapping burns and natural reforestation using thematic Mapper data. Geocarto Int. 6 (1), 31–37. Loveland, T.R. and Dwyer, J.L., 2012. Landsat: Building a strong future. Remote Sens. Environ. 122, 22–29. Mack, M.C., Treseder, K.K., Manies, K.L., Harden, J.W., Schuur, E. a. G., Vogel, J.G., Randerson, J.T. and Chapin, F.S., 2008. Recovery of Aboveground Plant Biomass and Productivity After Fire in Mesic and Dry Black Spruce Forests of Interior Alaska. Ecosystems. 11, 209–225. Magnussen, S. and Boudewyn, P., 1998. Derivations of stand heights from airborne laser scanner data with canopy-based quantile estimators. Can. J. For. Res. 28, 1016–1031. Magnussen, S. and Wulder, M.A., 2012. Post-fire canopy height recovery in Canada’s boreal forests using Airborne Laser Scanner (ALS). Remote Sens. 4, 1600–1616. Mansuy, N., Gauthier, S., Robitaille, A.A. and Bergeron, Y., 2012. Regional patterns of postfire canopy recovery in the northern boreal forest of Quebec: interactions between surficial deposit, climate, and fire cycle. Can. J. For. Res. 42 (7), 1328–1343. Masek, J.G., Vermote, E.F., Saleous, N.E., Wolfe, R., Hall, F.G., Huemmrich, K.F., Gao, F., Kutler, J. and Lim, T.-K., 2006. A Landsat surface reflectance dataset for North America, 1990-2000. IEEE Geosci. Remote Sens. Lett. 3 (1), 68–72. Meigs, G.W., Kennedy, R.E. and Cohen, W.B., 2011. A Landsat time series approach to characterize bark beetle and defoliator impacts on tree mortality and surface fuels in conifer forests. Remote Sens. Environ. 115, 3707–3718. Monteith, J.L., 1972. Solar radiation and productivity in tropical ecosystems. J. Appl. Ecol. 9 (3), 747–766. Montesano, P.M., Rosette, J., Sun, G., North, P., Nelson, R.F., Dubayah, R.O., Ranson, K.J. and Kharuk, V., 2015. The uncertainty of biomass estimates from modeled ICESat-2 returns across a boreal forest gradient. Remote Sens. Environ. 158, 95–109. 158  Myneni, R., Knyazikhin, Y. and Shabanov, N., 2011. Leaf area index and fraction of absorbed PAR products from Terra and Aqua MODIS sensors: analysis, validation, and refinement. Land remote sensing and global environmental change. Remote Sensing and Digital Image Processing.(ed. by B. Ramachandran,C.O. Justice,and M.J. Abrams). pp. 603–633. Springer, New York. Næsset, E., 2004. Practical large-scale forest stand inventory using a small-footprint airborne scanning laser. Scand. J. For. Res. 19 (2), 164–179. NFI, 2013. Canada’s National Forest Inventory. Natural Resources Canada, Victoria, B.C. https://nfi.nfis.org. Oliver, C. and Larson, B., 1990. Forest stand dynamics. John Wiley & Sons, Inc., New York. Pan, Y., Birdsey, R.A., Fang, J., Houghton, R., Kauppi, R.E., Kurz, W.A., Phillips, O.L., Shvidenko, A., Lewis, S.L., Canadell, J.G., Ciais, P., Jackson, R.B., Pacala, S.W., McGuire, A.D., Piao, S., Rautiainen, A., Sitch, S. and Hayes, D., 2011. A large and persistent carbon sink in the world’s forests. Science. 333, 988–993. Pang, Y., Lefsky, M.A., Sun, G. and Ranson, J., 2011. Impact of footprint diameter and off-nadir pointing on the precision of canopy height estimates from spaceborne lidar. Remote Sens. Environ. 115, 2798–2809. Pare, D. and Bergeron, Y., 1995. Above-ground biomass accumulation along a 230-year chronosequence in the southern portion of the Canadian boreal forest. J. Ecol. 83 (6), 1001–1007. Patterson, P. and Healey, S., 2015. Global ecosystem dynamics investigation (GEDI) LiDAR sampling strategy. Pushing boundaries: new directions in inventory techniques and applications: Forest Inventory and Analysis (FIA) symposium 2015. p. 245. Portland, Oregon. Payette, S. and Delwaide, A., 2003. Shift of Conifer Boreal Forest to Lichen-Heath Parkland Caused by Successive Stand Disturbances. Ecosystems. 6 (6), 540–550. Peng, C., Ma, Z., Lei, X., Zhu, Q., Chen, H., Wang, W., Liu, S., Li, W., Fang, X. and Zhou, X., 2011. A drought-induced pervasive increase in tree mortality across Canada’s boreal forests. Nat. Clim. Chang. 1 (9), 467–471. Pflugmacher, D., Cohen, W.B. and E. Kennedy, R., 2012. Using Landsat-derived disturbance history (1972–2010) to predict current forest structure. Remote Sens. Environ. 122, 146–165. Pflugmacher, D., Cohen, W.B., Kennedy, R.E. and Yang, Z., 2014. Using Landsat-derived disturbance and recovery history and lidar to map forest biomass dynamics. Remote Sens. Environ. 151, 124–137. Pickell, P.D., Hermosilla, T., J. Frazier, R., Coops, N.C. and Wulder, M.A., 2016. Forest recovery trends derived from Landsat time series for North American boreal forests. Int. J. 159  Remote Sens. 37 (1), 138–149. Pregitzer, K.S. and Euskirchen, E.S., 2004. Carbon cycling and storage in world forests: Biome patterns related to forest age. Glob. Chang. Biol. 10 (12), 2052–2077. Price, J.S. and Waddington, J.M., 2000. Advances in Canadian wetland hydrology and biogeochemistry. Hydrol. Process. 14, 1579–1589. Le Quéré, C., Raupach, M.R., Canadell, J.G. and Al., G.M., 2009. Trends in the sources and sinks of carbon dioxide. Nat. Geosci. 2 (12), 831 – 836. R Core Team, 2015. R: A language and environment for statistical computing. Vienna, Austria. http://www.r-project.org/. Rosenberg, M.S. and Anderson, C.D., 2011. PASSaGE: Pattern analysis, spatial statistics and geographic exegesis. Version 2. Methods Ecol. Evol. 2 (3), 229–232. Running, S.W., Nemani, R., Heinsch, F.A., Zhao, M., Reeves, M. and Hashimoto, H., 2004. A continuous satellite-derived measure of global terrestrial primary production. Bioscience. 54 (6), 547–560. Ryan, K.C., 2002. Dynamic Interactions between Forest Structure and Fire Behavior in Boreal Ecosystems. Silva Fenn. 36 (1), 13–39. Ryan, M., Binkley, D. and Fownes, F., 1997. Age-related decline in forest productivity: pattern anasdadad process. Adv. Ecol. Res. 27, 213–262. Schroeder, T.A., Wulder, M.A., Healey, S.P. and Moisen, G.G., 2011. Mapping wildfire and clearcut harvest disturbances in boreal forests with Landsat time series data. Remote Sens. Environ. 115, 1421–1433. Schulze, E.D., Beck, E. and Müller-Hohenstein, K., 2002. Plant Ecology. Springer-Verlag, Berlin. Simard, M., Pinto, N., Fisher, J.B. and Baccini, A., 2011. Mapping forest canopy height globally with spaceborne lidar. J. Geophys. Res. 116 (G04021), doi:10.1029/2011JG001708. Sirois, L. and Payette, S., 1989. Postfire black spruce establishment in subarctic and boreal Quebec. Can. J. For. Res. 19, 1571 – 1580. Solberg, S., Næsset, E., Hanssen, K.H. and Christiansen, E., 2006. Mapping defoliation during a severe insect attack on Scots pine using airborne laser scanning. Remote Sens. Environ. 102 (3-4), 364–376. Song, C. and Woodcock, C.E., 2003. Monitoring Forest Succession With Multitemporal Landsat Images : Factors of Uncertainty. IEEE Trans. Geosci. Remote Sens. 41 (11), 2557–2567. Soverel, N.O., Perrakis, D.D.B. and Coops, N.C., 2010. Estimating burn severity from Landsat dNBR and RdNBR indices across western Canada. Remote Sens. Environ. 114, 1896–1909. Stanojevic, Z., Lee, P. and Gysbers, J.D., 2006a. Recent Anthropogenic Changes within the 160  Boreal Plains Ecozone of Saskatchewan and Manitoba: Interim Report. Global Forest Watch Canada, Edmonton, Alberta. Stanojevic, Z., Lee, P. and Gysbers, J.D., 2006b. Recent Anthropogenic Changes within the Northern Boreal, Southern Taiga, and Hudson Plains Ecozones of Québec: A Global Forest Watch Canada Report. Global Forest Watch Canada, Edmonton, Alberta. http://www.globalforestwatch.ca/publications/20060209A. Stinson, G., Kurz, W.A., Smyth, C.E., Neilson, E.T., Dymond, C.C., Metsaranta, J.M., Boisvenue, C., Rampley, G.J., Li, Q., White, T.M. and Blain, D., 2011. An inventory-based analysis of Canada’s managed forest carbon dynamics, 1990 to 2008. Glob. Chang. Biol. 17, 2227–2244. Stocks, B.. J., Mason, J.A., Todd, J.B., Bosch, E.M., Wotton, B.. M., Amiro, B.D.B., Flannigan, M.D., Hirsch, K.G., Logan, K.A., Martell, D.L. and Skinner, W.R., 2002. Large forest fires in Canada, 1959–1997. J. Geophys. Res. 107 (8149), doi:10.1029/2001JD000484. Sun, G., Ranson, K.J., Kimes, D.S., Blair, J.B. and Kovacs, K., 2008. Forest vertical structure from GLAS: An evaluation using LVIS and SRTM data. Remote Sens. Environ. 112, 107–117. Taylor, A.R. and Chen, H.Y.H., 2011. Multiple successional pathways of boreal forest stands in central Canada. Ecography. 34 (2), 208–219. Tucker, C.J., 1979. Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens. Environ. 8, 127–150. Turner, M.G., Romme, W.H., Gardner, R.H. and Hargrove, W.W., 1997. Effects of fire size and pattern on early succession in Yellowstone National Park. Ecol. Monogr. 67 (4), 411–433. USDA Forest Service, 2003. Field Procedures for the Coastal Alaska Inventory. Pacific Northwest Station. http://www.fs.fed.us/pnw/fia/local-resources/pdf/field_manuals/ak/2003_coak_field_manual.pdf. Venier, L.A., Thompson, I.D., Fleming, R., Malcolm, J., Aubin, I., Trofymow, J.A., Langor, D., Sturrock, R., Patry, C., Outerbridge, R.O., Holmes, S.B., Haeussler, S., Grandpré, L. De, Chen, H.Y.H., Bayne, E., Arsenault, A. and Brandt, J.P., 2014. Effects of natural resource development on the terrestrial biodiversity of Canadian boreal forests. Environ. Rev. 22, 457–490. Viereck, L.A., 1983. The effects of fire in black spruce ecosystems of Alaska and northern Canada. The Role of Fire in Northern Circumpolar Ecosystems.(ed. by R. Wein and D. MacLean). pp. 201–220. John Wiley and Sons, New York, New York, USA. Vitt, D.H., Halsey, L.A., Bauer, I.E. and Campbell, C., 2000. Spatial and temporal trends in carbon storage of peatlands of continental western Canada through the Holocene. Can. J. Earth Sci. 37 (5), 683–693. Vogelmann, J.E. and Rock, B.N., 1988. Assessing forest damage in high-elevation coniferous 161  forests in Vermont and New Hampshire using Thematic Mapper data. Remote Sens. Environ. 24, 227–246. Wehr, A. and Lohr, U., 1999. Airborne laser scanning-an introduction and overview. ISPRS J. Photogramm. Remote Sens. 54, 68–82. White, J.C., Wulder, M.A., Hobart, G.W., Luther, J.E., Hermosilla, T., Griffiths, P., Coops, N.C., Hall, R.J., Hostert, P., Dyk, A. and Guindon, L., 2014. Pixel-based image compositing for large-area dense time series applications and science. Can. J. Remote Sens. 40 (3), 192–212. Wilson, E.H. and Sader, S.A., 2002. Detection of forest harvest type using multiple dates of Landsat TM imagery. Remote Sens. Environ. 80, 385–396. Wulder, M.A., Bater, C.W., Coops, N.C., Hilker, T. and White, J.C., 2008a. The role of LiDAR in sustainable forest management. For. Chron. 84 (6), 807–826. Wulder, M.A., Franklin, S.E. and Lavigne, M., 1996. High Spatial Resolution Optical Image Texture for Improved Estimation of Forest Stand Leaf Area Index. Can. J. Remote Sens. 22 (1), 441–449. Wulder, M.A., Masek, J.G., Cohen, W.B., Loveland, T.R. and Woodcock, C.E., 2012a. Opening the archive: How free data has enabled the science and monitoring promise of Landsat. Remote Sens. Environ. 122, 2–10. Wulder, M.A., White, J.C., Alvarez, F., Han, T., Rogan, J. and Hawkes, B., 2009. Characterizing boreal forest wildfire with multi-temporal Landsat and LIDAR data. Remote Sens. Environ. 113, 1540–1555. Wulder, M.A., White, J.C., Bater, C.W., Coops, N.C., Hopkinson, C. and Chen, G., 2012b. Lidar plots — a new large-area data collection option: context, concepts, and case study. Can. J. Remote Sens. 38 (5), 600–618. Wulder, M.A., White, J.C., Cranny, M., Hall, R.J., Luther, J.E., Beaudoin, A., Goodenough, D.G. and Dechka, J.A., 2008b. Monitoring Canada’s forests. Part 1: Completion of the EOSD land cover project. Can. J. Remote Sens. 34 (6), 549–562. Xing, Y., de Gier, A., Zhang, J. and Wang, L., 2010. An improved method for estimating forest canopy height using ICESat-GLAS full waveform data over sloping terrain: A case study in Changbai mountains, China. Int. J. Appl. Earth Obs. Geoinf. 12, 385–392. Zald, H.S.J., Wulder, M.A., White, J.C., Hilker, T., Hermosilla, T., Hobart, G.W. and Coops, N.C., 2016. Integrating Landsat pixel composites and change metrics with lidar plots to predictively map forest structure and aboveground biomass in Saskatchewan, Canada. Remote Sens. Environ. 176, 188–201. Zhao, M., Heinscha, F.A., Nemani, R.R. and Running, S.W., 2005. Improvements of the MODIS terrestrial gross and net primary production global data set. Remote Sens. Environ. 95, 164–176. Zhao, M. and Running, S.W., 2010. Drought-induced reduction in global terrestrial net primary 162  production from 2000 through 2009. Science. 329 (5994), 940–943. Zhu, Z. and Woodcock, C.E., 2012. Object-based cloud and cloud shadow detection in Landsat imagery. Remote Sens. Environ. 118, 83–94. Zhu, Z., Woodcock, C.E. and Olofsson, P., 2012. Continuous monitoring of forest disturbance using all available Landsat imagery. Remote Sens. Environ. 122, 75–91. Zimble, D.A., Evans, D.L., Carlson, G.C., Parker, R.C., Grado, S.C. and Gerard, P.D., 2003. Characterizing vertical forest structure using small-footprint airborne lidar. Remote Sens. Environ. 87, 171–182. Zwally, H.J., Schutz, B., Abdalati, W., Abshire, J., Bentley, C., Brenner, A., Bufton, J., Dezio, J., Hancock, D., Harding, D., Herring, T., Minster, B., Quinn, K., Palm, S., Spinhirne, J. and Thomas, R., 2002. ICESat’s laser measurements of polar ice, atmosphere, ocean, and land. J. Geodyn. 34, 405–445.   

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0315350/manifest

Comment

Related Items