UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Employing advanced airborne remotely sensed data to improve terrestrial ecosystem mapping Jones, Trevor Gareth 2011

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

Notice for Google Chrome users:
If you are having trouble viewing or searching the PDF with Google Chrome, please download it here instead.

Item Metadata

Download

Media
24-ubc_2011_spring_jones_trevor.pdf [ 10.75MB ]
Metadata
JSON: 24-1.0071720.json
JSON-LD: 24-1.0071720-ld.json
RDF/XML (Pretty): 24-1.0071720-rdf.xml
RDF/JSON: 24-1.0071720-rdf.json
Turtle: 24-1.0071720-turtle.txt
N-Triples: 24-1.0071720-rdf-ntriples.txt
Original Record: 24-1.0071720-source.json
Full Text
24-1.0071720-fulltext.txt
Citation
24-1.0071720.ris

Full Text

EMPLOYING ADVANCED AIRBORNE REMOTELY SENSED DATA TO IMPROVE TERRESTRIAL ECOSYSTEM MAPPING by Trevor Gareth Jones B.A., Clark University, 2005 M.A., Clark University, 2006  A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY  in  The Faculty of Graduate Studies (Forestry)  THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver)  April 2011 © Trevor Gareth Jones, 2011  ABSTRACT  Information representing the species composition and structural configuration of forested ecosystems is critical for effective, sustainable management. In Canada, the methods employed to map forest species and structure vary, however, they conventionally include photogrammetric techniques. Despite common use, aerial photograph delineation and interpretation is time consuming and laborious, often yielding subjective results which cannot be easily updated, and is thus not well suited for quantitative mapping over extensive areas. In contrast, advanced methods for remotely quantifying forest characteristics show promise for improving conventional approaches. Two data sources of particular interest are hyperspectral and light detection and ranging (LiDAR). Hyperspectral sensors acquire data simultaneously in upwards of hundreds of narrow spectral channels, providing an unprecedented tool for differentiating between vegetation species. LiDAR systems directly measure the vertical distribution of foliage, providing detailed information on height, cover, and structure. This thesis integrated new generation remote sensing technologies with field data to improve forest species and structural information in the British Columbian southern Gulf Islands (SGI). Results indicate the objective was met, providing a stateof-the-art, step-by-step protocol for forest managers and ecologists to undertake detailed and accurate species and structural mapping of protected areas, while decreasing associated labor, time and subjectivity, and increasing repeatability, at a cost comparable, if not less, than conventional aerial photography. The unique outcomes of this thesis include the first spectral library of dominant tree species in Canada’s coastal Pacific Northwest, the first SGI inventory of LiDAR-metrics able to characterize and differentiate forest structure, significantly improved data for rare Garry oak habitat, markedly more detailed and accurate distribution information for 11 dominant tree species derived using an innovative classification approach and newly developed LiDAR metrics, and the first assessment in any environ of hyperspectral metrics for describing and differentiating avifaunal guilds based on diversity. In addition, results provide the first tree species heterogeneity predictions for the SGI, yielded through an object-based classification incorporating airborne hyperspectral data and space-borne multispectral data. The innovative methods described are not limited to the SGI, and can be replicated where targeted species/structural characteristics can be defined and differentiated based on hyperspectral-derived and/or LiDAR-derived metrics.  ii  PREFACE  This thesis consists of six scientific papers of which I am the first author. The initial project overview was proposed by my supervisor, Dr. Nicholas Coops. Airborne hyperspectral and LiDAR data and logistical support for field campaigns were provided by Parks Canada (Gulf Islands National Park Reserve). For scientific journal submissions, I performed all research, data analyses, and interpretation of results, and prepared the final manuscripts. Co-authors provided advice on methodology and made editorial comments as required. Publications arising from this thesis thus far include (reprinted with permission from publishers): •  Chapter 2: Jones, T. G., Coops, N. C., and Sharma, T. (2010a). Employing ground-based spectroscopy for tree species differentiation in the Gulf Islands National Park Reserve. International Journal of Remote Sensing 31, 1121−1127.  •  Chapter 3: Jones, T.G., Coops, N.C., and Sharma, T. (in press). Assessing the utility of LiDAR to differentiate among vegetation structural classes. Remote Sensing Letters.  •  Chapter 4: Jones, T.G., Coops, N.C. & Sharma, T. (2011). Exploring the utility of hyperspectral imagery and LiDAR data for predicting Quercus garryana ecosystem distribution and aiding in habitat restoration. Restoration Ecology 19, 245-256.  •  Chapter 5: Jones, T.G., Coops, N.C. and Sharma, T. (2010b). Assessing the utility of airborne hyperspectral and LiDAR data for species distribution mapping in the coastal Pacific Northwest, Canada. Remote sensing of Environment 114, 2841-2852.  In addition, Chapters 6 and 7 are currently under review and/or in revision in leading scientific journals. For journal details please consult Chapter 1 and/or the Bibliography.  iii  Table of Contents Abstract .......................................................................................................................................... ii Preface ........................................................................................................................................... iii Table of Contents ......................................................................................................................... iv List of Tables................................................................................................................................ vii List of Figures ............................................................................................................................... ix Glossary of Acronynms ............................................................................................................... xi Acknowledgements .................................................................................................................... xiv Dedication .................................................................................................................................... xv Introduction .................................................................................................................................... 1 1.1 Forest species and structure ........................................................................................... 1 1.2 Conventional terrestrial ecosystem mapping in Canada ................................................. 2 1.3 Terrestrial ecosystem mapping in British Columbia, Canada ......................................... 2 1.4 Methods for improving terrestrial ecosystem mapping .................................................... 4 1.4.1 Advanced remotely sensed data ............................................................................. 4 1.4.2 Hyperspectral data ................................................................................................... 4 1.4.3 LiDAR data .............................................................................................................. 6 1.5 Fusing advanced airborne remotely sensed data ........................................................... 7 1.6 Applying improved terrestrial ecosystem information ...................................................... 8 1.7 Spatially extending detailed and accurate ecosystem information ................................ 10 1.8 Approach and objectives ............................................................................................... 11 2 Employing ground-based spectroscopy for tree-species differentiation in the gulf islands national park reserve ..................................................................................................... 15 2.1 Introduction .................................................................................................................... 15 2.2 Methods ......................................................................................................................... 16 2.2.1 Study area ............................................................................................................. 16 2.2.2 Sampling protocol .................................................................................................. 18 2.2.3 Ground-based collection of spectra ....................................................................... 19 2.2.4 Statistical analysis ................................................................................................. 19 2.3 Results ........................................................................................................................... 20 2.3.1 Statistical analysis ................................................................................................. 20 2.3.2 Assessing wavelength significance ....................................................................... 24 2.4 Discussion ..................................................................................................................... 24 2.4.1 Performance of data sets....................................................................................... 24 2.4.2 Significant discriminatory variables ....................................................................... 25 2.5 Conclusions ................................................................................................................... 26 3 Assessing the utility of LiDAR to differentiate among vegetation structural classes .......................................................................................................................................... 27 3.1 Introduction .................................................................................................................... 27 3.2 Methods ......................................................................................................................... 28 3.2.1 Study area ............................................................................................................. 28 3.2.2 TEM data ............................................................................................................... 29 3.2.3 LiDAR data ............................................................................................................ 29 3.2.4 Derived metrics ...................................................................................................... 30 3.2.5 Statistical analysis ................................................................................................. 32 3.3 Results ........................................................................................................................... 33 3.4 Discussion ..................................................................................................................... 35 iv  3.5  Conclusion ..................................................................................................................... 37  4 Exploring the utility of hyperspectral imagery and LiDAR data for predicting Quercus garryana ecosystem distribution and aiding in habitat restoration ....................... 39 4.1 Introduction .................................................................................................................... 39 4.2 Methods ......................................................................................................................... 43 4.2.1 Study area ............................................................................................................. 43 4.2.2 Remotely sensed data ........................................................................................... 45 4.2.3 Remotely sensed data pre-processing .................................................................. 46 4.2.4 Field program ......................................................................................................... 48 4.2.5 Classification .......................................................................................................... 49 4.2.6 Accuracy assessment ............................................................................................ 50 4.3 Results ........................................................................................................................... 50 4.4 Discussion ..................................................................................................................... 57 4.5 Conclusion ..................................................................................................................... 61 4.5.1 Implications for practice ......................................................................................... 63 5 Assessing the utility of airborne hyperspectral and LiDAR data for species distribution mapping in the coastal Pacific Northwest, Canada ............................................ 64 5.1 Introduction .................................................................................................................... 64 5.2 Methods ......................................................................................................................... 67 5.2.1 Study area ............................................................................................................. 67 5.2.2 Remotely sensed data ........................................................................................... 68 5.2.3 Field program ......................................................................................................... 70 5.2.4 Pre-processing of hyperspectral data .................................................................... 71 5.2.5 Pre-processing of LiDAR data ............................................................................... 71 5.2.6 Classification .......................................................................................................... 72 5.3 Results ........................................................................................................................... 75 5.3.1 Differentiation of species based on spectral properties ......................................... 75 5.3.2 Differentiation of species based on mean canopy surface height characteristics . 78 5.3.3 Differentiation of species based on volumetric canopy height values ................... 79 5.3.4 Support vector machine classification results ....................................................... 80 5.3.5 Significance tests based on proportion correct ..................................................... 84 5.3.6 Area occupied ........................................................................................................ 85 5.4 Discussion ..................................................................................................................... 86 5.5 Conclusion ..................................................................................................................... 91 6 Describing avifaunal richness with functional and structural bio-indicators derived from advanced airborne remotely sensed data .......................................................... 93 6.1 Introduction .................................................................................................................... 93 6.2 Materials and methods .................................................................................................. 97 6.2.1 Study area ............................................................................................................. 97 6.2.2 Hyperspectral data ................................................................................................. 97 6.2.3 LiDAR data ............................................................................................................ 98 6.2.4 Avian survey .......................................................................................................... 99 6.2.5 Statistical analyses .............................................................................................. 100 6.3 Results ......................................................................................................................... 104 6.3.1 Species richness .................................................................................................. 104 6.3.2 Guild-based discrimination .................................................................................. 105 6.3.3 Assessing richness .............................................................................................. 108 6.4 Discussion ................................................................................................................... 115 6.4.1 Ecosystem function, bird species richness and hyperspectral imagery .............. 115 6.4.2 Tree species diversity and bird species richness ................................................ 116 6.4.3 Vegetation structure, bird species richness and LiDAR data .............................. 117 6.4.4 Combining data .................................................................................................... 119 6.5 Management Implications ............................................................................................ 121 v  6.6  Conclusion ................................................................................................................... 122  7 Extrapolation of tree species mapping using measures of heterogeneity and object-based classification....................................................................................................... 124 7.1 Introduction .................................................................................................................. 124 7.2 Methods ....................................................................................................................... 129 7.2.1 Study area ........................................................................................................... 129 7.2.2 Data ..................................................................................................................... 129 7.2.3 Segmentation of Landsat image .......................................................................... 135 7.2.4 Selection of segment features (independent variables) ...................................... 135 7.2.5 Extraction of segment-level heterogeneity values (dependent variables) ........... 136 7.2.6 Model creation and validation .............................................................................. 137 7.2.7 Extrapolation of species heterogeneity through object-based classification ....... 138 7.3 Results ......................................................................................................................... 138 7.3.1 Tree species dominance...................................................................................... 138 7.3.2 Tree species heterogeneity (pixel-level) .............................................................. 140 7.3.3 Landsat segmentation ......................................................................................... 142 7.3.4 Tree species heterogeneity: comparison of pixel to segment-level..................... 142 7.3.5 Model definition and validation (regression tree analysis)................................... 144 7.3.6 Object-based classification (extrapolation) .......................................................... 147 7.4 Discussion ................................................................................................................... 149 7.4.1 Comparison of coarsening techniques ................................................................ 149 7.4.2 Comparison of pixel versus segment-level heterogeneity ................................... 149 7.4.3 Regression tree results ........................................................................................ 150 7.4.4 Management implications for extrapolated heterogeneity ................................... 151 7.5 Conclusion ................................................................................................................... 153 Conclusions ............................................................................................................................... 155 Bibliography ............................................................................................................................... 165 Appendices ................................................................................................................................ 180 Appendix A: Data collection/pre-processing ................................................................ 180 Appendix B: Bird species stratified by guild ................................................................ 192  vi  LIST OF TABLES  Table 2.1: Species analyzed and per species allocation of samples.............................................................. 18 Table 2.2: Wavelengths deemed statistically significant by FSDA organized by data set and spectral partition ............................................................................................................................................... 23 Table 2.3: Average per species NDA cross-validation classification producer's accracies .......................... 24 Table 3.1: Terrestiral Ecosystem Mapping (TEM) structural stage descriptions based on definitions provided by Hamilton (1988), Oliver and Larson (1990), Weetman et al. (1990), Resource Inventory: Vegetation Inventory Working Group (1995) and, BC Ministry of Forests and BC Ministry of Environment (1998) ......................................................................................................... 29 Table 3.2: LiDAR metrics (organized by type) able to differentiate among 15 possible combinations of TEM structural classes, including: herbaceous (HB), shrub/herbaceous (SH), pole/sapling (PS), young forest (YF), mature forest (MF), and old forest (OF) ...................................................... 34 Table 3.3: The ability (gray cells) or lack thereof (black cells) of LiDAR metrics (organized by type) to significantly differentiate (p≤0.01) among 15 possible combinations of TEM structural classes, including: herbaceous (HB), shrub/herbaceous (SH), pole/sapling (PS), young forest (YF), mature forest (MF), and old forest (OF) .....................................................................................35 Table 4.1: Proportion of calibration and validation data per species in pixels ............................................. 49 Table 4.2: Confusion matrix for support vector machine (SVM) classification without the inclusion of LiDAR canopy volume profiles (CVP) as classification input. Excepting Kappa Index of Agreement (KIA), accuracy statements are provided in percentage ................................................... 53 Table 4.3: Confusion matrix for support vector machine (SVM) classification with the inclusion of LiDAR canopy volume profiles (CVP) as classification input. Excepting Kappa Index of Agreement (KIA), accuracy statements are provided in percentage ................................................... 53 Table 5.1: Common tree species found in and around the Gulf Islands National Park Reserve (GINPR), British Columbia (BC), Canada. The quantity of reference tree/tree clusters and the number of pixels comprising them (shown in parentheses) are shown per-species. Furthermore, the partition of reference data into training and validation data is shown ........................................................................ 71 Table 5.2: Comparison of support vector machine (SVM) classification accuracies resulting from using purely spectral variables as input versus pixel-level fusion with LiDAR-derived canopy height and canopy volume information ................................................................................................................... 82 Table 5.3: Pair-wise comparisons of │z│values resulting from McNemar’s tests .......................................85 Table 6.1: Narrow-band vegetation indices considered for analysis ............................................................ 98 Table 6.2: Variables considered for analysis and associated targeted information. For hyperspectralderived bio-indicators, spatial resolution and the wavelengths used for calculating vegetation indices are provided. For LiDAR-derived indicators, the window of analysis within which variabes were originally calculated and the resultant spatial resolution are supplied. For all variables, the plot-based mean and standard deviation were originally considered. Greyed bioindicators were eliminated based on correlation analyses ................................................................. 103 Table 6.3: Avifaunal richness related to sampling effort ............................................................................ 105 Table 6.4: Bio-indicators found through discriminant analyses to usefully discriminate between plots based on the guild type most often observed (i.e., open-country/shrubland/grassland vs. mature forest/woodland). Significant bio-indicators are organized by data-origin (i.e., hyperpsectral vs. LiDAR-derived) and whether significance (p-value) was assocaited with the mean and/or standard deviation ............................................................................................................................. 108 Table 6.5: Bio-indicators selected using the Akaike information criterion organized by guild type, data type (i.e., hyperspectral vs. LiDAR-derived) and nature of summary (i.e., mean vs. standard deviation) .......................................................................................................................................... 110  vii  Table 6.6: The results of generalized linear models built using the most parsimonious bio-indicators as selected using the Akaike information criterion organized by guild and metric type ....................... 111 Table 7.1: Aspects of segment geometry and segment-level spectral (i.e., Landsat) features orginally considered as independent predictor variables for building models to extrapolate fine tree species heterogeneity (i.e., richness, diversity, and evenness) data derived from airborne hyperspectral/LiDAR data to the broader extent of a Landsat scene ................................................ 136 Table 7.2: The effect of coarsening spatial resolution (i.e., pixel size in meters) on three measures of tree species hetergeneity (i.e., richness, diversity, and evenness) derived from airborne hyperspectral/LiDAR data within and limited to the extent of flightlines ........................................ 141 Table 7.3: Comparison of descriptive statistics for tree species heterogeneity (i.e., richness, diversity, and evenness) values calculated from and within the flightlines of airborne hyperspectral/LiDAR data, within 30 m image pixels versus variably sized image segments ......... 143 Table 7.4: Independent variables representing aspects of segment geometry and segment-level spectral (i.e., Landsat) statistics considered for and selected in regression tree analyses. Selected variables explined variance in tree species richness, diversity and/or evenness. An 'x' represents variables considered for but not selected in analyses, whereas an emboldened 'xx' represents selected variables. Grayed out cells represent variables which were not considered for anlysis, based on pair-wise Pearson's correlation coeficients (p<0.05), a lack of ecological significance, and/or preliminary relationship assessment .......................................................................................145  viii  List of figures Figure 1.1: Spatial scales used for analysis and corresponding chapters ...................................................... 14 Figure 2.1: Baseline reflectance (top) and its first (middle) and second (bottom) derivatives ......................22 Figure 4.1: The AOI includes portions of two islands in the SGI (latitude 48.76o, longitude −123.18o) totaling approximately 600 ha, over which hyperspectral and LiDAR data were collected in transects during July 2006. Transects were collected both in and around GINPR properties. Within transect boundaries, plot centers and polygon centroids representing the location of approximately 200 tree/tree clusters used as reference (i.e., calibration and validation), data are shown. The background image is Landsat TM band 4 (NIR) collected in July, 2006. ........................44 Figure 4.2: Mean hyperspectral reflectance of reference (i.e., calibration and validation) data for all species, collected by the Airborne Imaging Spectrometer for Applications (AISA) Dual sensor in July, 2006. ........................................................................................................................................52 Figure 4.3: Species primarily responsible for Gary oak classification error without the inclusion of LiDAR canopy volume profiles (CVPs) as classification input. ..........................................................54 Figure 4.4: Species primarily responsible for Gary oak classification error with the inclusion of LiDAR canopy volume profiles (CVPs) as classification input. ..........................................................54 Figure 4.5: With focus on the area of interest: the total number of polygons inside and outside of Gulf Island National Park Reserve (GINPR) boundaries occupied by Garry oak according to existent ecological data (terrestrial ecosystem mapping (TEM)) versus support vector machine (SVM) classification results, with and without canopy volume profiles (CVPs) as input ...............................55 Figure 4.6: With focus on the area of interest: the approximate total hectares inside and outside of Gulf Island National Park Reserve (GINPR) boundaries occupied by Garry oak according to existent ecological data (terrestrial ecosystem mapping (TEM)) versus support vector machine (SVM) classification results, with and without canopy volume profiles (CVPs) as input ...................55 Figure 4.7: Comparison of existent Garry oak locations (terrestrial ecosystem mapping (TEM) polygons) with support vector machine (SVM) classification results with and without canopy volume profiles (CVPs) as input for three distinct Garry oak habitat types (i.e., rocky bluff with woodland patch, steep slope woodland, and field with woodland patch) ............................................56 Figure 5.1: The area of interest (AOI) includes ~2800 ha encompassed by 23 transects of hyperspectral and LiDAR data, concurrently collected in July, 2006 in and around the Gulf Islands National Park Reserve (GINPR), British Columbia (BC), Canada (lat 48.76°, long −123.18°). Within the extent of transects the location of 411 tree/tree clusters representing 11 species are shown. Tree/tree cluster information was used to train and assess accuracy for a series of support vector machine (SVM) classifications. The background image is the nearinfrared band (channel 4) from a Landsat Thematic Mapper (TM) image acquired in July, 2006. .....69 Figure 5.2: Hyperspectral imagery representing 40 targeted AISA bands and rasterized LiDARderived height and volume information were masked and fused at the pixel-level, resulting in four layer stacks used as support vector machine (SVM) classification input. ....................................74 Figure 5.3ab: Examples of the mean reflectance value of trees/tree clusters for all broad-leaved (top) and coniferous (bottom) species at 453 wavelengths corresponding to analyzed AISA channels. ......76 Figure 5.4ab: The mean spectral value (±1 standard deviation) of broad-leaved (t) and coniferous (b) species at 40 targeted AISA wavelengths, identified through discriminant analyses as significant (p<0.05) in their ability to differentiate between species of interest. ...................................................77 Figure 5.5: Mean canopy surface heights (±1 standard deviation) for 11 species based on a LiDARderived canopy height model (CHM)...................................................................................................78 Figure 5.6: The average proportion of total volume occupied by open gap, euphotic, oligophotic and closed gap zones for 11 species based on LiDAR-derived canopy volume profiles (4CVPs). ............80 Figure 5.7ab: Using the categorical classification accuracies resulting from purely spectral input variables as the frame of reference (i.e., the x-axis), changes in producer’s (top) and user’s accuracies (bottom) are provided .........................................................................................................83  ix  Figure 5.8: The total area in hectares (ha) occupied by each species in relation to four different iterations of support vector machine (SVM) classification using, 1) spectral only, 2) spectral and CHM, 3) spectral and 4CVPs and 4) spectral and 2CVPs as input ......................................................86 Figure 6.1: The difference in open-country/shrubland/grassland vs. mature forest/woodland guilds (±1 stdev) exhibited by the mean normalized difference wetness index (a) and midstory cover (b) values, and variation within the volume of the euphotic strata of vegetation (c) ...............................107 Figure 6.2: The frequency of hyperspectral (a) and LiDAR-derived (b) variables selected using the Akaike Information Criterion .............................................................................................................109 Figure 6.3: Explained variance in species richness. ....................................................................................112 Figure 6.4: Observed vs. predicted richness for all guilds, independently and in combination...................113 Figure 7.1: For a portion of the study area: (a) 2 m tree species distribution data derived from fused airborne hyperspectral/LiDAR data (Chapter 5) shown for and limited to the extent of three airborne transects. The background is area outside of the flightlines, represented by Landsat band 4 (NIR), (b) tree species richness calculated at a 30 m grain within and limited to the extent of airborne hyperspectral/LiDAR flightlines. The background is area outside of the flightlines, represented by Landsat band 4 (NIR), (c) Landsat-5 TM objects/segments falling within appropriate size thresholds and within the extent of airborne hyperspectral/LiDAR flightlines, and (d) 30 m tree species richness extrapolated beyond the extent of flightlines, based on rules defined through regression tree analysis and applied using object-based classification, wherein eight richness classes range from low to high ...............................................131 Figure 7.2: The study area is two-fold, including: 1) the extent of hyperspectral/LiDAR flightlines shown in green/purple above, and 2) the British Columbian southern Gulf Islands: the extent of the figure as represented by landsat-5 TM band 4 (near infrared) (30 km east/west by 35 km north/south, centered at latitude 48.76o and longitude -123.18o) .......................................................133 Figure 7.3: The effect of coarsening spatial resolution on the amount (%) of forested land occupied by each species ........................................................................................................................................140 Figure 7.4: Comparison of frequency distribution for tree species heterogeneity values (i.e., richness, diversity, and evenness) calculated from and within the flightlines of airborne hyperspectral/LiDAR data, within 30 m image pixels versus variably sized image segments. The distribution of richness values is represented by (a) 30 m pixels and (d) variably sized segments. Similarly, diversity and evenness are represented by (b) 30 m pixels and (e) variably sized segments, and (c) 30 m pixels and (f) variably sized segments, respectively ....................................144 Figure 7.5: Decision rules generated through regression tree analysis, wherein tree species richness was the dependent response variable, and segment-level spectral properties of Landsat data were selected as independent predictor variables. Generated rules identify spectral thresholds which define distinct richness classes, while explaining nearly 50% of variance ........................................147 Figure 7.6: Tree species richness, derived from 2 m spatial resolution species distribution maps (from and within the flightlines of airborne hyperspectral/LiDAR data) extrapolated based on relationships established with specific spectral properties of Landsat segments. Species richness has eight distinct classes, ranging from low to high values. The background image is Landsat band 4 (NIR) coverage. ......................................................................................................................148  x  GLOSSARY OF ACRONYMS  AIC: Akaike’s information criterion AISA: airborne imaging spectrometer for applications AOI: area of interest API: aerial photo interpretation ASD: Analytical Spectral Devices BC: British Columbia CDF: Coastal Douglas-fir CG: closed gap CHDs: canopy height descriptors CHM: canopy height model CVPs: volume profiles D: tree species diversity DEM: digital elevation model E: species evenness FR: full range FSDA: forward stepwise discriminant analysis fwhm: full width half maximum GINPR: Gulf Islands National Park Reserve GIS: geographic information system GLM: general linearized models GPS: global positioning system ha: hectare xi  HB: herbaceous HPs: height percentiles KIA: Kappa index of agreement LiDAR: light detection and ranging MDA: MacDonald, Dettwiler and Associates Ltd. MF: mature forest mm: moist maritime NDA: normal discriminant analyses NIR: near infrared nm: nanometres OF: old forest OG: open gap PS: pole/sapling R: species richness RBF: radial basis function SB: sparse/bryoid SGI: southern Gulf Islands SH: shrub/herbaceous SID: Shannon’s index of diversity SVM: support vector machine SWIR: short-wave infrared TC: total canopy cover TEM: terrestrial ecosystem mapping  xii  TIN: triangulated irregular network TRSI: Terra Remote Sensing, Inc. USD: United States Dollar UTM: universal transverse Mercator VIs: vegetation indices VRI: vegetation resource inventory WGS: world geodetic system YF: young forest (µm): micrometres 10N: zone 10 north  xiii  ACKNOWLEDGEMENTS  Components of this research were supported by Parks Canada Ecological Integrity Funding, an NSERC discovery grant, a University of British Columbia University Graduate Fellowship, and a Canadian Remote Sensing Society grant. LiDAR and hyperspectral data were acquired by Parks Canada through the University of Victoria. Logistical support for field campaigns, including expert advice for finding appropriate plot locations, and transportation and access throughout the Southern Gulf Islands, was provided by Parks Canada staff (specifically Todd Golumbia (Park Ecologist), Nathan Cardinal (Park Warden), Cameron Sanjivi (Park Warden), and Rundi Anderson (Park Warden)). I would like to thank my advisor Nicholas Coops. In addition, I would like to thank my committee, including Drs. Tara Shamra, Sarah Gergel and Peter Arcese, for their collective support and invaluable insights. I am also immensely grateful to Drs. Thomas Hilker and Nicholas Goodwin, as well as Christopher Bater (M.Sc.), for critical assistance with processing both LiDAR and hyperspectral data sets. Special thanks to Christopher Bater (M.Sc.) and Shanley Thompson (M.Sc.); and Drs. Samuel Coggins, Robert (“Robbie”) Hember, Thomas Hilker, and Rachel Gaulton for field assistance. Additional thanks to Richard Schuster for summarizing bird survey data. Furthermore, loud thanks to all my brethren within the Integrated Remote Sensing Studio (IRSS) for their help and assistance in association with and, unrelated to academic endeavours. Similarly, a sincere and extensive series of thank yous (and apologies?) to all of my friends and family. A vintage thanks to John Rogan and the folks in the HERO team at Clark University, especially the MaFoMP crew. Also, last but by no means least: Merike! xiv  DEDICATION  To Abu Abd Allah Muhammad al-Idrisi al-Qurtubi al-Hasani al-Sabti Geographer, writer, scientist, cartographer, traveler, artist, cultural fuser ~1100 (Ceuta) – ~1160/1166/1180 (Sicily/Sabtah)  And my parents: Thanks!  IN LOVING MEMORY OF  Chris Topping Steve Conway Evan Paye You live on in a rich and varied bank of fond and formative memories 12:51, V.O.C. xv  1  1.1  Introduction  Forest species and structure  Information characterizing the species composition and structural configuration of forested ecosystems is fundamental for effective, sustainable management (Gong et al., 1997, Lefsky et al., 2002, Plourde et al., 2007). Species data are utilized for many purposes, including inventory (Chubey et al., 2006), pest and invasive species mitigation (Everitt et al., 1996, Peterson, 2005, Morisette et al., 2006, White et al. 2006, and White et al., 2007), carbon sequestration calculations (van Aardt and Wynne, 2007), faunal habitat identification and characterization (Coops and Catling, 1997, Scarth et al., 1999), and biodiversity assessments (Nagendra, 2001, Turner et al., 2003). Similarly, forest structural information helps to characterize wildlife habitat (Morrison et al., 1987), species diversity (Kimmins, 1997) and/or overall biodiversity (Freemark and Merriam, 1986), provides insight into growth processes (Coops et al., 2007), facilitates future growth predictions (Staudhammer and LeMay, 2001), helps explain tree competition (Biging and Dobbertin, 1992), and elucidates the response of forests to disturbance (Parker et al., 2004, Rhoads et al., 2004). Individually and in concert, species and structural information can also influence policy, guide subsequent implementation, and provide reference from which change can be quantified and management decisions evaluated (Innes and Koch, 1998, Dalponte et al., 2008, Voss and Sugumaran, 2008).  1  1.2  Conventional terrestrial ecosystem mapping in Canada  In Canada, methods employed to map forest species and structure vary, but traditionally include well established photogrammetric techniques augmented by field plots measurements (Leckie and Gillis, 1995, Gillis et al., 2005). In particular, the manual definition of homogeneous forest units on aerial photographs based on attributes (e.g., species and/or structure) remains the foundation of forest inventory (Wulder et al., 2008a). Despite its common use, aerial photograph delineation and interpretation is a time consuming, laborious process, yielding subjective results which lack detail, accuracy and cannot be easily updated, and are thus not well suited for extensive areas (Anderson et al., 1993, Gong et al., 1997, Gillis et al., 2005, Leckie et al., 2005, Evans et al., 2006, Lucas et al., 2008). Specific to species information, aerial photograph units are often unable to recognize the presence of all species, instead often focusing on the major dominant species. Furthermore, even if a unit is said to contain a certain species, the precise amount and location remains unknown. Similarly, structural attributes are limited to a single label for each unit which fails to describe internal variability. Despite these limitations, delineated aerial photographs remain the most common information source for forest managers.  1.3  Terrestrial ecosystem mapping in British Columbia, Canada  In British Columbia (BC), Canada, for economic/industrial applications, forests are characterized through Provincial Government Vegetation Resource Inventory (VRI). In  2  contrast, baseline forest species and/or structural data used for ecological applications are often provided through the Terrestrial Ecosystem Mapping (TEM) system which uses aerial photographs to hierarchically stratify the landscape into units based on aspects of climate, physiography, surficial material, geology, soil and vegetation. To enhance polygon-level ecosystem characterizations, 11.3 m radial field plots are commonly established within targeted polygon boundaries. The composition and abundance of tree species are estimated within each TEM unit, which at the polygon-level incorporates the three species interpreted as most dominant. In contrast, the finer scale of plot-level inventory record all dominant species present (ranging from 1-10), representing the percentage of occupation for each species divided by 10. Each TEM unit is also assigned to a structure class based on structural features and age criteria. TEM structural definitions partition vegetation into seven categories, including sparse/bryoid (SB), herbaceous (HB), shrub/herbaceous (SH), pole/sapling (PS), young forest (YF), mature forest (MF) and old forest (BCEDC, 2000). Although TEM facilitates standardized mapping protocols and supplies critical baseline information, it remains time-consuming, laborious and expensive, and yields subjective results with species and structural attributes that are not easily updated and lack the detail and accuracy required by certain management initiatives and mandates. Long-term management of forested ecosystems within BC demands reliable and repetitive ecosystem observations obtainable at a reasonable cost and in a timely manner.  3  1.4  Methods for improving terrestrial ecosystem mapping  1.4.1  Advanced remotely sensed data  In contrast to manually delineating aerial photographs, advanced techniques for remotely quantifying forest characteristics show great promise for improving terrestrial ecosystem mapping. Digital remotely sensed data are increasingly becoming one of the most important information sources for forest assessment and inventory (van Aardt and Wynne, 2001). Over the last 35 years, technological and methodological advancements within the field of remote sensing have resulted in a new generation of sensors operating at a wide range of spatial scales (Rogan and Chen, 2004). In particular, employing advanced airborne and ground-based data to derive ecological products has proven especially promising and beneficial to resource managers. Two promising remote sensing information sources for improving terrestrial ecosystem mapping are hyperspectral and light detection and ranging (LiDAR) data. In addition to improving the detail and accuracy of species and structural characterizations while decreasing subjectivity, the acquisition, processing and interpretation of these advanced data types involves less time and labor and the costs are comparable with expenses associated with aerial photographs (i.e., ~$5.00 USD per hectare) (Wulder et al., 2008b).  1.4.2  Hyperspectral data  Hyperspectral sensors acquire data simultaneously in upwards of hundreds of narrow, adjacent spectral channels, providing an unprecedented tool for the detailed analysis of vegetation distributions (Okin et al., 2000, Smith, 2006). Hyperspectral data have 4  demonstrated their utility for species differentiation (e.g., Cochrane, 2000, Goodwin et al., 2005, van Aardt and Wynne, 2007), owing to their fine spectral and spatial resolutions, which enable measurements of subtle bio-chemical attributes (Blackburn, 1998, Ustin et al., 2004, Kumar et al., 2006). Spectral curves representing bio-chemical properties illuminate discrepancies in the variable presence of pigments, water, and other organic compounds, and reflective differences associated with the internal cell structure of leaves, thus facilitating differentiation (Sinclair et al., 1971, van der Meer, 2006). Despite the ability to discern species-level information, hyperspectral data have inherently high dimensionality and are often extremely correlated and/or noisy, which increases within-species variance and decreases between-species seperability (Clark et al., 2005, Smith, 2006). Therefore, reducing data volume through identifying optimal wavelengths is a fundamental pre-processing step. As such, ground-based spectrometers provide a logical first step for species mapping endeavours by establishing species detectablity,  and  subsequently  identifying  wavelengths  optimal  for  species  differentiation, thus reducing data dimensionality and noise (van Aardt and Wynne, 2001, Clark et al., 2005). Ground-based spectrometers, either hand-held or supported by various means (i.e., tripods, cherry pickers), permit comparatively finer scale and less noisy measurements than air or spaceborne devices. This finer scale enables users to distinguish between subtle spectral properties, such as using leaf-level biochemical attributes to characterize and differentiate among tree species. Wavelengths identified at the leaflevel, which minimize within-species variance while maximizing between-species variance, can then be targeted in airborne hyperspectral data sets, which can then be used to map species distribution for the extent of their coverage. The utility of airborne  5  hyperspectral data for species mapping has previously been demonstrated within tropical (e.g., Carlson et al., 2007), sub-tropical (e.g., Yang et al., 2009) and temperate (e.g., Boschetti et al., 2007) contexts.  1.4.3  LiDAR data  While hyperspectral sensors can provide detailed species information, they are passive (i.e., measuring reflectance) and their measurements are confined to the horizontal, 2dimensional plane. In contrast, airborne LiDAR sensors generate active pulses of near infrared (NIR) light, providing highly accurate horizontal and vertical measurements (Lim et al., 2003). LiDAR pulses can penetrate forest canopies and retrieve information on both vegetation and the ground surface, which has important implications for the characterization of vegetation. Specifically, pulses can directly measure the vertical distribution of foliage and provide detailed information on vegetation height, cover, and structure (Lefsky et al., 2002, Wulder et al., 2008b). LiDAR systems are categorized as either ‘discrete return’ or ‘full waveform’, based on their measurement characteristics (Lim et al., 2003). Discrete systems record large peaks which are interpreted to represent distinct objects in the path of the beam (e.g., the forest canopy, understory, and ground). The sensor then records these peaks as discrete points in three-dimensional space. With full waveform systems, the comparatively higher sampling rate records the full height distribution of the surfaces illuminated by the laser. Therefore, within a forest canopy, discrete return systems yield clouds of points representing intercepted surfaces, whereas full waveform sensors record the entire reflected signal for analysis (Lefsky et al., 2002). For both system types, information 6  content is a function of the horizontal circular sampling area (i.e., footprint) which typically ranges from 0.2-2 m, whereas large footprint, full waveform LiDAR systems range from 8-100 m (Means, 1999, Lim et al., 2003, Wulder et al., 2008b). While the use of full waveform systems remains comparatively rare, small footprint discrete return sensors are increasingly being employed for forestry applications (Lim et al., 2003, Wulder et al., 2008b).  1.5  Fusing advanced airborne remotely sensed data  Given that passive and active sensors acquire information on fundamentally different aspects of vegetation (Hudak et al., 2006), when used alone, their data are subject to inherent limitations. Therefore, integrating passive imagery with LiDAR data to map forest attributes holds significant promise for improving comprehensive canopy characterization (Lefsky et al., 1999a, Hudak et al., 2002, Gillespie et al., 2004, Hill and Thomson, 2005, Hudak et al., 2006). While several studies have investigated the indirect fusion of airborne hyperspectral and LiDAR data (i.e., Blackburn, 2002, Hill and Thomson, 2005, Mundt et al., 2006, Asner et al., 2008, Lucas et al., 2008, Voss and Sugumaran, 2008, Andrew and Ustin, 2009, Geerling et al., 2009, Hall et al., 2009), few have investigated direct fusion to classify vegetation type (i.e., Geerling et al., 2007, Koetz et al., 2008) or species-level distributions (i.e., Dalponte et al., 2008). In addition to confirming the utility of hyperspectral for vegetation mapping, Geerling et al. (2007), Koetz et al. (2008) and Dalponte et al. (2008) demonstrated that using LiDAR-derived height metrics directly as classification input increased accuracies. In accordance with these results, the scantly explored fusion of hyperspectral and LiDAR data for species7  level mapping merits further investigation. In addition, exploring the incorporation of LiDAR metrics beyond those representing height remains an area in need of investigation.  1.6  Applying improved terrestrial ecosystem information  While there are numerous beneficial end-users for improved terrestrial ecosystem information, Parks Canada provides one concrete example. Broadly speaking, National Parks are mandated to restore, conserve and maintain ecological integrity, defined as the condition of an ecosystem or assemblage of ecosystems wherein structure and function are unhindered by anthropogenic processes and inherent biological diversity and supporting processes are likely to persist (Woodley, 1993, Parks Canada Agency, 1997). Increasing the detail and accuracy of species and structural attributes directly aids this mandate, permitting pertinent yet potentially unexplored ecological and managerial concerns to be considered and addressed. While there are seven National Parks in BC (i.e., Pacific Rim, Gulf Islands, Gwaii Haanas, Mount Revelstoke, Kootenay, Glacier and Yoho), some examples of important applications for the Gulf Islands National Park Reserve (GINPR), in the southern Gulf Islands (SGI) of south-western BC are listed below, however, the list is not exhaustive and all examples have pertinence within other Park systems:  1-Rare habitat identification: Based on tree species presence, habitats associated with rare and/or at-risk floral and/or faunal species can be identified. Locating previously  8  unidentified habitats helps restore their landscape connectivity and permits finer scale within ecosystem restoration activities to commence and/or continue. For example, within the GINPR, Garry oak habitats have been degraded and reduced to 1-5% of their pre-European range and are now affiliated with >100 at-risk floral and faunal species (Fuchs, 2001, GOERT, 2003). While restoring and conserving rare and at-risk Garry oak habitat is a management priority, accurate and detailed distribution information remains a primary knowledge gap (AXYS EC, 2004, Green, 2007). Therefore, effectively identifying Garry oak distribution would facilitate restoration within remnant ecosystems and serves to re-establish a network of ecosystem connectivity, helping to reverse the rare and/or at-risk status of species. 2-Habitat acquisition: Within a broader landscape context, protected areas often lack the size and connectivity required to meet restoration and conservation goals (Hobbs and Harris, 2001). Based on tree species presence, areas of high conservation value can be also be targeted for acquisition to expand Park properties and/or private conservation easements. Identifying Garry oak distribution could locate unprotected components of the ecosystem mosaic which could then be targeted for acquisition, meeting the primary management goal of increasing the total area and connectivity of protected ecosystems (AXYS EC, 2004, Green, 2007). 3-Restoration  assessment:  Based  on  tree  species  presence  and/or  habitat  characterizations, the effectiveness of past restoration activities can be assessed. For instance, Garry oak distribution could provide baseline contemporary reference data. Contemporary reference data can be compared with historical archives to assess the  9  effectiveness of past and/or ongoing management restoration interventions (Yates et al., 1994, Noss, 1996, Hobbs and Harris, 2001). 4-Habitat characterization: Using hyperspectral and/or LiDAR metrics, the habitats of targeted floral and/or faunal species can be characterized based on known associations with  certain  tree  species  assemblages  and/or  structural  attributes.  Habitat  characterizations calculate diversity (e.g., richness, evenness, overall diversity) and /or predict suitability. For example, because birds are widespread, easily identified with welldescribed taxonomy, and sensitive to land-cover modification, avian diversity provides a powerful surrogate for ecosystem condition from which habitat conservation value and biodiversity can be assessed (Turner et al., 2003, Gottschalk et al., 2005, Leyequien et al., 2007). Within the SGI a rich network of bird census data has been collected, providing a prime opportunity to assess the utility of advanced airborne metrics for describing bird species diversity.  1.7  Spatially extending detailed and accurate ecosystem information  While the new methods discussed present unprecedented research opportunities, potential applications remain confined to the spatial coverage of airborne surveys. As with aerial photographs, advanced airborne data are typically collected in transects with narrow swath widths (e.g., ~1 km). In addition, all airborne data, conventional or advanced, remain expensive to acquire, process and interpret (i.e., ~$5.00 USD per ha) (Wulder et al.,  2008b).  Therefore,  an  important  management  goal  involves  developing  methodologies to extend the detailed information derived from airborne sensors to the  10  spatial extent provided by satellite imagery (Vitousek et al., 1987, Huang and Asner, 2009). While generally unable to provide species-level and/or forest structural information, medium spatial resolution multispectral satellite imagery supplies contiguous wall-to-wall landscape-level coverage. In addition, these data are readily available at little to no cost and provide unparalleled temporal coverage. For instance, since 2008, moderate spatial resolution Landsat data have been available at no cost, which for most terrestrial areas equates to a free data source extending back nearly 40 years (Cohen and Goward, 2004, Woodcock et al., 2008).  1.8  Approach and objectives  New methodologies for improving tree species and forest structure characterization should exploit multi-scale analysis, combining field-collected data with ground-based spectrometer measurements, airborne hyperspectral and LiDAR data, and spaceborne multispectral satellite imagery. The objective of this thesis was to examine how new generation remote sensing technologies could be integrated with field data to improve conventional forest species and structural information in and around the newly established Gulf Islands National Park Reserve (GINPR), in south-western BC, Canada. To meet this objective, six questions were posed:  1) Can ground-based spectrometry distinguish between 11 dominant tree species? 2) Which, if any, of three suites of LiDAR metrics can differentiate among TEM defined structural classes?  11  3) For a subset of the GINPR, can leaf-level spectral measurements be scaled to airborne hyperspectral data to identify wavelengths optimal for species differentiation, and can optimal airborne channels be fused with targeted LiDAR metrics to predict the distribution of rare and/or at-risk Garry oak habitat? 4) Can optimal hyperspectral channels be fused with targeted LiDAR metrics to predict the distribution of 11 tree species in and around the extent of the GINPR? 5) Can ecosystem function metrics and tree species diversity derived from hyperspectral data and aspects of vegetation structure derived from LiDAR be used to describe the richness of and differentiate among avian guilds? 6) Can detailed and accurate species information be spatially extended beyond limited survey coverage to the extent of a multispectral, medium spatial resolution satellite image?  Answering the first question involved first establishing a network of field plots, wherein species-specific foliage samples were collected. An Analytical Spectral Devices Full Range spectrometer was then used to measure the spectral properties of vegetation samples and determine which wavelengths could most significantly characterize and subsequently discriminate among 11 dominant tree species, and why (Jones et al., 2010a (International Journal of Remote Sensing): Chapter 2). To answer the second question, within established plots, the utility of three suites of metrics derived from small footprint, discrete-return LiDAR data for differentiating among TEM defined structural classes was assessed (Jones et al., in press (Remote Sensing Letters): Chapter 3). To address the third question, wavelengths found to be significant in their ability to discriminate between  12  species at the leaf-level (i.e., Chapter 2) were used to isolate corresponding airborne hyperspectral channels, extending the scale of spectral analysis to the canopy-level. For a subset of hyperspectral/LiDAR transects, using species-specific trees and/or tree clusters identified through field work on hyperspectral imagery as classification reference, targeted hyperspectral channels were then fused with important LiDAR metrics (i.e., Chapter 3) to assess the utility of mapping nine tree species based on their spectral and structural characteristics (Jones et al., 2011 (Restoration Ecology): Chapter 4). In addition to assessing the utility of scaling from the leaf to the canopy-level, and fusing hyperspectral/LiDAR to classify the distribution of tree species, Chapter 4 also quantified the precise amount and location of rare and/or at-risk Garry oak habitats, providing an example of a direct application which benefited several managerial mandates and ongoing initiatives. Building on the results of Chapter 4, and using a wider network of species-specific tree/tree clusters identified through fieldwork as classification reference, answering the fourth question involved assessing the utility of targeted airborne hyperspectral channels fused with selected LiDAR metrics for mapping the distribution of 11 tree species over the full extent of airborne surveys (Jones et al., 2010b (Remote Sensing of Environment): Chapter 5). Using the final species maps generated through Chapter 5 in concert with additional hyperspectral metrics and the LiDAR metrics generated through Chapters 3, 4 and 5, the fifth question involved assessing the utility for explaining richness in three distinct avifaunal guilds (stratified by habitat preference) based on their functional and structural habitat characteristics (Jones et al., in revision (Systematics and Biodiversity): Chapter 6). Lastly, to spatially extend the detailed and accurate species information resulting from Chapter 5 beyond the coverage of airborne  13  transects, the sixth question assessed the utility of object-based classification and regression methods for extrapolation to the extent of a Landsat-5 TM satellite image, which provided wall-to-wall coverage for the entire southern Gulf Islands, including the full extent of the GINPR (Jones et al., under review (Ecography): Chapter 7). Figure 1.1 shows the multiple spatial scales used for analysis and corresponding chapters.  Figure 1.1. Spatial scales used for analysis and corresponding chapters.  14  2  EMPLOYING GROUND-BASED SPECTROSCOPY FOR TREESPECIES DIFFERENTIATION IN THE GULF ISLANDS NATIONAL PARK RESERVE  2.1  Introduction  The Gulf Islands National Park Reserve (GINPR) represents some of British Columbia (BC), Canada’s most diverse, yet threatened, terrestrial ecosystems (BC Ministry of Forestry and Range, 2009). Effective Park management requires detailed and accurate tree-species distribution information. Due to an inherent ability to measure subtle absorption features related to biogeochemical properties, airborne hyperspectral data can be used to map species distribution (Martin et al., 1998, Cochrane, 2000, Ustin et al., 2004). However, hyperspectral data are highly correlated and certain bands may exhibit significant noise, which can increase within-class variance and decrease between-class seperability (Clark et al., 2005). Ground-based spectrometers measure the same spectral range as airborne hyperspectral sensors, but with enhanced spectral and spatial resolution, which enables detection of extremely subtle leaf-level differences between species reflective properties. An analysis of spectrometer reflectance spectra of pure vegetation foliage is an important first step in establishing (a) the detectability of and (b) the difference between key species. Likewise, the analysis can also be used to establish which wavelengths actually provide spectral predictive power. Feature selection using discriminant analysis has been shown to be a useful approach to reduce data dimensionality (van Aardt and Wynne, 2001, Clark et al., 2005).  15  This article presents the results of an investigation examining the spectral seperability of 11 important tree species in the GINPR. The objectives were (1) to determine the feasibility of species differentiation, (2) to compare the discretionary ability of baseline spectrometer data with its derivatives and (3) to determine which spectral regions and bands are best suited for separation.  2.2  Methods  2.2.1  Study area  The GINPR, established in 2003, encompasses 2832 ha (28.32 km2) of terrestrial lands distributed across 16 islands and numerous islets in the southern Gulf Islands (SGI) archipelago, 50 km south of Vancouver, BC, Canada (latitude 48.76o, longitude −123.18o). The SGI experiences a Mediterranean-like climate characterized by dry summers and wet winters. Based on the biogeoclimatic ecosystem classification system (Meidinger and Pojar, 1991), the majority of terrestrial ecosystems are categorized as Coastal Douglas-fir (CDF) zones located within moist maritime (mm) subzones (CDFmm). Although CDFmm zones are among the least common in BC, they comprise some of the province’s most diverse ecosystems (Ministry of Forests and Range, 2006). However, a complex disturbance history involving large-scale, wide-spread industrial logging, agricultural and residential land modification, historic First Nations burning and cultivation practices, contemporary fire suppression, invasive flora and extensive feralanimal related herbivory has significantly impacted the distribution of vegetation throughout the SGIs (Fuchs, 2001, Gedalof et al., 2006, Golumbia, 2006, Vellend et al., 16  2008). Throughout the SGI, the increasing range, frequency and magnitude of anthropogenic activities poses an increasing threat to a variety of terrestrial ecosystems, qualifying it as one of the most rare and diverse, yet degraded and at-risk regions in Canada (Ministry of Forests and Range, 2006). The majority of the terrestrial land-base (~85.0%) is dominated by forests and nonagricultural vegetation, intermixed with tracts of active and remnant agricultural lands (~10.0%), and residential and commercial buildings (~5.0%). Approximately 80.0% of the vegetated land-base is forested, with the rest dominated by herbaceous cover (~16.0%), mixtures of herbaceous and shrub (~2.0%) and/or bryoid vegetation (~2.0%). The majority of the region’s forests (~70.0%) are 40-80 years old, with significant remnants of mature forests 80-250 years old (~25.0%), a small amount of regenerating pole/sapling forest <40 years old (~4.0%), and isolated pockets of old-growth >250 years old (<1.0%) (Green, 2007). The dominant tree species is Pseudotsuga menziesii (Douglas-fir). Other common conifer tree species, listed in relative order of dominance include Thuja plicata (Western redcedar), Pinus contorta (Lodgepole pine), and Tsuga heterophylla (Western hemlock). Common broad-leaved tree species, also listed in relative order of dominance include Alnus rubra (Red alder), Arbutus menziesii (Arbutus), Acer macrophylum (Bigleaf maple), Garry oak, and Populus tremuloides (Trembling aspen) (Table 2.1).  17  Taxonomic group Broadleaf  Conifer  Species Black cottonwood Trembling aspen Red alder Bigleaf maple Garry oak Arbutus Grand fir Western redcedar Douglas-fir Western hemlock Lodgepole pine all species  Scientific name Populus balsamifera Populus tremuloides Alnus rubra Acer macrophylum Quercus garryana Arbutus menziesii Abies grandis Thuja plicata Pseudotsuga menziesii Tsuga heterophylla Pinus contorta  Samples 30 22 28 30 31 24 32 23 26 30 30  Total 165  141  306  Table 2.1. Species analyzed and per species allocation of samples.  2.2.2  Sampling protocol  To capture the spectral variability of tree species, a stratified random sampling scheme was implemented in July 2007, wherein 26 plot locations were initially stratified by islands and then selected randomly within overlapping strata based on land tenure and tree-species distributive information derived from interpreted aerial photographs. Within each plot, trees were climbed and branches collected using clippers attached to the end of a 6 m long pole. Branch selection was based on targeting unimpeded sunlit vegetation, and samples were collected as high as operational constraints permitted. As recommended by Foley et al., (2006), broadleaves and coniferous needle clusters were removed and placed in sealed plastic freezer bags. Table 2.1 displays the allocation of samples per species. For additional information please consult Appendix A.  18  2.2.3  Ground-based collection of spectra  An Analytical Spectral Devices (ASD) full range (FR) spectrometer (Analytical Spectral Devices, Boulder, CO, USA) was used to measure spectral reflectance from 350–2500 nanometers (nm) in a controlled indoor setting. For each measurement, broadleaves or needle clusters were stacked six layers deep to approximate an infinite optical thickness, therefore, simulating canopy reflectance and the maximum of near infrared (NIR) reflectance (Datt, 1998). To reduce noise and capture variability, each species specific spectral curve was calculated by averaging 10 measurements, resulting in 306 curves (Table 2.1). The original sampling interval of the ASD FR is approximately 2 nm, however, during spectral-curve acquisition, measurements are automatically interpolated to a 1 nm sampling interval. To mimic the original sampling interval, baseline reflectances were obtained by averaging values using 2 nm wavelength increments and a 2 nm full width half maximum (fwhm). First and second derivatives were also calculated from the baseline data, resulting in three datasets. Wavelengths 1350–1416 and 1796– 1970 nm, 350–429 and 2401–2500 nm, and 998–1002 and 1798–1802 nm were removed as they represent water absorption regions (van Aardt and Wynne, 2001), sensor extremes, and sensor transitional zones, respectively. For additional information please consult Appendix A.  2.2.4  Statistical analysis  Based on adaptations of methodologies presented in van Aardt and Wynne, (2001) and Clark et al., (2005), three datasets (i.e., baseline reflectance and its two derivatives) 19  served as input for three separate iterations of forward stepwise discriminant analysis (FSDA), wherein a Wilks Lambda test (level of significance p<0.005) selected up to 40 optimal wavelengths that minimize within-species variance while maximizing betweenspecies variance. To identify the most influential regions, the spectrum was broadly partitioned into 400–500 nm (pigment absorption), 501–550 nm (chlorophyll reflection), 551–680 nm (pigment absorption), 681–740 nm (red edge transition), 741–1100 nm (biogeochemical), 1101–1400 nm (transitional) and 1401–2400 nm (biogeochemical), based on Jensen (2007). Whilst general partition characteristics may overlap, they serve as a means to investigate specific regions based on functional attributes. Wavelengths identified in the FSDA were then used as input in normal discriminant analyses (NDA), wherein accuracy was assessed through cross-validation classification using independent samples. Classification results for all three datasets were compared to determine species seperability and to assess the difference in overall and per-species accuracies.  2.3  Results  2.3.1  Statistical analysis  Baseline reflectance and its derivatives are shown in Figure 2.1. FSDA identified 40 wavelengths capable of discriminating between all species for all datasets. Table 2.2 presents spectral regions and specific wavelengths found to be significant, organized by spectral partition and dataset. The partition containing the largest number of significant  20  wavelengths is 1401–2400 nm, whereas 1101–1400 nm contains the least. Specific wavelengths selected at multiple times included 452, 514, 522, 528, 530, 546, 550, 692, 694, 696, 702, 1388, 1656, 1730, 1736 and 1742 nm, with 501–550 nm containing the highest occurrence (six). Instances of consecutive wavelengths selected multiple times (±5 nm) occurred in partitions 501–500 nm (514–530, 546–550 nm) and 681–740 nm (692–720 nm).  21  Figure 2.1. Baseline reflectance (top) and its first (middle) and second (bottom) derivatives. 22  Baseline  1st derivative  2nd derivative  Baseline  400-500 nm  460  438  442  741-1100 nm  pigment  482  452  452  biogeochemical  absorption  494  480  478  1st derivative  2nd derivative  960  754  816  1020  764  988  1050  796  498  884  501-550 nm  504  502  546  904  chlorophyll  514  514  550  974  reflection  516  522  1044  520  528  1094  522  530  1096  528  1101-1400 nm  1272  1328  1146  530  transition zone  1358  1388  1152  534  1388  1376  546  1401-2400 nm  1418  1516  1650  550  biogeochemical  1656  1520  1652  551-680 nm  564  584  552  1674  1644  1654  pigment  572  634  556  1718  1648  1656  absorption  618  648  562  1730  1664  1658  642  666  628  1780  1668  1660  678  636  1794  1706  1662  646  2106  1722  1728  681-740 nm  692  692  684  2196  1726  1730  red edge  694  702  690  2274  1736  1732  transition  696  708  694  1734  706  720  696  1736  710  730  702  1738  714  704  1740  720  712  1742  724 728 Table 2.2. Wavelengths (nm) deemed statistically significant by FSDA, organized by data set and spectral partition.  23  2.3.2  Assessing wavelength significance  Cross-validation classifications confirmed that all datasets could differentiate among species with overall accuracies ≥98% and producer’s accuracies >85% when employing 40 targeted wavelengths per dataset. Baseline data resulted in slightly higher accuracy than its derivatives. Table 2.3 presents cross-validation producer’s accuracies averaged across all datasets organized by species.  Species  Average accuracy  Black cottonwood  100  Trebling aspen  85.7  Grand fir  96.7  Western redcedar  100  Red alder  100  Douglas fir  100  Western hemlock  96.3  Bigleaf maple  100  Lodgepole pine  100  Garry oak  100  Arbutus  100  Table 2.3. Average per species NDA cross-validation classification producer’s accuracies.  2.4  Discussion  2.4.1  Performance of data sets  Highest overall accuracy was observed using baseline reflectance, indicating a potential reduction in accuracy associated with derivatives. Species with lower accuracies included  24  Trembling aspen (confused as Black cottonwood), Grand fir (confused as Douglas fir) and Western hemlock (confused as Lodgepole pine).  2.4.2  Significant discriminatory variables  Discriminant analysis indicated that spectral regions 501–550 nm and 681–740 nm were the most important for this study. More than any other partition, 501–550 nm contained the largest number of wavelengths selected multiple times, which is likely explained by influential differences in reflective interactions related to chlorophyll pigments. Region 681–740 nm contained the highest frequency of consecutive wavelengths selected multiple times, which is likely attributable to the importance of the NIR, as emphasized in most studies employing hyperspectral imagery to measure vegetative reflective properties (Gong et al., 1997, Martin et al., 1998, van Aardt and Wynne, 2001) due to species-specific differences related to the transitional red edge, internal vegetative structure and various biogeochemical constituents. In addition, the wavelength region 1401–2400 nm consistently contained the largest number of significant wavelengths. A wide variety of biogeochemical foliar constituents are responsible for influence in the short-wave infrared (SWIR), including oil, water, cellulose, lignin, starch, protein, sugar and nitrogen (Williams and Norris, 1987, Curran, 1989, Kumar et al., 2006, van der Meer, 2006). It is acknowledged that 1400–2400 nm also spans the largest spectral range, however, with the exception of three wavelengths in the baseline reflectance dataset (i.e., 2106, 2196 and 2274 nm), all wavelengths identified in this partition were within 1418–  25  1794 nm, highlighting the comparative lack of influence associated with wavelengths >1794 nm.  2.5  Conclusions  The identification of spectral regions and specific wavelengths capable of leaf-level species differentiation is a logical first step for scaling up to airborne hyperspectral data. At the canopy scale, although specific wavelengths found to be most significant may differ due to other influences, such as non-photosynthetic vegetation, shadowing effects and understorey presence (among others), these results provide the basis for an initial decision regarding specific regions of the spectrum warranting investigation with airborne data. Furthermore, species-specific spectral curves collected by ground-based spectrometers permit the identification of pure endmembers. This is information that is useful for canopy-scale investigations that are subjected to mixed pixels and a higher signal-to-noise ratio. Our results indicate spectral regions 501–550 nm and 681–740 nm were the most influential, providing a key focus for ongoing work with the airborne hyperspectral airborne imaging spectrometer for applications (AISA) data currently being used to map tree-species distribution in the GINPR. Beyond the GINPR, the methodology outlined is an important approach that can be applied to other studies where the identification of discriminatory spectral regions is required.  26  3  ASSESSING THE UTILITY OF LIDAR TO DIFFERENTIATE AMONG VEGETATION STRUCTURAL CLASSES  3.1  Introduction  Representations of vegetation structure are a critical component of effective forest ecosystem management (Lefsky et al., 2002, Yu et al., 2004, Coops et al., 2007). In Canada, structural information is typically derived from aerial photographs and augmented through field measurements (Leckie and Gillis, 1995, Gillis et al., 2005, Wulder et al., 2008a), however, resulting data often lack the detail required for certain management goals and are not easily updated (Lucas et al. 2008, Chapter 4, Chapter 5). In contrast, light detection and ranging (LiDAR) sensors measure the three-dimensional distribution of plant canopies and sub-canopy topography, providing highly accurate and detailed height, cover, and canopy structure estimates and are therefore a promising tool for improving conventional forest characterization techniques (Lefsky et al., 2002, Wulder et al., 2008b). In British Columbia (BC), Canada, Terrestrial Ecosystem Mapping (TEM) data provide baseline structural information. TEM is a hierarchical aerial photograph delineation system, augmented through field measurements, which stratifies the landscape into units based on climate, physiography, surficial material, geology, soil and vegetation. TEM structural data provide important and necessary baseline information, enabling the characterisation of floral and faunal habitat and facilitating biodiversity assessments (Green, 2007, Ecological Data Committee, 2000), however, TEM data lack detail, remain costly and subjective, and cannot be easily updated. This research assesses if three types of metrics derived from small footprint, discrete return  27  LiDAR data can differentiate among TEM defined structural classes in and around the Gulf Islands National Park Reserve (GINPR) in south-western BC. LiDAR metrics able to distinguish among structural classes would improve current forest characterization techniques, increasing detail, and reducing the time, labour and subjectivity associated with conventional measurements.  3.2  Methods  3.2.1  Study area  For a complete study area description, please consult section 2.2.1. For this chapter, TEM structural definitions partition vegetation into seven classes based on structural features and age criteria, including sparse/bryoid (SB), herbaceous (HB), shrub/herbaceous (SH), pole/sapling (PS), young forest (YF), mature forest (MF) and old forest (OF) (Table 3.1) (Ecological Data Committee, 2000). Resulting from multi-faceted and intensive anthropogenic influence, YF dominates the region, followed by MF, which in tandem represent the majority (i.e., >80%) of vegetation in the SGI (Green, 2007). HB is the third most prevalent vegetation type (>15%), followed by PS (>3.5%), SH (>2%) and OF (0.5%) (Green, 2007).  28  Structural stage herbaceous  Code HB  Tree cover <10%  Tree age (yrs) X  Stand appearance herb dominant  shrub/herbaceous  SH  <10%  X  pole/sapling  PS  dominant  20-40  very dense stands  young forest  YF  dominant  40-80  more open than PS, self-thinning evident  mature forest  MF  dominant  80-250  main canopy mature, well developed understory, advance regeneration  old forest  OF  dominant  >250  shrub dominant, regeneration often abundant  structurally complex, snags present, multi-storied  Table 3.1: Terrestrial Ecosystem Mapping (TEM) structural stage descriptions based on definitions provided by Hamilton (1988), Oliver and Larson (1990), Weetman et al. (1990), Resource Inventory: Vegetation Inventory Working Group (1995) and, BC Ministry of Forests and BC Ministry of Environment (1998).  3.2.2  TEM data  From 2003-2006, >8,000 1:10,000 (or coarser) scale polygons were derived from fine spatial resolution (i.e., 1.6 m) aerial photography providing wall-to-wall coverage of the SGI. During April-July, 2006, 704 11.3 m radial field plots were established throughout the SGI to enhance polygon-level ecosystem characterizations. Plots completely within GINPR boundaries and LiDAR surveys were targeted for analysis, isolating 141 plots representing the full range of structural classes. For additional information please consult Appendix A.  3.2.3  LiDAR data  23 LiDAR surveys encompassing ~2800 ha were acquired by Terra Remote Sensing, Inc., during July, 2006, using a TRSI Mark II two-return sensor onboard a fixed-wing platform. A mean flying height above the ground of 1600 m and a beam divergence of 0.05 mrad yielded a footprint diameter of 0.8 m. A swath width of approximately 980 km 29  resulted from the flying height (i.e., 1600 m) and a maximum scan angle of 17°. Pulse frequency (50 kilohertz (kHz)), flight speed, and altitude were optimized to achieve a nominal return spacing of one laser pulse every 1.6 m or 0.4 returns/m2. LiDAR point clouds were labeled using Terrascan v 4.006 (Terrasolid, Helsinki, Finland),which employs iterative algorithms combining filtering and thresholding methods (Axelsson, 1999, Kraus and Pfeifer, 1999) to separate the ground from objects (e.g., vegetation) and subsequently classify returns as either ground or non-ground. As described by Kraus and Pfeifer (1999), this involves: 1) generating a terrain surface to represent all points, equally weighted and forming a triangulated irregular network (TIN), 2) using the surface for averaging to determine the residuals of all z coordinates, and 3) using the residuals to adjust weights and establish thresholds to differentiate between terrain and vegetation. Categorized returns were georeferenced to Universal Transverse Mercator (UTM) zone 10 north (10N) and World Geodetic System (WGS) 84. Based on comparisons with BC Government elevation benchmarks, average horizontal and vertical accuracies of 0.5 and 0.3 m, respectively, were exhibited. For additional information please consult Appendix A.  3.2.4  Derived metrics  While a wide variety of metrics can be derived from LiDAR data, three metric types are common to the literature, 1) canopy height descriptors (CHDs), 2) height percentiles (HPs), and 3) volume profiles (CVPs) (Lim et al., 2003, Wulder et al., 2008b). For this analysis, plot-level calculations of CHDs included the mean, standard deviation, variance,  30  skewness (asymmetrical measure of LiDAR return distribution), maximum height, kurtosis (peakedness of distribution), coefficient of variation, and the harmonic mean (number of variables divided by the sum of variable’s reciprocals). HPs summarizing the distribution of all returns within pre-determined strata (Wulder et al., 2008b) provided height estimates in 5-10% intervals. CVPs were calculated based on a methodology developed by Lefsky et al. (1999b) and modified for discrete return data by Coops et al. (2007), wherein the entire volume and spatial orientation of vegetative material and related empty space found within forest canopies is characterized. This volumetric method used a predefined grid to categorize forest canopy into a three-dimensional matrix representing either “filled” or “empty” volume. Filled elements were further stratified into “euphotic” and “oligophotic” zones, with the former representing extant profile elements located within the uppermost 65% of the total energy returned from a canopy, and the latter embodying the remainder of the filled elements of the profile (Richards, 1983). Areas classified as non-filled were further separated into empty volume within the canopy (i.e., closed gap (CG) space), and empty volume above the canopy (i.e., open gap (OG) space) (Lefsky et al., 1999b). In addition, filled zones were simultaneously considered permitting an estimate for total canopy cover (TC), which was subsequently divided by closed gap (TC/CG), resulting in 25 LiDAR metrics partitioned into three suites considered for analysis.  31  3.2.5  Statistical analysis  Non-parametric Mann-Whitney U tests established if any LiDAR metrics could significantly distinguish among TEM defined structural classes. Mann-Whitney U tests iteratively compared each pair of TEM defined structural classes to each LiDAR variable, thus testing the null hypothesis that there is no difference in LiDAR metric value among structural classes. For all iterations, the significance of a LiDAR metric’s ability to differentiate among two TEM classes was assessed based on a p-value ≤0.01.  32  3.3  Results  All LiDAR variables belonging to each metric suite type (i.e., HPs, CHDs and CVPs) significantly distinguished among certain structural classes, thus rejecting the null hypothesis. Out of 15 possible combinations of structural TEM classes, certain HPs, CHDs, and CVPs permitted 12, 14 and 14 separations, respectively (Table 3.2). The importance of each metric type and the ability of individual metrics varied widely with the stage differentiation under consideration (Table 3.3). Specifically: •  All three metric types could separate HB, SH and PS from all other structural classes.  •  Only one CHD (i.e., kurtosis) and certain CVPs (i.e., CG, oligophotic, TC, TC/CG) could distinguish among YF and MF  •  Only certain CHDs (i.e., standard deviation, variance, coefficient of variation, kurtosis) could distinguish among YF and OF  •  Only one CVP (i.e., euphotic) could distinguish among MF and OF  33  HB HB  SH  PS  YF MF  SH HPs CHDs CVPs  PS HPs CHDs CVPs  YF HPs CHDs CVPs  MF HPs CHDs CVPs  OF HPs CHDs CVPs  HPs CHDs CVPs  HPs CHDs CVPs  HPs CHDs CVPs  HPs CHDs CVPs  HPs CHDs CVPs  HPs CHDs CVPs  HPs CHDs CVPs  CHDs CVPs  CHDs CVPs  OF Table 3.2: LiDAR metrics (organized by type) able to differentiate among 15 possible combinations of TEM structural classes, including: herbaceous (HB), shrub/herbaceous (SH), pole/sapling (PS), young forest (YF), mature forest (MF), and old forest (OF).  34  HB vs. SH  HB vs. PS  HB vs. YF  HB vs. MF  HB vs. OF  SH vs. PS  SH vs. YF  SH vs. MF  SH vs. OF  PS vs. YF  PS vs. MF  PS vs. OF  YF vs. MF  Height percentiles (HPs) 10th percentile 20th percentile 30th percentile 40th percentile 50th percentile 60th percentile 70th percentile 80th percentile 90th percentile 95th percentile 99th percentile Canopy height descriptors (CHDs) Mean standard deviation Variance coefficient of variation max height Kurtosis Skewness harmonic mean Canopy Volume Profiles (CVPs) Closed gap (CG) Oligophotic Euphotic Open gap (OG) Total cover (TC) TC/CG Table 3.3: The ability (gray cells), or lack thereof (black cells), of LiDAR metrics (organized by type) to significantly differentiate (p≤0.01) among 15 possible combinations of TEM structural classes, including: herbaceous (HB), shrub/herbaceous (SH), pole/sapling (PS), young forest (YF), mature forest (MF), and old forest (OF).  3.4  Discussion  Results indicate that all structural classes can be differentiated, but that the number and types of LiDAR metrics able to distinguish between particular combinations decreases with a stand’s age and complexity. As would be expected, non-forested structural classes  35  YF vs. OF  MF vs. OF  (i.e., HB, SH) were consistently differentiated as they typically exhibit uniform height ranges. Similarly, younger and less complex forest stands (i.e., PS) are generally uniform with single over or mid-stories which have representative and distinct average heights. Therefore, HB, SH and PS classes could be consistently differentiated among all structural classes using HPs. Even though PS stands exhibit variability in the range of their height values, percentiles such as the 50th, which represents close to average height, proved consistently significant. In contrast, the average heights of older, more complex forested stands are often not similar and therefore cannot be generalized, making their discernment with height percentiles a difficult task. Metrics which can better explain and capture inherent variability are more appropriate for differentiating among comparatively more complex classes. However, while certain metrics can distinguish between one pair of comparatively complex classes, that same metric often cannot distinguish among other pairs. While older and more structurally complex forested classes can only be differentiated using a few metrics, the successful differentiation among all classes is promising and has important ramifications. For instance, based on relationships with certain metrics, structural stage could be predicted for the extent of LiDAR surveys. In addition, the ability to distinguish among classes and predict them facilitates augmenting TEM data. At present, the structural detail associated with TEM polygons and/or plots is limited to a structural stage label, meaning aspects of structure within the unit are unknown. Furthermore, field work indicates that at the polygon level the structural label is sometimes partially or entirely incorrect. While certain metrics can distinguish among and predict TEM classes, other metrics could further be used to enhance within defined  36  units, providing detailed and accurate characterizations of horizontal and vertical vegetation structure. Using LiDAR surveys as samples from which to characterize forest structure would decrease the time, labor and subjectivity associated with interpretation and provide a more easily updated approach. While the coverage of LiDAR is limited to surveys, the acquisition and processing costs of small footprint, discrete return LiDAR data are estimated to be approximately the same as for aerial photographs (i.e., ~$5.00 USD per ha) (Wulder et al., 2008b). In addition, in the SGI, the costs associated with LiDAR are half as much as aerial photographs (Chapter 5). Furthermore, recent studies have demonstrated the propensity of Landsat-like data to spatially extend the coverage of geographically limited yet detailed airborne data (e.g., Wulder and Seeman, 2003). Therefore, using airborne data surveys as samples and spatially extending them with multispectral satellite imagery (e.g., Landsat data (free since 2008)) provides a cheap alternative to wall-to-wall acquisition, processing, and interpretation of airborne data.  3.5  Conclusion  While vegetation structure is critical to managing forested ecosystems, conventional methods remain unable to facilitate all management goals. This study shows the utility of LiDAR-derived HPs, CHDs, and CVPs for differentiating HB, SH, and PS from other structural types. In addition, the ability of using CHDs and CVPs for differentiating among older, more complex structural stages is demonstrated. We believe similar results can be emulated in other forested environments, however, recognize that statistical relationships generated from LiDAR data are often sensitive to the area under 37  investigation and the data parameters used (Wulder et al., 2008b). Limitations to this study included potential survey and interpretation (structural label assignment) errors associated with TEM plots, which could have been responsible for instances of misrepresented structural stage.  38  4  EXPLORING THE UTILITY OF HYPERSPECTRAL IMAGERY AND LIDAR DATA FOR PREDICTING QUERCUS GARRYANA ECOSYSTEM DISTRIBUTION AND AIDING IN HABITAT RESTORATION  4.1  Introduction  Quercus garryana (Garry oak) habitats are among the most diverse yet simultaneously rare, at-risk and degraded terrestrial ecosystems in Canada (Fuchs, 2001, GOERT, 2003). Over the last 150 years, agricultural, residential, and industrial land modifications, in conjunction with fire suppression, invasive flora, and herbivory have reduced the distribution of near natural Garry oak ecosystems to 1–5% of their pre-European settlement range and have significantly degraded remnant habitat (Fuchs, 2001, GOERT, 2003). According to the Committee on the Status of Endangered Wildlife in Canada (COSEWIC) and specialists from the British Columbia (BC) Conservation Data Centre (BCCDC), over 100 floral and faunal species are currently at-risk and several have been extirpated (Fuchs, 2001). To reverse the at-risk status of these species and avoid future extirpation, restoration efforts are critically required to restore remaining Garry oak ecosystems. Numerous definitions of effective ecosystem restoration exist. According to Hobbs and Norton (1996) and Hobbs and Harris (2001), effective restoration must consider the composition, structure, function, heterogeneity, and resilience of ecosystems. Therefore, a clear decipherable logic is required for setting goals which consider the nature of the systems targeted for restoration, the reasons for degradation, and the actions necessary for successful restoration of important attributes (Hobbs and Harris, 2001).  39  The newly formed BC Gulf Islands National Park Reserve (GINPR) and its surrounding lands are one of few places where near-natural Garry oak habitats remain (Fuchs, 2001, GOERT, 2003). The general mandate of the GINPR is to restore, conserve, and maintain ecological integrity, defined as the condition of an ecosystem or assemblage of ecosystems wherein structure and function are unhindered by anthropogenic processes and inherent biological diversity and supporting processes are likely to persist (Woodley, 1993, Parks Canada Agency, 1997). Within this mandate is the specific primary management goal of restoring degraded ecosystems and conserving rare and at-risk ecosystems (AYXS EC, 2004, Green, 2007). According to Hobbs and Harris (2001), the restoration of degraded ecosystems requires broader contextual placement within sustainable land-use and conservation. Therefore, because restoring remaining habitats is a fundamental component of conservation, restoration and conservation are inextricably linked.  Within the GINPR and its surrounding lands, Garry oak habitat has been recognized as degraded, rare and at-risk and as a result is a priority for restoration and conservation. Single ecosystems do not represent the full range of regional habitat variation, nor can they be used to assess regional dynamics (Meinke et al., 2008). In contrast, focusing on an assemblage of ecosystems elucidates the significance of broad-scale processes and interactions (Hobbs, 2002). Management decisions made within one ecosystem impact others and as habitat assemblages become increasingly modified and fragmented, a loss of biotic connectivity occurs, significantly altering broad-scale processes and interactions. To meet restoration and conservation goals, efforts need to ensure that not only are individual degraded ecosystems targeted for restoration, but that restoration also 40  occurs within a broader landscape context which reestablishes adequate ecosystem mosaic size and connectivity (Hobbs and Harris, 2001, AXYS EC, 2004). To meet these goals, the extent of Garry oak distribution needs to be understood, which requires appropriate contemporary reference data, knowledge of past conditions (i.e., historic reference data) and a practical vision for future conditions. It is therefore critical that pertinent reference information is available as it is fundamental for effective restoration (White and Walker, 1997).  Suitable contemporary reference data representing Garry oak distribution have been identified by park managers as a primary knowledge gap (AXYS EC, 2004). Distributional information (circa 2003–2006) is currently provided by 1:10,000 scale (or coarser) terrestrial ecosystem mapping (TEM) data derived from a hierarchical classification system that stratifies aerial photographs based on physiography, climate, geology, surficial material, soil and vegetation. Despite the importance of this baseline information, due to its minimum mapping unit (mmu) of 0.04 ha (i.e., 400 m2), the final product lacks spatial detail and accuracy. As a supplement to baseline data, advanced geotechnologies can provide managers with increasingly detailed and accurate species distributive information (i.e., contemporary reference information).  Two examples of advanced remotely sensed data types gaining widespread use in ecological applications include hyperspectral and light detection and ranging (LiDAR) data. Hyperspectral remote sensors acquire image data simultaneously in upward of hundreds of narrow, adjacent spectral channels with each picture element containing a complete reference spectrum (Goetz, 1992, Smith, 2006). Due to their fine spatial and  41  spectral resolution, hyperspectral sensors can measure subtle absorption features related to various biogeochemical properties which enable vegetative discrimination (Cochrane, 2000, Ustin et al., 2004) allowing effective mapping of species distributions (Martin et al., 1998, Clark et al., 2005, van Aardt and Wynne, 2007, Carlson et al. 2007, Yang et al., 2009). LiDAR, a separate yet equally important technology generates active pulses of near infrared (NIR) light which measures the vertical distribution of vegetation and can characterize vertical forest structure (Wulder et al., 2008b). Detailed forest structural information can complement passive spectral information and facilitate comprehensive forest characterization (Gillespie et al., 2004, Hill and Thomson, 2005, Koetz et al., 2007).  This research aims to provide GINPR managers with appropriate contemporary reference data depicting Garry oak distribution (circa 2006) which can be used to meet the primary management goal of restoring individual degraded ecosystems and works toward reestablishing and conserving the greater ecosystem mosaic. In doing so, this research establishes whether classified hyperspectral imagery can provide Park managers with predictions of Garry oak distribution that are more detailed and accurate than those resulting from conventional aerial photograph interpretation. In addition, this research determines if the combination of spectral and structural information derived from hyperspectral and LiDAR data, respectively, results in more accurate predictions than when relying solely on Garry oak spectral properties.  42  4.2  Methods  4.2.1  Study area  For a complete study area description, please consult section 2.2.1. The area of interest (AOI) for this chapter is a methodological test site including portions of two islands totaling approximately 600 ha for which hyperspectral and LiDAR data were collected in transects (Figure 4.1). On these islands, Garry oak are known to occur within three types of areas: rocky bluffs with woodland patches, fields with woodland patches, and steep slope woodlands (Green, 2007).  43  Figure 4.1: The AOI includes portions of two islands in the SGI (latitude 48.76o, longitude -123.18o) totalling approximately 600 ha, over which hyperspectral and LiDAR data were collected in transects during July, 2006. Transects were collected both in and around GINRP properties. Within transect boundaries, plot centers and polygon centroids representing the location of approximately 200 tree/tree clusters used as reference (i.e., calibration and validation) data are shown. The background image is a Landsat TM NIR band collected in July, 2006.  44  4.2.2  Remotely sensed data  The Airborne hyperspectral data were collected in 23 transects (Figure 4.1) in mid July, 2006, close to solar noon, by Terra Remote Sensing, Inc. (Sidney, BC, Canada), using an Airborne Imaging Spectrometer for Applications (AISA) Dual push-broom sensor (Spectral Imaging Ltd.) on a fixed-wing platform flown at an altitude of 1600 m yielding a spatial resolution of 2×2 m pixels. The AISA Dual sensor facilitates the simultaneous acquisition of hyperspectral data from 350 to 2500 nm captured in 492 narrow spectral channels. Data were delivered georeferenced to Universal Transverse Mercator (UTM) zone 10 north, World Geodetic System (WGS) 1984 using a 2-m LiDAR derived digital elevation model (DEM) and have an estimated horizontal accuracy of ±10 m. LiDAR were collected concurrently with hyperspectral data from the same platform. For complete details concerning LiDAR data, consult section 3.2.3. The AISA data were delivered by the vendor georectified to the LiDAR. In order to assess and verify the accuracy of the co-registered, simultaneously collected hyperspectral/LiDAR data, 50 control points, randomly located along roads and/or the coast line, were selected on both the AISA imagery and a LiDAR canopy height model (CHM). A simple least squares adjustment indicated averaged error between the two was within 3 m, equal to 1.5 pixels. This level of co-registration was deemed to be suitable and was well within the smallest average crown size. For this chapter, the area of interest concerned portions of three transects (Figure 4.1). For additional information please consult Appendix A.  45  4.2.3  Remotely sensed data pre-processing  Effective utilization of remotely sensed data requires atmospheric correction, which minimizes atmospheric effects associated with particulate and molecular scattering and rescales raw radiance data to reflectance values (Felde et al., 2003, van der Meer et al., 2006). The Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) atmospheric correction code derived from the MODTRAN4 radiative transfer code was utilized to perform atmospheric correction for all 23 hyperspectral transects. FLAASH input parameters included the wavelength and full-width-half-maximum (FWHM) sensor specifications, image acquisition information, and sensor altitude. Site specifications included geographic location, elevation, and atmospheric model. For this dataset, local visibility was determined to be good and a maritime atmospheric model was applied. In addition, known water absorption bands (i.e., 1350-1416 and 1796-1970nm (PalaciosOrueta & Ustin, 1996, Price, 1998, van Aardt and Wynne, 2001)) and regions exhibiting extreme levels of spectral noise (i.e., 395-429 and 2401-2500nm) were removed, reducing the AISA dataset to 453 channels. For additional information please consult Appendix A. Even after atmospheric correction has occurred, the large number of spectral bands in hyperspectral datasets results in highly correlated spectral channels (Smith, 2006). Therefore, when utilizing hyperspectral imagery for mapping purposes, reducing data volume through the identification of task specific optimal wavelengths is a fundamental preprocessing step. For this chapter, the set of 453 spectral channels was further reduced to 40 key wavelengths, which represented the major portions of the electromagnetic  46  spectrum based on findings in Chapter 2. Spatial masking was used to remove scattered non-vegetative elements from the imagery (i.e., buildings). To combine the hyperspectral imagery with vertical forest information, LiDAR derived products were employed. For all 23 LiDAR transects, returns classified as ground returns were used to develop a 2 m spatial resolution DEM using the natural neighbor algorithm (Sibson, 1981, Sambridge et al., 1995, Abramov and McEwen, 2004). Once the DEM was derived, ground elevation values were subtracted from the height values of the nonground returns to model canopy height. The CHM was used to remove AISA pixels <5 m as this height represented a threshold for sunlit canopy based on average field measurements and local expert opinion. The spectral pixels remaining after masking were assumed to characterize the reflectance signatures of unobstructed tree canopies, which in theory should most accurately spectrally represent targeted species and are best suited for species discrimination (Asner et al., 2008, Blackburn, 2002). In addition to metrics describing height and its variation, three-dimensional volumetric canopy profiles were derived. Based on a method originally introduced for full waveform LiDAR data by Lefsky et al. (1999b) and successfully applied to discrete-return data by Coops et al. (2007), Chapter 4, and Chapter 5, vegetation volume profiles characterize the volume and spatial orientation of both canopy vegetation and open space. Vegetation is categorized based on three-dimensional matrices, comprised of 2 m cells within 6 m windows and 50 m vertical bins. For each cell, the non-ground LiDAR returns within the window and vertical bin were categorized as either “filled” or “empty” volume, wherein filled elements were further separated into “euphotic” and “oligophotic” zones. Euphotic zones represent the totality of elements found within the uppermost 65.0% of the total energy  47  returned from a canopy, whereas oligophotic zones embody the remainder of filled elements within a profile (Richards, 1983). Non-filled areas were further distinguished as unoccupied canopy (i.e., closed gap space) and vacant non-canopy (i.e., open gap space).  4.2.4  Field program  Forest inventory data were employed to produce a stratified random sampling design. Plot locations spanned the full range of tree species and forest age and structural complexity. A total of forty-nine 20 m radius plots were established. In each plot, the location of species-specific trees and/or tree clusters ≥5 m in height were recorded using Trimble global positioning system (GPS) and delineated on masked hyperspectral AISA imagery with a global accuracy of ±1.5 m. TEM data definitively representing the location of particular species (i.e., geographic information system (GIS) polygons composed of 100% pure species), although small in number, were isolated, rasterized, and masked to conform to AISA imagery. Species-specific identifiable trees/tree clusters were then selected from targeted TEM polygons, which in combination with plot data resulted in approximately 200 tree/tree clusters representing nine species distributed throughout the AOI (Figure 4.1). For each species, pixels comprising two-thirds of tree/tree clusters were randomly selected as calibration data, with the remaining one-third of pixels put aside as independent validation data (Table 4.1). For additional information please consult Appendix A.  48  Taxonomic group Broadleaf  Conifer  Species Trembling aspen Red alder Bigleaf maple Garry oak Arbutus Western redcedar Douglas-fir Western hemlock Lodgepole pine  Scientific name Populus tremuloides Alnus rubra Acer macrophylum Quercus garryana Arbutus menziesii Thuja plicata Pseudotsuga menziesii Tsuga heterophylla Pinus contorta  Abbreviation At Dr Mb Qg Ra Cw Fd Hw Pl  Calibration 72 2300 55 353 54 475 3520 68 19  Validation 45 963 31 213 32 312 2634 51 10  Table 4.1: Proportion of calibration and validation data per species in pixels.  4.2.5  Classification  A support vector machine (SVM) classifier was applied, which is a nonparametric algorithm that employs a kernel function to map data from spectral space into higher dimensional feature space wherein an N-dimensional hyperplane is used to separate continuous predictor variables into predefined categorical target variables (Wu et al., 2004, Hsu et al., 2010). Using SVMs to classify hyperspectral imagery has seen increasing interest, as studies have shown improvements in classification accuracies, computational time, and stability as compared with results yielded by conventional parametric classification algorithms (e.g., minimal-distance-to-means) as well as nonparametric algorithms (e.g., neural networks, Knn classifier) (Melgani and Bruzzone, 2004, Bruzzone and Carlin, 2006). Two rounds of SVM classification were undertaken, one without and one with the inclusion of LiDAR CVPs as input.  49  4.2.6  Accuracy assessment  For both rounds of classification, accuracy was assessed using a confusion matrix which is a simple cross-tabulation of the mapped class label against validation data for an independent sample of cases from specified locations (Foody, 2002). This matrix measures the percentage of cases correctly allocated overall (i.e., overall accuracy), the Kappa index of agreement (KIA), and the percentage of cases correctly allocated per class (i.e., producer’s and user’s accuracies) (Congalton and Green, 1999).  4.3  Results  Figure 4.2 shows the mean reflectance of reference data for all species. All species curves exemplify typical reflectance properties for healthy green vegetation, experiencing a slight peak at approximately 550 nm (chlorophyll pigment reflection), a sharp peak from approximately 700–1000 nm (red edge transitional zone), and multiple deep troughs in the short wave infrared (water related absorption). Garry oak exhibits the third highest reflectance at 550 nm, surpassed by Bigleaf maple and Arbutus. In the red edge transitional region, Garry oak reflectance is generally fifth highest behind all other broadleaved species. Throughout the short-wave infrared (SWIR), with the exception of Trembling aspen, Garry oak typically has the lowest reflectance of broad-leaved species. Classification without the inclusion of LiDAR CVPs resulted in an overall accuracy of 86.4% with a 0.8 KIA (Table 4.2). Six out of nine species were classified with reasonable to high producer’s (i.e., ≥65.1%) and user’s accuracies (i.e., ≥70.2%), whereas Western 50  hemlock, Bigleaf maple, and Lodgepole pine were not classified with acceptable producer’s (i.e., 49, 6.5, and 0%, respectively) or user’s accuracies (23.8, 28.6, and 0%, respectively). In contrast, classification including LiDAR CVPs resulted in an overall accuracy of 87.2% with a 0.8 KIA (Table 4.3). Six out of nine species were also classified with reasonable to high producer’s (i.e., ≥67.9%) and user’s accuracies (i.e., ≥74.7%), excluding Western hemlock, Bigleaf maple, and Lodgepole pine (i.e., 54.9, 6.5, and 10% producer’s and 45.9, 50, and 8.3% user’s). Focusing on Garry oak, classification without inclusion of LiDAR CVPs resulted in a producer’s accuracy of 81.7% and user’s accuracy of 92.1% (Table 4.2). The species primarily responsible for confusion were Lodgepole pine and Arbutus (commission error) and Douglas-fir and Red alder (omission error) (Figure 4.3). In contrast, classification with the inclusion of LiDAR CVPs resulted in a producer’s accuracy of 86.9% and user’s accuracy of 81.5% (Table 4.3). Lodgepole pine and Bigleaf maple were primarily responsible for commission error whereas Douglas-fir and Arbutus were the main source for omission error (Figure 4.4). Focusing on the AOI, comparisons with TEM data elucidate a discrepancy in the total number of ecological units (i.e., TEM polygons) said to contain Garry oak (Figure 4.5) and the total area in hectares occupied by Garry oak within these polygons (Figures 4.6 and 4.7). Within GINPR boundaries, existent TEM data inventory 21 polygons containing approximately 2.9 ha of Garry oak. In contrast, SVM classification without LiDAR CVPs indicates Garry oak occurs within 96 polygons wherein 1.4 ha comprises total coverage. Classification results with LiDAR CVPs indicate 113 polygons containing 2.2 ha of Garry oak. Outside the boundaries of the GINPR, TEM data inventory 16  51  polygons on private lands containing approximately 0.95 ha of Garry oak, whereas SVM classification without LiDAR CVPs indicates 255 polygons containing 3.5 ha of Garry oak, and classification with LiDAR CVPs indicates 217 polygons with 3.1 ha of coverage.  Figure 4.2: Mean hyperspectral reflectance of reference (i.e., calibration and validation) data for all species, collected by the AISA Dual sensor in July, 2006.  52  At Cw Dr Fd Hw Mb Pl Qg Ra Producer's accuracy Omission error  At 71.1 0 26.7 2.2 0 0 0 0 0 71.1 28.9  Cw 0 65.1 8.7 26 0 0 0 0.3 0 65.1 34.9  Dr 0.4 2.3 87.7 6.5 2.4 0 0.1 0.5 0 87.7 12.3  Fd 0.2 2.4 3.4 91.2 2.2 0.2 0 0.2 0.2 91.2 8.8  Hw 0 2 0 49 49 0 0 0 0 49 51  Mb 0 0 29 64.5 0 6.5 0 0 0 6.5 93.5  Pl 0 0 40 40 0 0 0 20 0 0 100  Qg 0 0 3.8 14.1 0 0 0 81.7 0.5 81.7 18.3  Ra 0 0 3.1 25 0 0 0 3.1 68.8 68.8 31.3  User's accuracy 76.2 70.2 84.8 91.2 23.8 28.6 0 92.1 81.5 Overall accuracy KIA  Commission error 23.8 29.8 15.2 8.8 76.2 71.4 100 7.9 18.5 86.4 0.8  Table 4.2: Confusion matrix for SVM classification without the inclusion of LiDAR CVPs as classification input. Excepting KIA, accuracy statements are provided in percentage.  At Cw Dr Fd Hw Mb Pl Qg Ra Producer's accuracy Omission error  At 77.8 0 0 22.2 0 0 0 0 0 77.8 22.2  Cw 0 67.9 1.6 30.4 0 0 0 0 0 68 32.1  Dr 0.3 3.3 84.5 8.3 2 0.2 0 1.3 0 84.5 15.5  Fd 0.2 1.5 3.9 92.7 0.5 0 0.4 0.8 0.1 92.7 7.3  Hw 0 0 0 45.1 54.9 0 0 0 0 54.9 45.1  Mb 0 0 6.5 64.5 3.2 6.5 0 19.4 0 6.5 93.6  Pl 10 0 0 50 0 0 10 20 10 10 90  Qg 0 0 0 11.3 0 0 0.5 86.9 1.4 86.9 13.2  Ra 0 0 0 21.9 0 0 0 0 78.1 78.1 21.9  User's accuracy 81.4 74.7 88.1 90.2 45.9 50 8.3 81.5 80.7 Overall accuracy KIA  Commission error 18.6 25.4 11.9 9.8 54.1 50 91.7 18.5 19.4 87.2 0.8  Table 4.3: Confusion matrix for SVM classification with the inclusion of LiDAR CVPs as classification input. Excepting KIA, accuracy statements are provided in percentage.  53  Figure 4.3: Species primarily responsible for Garry oak classification error without the inclusion of LiDAR CVPs as classification input.  Figure 4.4: Species primarily responsible for Garry oak classification error with the inclusion of LiDAR CVPs as classification input.  54  Figure 4.5: With focus on the AOI, the total number of polygons inside and outside of GINPR boundaries occupied by Garry oak according to existent ecological data (TEM) versus SVM classification results, with and without CVPs as input.  Figure 4.6: With focus on the AOI, the approximate total hectares inside and outside of GINPR boundaries occupied by Garry oak according to existent ecological data (TEM) versus SVM classification results, with and without CVPs as input.  55  Figure 4.7: Comparison of existent Garry oak locations (TEM polygons) with SVM classification results with and without CVPs as input for three distinct Garry oak habitat types (i.e., rocky bluff with woodland patch, steep slope woodland, and field with woodland patch).  56  4.4  Discussion  Although tree species classes cannot be fully described by a single reflectance spectrum due to within-class variation, the spectral characteristics indicate Garry oak can be spectrally distinguished from all other species at numerous wavelengths. Due to unique spectral characteristics, visual comparisons of the classified maps with orthoimagery and existing TEM data supplemented by expert opinion from Parks staff confirm SVM classifications provide significantly improved distribution predictions for all three Garry oak habitat types. A fundamental difference in detail and accuracy between TEM and SVM results relates to the difference between the spatial resolution of the hyperspectral and LiDAR imagery versus the mapping scale inherent to TEM data. TEM data consist of 1:10,000 scale (or coarser) polygons with a mmu of 0.04 ha (i.e., 20×20 m), whereas SVM classifications provide 2 m spatial resolution. It is clear that TEM data often fail to acknowledge Garry oak presence inside polygons where it is known and/or expected to exist. Furthermore, even if a polygon is said to contain Garry oak, the precise amount and location of trees and/or tree clusters are unknown. Therefore, SVM results can be used to locate new areas within the GINPR and outside Park boundaries that contain Garry oak and increase the spatial explicitness of all polygons containing Garry oak. Despite the overall success of both rounds of SVM classification, their differences are apparent. The primary source of classification error without the inclusion of LiDAR CVPs is a discrepancy in the quality and quantity of species-specific reference data. For Garry oak, numerous isolated trees and/or tree clusters were easily located and discernable in the hyperspectral imagery and thus delineated as reference with a high level of confidence. For other species, chiefly Lodgepole pine and Arbutus, single trees 57  and/or tree clusters were often found in overlapping proximity to other species, and thus, delineating pure crowns was more challenging. These comparatively poorly defined species (e.g., Lodgepole pine and Arbutus) are the main source of classification error for well defined classes (e.g., Garry oak). The fusion of hyperspectral imagery with LiDAR CVPs results in increased overall classification accuracy, increased producer’s accuracies for seven species (excepting Red alder and Bigleaf maple) and higher user’s accuracies for six species (excepting Douglasfir, Arbutus, and Garry oak). Although the majority of per-class accuracies are higher, for Garry oak, despite raising producer’s accuracy, the inclusion of vertical forest characterization lowered user’s accuracy. When analyzing the sources of commission error (i.e., user’s accuracy) associated with Garry oak for both rounds of classification, the main difference is attributable to Bigleaf maple. Without LiDAR CVPs, Bigleaf maple does not contribute to Garry oak commission error, whereas with LiDAR CVPs, commission error associated with Bigleaf maple is second only to Lodgepole pine and clearly solely responsible for the reduction in user’s accuracy. Garry oak, although known to occur in close proximity to other species, are typically isolated trees or tree clusters. Similarly, Bigleaf maple are sometimes found as isolated trees/tree clusters growing in meadows where they are easily discernible, however, they are typically found in mixed, overlapping forest stands which are not easily discernable in high spatial resolution hyperspectral imagery. Furthermore, when found in meadows, the canopy architecture of Garry oak and Bigleaf maple shares more similarities than any other two broad-leaved species found in the study area. Structural similarities with Garry oak in combination with an underrepresentation of the range of structural conditions resulted in  58  numerous instances of Bigleaf maple being misclassified as Garry oak. This is not a reflection on the inability of the methodology to accurately locate Garry oak (as indicated by producer’s accuracy), but rather, on the comparative inability of the methodology to differentiate Bigleaf maple from Garry oak in instances when structural characteristics were too similar, the full range of structural conditions had not been fairly represented, and a disparity existed in the amount of and quality of Bigleaf maple specific calibration and validation data. With or without the inclusion of LiDAR CVPs, SVM results directly facilitate filling the fundamental knowledge gap of understanding Garry oak ecosystem distribution in an accurate and spatially explicit manner by serving as contemporary Garry oak habitat reference. It has been argued that using contemporary reference data as a basis to revert habitats to their pre-European state is potentially not feasible due to the dynamism of measurable ecosystem characteristics which vary on a variety of spatiotemporal scales (Hobbs and Harris, 2001). Furthermore, the concept of alternative stable states implies that there may not even exist one correct state to revert back to (Hobbs and Harris, 2001, Hobbs, 2007), however, contemporary reference data can be combined with existent ecological data (i.e., TEM) and historical archives to provide inference into management interventions that can most effectively restore degraded systems (Yates et al., 1994, Noss 1996, Hobbs and Harris, 2001) and establish a foundation from which pertinent restoration activities can be measured against (White and Walker, 1997) and as a result, its use is practical and necessary to guide restoration. Within the context of providing distributive information, contemporary reference data address the key conservation goal of identifying unprotected components of ecosystem  59  mosaics which works toward the primary management goal of increasing the total area and connectivity of protected Garry oak ecosystems through land acquisition (AYXS EC, 2004, Green, 2007). In total, >3 ha of additional private lands occupied exclusively by Garry oak spread out within >216 TEM polygons have been identified through SVM results. This spatially explicit information serves Park managers directly by identifying geographically specific ecological units (i.e., TEM polygons) to target for acquisition. If private lands are not available, but are adjacent to GINPR properties, conservation easement arrangements can be explored. When landscape scale habitat connectivity is reestablished, finer scale within ecosystem restoration activities can commence and continue in a more informed manner. Before restoration activities commence, the current state of a system needs be rigorously assessed with underlying factors understood (Hobbs and Harris, 2001). Although not able to provide appropriate Garry oak distributive information, work by AYXS EC (2004) and Green (2007) and TEM data present a clear picture of necessary restoration activities, many of which are ongoing. The types of restorative measures in Garry oak habitats typically involve relatively minute alterations of management practices or species composition, including removal of non-native plants, reintroduction of native plants, and prevention of herbivory (Whisenant, 1999, GOERT, 2003, AXYS EC, 2004, Green, 2007). In addition, due to prolonged fire suppression, many habitats have experienced a change in abiotic attributes, and as a result, prescribed burns are being integrated into restoration practices. In aggregate, restorative measures aim to represent extant Garry oak habitat types and in doing so enable reestablishment of a landscape scale mosaic. It is accepted that within this range of restorative measures, there exist a variety of potential  60  short and long-term outcomes, but focus, within a historical context, must remain on desired attributes for systems in subsequent years rather than dwelling on specifically what existed in the past (Pfadenhauer and Grootjans, 1999). SVM results serve as the missing contextual backdrop to continue restoration activities in a more informed manner and allow the targeting of as yet unmanaged individual ecosystems. This methodology has pertinence for any degraded habitat which is associated with a floral species identifiable in high spatial and spectral resolution airborne hyperspectral imagery. The methodology presented is not limited to the SGI or to tree species of the Pacific Northwest. The approach outlined can be replicated so long as the species of interest are directly visible to a hyperspectral sensor and are spectrally discernable from one another. Given the dearth of reliable, accurate, and detailed contemporary reference data associated with countless degraded habitats worldwide as a result of reliance on conventional aerial photograph interpretation techniques, this approach has serious implications to land managers tasked with restoring degraded ecosystems. The caveat to this statement is that at present, advanced geotechnologies such as hyperspectral and LiDAR data remain prohibitively expensive for many interested parties, however, as costs go down and availability goes up, access to these data will increase and this will benefit restoration programs in a variety of degraded ecosystems throughout the world.  4.5  Conclusion  Regardless of whether LiDAR CVPs are incorporated into classification, SVM results provide a vastly improved inventory of the amount and location of Garry oak trees and  61  tree clusters in the AOI. This contemporary reference data can be combined with existent ecological data and historical archives to establish baseline reference sites from which restoration activities are themed around and assessed in relation to. Furthermore, distributional Garry oak data act as a surrogate for 100+ species of at-risk flora and fauna. Based on the single date success of this methodology, periodic acquisition of fine spatial and spectral resolution airborne remotely sensed data can provide a means to update species distributive maps. For Garry oak, the difference in SVM results is attributable to structural similarities shared with Bigleaf maple, as well as a lack in the quality and quantity of Bigleaf maple calibration and validation data. Because overall accuracy, Garry oak producer’s accuracy, and most per-class accuracies (producer’s and user’s) increase with the addition of LiDAR CVPs as classification input their use is recommended, however, special care must be taken to ensure an adequate amount of high quality reference data is collected for species which are spectrally or structurally similar to the target species. Advanced geotechnologies can characterize Garry oak habitat with more accuracy and spatial explicitness than is possible using conventional aerial photograph interpretation (i.e., TEM). The information supplied by SVM results serves as invaluable contemporary reference, informing and providing a model for ecosystem level restoration which facilitates the reestablishment of broader scale habitat connectivity and subsequent conservation of the landscape scale mosaic. This approach can be replicated for other species and in other geographic locations, so long as the surrogate species (e.g., Garry oak) are identifiable in hyperspectral imagery (i.e., canopy species) and special attention is paid to spectral and structural similarities exhibited between species.  62  4.5.1  •  Implications for practice The use of advanced geotechnologies results in detailed and accurate species distribution maps applicable to a wide variety of conservation and restoration oriented managerial goals.  •  Species distribution maps guide land acquisition and/or establishment of conservation easements in as of yet unmanaged ecological units.  •  Species distribution maps serve as contemporary reference data, which help to establish baseline reference from which restoration activities can be themed and assessed in relation to.  •  Baseline contemporary reference allows for restoration of landscape scale connectivity and permits finer scale within ecosystem restoration activities to commence and/or continue.  •  Restorative measures on a per-ecosystem basis in aggregate serve to reestablish the range of ecosystem types and directly impact numerous species of flora and fauna.  •  The methodology outlined in this research is replicable for land managers tasked with restoring a wide variety of degraded ecosystems.  63  5  ASSESSING THE UTILITY OF AIRBORNE HYPERSPECTRAL AND LIDAR DATA FOR SPECIES DISTRIBUTION MAPPING IN THE COASTAL PACIFIC NORTHWEST, CANADA  5.1  Introduction  Accurately characterizing tree species distribution is critical for the effective management of forested ecosystems (Gong et al., 1997, Plourde et al., 2007). Species-level maps directly influence policy, guide subsequent implementation, and provide a reference point from which change can be quantified and management decisions evaluated (Dalponte et al., 2008, Innes and Koch, 1998, Voss and Sugumaran, 2008). Two promising remote sensing techniques for species mapping involve hyperspectral and Light Detection and Ranging (LiDAR) sensor technology. Owing to an extremely fine spectral and spatial resolution, hyperspectral sensors can measure subtle absorption features related to biogeochemical properties, enabling improved vegetative discrimination as compared with established multispectral methods (Cochrane, 2000, Ustin et al., 2004). As a result, hyperspectral data have successfully been used to map tree species within tropical (e.g., Carlson et al., 2007, Clark et al., 2005), sub-tropical (e.g., Dennison and Roberts, 2003, Lucas et al., 2008, Yang et al., 2009), and temperate (e.g., Boschetti et al., 2007, Goodwin et al., 2005, Martin et al., 1998, Plourde et al., 2007, Xiao et al., 2004) ecozones. However, hyperspectral imagery is generally restricted to the horizontal plane, providing limited insight pertaining to the vertical distribution of forest structure. In contrast, terrestrial LiDAR sensors generate active pulses of near-infrared (NIR) light which can measure ground elevation (Ackermann, 1999, Wehr and Lohr, 1999) and  64  vertical vegetation distribution (Drake et al., 2002, Lefsky et al., 1999a). With many authors demonstrating the value of LiDAR measurements for representing detailed forest canopy structural properties, see reviews by Lim et al. (2003) and Wulder et al. (2008b). Combining information acquired by these disparate sensors is thought to hold great promise for enhancing comprehensive canopy characterization (Gillespie et al., 2004, Hill and Thomson, 2005) and improving forest inventory (Anderson et al., 2008), particularly at the species-level (Plourde et al., 2007). To date, the synergistic use of hyperspectral/LiDAR has proven successful for a variety of applications, including shadow, height, and gap related masking techniques (Asner et al., 2008, Blackburn, 2002, Voss and Sugumaran, 2008), crown identification (Asner et al., 2008), illumination geometry calculations (Asner et al., 2008), above-ground biomass estimates (Lucas et al., 2008), quantifying riparian habitat structure (Hall et al., 2009), and fuel type mapping (Koetz et al., 2008). In addition, the combination of hyperspectral/LiDAR has been shown to enhance image segmentation capabilities and subsequently guide object-based classification (e.g., Geerling et al., 2009, Hill and Thomson, 2005, Voss and Sugumaran, 2008). Furthermore, numerous studies specifically investigating vegetation type classification (e.g., Geerling et al., 2007, Hill and Thomson, 2005, Koetz et al., 2008) and/or species-level discrimination (e.g., Andrew and Ustin, 2009, Asner et al., 2008, Lucas et al., 2008, Mundt et al., 2006, Voss and Sugumaran, 2008) have benefitted from hyperspectral/LiDAR fusion. The fusion of hyperspectral and LiDAR data can occur using several approaches. Pohl and van Genderen (1998) define three fusion methods: 1) decision-, 2) feature-, and 3) pixel-level. At the decision-level, datasets are processed independently, with end results  65  integrated in a GIS. At the feature-level, characteristics are identified and represented through combining information from multiple datasets. At the pixel-level, datasets are directly fused and processed simultaneously for end results (Geerling et al., 2007, Pohl and van Genderen, 1998). To date, most fusion studies have occurred at the decision and/or feature-level, with few explicitly involving pixel-level fusion of hyperspectral data with rasterized layers representing vertical vegetative characteristics, available directly to classify vegetation cover-type (e.g., Geerling et al., 2007, Koetz et al., 2008) and/ or species-level distributions (e.g., Dalponte et al., 2008). These studies confirmed that using LiDAR-derived height information directly as classification input increases accuracy for classes exhibiting different average heights (Dalponte et al., 2008, Geerling et al., 2007, Koetz et al., 2008). In contrast to metrics representing height strata and/or summarizing variation in height, three-dimensional canopy volume profiles (CVPs) characterize volumetric canopy architecture (Lefsky et al., 1999b). Originally introduced by Lefsky et al. (1999b), CVPs can distinguish species specific volumetric canopy architecture, potentially acting as strong explanatory variables. To the best of our knowledge, no studies have assessed the utility of using volumetric variables as direct classification input. In this study we assess the capacity of using targeted Airborne Imaging Spectrometer for Applications (AISA) Dual airborne hyperspectral bands for mapping 11 tree species within the coverage of 23 transects covering ~2800 hectares (ha) in and around the Canadian Gulf Islands National Park Reserve (GINPR). In addition, the impact of fusing AISA data at the pixel-level with forest height and volume information derived from concurrently collected, small footprint, multi-return LiDAR data is assessed. Accuracy  66  metrics of classifications with and without LiDAR-derived inputs are compared. Lastly, the significance of observed differences exhibited between classifications is quantified, and the costs and benefits of this new approach in relation to an existing forest inventory method are considered.  5.2  Methods  5.2.1  Study area  For a complete study area description, please consult section 2.2.1. Resulting from multifaceted and intensive anthropogenic influence, numerous forested ecosystem types have become either rare and/or at-risk. For most tree species, accurate and up-to-date distributional information remains incomplete. Despite increasing anthropogenic influences, many islands, although once heavily logged, remain undeveloped, inaccessible by road, and difficult to traverse on foot. As a result, conventional forest inventory has relied primarily on aerial photograph interpretation (API) which provides 1:10,000 (or coarser) scale polygons mapped with a minimum unit of 0.04 ha (i.e., 400 m2) (Green, 2007). Despite its common use, API is time consuming and yields subjective outcomes with limited applicability for rapid assessment over extensive areas (Anderson et al., 1993, Evans et al., 2006, Gillis et al., 2005, Gong et al., 1997, Leckie et al., 2005, Lucas et al., 2008). To provide managers with comparably priced yet more detailed and accurate baseline species distribution information, the development and implementation  67  of alternative approaches involving advanced techniques for remotely quantifying forest characteristics is critical.  5.2.2  Remotely sensed data  Hyperspectral data were collected in 23 transects (Figure 5.1) according to specifications outlined in section 4.2.2. LiDAR were collected concurrently with hyperspectral data from the same platform. For details concerning LiDAR acquisition, consult section 3.2.3. For additional information please consult Appendix A.  68  Figure 5.1: The area of interest (AOI) includes ~2800 ha encompassed by 23 transects of hyperspectral and LiDAR data, concurrently collected in July, 2006 in and around the Gulf Islands National Park Reserve (GINPR), British Columbia (BC), Canada (lat 48.76o, long -123.18o). Within the extent of transects the location of 411 tree/tree clusters representing 11 species are shown. Tree/tree cluster information was used to train and assess accuracy for a series of support vector machine (SVM) classifications. The background image is the near-infrared band (channel 4) from a Landsat Thematic Mapper (TM) image acquired July, 2006. 69  5.2.3  Field program  In July, 2007 a stratified random sampling design was implemented wherein suitable plot locations were initially stratified by island and filtered by land tenure. Within island strata plots were further stratified based on API-derived tree species distribution supplied by Parks Canada. Within overlapping strata plots were randomly selected, factoring in island size and aiming to relatively represent the proportion of the landscape estimated by Parks Canada to be occupied by each species. A total of 150 20 m radial plots were established. In each plot the perimeters of all trees/tree clusters observable on the AISA imagery and ≥5 m in height were delineated. The coordinates of all trees/tree clusters were recorded using post-differentially corrected Trimble GPS measurements, accurate within ±1.5 m. In a subset of plots, vegetation samples were collected for all species. In addition to the 150 plots, polygons composed purely of single species were selected from the API-derived inventory. Factoring in the same strata, and in accordance with plot dimensions, segments of species-specific polygons were used to augment field-collected reference. As a result, a total of 411 tree/tree clusters representing 11 species were located (Figure 5.1). For each species, two-thirds of trees/tree clusters were randomly selected as training data, with the remaining reference used to independently assess accuracy (Table 5.1). For additional information please consult Appendix A.  70  Taxonomic group Broadleaf  Conifer  Common name black cottonwood trembling aspen red alder bigleaf maple Garry oak arbutus grand fir Western redcedar Douglas-fir Western hemlock lodgepole pine All species  Scientific name Populus balsamifera Populus tremuloides Alnus rubra Acer macrophylum Quercus garryana Arbutus menziesii Abies grandis Thuja plicata Pseudotsuga menziesii Tsuga heterophylla Pinus contorta  Training 11 (513) 7 (121) 14 (2,570) 8 (331) 12 (418) 43 (668) 9 (96) 51 (1,025) 76 (8,801) 13 (105) 13 (256)  Validation 5 (283) 3 (83) 15 (1,210) 6 (186) 8 (225) 28 (570) 8 (58) 40 (766) 50 (4,092) 4 (70) 14 (138)  Total 16 (796) 10 (204) 29 (3,780) 14 (517) 20 (643) 71 (1,238) 17 (154) 91 (1,791) 126 (12,893) 17 (175) 27 (394) 411 (22,585)  Table 5.1: Common tree species found in and around the Gulf Islands National Park Reserve (GINPR), British Columbia (BC), Canada. The quantity of reference tree/tree clusters and the number of pixels comprising them (shown in parentheses) are shown per-species. Furthermore, the partition of reference data into training and validation data is shown.  5.2.4  Pre-processing of hyperspectral data  For processing details, consult the specifications outlined in 4.2.3, which reduced the number of airborne hyperspectral variables used as classification input to 40 channels, based on findings in Chapter 2. For additional information please consult Appendix A.  5.2.5  Pre-processing of LiDAR data  For details regarding LiDAR pre-processing and/or derivation of LiDAR digital terrain model (DTM), canopy height model (CHM), and/or canopy volume profiles (CVPs), please consult section 4.2.3. The four raster layers produced to represent euphotic, oligophotic, open gap, and closed gap canopy zones, are herein referred to as 4CVPs. To represent a total “filled” canopy, euphotic and oligophotic zones were considered  71  simultaneously and independent of open and closed gap zones, herein referred to as 2CVPs.  5.2.6  Classification  Conventional parametric classification approaches such as maximum likelihood have been shown to be limited in their ability to classify high dimensional, multi-source data (Benediktsson et al., 1990). Over the past two decades, the utility of numerous advanced non-parametric classification algorithms has been investigated, including support vector machines (SVMs) (Foody and Mathur, 2004, Huang et al., 2002). SVMs are a form of non-parametric, supervised classification, which map data from spectral space into feature space, wherein continuous predictor variables are partitioned into binary categories by an optimal n-dimensional hyperplane (Vapnik, 1998). A kernel function maps inseparable cases into a higher dimensional space (Hsu et al., 2010, Wu et al., 2004) using parameters which permit the best possible fit for the hyperplane (Koetz et al., 2008). These hyperplanes minimize a cost function, which maximizes the distance from the closest training samples referred to as support vectors (Bruzzone and Carlin, 2006), with a penalty parameter controlling the permitted degree of misclassification for non separable cases. The defined rigidity of the penalty parameter has a strong effect on model accuracy and the potential for generalization. Using SVMs to classify high dimensionality, multi-source data representing complex environments has yielded results which match or outperform conventional approaches (Dalponte et al., 2008, 2009, Huang et al., 2002, Koetz et al., 2008, Melgani and 72  Bruzzone, 2004, Pal and Mather, 2006). Specific to hyperspectral data, SVMs have been shown to improve classification accuracy, computational time, and stability, as compared with conventional parametric classification algorithms (e.g., minimum distance to means, maximum likelihood) and other non-parametric algorithms (e.g., neural networks, K-nn) (Bruzzone and Carlin, 2006, Dalponte et al., 2008, 2009, Koetz et al., 2008, Melgani and Bruzzone, 2004). For this study, a multiclass SVM classification was used, where a pairwise strategy represented all possible combinations of classes. A Gaussian radial basis function (RBF) kernel, controlled by parameter γ, was employed to solve linearly inseparable cases (Vapnik, 1998). Penalty parameter C controlled the degree of misclassification for non-separable cases. As described by Hsu et al. (2010) and Wu et al. (2004), a “grid-search” was used to determine the best (C, γ), resulting in a C of 1000 and γ of 0.075. SVMs were trained for four sets of different inputs: 1) AISA data (40 bands), 2) multi-source AISA and LiDAR height data (40 AISA bands+CHM), 3) multi-source AISA and full LiDAR volume data (40 AISA bands +4 CVPs), and 4) multi-source AISA data and LiDAR-derived total filled canopy volume (40 AISA bands +2CVPS) (Figure 5.2). Accuracy was determined using confusion matrices, which provide a crosstabulation of mapped class labels against validation data, corresponding to an independent sample of cases representing specified locations (Foody, 2002). A McNemar's test (Agresti, 1996, Bradley, 1968) was used to assess the significance of the difference between the proportions of correctly allocated cases resulting from SVM iterations. This non-parametric test determines the significance of the difference between two proportions by comparing the calculated value of z against known, tabulated values (Foody, 2004).  73  Figure 5.2: Hyperspectral imagery representing 40 targeted AISA bands and rasterized LiDARderived height and volume information were masked and fused at the pixel-level, resulting in four layer stacks used as support vector machine (SVM) classification input.  74  5.3  Results  5.3.1  Differentiation of species based on spectral properties  Figure 5.3 shows the mean reflectance value of trees/tree clusters for all broad-leaved (top) and coniferous (bottom) species at the 453 analyzed AISA wavelengths, and Figure 5.4 shows the average reflectance of ±1 standard deviation for all broad-leaved (top) and coniferous (bottom) species at the 40 wavelengths corresponding to targeted AISA channels used as classification input. Focusing on the 40 targeted wavelengths, for broadleaved species, increased seperability was typically exhibited at wavelengths >600 nm, with maximum spectral distinction commonly occurring in the NIR and shortwave infrared (SWIR). Bigleaf maple, Garry oak, and arbutus were particularly discernable from one another and all other species, most notably in the SWIR. In contrast, black cottonwood, trembling aspen, and red alder were commonly similar throughout the spectrum. For coniferous species, as with broad-leaved species, the NIR and SWIR provided optimal spectral contrast. At most wavelengths, spectral characteristics were similar for grand fir and Western red cedar, and likewise for Douglas-fir, Western hemlock and lodgepole pine. However, these taxonomic sub-groupings were typically spectrally distinct from one another.  75  Figure 5.3: Examples of the mean reflectance value of trees/tree clusters for all broad-leaved (top) and coniferous (bottom) species at 453 wavelengths corresponding to analyzed AISA channels.  76  Figure 5.4: The mean spectral value (±1 standard deviation) of broad-leaved (a) and coniferous (b) species at 40 targeted AISA wavelengths, identified through discriminant analyses as significant (p<0.005) in their ability to differentiate between species of interest.  77  5.3.2  Differentiation of species based on mean canopy surface height characteristics  Figure 5.5 presents per-species differences in mean canopy surface heights of ±1 standard deviation. Within the broad-leaved taxonomic grouping, black cottonwood had an average canopy surface height of 36.8 m (±5.3 m), occupying the highest strata of all species. The height ranges occupied by all other broad-leaved species were less distinct and overlapped, ranging from 13.4 to 20.6 m (±≤7.5 m). Within the coniferous taxonomic grouping, lodgepole pine exhibited an average canopy surface height of 11.2 m (±4.1 m), occupying a relatively unique strata, narrower in its distributional range and shorter. Average surface canopy heights for all other coniferous species exhibited overlap and ranged from 15.5 m (Western red cedar (±6.3 m)) −22.3 m (Western hemlock (±9 m)).  Figure 5.5: Mean canopy surface heights (±1 standard deviation) for 11 species based on a LiDARderived canopy height model (CHM). 78  5.3.3  Differentiation of species based on volumetric canopy height values  Figure 5.6 displays mean CVP values for all species, depicting the average proportion of total volume occupied in each zone. For broad-leaved species below the open gap zone, black cottonwood and to a lesser degree red alder exhibited the most distinct volume profiles, except within the euphotic zone which demonstrated minimal difference amongst species (i.e., ±3.0 ranging from 5.8 to 8.8%). In the closed gap zone, 47.4 and 18.1% of total volume were occupied for black cottonwood and red alder, respectively, whereas for all other species, <9.0% of total volume was occupied. In the oligophotic zone, 15.1 and 7.7% of total volume were occupied for black cottonwood and red alder, respectively, whereas ≤6.0% of volume was occupied for all other broad-leaves. There was a clear difference within the open gap zone associated with black cottonwood, wherein 28.8% of total volume was accounted for, as compared with >67.0% for all other broad-leaves. For coniferous species within the closed gap zone, 18.6–20.1% of total volume was filled for grand fir, Douglas-fir, and Western hemlock, whereas 10.3 and 4.0% were occupied for Western red cedar and lodgepole pine, respectively. Lodgepole pine was distinct for having a markedly low proportion occupied in the oligophotic and euphotic zones (2.0 and 5.0% respectively), as compared with all other conifers (6.5–12.1 and 8.9–12.0%, respectively). In the open gap zone coniferous species had <73.0% of their total volume filled, excepting lodgepole pine with >88.0%.  79  Figure 5.6: The average proportion of total volume occupied open gap, euphotic, oligophotic and closed gap zones for 11 species based on LiDAR-derived canopy volume profiles (4CVPs).  5.3.4  Support vector machine classification results  Using only spectral variables in the SVM classification resulted in an overall accuracy of 72.3% (KIA 0.6). Broad-leaved species were mapped with accuracies ≥52.0–95.0 (producer's) and 63.0–75.0% (user's), excepting trembling aspen. Results for coniferous species were generally lower, with only three class’s producer's accuracies ≥60% (i.e., Douglas-fir (84.9), Western hemlock (61.4) and lodgepole pine (65.9)), and two with user's accuracies ≥79% (i.e., Douglas-fir (79.4) and Western hemlock (87.8)).  80  Adding height information into the SVM classifier increased producer's accuracies for trembling aspen (+12.1%) and lodgepole pine (+6.5%) (Table 2, Figure 7(top)). Increases in user's accuracies were observed for black cottonwood (+12.5%), trembling aspen (+7.8%), grand fir (+20.0%), and lodgepole pine (+13.8%). A notable reduction in user's accuracy occurred for Western hemlock (i.e., −23.2%) (Table 2, Figure 7(bottom)). Including 4CVPs as classification input increased producer's accuracies for trembling aspen (+26.5%), Western hemlock (+8.6%), bigleaf maple (+5.4%), lodgepole pine (+5.1%), and arbutus (+7.8%). Increases in user's accuracies were observed for black cottonwood (+8.4%), trembling aspen (10.8%), grand fir (+6.5%), and lodgepole pine (+14.6%), but there was a 9.1% reduction in user's accuracy for bigleaf maple. In comparison to the spectral/CHM results, incorporating 4CVPs increased accuracies for trembling aspen (producer's +14.4%) and Western hemlock (producer's and user's+12.9 and 24.6%, respectively). However, 4CVPs were responsible for notable reductions in user's accuracies for grand fir (−13.5%) and bigleaf maple (−9.1%).  81  Species black cottonwood trembling aspen red alder  Overall  Garry oak arbutus  40 spectral + CHM  40 spectral + 4CVPs  40 spectral + 2CVPs  72.3  72.9  73.2  73.5  Kappa  0.6  0.6  0.6  0.6  Producer's  95.4  97.9  97.9  97.5  User's  69.2  81.7  77.6  73.8  Producer's  7.2  19.3  33.7  22.9  User's  19.4  27.1  30.1  26.8  Producer's  67.7  65.7  64.6  66.5  73  74.4  75.3  74.5  Producer's  52.2  55.4  57.5  54.3  User's  75.2  75.2  66.1  73.7  Producer's  82.7  84  82.2  81.8  User's  75.3  74.7  76.1  79.3  Producer's  55.4  59.5  63.2  57.4  User's  User's bigleaf maple  40 spectral  63.8  66.2  67.5  68.1  Producer's  0  3.5  3.5  6.9  User's  0  20  6.5  66.7  Western redcedar  Producer's  41  40.6  42.2  42.7  48  43.5  44.4  47  Douglas-fir  Producer's  84.9  85.2  84.5  85.8  User's  79.4  79.8  80.9  79.7  Western hemlock  Producer's  61.4  57.1  70  64.3  User's  87.8  64.5  89.1  84.9  lodgepole pine  Producer's  65.9  72.5  71  77.5  grand fir  User's  User's 42.7 56.5 57.3 61.5 Table 5.2: Comparison of support vector machine (SVM) classification accuracies resulting from using purely spectral variables as input versus pixel-level fusion with LiDAR-derived canopy height and canopy volume information.  82  Figure 5.7: Using the categorical classification accuracies resulting from purely spectral input variables as the frame of reference (i.e., the x-axis), changes in producer’s (top) and user’s (accuracies (bottom) are provided.  83  Contrasting the use of spectral variables on their own with integrating 2CVPs, increases in producer's and user's accuracies were noted for trembling aspen (15.7 and 7.5%), grand fir (6.9 and 66.7%), and lodgepole pine (11.6 and 18.8%). Compared with the results of combined spectral/CHM, fusion with 2CVPs raised producer's and user's accuracies for Western hemlock (7.2 and 20.4%) and lodgepole pine (5.0 and 5.0%). In addition, the user's accuracy for grand fir increased by 46.7%. As compared with including 4CVPs, incorporating 2CVPs raised producer's accuracies for lodgepole pine (+6.5%) and user's accuracies for grand fir (+60.2%) and bigleaf maple (+7.6%). In contrast, there were reductions in producer's accuracies for trembling aspen (−10.8%), Western hemlock (−5.7%), and arbutus (−5.8%).  5.3.5  Significance tests based on proportion correct  When comparing spectral variables with spectral/CVP fusion (both 2 and 4CVP), the observed differences were found to be the most significant (p<0.001). The difference exhibited between data sets incorporating height (CHM) versus full volume (4CVP) information was also found to be significant, although less (p<0.01), as was the difference between purely spectral variables and fused hyperspectral/CHM (p<0.01). The least significant differences observed, although still greater than the 5% level of significance, were between comparisons of height and volume-based LiDAR metrics. All pairwise comparisons of observed differences in accuracies are presented in Table 5.3.  84  Classification 1 40 spectral 40 spectral 40 spectral 40 spectral + CHM 40 spectral + CHM 40 spectral + 4CVPs  Classification 2 40 spectral + CHM 40 spectral + 4CVPs 40 spectral + 2CVPs 40 spectral + 4CVPs 40 spectral + 2CVPs 40 spectral + 2CVPs  IzI 2.99 4.43 3.82 3.15 2.38 2.06  Significance p<0.01 p<0.001 p<0.001 p<0.01 p<0.05 p<0.05  Table 5.3: Pairwise comparison of IzI values resulting from McNemar’s tests.  5.3.6  Area occupied  A comparison of the total area occupied by each species based on different iterations of SVM reveals relatively minimal variation (Figure 5.8). For broad-leaved species, average occupied hectare ranged from 5.6 (±1.7) (i.e., trembling aspen) to 282 (±11.6 ha) (i.e., red alder). Red alder occupied nearly as much land on average as all other broad-leaved species combined (i.e., 234.7 ha). In aggregate, broad-leaved species occupied an average 516.7 ha of the study area. For coniferous species, Douglas-fir represented the vast majority of occupied land, with a mean of 2017 ha (±4.1). The second most wide-spread coniferous species was Western red cedar, occupying an average 222.2 ha (±6.6). All other coniferous species occupied b35 average ha individually and <63 ha in combination. A total of 2302.1 ha were occupied on average by coniferous species within the AOI.  85  Figure 5.8: The total area in hectares (ha) occupied by each species in relation to four different iterations of support vector machine (SVM) classification using, 1) spectral only, 2) spectral and CHM, 3) spectral and 4CVPs and 4) spectral and 2CVPs as input.  5.4  Discussion  The potential explanatory power of LiDAR-derived structural information relates to the ability of a particular metric to characterize a distinct growth stage, observable as a species-specific average canopy height and/or architectural arrangement. If a species is dominated by a particular structural stage, then mapping accuracy improved over spectral information on its own. Confirming the findings of Anderson et al. (2008) and Dalponte et al. (2008), incorporating height information increased accuracy for classes exhibiting similar spectral properties, but associated with different mean canopy heights. For broadleaved species, height was particularly important for black cottonwood, substantially 86  reducing commission error associated with trembling aspen and red alder. Although the spectral characteristics of black cottonwood were very similar to trembling aspen and red alder, occupying a unique canopy surface height range rendered it distinct and raised the user's accuracy to 81.7% (+12.5). For coniferous species, height information further distinguished lodgepole pine, reducing omission error associated with Douglas-fir and raising producer's accuracy to 72.5% (+6.5). Due to insignificant additional commission error, user's accuracy increased to 56.5% (+13.8). While canopy height does prove very useful for differentiating certain species, overlaps in height ranges occupied by spectrally similar species indicate it is often not the most relevant structural surrogate. In some instances, including height information into the SVM increased confusion between spectrally similar classes as compared with using purely spectral variables. For example, due to accentuated confusion with Douglas-fir (omission error), the CHM decreased producer's (−4.3%) and user's (−23.2%) accuracies for Western hemlock. In contrast, comparatively distinct canopy architectural properties facilitated differentiating these two species, resulting in higher accuracies than when using spectral variables independently. Specifically, including 4CVPs decreased omission error with Douglas-fir, raising producer's accuracy to 70% and increasing user's accuracy to 89.1%. For broad-leaved species, 4CVPs improved accuracy for arbutus over spectral variables on their own and spectral/CHM fusion. In addition, there was a notable increase in user's accuracy for black cottonwood as compared with stand-alone spectral variables. However, in this case canopy architecture was not as important to differentiation as the CHM. While incorporating 4CVPs proves useful for characterizing certain species, the use of full canopy volume profiles is not optimal for others. Isolating total filled canopy  87  (2CVPs) provided additional explanatory power for lodgepole pine when compared to all other combinations of input variables. Commission error with Western red cedar and Douglas-fir decreased, increasing producer's accuracy to 77.5% and raising user's accuracy to 61.5%. An in-depth look at the variability associated with filled canopy volumes indicated the ability of these zones to distinguish lodgepole pine from other coniferous species. Confusion between lodgepole pine and Western red cedar in the closed gap zone explains why exclusion of this information increased accuracy. Red alder, Douglas-fir, and Garry oak were spectrally distinguishable and additional structural layers did not provide notable added discrimination. These species exhibited relatively unique spectral properties but also spanned multiple age classes. They provide prime examples of species for which spectral data alone sufficed and added structural information provided no additional benefit. In contrast, for trembling aspen, grand fir, and to a lesser degree Western red cedar, regardless of variable input, classification accuracies were markedly low. Spectrally these species exhibited confusion, and structurally these species spanned a range of age classes. As a result, these species were consistently mapped with accuracies <50%. As indicated in Chapter 2, trembling aspen and grand fir were spectrally distinct at the leaf-level, however, it is apparent that simultaneously considering all of the elements present in open crowns results in other reflectance responses, such as from non-photosynthetic and understory vegetation, shadows, variable illumination, and atmospheric effects, which can all contribute to reducing spectral uniqueness. In conjunction with additional crown materials impacting spectral measurements, geometric error may have contributed to the error. Owing to the  88  comparatively smaller average crown dimensions of these species, residual positional error may have substantially increased confusion. In the southern Gulf Islands, as detailed in Chapter 4, conventional forest species distribution information provided by API lacks the detail, accuracy, and spatial explicitness vital to a variety of managerial tasks. In comparison to 1:5000 scale polygons, our methodology provides an improved approach for mapping canopy level tree species distribution. Furthermore, all classified maps were well within the expected error range of operational API, defined by Leckie and Gillis (1995) as 70–85% for the main species, i.e., Douglas-fir. Based on our results, using combined spectral and 4CVPs provided the most explanatory power for the most species. As compared with the results achieved using all other combinations of input variables, this combination yielded the highest accuracies for arbutus and second highest for black cottonwood. For all combinations there were insignificant differences in accuracies exhibited for red alder, bigleaf maple and Garry oak, typically <5.0%. For coniferous species, spectral fused with 4CVPs yielded the most accurate results for Western hemlock, and amongst the most accurate results for lodgepole pine. When comparing all combinations, differences in accuracies of <2.0% were exhibited for Douglas-fir. Since the mid 1990's, LiDAR acquisition costs have decreased as the technology has moved into the commercial terrain (i.e., geospatial) mapping sector (Reutebuch et al., 2005). Although commonly employed for geospatial mapping, LiDAR biospatial data (e.g., forest structural data) have only been embraced by natural resource scientists within the last decade (Olsson and Næsset, 2004, Wulder et al., 2003). Cost-effective strategies for acquiring LiDAR need to involve multiple stakeholders representing a range of 89  objectives (Reutebuch et al., 2005), thus making vegetation studies more affordable and realistic. While dependent on many variables, small footprint, discrete return LiDAR data are estimated to cost approximately $5.00 USD per ha, including acquisition and basic processing (Wulder et al., 2008b). These costs are comparable with the expenses for acquiring and initially processing remotely sensed data conventionally used in forest inventory, such as aerial photographs (Wulder et al., 2008b). For our study the initial costs were approximately $2.00 USD per ha, including both hyperspectral and LiDAR data. Assuming processing and interpretation are each as expensive as acquisition costs, final products such as our tree species distribution maps equate to approximately $6.00 USD per ha. In comparison, the 1:5000 scale baseline ecological data resulting from a conventional API campaign cost an estimated $12.00 USD per ha. The presented methodology is not limited to the SGI, nor to tree species of the Pacific Northwest, and can be replicated anywhere canopy species of interest are directly measureable from air and can be spectrally and/or structurally discerned. Using key spectral regions to facilitate feature selection and species differentiation can be geographically transferable, however, this depends on whether or not species inhabit spectral regions unique beyond specific locations (Cochrane, 2000). In contrast, due to local and regional variability of forest age make-up, transferring species-specific structural characteristics can be problematic. A vital avenue for future research could involve defining forest canopy classes based on their species composition and their structural stage. This daunting proposition mandates laborious and accurate field work to characterize representative examples of speciesspecific, structural stage combinations, a process which may result in too many classes 90  for a realistic mapping endeavor. However, further division of species categories based on age would most likely extend the utility of LiDAR-derived metrics and increase mapping accuracies. Furthermore, such stratification could intensify the potential of LiDAR/hyperspectral fusion for species mapping. Lastly, the relatively small swath width of imagery collected by aircraft limits these data to relatively small scale investigations. Therefore, another increasingly important avenue for future research relates to upscaling detailed yet comparatively small scale measurements to wider regional coverage via satellite data (Huang and Asner, 2009, Vitousek et al., 1987).  5.5  Conclusion  To effectively manage forested ecosystems species distribution must be understood. Advanced geo-technologies, such as hyperspectral and LiDAR data, permit the exploration of cutting-edge mapping approaches. In the SGI, this study has found that most tree species considered can be mapped with accuracies ranging from >60 to ~90% using only spectral variables, reinforcing the proven ability of airborne hyperspectral data to distinguish between tree species and map their distribution. For certain species, accuracies improved with hyperspectral/LiDAR fusion, reflecting clearly definable mean canopy surface heights and/or unique architectural characteristics, most likely associated with singularly dominant age classes. When per-species structural characteristics are dominated by multiple age classes, CHMs do not provide additional explanatory power. Furthermore, because canopy architecture is typically dynamic in relation to age class, volume information does not prove useful for multi-structured species classes. Although 91  structural information does not provide additional explanatory power for all species, both height and volumetric information result in many increased per-class accuracies and higher overall accuracies than with spectral data alone, consistent with results presented by Dalponte et al. (2008), Geerling et al. (2007, 2009), Koetz et al. (2008), and Mundt et al. (2006), who all noted improvements in overall accuracy and certain class accuracies when incorporating LiDAR-derived structural data. Although there were only slight differences in overall accuracy resulting from using AISA data alone compared with including structural information, the higher overall accuracies associated with using LiDAR data were significantly different than purely spectral approaches (p<0.05). In the SGI, both conventional API campaigns and hyperspectral/LiDAR fusion approaches have provided information that the other has not. However, acquiring and processing advanced remotely sensed data costs approximately half as much as conventional API. Furthermore, pixel-level fusion of hyperspectral/LiDAR data can significantly augment and improve current forest inventory approaches, providing detailed information pertinent to a wide range of forest management tasks, including the full scope of information conventionally derived from API. Therefore, these data are a viable replacement technology. To take full advantage of these disparate and advanced remotely sensed data types in as cost-effective and efficient a manner as possible, simultaneous acquisition and accurate coregistration are required.  92  6  DESCRIBING AVIFAUNAL RICHNESS WITH FUNCTIONAL AND STRUCTURAL BIO-INDICATORS DERIVED FROM ADVANCED AIRBORNE REMOTELY SENSED DATA  6.1  Introduction  Anthropogenic and naturally induced changes to vegetation structure and ecosystem function increasingly impact species abundance and habitat availability worldwide, emphasizing the need for cost-effective tools to identify and manage species and ecosystems of high conservation value (Wilson, 1988; Sala et al., 2000; Nagendra, 2001; Kerr & Ostrovsky, 2003; Bergen et al., 2009). Developing such tools requires applicable indicators of ecosystem condition from which conservation value can be estimated (Turner et al., 2003; Leyequien et al., 2007). Meeting this need, avian diversity studies have received disproportionate attention, as birds are widespread, easily identified with well-described taxonomy and sensitive to land-cover modification, making them attractive for characterizing complex ecosystems (Gottschalk et al., 2005). Here, we ask if bird species richness can be predicted using biological indicators (bio-indicators) derived from advanced airborne remotely sensed data and if predictions can help advance methodologies which identify, characterize and conserve valued ecosystems. To effectively predict species richness, pertinent habitat characteristics must be considered. It is well established that faunal richness and environmental heterogeneity are inextricably linked (Stoms & Estes, 1993), with areas of higher heterogeneity typically supporting richer species assemblages (Simpson, 1949; Lack, 1969; Huston, 1994; Tews et al., 2004). The literature implies that environmental heterogeneity is primarily driven 93  by aspects of ecosystem function (MacArthur, 1972) and/or vegetation structure (MacArthur & MacArthur, 1961; Wilson, 1974, Roth, 1976). In this context, function includes collective ongoing processes and interactions (e.g., nutrient cycling; energy flow), associated bio-chemical constituents (e.g., nutrients; pigments) and their concentrations, whereas structure refers to the totality and spatial arrangement of physical elements within an ecosystem (Hobbs & Mooney, 1990; Grimm, 1995; Wallace & Gray, 2002). Directly inventorying the functional and structural aspects of faunal habitat at or beyond the landscape level (defined by Bradbury et al. (2005) as ≥10 hectare (ha)) through traditional field surveys is challengingly time and resource intensive, if not impossible (Turner et al., 2003; Gottschalk et al., 2005; Bergen et al., 2009; Seavy et al., 2009) and therefore, diversity studies often necessitate proxies (Leyequien et al., 2007). Fulfilling this requirement, digital remotely sensed data permit the collection and comparatively rapid derivation of functional and structural aspects of vegetated ecosystems known to have associations with fauna at a range of spatial scales (Rock et al., 1986; Mooney & Hobbs, 1990; Bradbury et al., 2005). For nearly 40 years, broadband, multispectral sensors (e.g., Landsat) have offered the primary source of faunal habitat data for many parts of the world (Kerr & Ostrovsky, 2003; Murthy et al., 2003; Gottschalk et al., 2005). However, their moderate spatial and spectral resolution limits the scope of measurable ecosystem functional information, typically quantified using vegetation indices (VIs). In addition, while some studies have exhibited success with species-level investigations (e.g., Hill et al., 2010), multispectral sensors typically prohibit mapping vegetation with floristic detail. Furthermore, passive multispectral satellite imagery lacks vertical structural information representing  94  vegetation height, cover and volume (Vierling et al., 2008; Bergen et al., 2009). In contrast, advanced airborne remotely sensed data can characterize and predict faunal habitat with an increased level of detail and accuracy, providing the next generation of vegetation products available as surrogates for faunal distribution (Leyequien et al., 2007; Seavy et al., 2009). The extremely fine spectral and spatial resolutions of hyperspectral sensors permit measuring absorption features in upwards of hundreds of adjacent, narrow channels, facilitating the quantification of extremely subtle aspects of plant bio-chemistry and species-level mapping. Hyperspectral data have been used successfully for species-level mapping within tropical (e.g., Carlson et al., 2007; Clark et al., 2005), sub-tropical (e.g., Lucas et al., 2008; Yang et al., 2009) and temperate (e.g., Goodwin et al., 2005; Jones et al. 2010) systems, and the potential and realized applications of derivable narrow-band VIs have been explored and/or demonstrated within variable contexts (see Gao, 1996; Gitelson et al., 2001; Gitelson et al., 2002; Serrano et al., 2002; Sims & Gamon, 2002). While tree species composition has previously been demonstrated as an aspect of habitat preference for avifauna (e.g., Holmes & Robinson, 1981; Peck, 1989), until recently, detailed floristic composition has received little attention at or above the landscape level. In addition, although aspects of foliar biochemistry quantifiable through narrow-band bio-indicators are associated with ecosystem functions linked with faunal habitat suitability, no studies have assessed their propensity for describing faunal diversity. In contrast to hyperspectral sensors, light detection and ranging (LiDAR) sensors emit active pulses of near-infrared energy which simultaneously measure ground elevation (Ackermann, 1999; Wehr & Lohr, 1999) and vertical vegetative structure (Lefsky et al., 95  1999; Drake et al., 2002). The signal returned from active LiDAR pulses represents varying depths from which a variety of continuous, fine spatial resolution grids can be interpolated (Bradbury et al., 2005). LiDAR sensors record either the entirety of a return signal (i.e., full waveform), or information from major peaks within each pulse (i.e., discrete-return) (Bradbury et al., 2005; Wulder et al., 2008). For both system types, information content is a function of the horizontal circular sampling area (i.e., footprint) (Means, 1999; Lim et al., 2003; Wulder et al., 2008). Providing an unprecedented data source, applications of LiDAR to ecological investigations have tremendous potential (Lefsky et al., 2002; Turner et al., 2003; Bergen et al., 2009), and while numerous studies discuss the encouraging prospects for advancing our understanding of animal-habitat associations, few have attempted to quantify these relationships (see Bradbury et al., 2005; Clawges et al., 2008; Vierling et al. 2008). We aimed to quantify how much variance in richness could be explained for three distinct guilds of native birds, stratified by habitat preference, in and around the Gulf Islands National Park Reserve in south-western British Columbia (BC), Canada, based on pertinent bio-indicators derived from hyperspectral and LiDAR data. We assess the explanatory power of metrics pertaining to ecosystem function and tree species diversity derived from Airborne Imaging Spectrometer for Applications (AISA) Dual hyperspectral data and structural metrics derived from small footprint, discrete-return LiDAR data for describing overall and per-guild richness. In addition, while traditionally considered in tandem (Wallace & Gray, 2002), owing largely to time-scale discrepancies, resource managers employing remotely sensed data have typically modeled ecosystem condition based on aspects of structure or function (Hobbs & Mooney, 1990). In contrast,  96  we determine if there is added benefit in considering aspects of ecosystem function and vegetation structure simultaneously to characterize avifaunal richness. Lastly, we consider the pertinence of bio-indicators in accordance with localized management criteria, and more importantly, their transferability to other systems.  6.2  Materials and methods  6.2.1  Study area  For a complete study area description, please consult section 2.2.1. Owing to mounting anthropogenic influence various fauna, including birds, are currently federally classified by the Committee on the Status of Endangered Wildlife in Canada as threatened (e.g., Olive-sided flycatcher (Contopus cooperi Nuttall)) and/or provincially listed by the BC Conservation Data Centre as at-risk (e.g., Purple Martin (Progne subis Linnaeus)) (BC Conservation Data Centre, 2010). Ensuring that the structure and function of at-risk ecosystems are unhindered by anthropogenic activities receives managerial emphasis in and around the Gulf Islands National Park Reserve (Woodley, 1993; Parks Canada Agency, 1997).  6.2.2  Hyperspectral data  For complete information on AISA data and pre-processing, consult 4.2.2 and 4.2.3, respectively. For additional information please consult Appendix A. To investigate foliar  97  biochemical properties relatable to ecosystem functions, narrow-band VIs pertaining to foliar nitrogen (i.e., normalized difference nitrogen index (Serrano et al., 2002)), lignin (i.e., normalized difference lignin index (Serrano et al., 2002)), anthocyanin (i.e., anthocyanin reflectance index (Gitelson et al., 2001), carotenoid (i.e., carotenoid reflectance index (Gitelson et al., 2002)) and foliar water content (i.e., normalized difference water index (Gao, 1996)) and greenness (i.e., red edge normalized difference vegetation index (Sims and Gamon, 2002)) were derived (Table 6.1). In addition, 2m spatial resolution land-cover products representing tree species richness and diversity (i.e., Shannon’s index of diversity (SID)) (Shannon, 1948)) were derived from previously classified AISA imagery (discussed in Jones et al. (2010)) and included in the analysis.  Narrow-band vegetation index (VI)  Equation  Reference  Red edge normalized difference vegetation index (RENDVI)  RENDVI =  R750 - R705 R750 + R705  Normalized difference nitrogen index (NDNI)  NDNI =  [log (1/R1510) - log (1/R1680)] [log (1/R1510) + (1/R1680)]  Serrano et al., 2002  Normalized difference lignin index (NDLI)  NDLI =  [log (1/R1754) - log (1/R1680)] [log (1/R1754) + (1/R1680)]  Serrano et al., 2002  Anthocyanin reflectance index (ARI)  ARI =  Carotenoid reflectance index (CRI)  CRI =  Normalized difference water index (NDWI)  NDWI =  (R550)-1 - (R700)-1 (R510)-1 - (R700)-1  R857 - R1241 R857 + R1241  Sims and Gamon, 2002  Gitelson et al., 2001 Gitelson et al., 2002  Gao, 1996  Table 6.1: Narrow-band vegetation indices considered for analysis.  6.2.3  LiDAR data  LiDAR were collected concurrently with hyperspectral data from the same platform. For complete details concerning LiDAR data acquisition, consult section 3.2.3. For details  98  regarding LiDAR data pre-processing, consult 4.2.3. For additional information please consult Appendix A. In addition to canopy surface heights, and canopy volume profiles (CVPs), non-ground returns located within predefined height strata were used to interpolate cover layers representing the percentage of under-story (0.5-5 m), mid-story (5-25 m) and over-story (>25 m) vegetation, and aggregate vegetation encountered within all height-based strata. To represent the heterogeneity of height, the coefficient of variation of non-ground returns was calculated within 6 m windows.  6.2.4  Avian survey  Bird data were collected within plots from early May to mid June, 2005-07, following methods described in detail Jewell et al. (2007) and applied at landscape scales (e.g., Jewell and Arcese 2008, Martin et al. 2011). Briefly, 50m radius plots were established across a range of forested, open-country and human-dominated habitats for the purpose of mapping species occurrence and estimating species richness. At each visit to plots a trained observer recorded all species detected by sight or sound within a 10 min period inside the plot. All observations were made between sunrise and 11AM in the absence of audible rainfall or wind to minimize variation in detection due to time of day, weather conditions or season. During counts, observers moved slowly throughout the plot area to maximize the likelihood of detecting cryptic species. Suitable plot locations were identified for this study using vegetation cover information derived from aerial photographs provided by Parks Canada. 136 plots distributed across 11 islands were identified as: 1) completely within the confines of airborne transects, 2) free of water,  99  intertidal and/or beach and rock and 3) almost entirely (i.e., > 95.0%) comprised of nonagricultural vegetation. The size of islands upon which plots were established varied from 12 to 3000 ha averaging 1170 ha. The combined size of all islands was ~12900 ha, 107 of which were represented by plots. Plots were visited twice on average, but up to 6 times in total. Richness is defined here as the cumulative number of species detected. For additional information please consult Appendix A. Uneven sampling effort across plots was accommodated in statistical analyses by weighting estimates based on their reliability (see below). To reliably assign each of 74 bird species detected to guilds, we asked 6 professional ornithologists with substantial experience in the coastal Douglas fir zone of BC to assign each species to one of three guilds below based on locally-observed habitat preferences/requirements including: 1) open-country/shrubland/grassland, 2) generalist and 3) mature forest/woodland (Appendix B; hereafter referred to as ‘open-country’, ‘generalist’ and ‘forest’ guilds, respectively). Birds were then distributed into distinct guilds using majority rule, and overall and per-guild richness values were then calculated for all plots. In addition, for each plot, the guild to which the largest number of observed species belonged to was used as a measure of guild dominance.  6.2.5  Statistical analyses  Two statistical approaches were undertaken to explore and describe relationships between avifaunal richness and hyperspectral and LiDAR-derived bio-indicators. First, discriminant analyses determined which variables were most useful in discerning between plots based on guild dominance. Using p-values, hyperspectral and LiDAR bio100  indicators were assessed independently. The second approach examined which bioindicators best described guild-specific richness. To determine which combinations of variables resulted in the most parsimonious models, we employed a stepwise modelselection procedure which initially consider all variables and subsequently removed or added terms, yielding the combination of predictors with the lowest Akaike’s (1973 ) Information Criterion (AIC) value (Venables & Ripley, 2002). The best combinations of hyperspectral and LiDAR bio-indicators were assessed independently. For each data type, all guilds were considered individually and in combination. Based on selected variables, 12 generalized linear models (GLMs) (McCullagh & Nelder, 1989) were built, representing each guild independently and all guilds at once, using hyperspectral, LiDAR-derived and combined bio-indicators deemed most important through AIC. To directly incorporate effort into the GLMs the square root of visits was included as a covariate. We employed the square root of number of visits to a site as our estimate of sampling effort because species richness asymptotes rapidly with repeated surveys, as expected in this binomial sampling process (i.e., species presence versus absence) (Gilbert 1979). In addition to standardizing for sampling effort, we also included island size as co-variate in models to account for the potential influence of island size on faunal richness and transformed island sizes by log10 to normalize these data before use in models. GLMs were built using 70% of available data, with remaining plots held back to assess goodness-of-fit based on the adjusted r2 and p-values. Prior to the analysis, an r2 threshold of 0.85 eliminated redundant predictor variables, reducing the original pool from 38 to 20 (9 hyperspectral and 11 LiDAR) (Table 6.2). Where two or more variables were highly correlated, we estimated the relevance of each to bird species occurrence and  101  ecological processes to decide which would remain in the analysis. For all remaining variables, the mean was extracted for the extent of each plot. In addition, to characterize the spatial variability of the predictors, standard deviations were calculated.  102  Hyperspectral-derived bio-indicators  Targeted information  Wavelengths (nm); spatial resolution (m)  Red edge normalized difference vegetation index  greenness, primary productivity, above-ground biomass (AGB)  705, 750; 2  Normalized difference nitrogen index  vegetation nitrogen content  1510, 1680; 2  Normalized difference lignin index  vegetation lignin content  1680, 1754; 2  Anthocyanin reflectance index  leaf anthocyanin content  550, 700; 2  Carotenoid reflectance index  leaf carotenoid content  510, 700; 2  Normalized difference water index  vegetation water content  857, 1241; 2  Tree species richness  number of tree species present  not applicable; 2  Shannon-Wiener diversity index (SID)  tree species diversity  not applicable; 2  LiDAR-derived bio-indicators  Targeted information  Analysis window (m); spatial resolution (m)  Digital elevation model  surface elevation  2; 2  Canopy surface model  average canopy height  2; 2  Maximum cell height  maximum canopy height  2; 2  Understory cover  % canopy cover from 0.5-5 m  24; 2  Midstory cover  % canopy cover from 5-25 m  24; 2  Overstory cover  % canopy cover occurring at >25 m  24; 2  Total cover  total vegetation cover from 0.5->25m  24; 2  Coefficient of variation of LiDAR non-ground returns  heterogeneity in vegetation height  6; 2  Closed gap  proportion of closed gap below the canopy  6; 2  Oligophotic vegetation volume  total vegetation volume occurring within the oligophotic stratum  6; 2  Euphotic vegetation volume  total vegetation volume occurring within the euphotic stratum  6; 2  Open gap  proportion of open gap above the canopy  6; 2  Table 6.2: Variables considered for analysis and associated targeted information. For hyperspectral-derived bio-indicators, spatial resolution and the wavelengths used for calculating vegetation indices are provided. For LiDAR-derived indicators, the window of analysis within which variables were originally calculated and the resultant spatial resolution are supplied. For all variables, the plot-based mean and standard deviation were originally considered. Greyed bio-indicators were eliminated based on correlation analyses.  103  6.3  Results  6.3.1  Species richness  Based on total visits to all plots avian species richness for all guilds ranged from 2 to 23 (mean [sd] = 10 [4]), wherein birds belonging to the forest guild accounted for 58.9% of species observed and open-country and generalist guilds represented 20.7 and 20.3% of species, respectively. Within the forest guild, plot richness ranged from 0-15 (mean [sd] = 6 [3]), whereas for open and generalist guilds, richness ranged from 0-7 (mean [sd] = 2 [2]) and 0-7 (mean [sd] = 2 [1]), respectively. For all guilds sampling effort increased richness (Table 6.3).  104  Opencountry  Generalist  Forest  All guilds  Mean  2  2  5  8  Minimum  0  0  0  2  Maximum Standard deviation  7  6  10  14  2  1  2  3  Mean  1  2  6  10  Visits 1  2  3  4  5  6  Minimum  0  0  2  5  Maximum Standard deviation  3  5  11  13  1  1  2  3  Mean  3  4  9  16  Minimum  1  1  4  12  Maximum Standard deviation  5  6  12  22  1  2  2  3  Mean  2  2  11  15  Minimum  0  1  7  11  Maximum Standard deviation  5  5  15  22  1  1  3  3  Mean  4  5  10  19  Minimum  1  3  9  14  Maximum Standard deviation  7  7  10  23  4  3  1  6  Mean  4  3  12  19  Minimum  3  3  11  17  Maximum Standard deviation  4  3  13  20  1  0  1  2  Table 6.3: Avifaunal richness related to sampling effort.  6.3.2  Guild-based discrimination  At the majority of plots (62.0%) forest guild birds were most often observed compared to 6.0% of plots dominated by open-country species. All other plots contained a mixture of guilds with no clear domination. From the pool of functional bio-indicators, narrow-band VIs representing carotenoid and foliar water content, and land-cover information representing tree species richness and diversity were each found to usefully discriminate 105  between plot types (p<0.01) (Table 6.4). Mean foliage water content was the most significant hyperspectral-derived bio-indicator (p<0.001) owing to distinctly wetter values for forest dominated than open-country plots (Figure 6.1(A)). Of structural bio-indicators, mean under, mid and overstory cover, mean surface elevation, and variability in vegetation volume occupying the euphotic zone were all significant in discriminating between guilds (p<0.05) (Table 6.4). Average cover from the midstory and variability in euphotic vegetation were the most informative of all LiDARderived structural variables (p<0.001). Open-country plots exhibited an average midstory cover value of 66.0% (±18) (Figure 6.1(B)) and an euphotic vegetation value of 0.4 (± 0.09) (Figure 6.1(C)). In contrast, average values for plots dominated by birds belonging to the forest guild were 7.0% (±7.0) for midstory cover and 0.16 (±0.14) for euphotic vegetation.  106  Figure 6.1: The difference in open-country/shrubland/grassland vs. mature forest/woodland guilds (± 1 stdev) exhibited by the mean normalized difference wetness index (a) and midstory cover (b) values, and variation within the volume of the euphotic strata of vegetation (c).  107  Mean  Standard deviation  Hyperspectral-derived bio-indicators Carotenoid reflectance index  p < 0.01  Normalized difference water index  p < 0.001  Tree species richness  p < 0.01  Shannon-Wiener diversity index (SID)  p < 0.01  p < 0.01  LiDAR-derived bio-indicators Digital elevation model  p < 0.01  Understory cover  p < 0.05  Midstory cover  p < 0.001  Overstory cover  p < 0.05  Euphotic vegetation volume  p < 0.001  Table 6.4: Bio-indicators found through discriminant analyses to usefully discriminate between plots based on the guild type most often observed (i.e., open country/shrubland/grassland vs. mature forest/woodland). Significant bio-indicators are organized by data-origin (i.e., hyperspectral vs. LiDAR-derived) and whether significance (p-value) was associated with the mean and/or standard deviation.  6.3.3  Assessing richness  Figure 6.2 presents the frequency of AIC selection for hyperspectral (A) and LiDARderived (B) bio-indicators, wherein tree species diversity, mean surface elevation, variability of heterogeneity in height and mean under/mid/overstory cover were selected most often. Based on indicators associated with biochemistry and land-cover, opencountry richness was best described by average foliar anthocyanin, lignin and water content, tree species richness and diversity, and variability in foliar water content (Table 6.5) (r2adj = 0.47, p<0.001 (Table 6.6)). Relationships between LiDAR-derived bioindicators and open-country richness were strongest for mean surface elevation, mid and overstory cover, and variability in the euphotic zone (r2 adj = 0.51, p<0.001). Combining hyperspectral and LiDAR-derived bio-indicators further improved the model, yielding an adjusted r2 of 0.59 (p<0.001) (Figures 6.3, 6.4).  108  Figure 6.2: The frequency of hyperspectral (a) and LiDAR-derived (b) variables selected using the Akaike Information Criterion.  109  Open-country/shrubland/grassland Mean  Hyperspectral-derived  Generalist  Mature forest/woodland  All guilds  foliar anthocyanin content foliar lignin content foliar carotenoid content  LiDAR-derived  foliar water content  foliar water content  tree species richness  tree species richness  tree species diversity  tree species diversity  surface elevation  surface elevation  tree species diversity  understory cover midstory cover overstory cover  understory cover  midstory cover overstory cover heterogeneity in vegetation height euphotic canopy volume  StDev  Hyperspectral-derived  foliar lignin content foliar carotenoid content foliar water content  LiDAR-derived  surface elevation  heterogeneity in vegetation height  heterogeneity in vegetation height  euphotic canopy volume  Table 6.5: Bio-indicators selected using the Akaike information criterion organized by guild type, data type (i.e., hyperspectral vs. LiDAR-derived), and nature of summary (i.e., mean vs. standard deviation).  110  Hyperspectral-derived  LiDAR-derived  Combined  Open-country/shrubland/grassland  Generalist  Mature forest/woodland  All guilds  n  8  5  5  3  adjusted r²  0.47  0.32  0.58  0.54  p-value  < 0.001  < 0.001  < 0.001  < 0.001  n  6  8  5  3  adjusted r²  0.51  0.19  0.64  0.5  p-value  < 0.001  < 0.001  < 0.001  < 0.001  n  12  11  8  4  adjusted r²  0.59  0.32  0.64  0.54  p-value  < 0.001  < 0.001  < 0.001  < 0.001  Table 6.6: The results of generalized linear models built using the most parsimonious bio-indicators as selected using the Akaike information criterion organized by guild and metric type.  111  Figure 6.3: Explained variance in species richness.  112  Figure 6.4: Observed vs. predicted richness for all guilds, independently and in combination.  For the generalist guild, hyperspectral bio-indicators associated with foliar lignin content and tree species diversity were also amongst the most important functional variables for describing richness, resulting in an adjusted r2 of 0.32 (p<0.001). In contrast to opencountry richness, the most influential structural bio-indicators for generalist richness included aspects of heterogeneity in vegetation height. Combining both types of bioindicators did not improve the model over using hyperspectral variables on their own (r2 adj =  0.32, p<0.001).  113  As with the open-country guild, tree species richness, mean foliar water content and average midstory cover were influential for describing the richness of forest birds. An additional bio-chemical indicator found to be useful was mean carotenoid concentration, which together with mean water content and tree species richness yielded an adjusted r2 of 0.58 (p<0.001). In addition to midstory cover, structural variables found to be most useful for describing forest bird species richness included average cover in the understory and variability of heterogeneity in vegetation height, yielding an adjusted r2 of 0.64 (p<0.001). A combined functional/structural model did not increase the overall explanatory power resulting from LiDAR variables on their own, yielding an adjusted r2 of 0.64 (p<0.001). When all guilds were considered simultaneously, all variables had been previously selected to predict the richness of a particular guild (Table 6.5). The hyperspectral and LiDAR-derived variables most capable of describing the richness of all guilds considered simultaneously were tree species diversity and mean understory cover, respectively. Combining bio-indicators resulted in an adjusted r2 value that was lower than models associated with specifically open-country or forest guilds, but 22.0% higher than the generalist guild (Table 6.6).  114  6.4  Discussion  6.4.1  Ecosystem function, bird species richness and hyperspectral imagery  The degree to which ecosystem processes are impacted by avifauna and vice versa remains an important ecological question. Applications of information pertaining to elements of various ecosystem functions have previously been demonstrated with multispectral VIs, primarily the normalized difference vegetation index (NDVI) (e.g., Jørgensen & Nøhr, 1996; Hurlbert & Haskell, 2003; Bailey et al., 2004; Seto et al., 2004). However, while encouraging relationships with avian diversity have been demonstrated, studies have found that broad-band NDVI is of limited use in regions exhibiting high topographical variation (Kerr & Ostrovsky, 2003) and saturates at moderate to high biomass density and leaf area index (Tucker, 1977; Gao et al., 2000). Furthermore, studies have shown that correlations with bird species diversity exhibit considerable variation, suggesting a functional link with diversity remains elusive (Leyequien, 2007). In contrast to studies employing broad-band VIs, we observed encouraging correlations between certain narrow-band VIs and bird species richness. For open-country bird species, a relatively strong relationship between foliar water, lignin and anthocyanin concentration was exhibited. Water content has an overall regulatory effect on numerous ecosystem processes, whereas lignin concentration is a central mechanism associated with vegetative growth and decomposition, both fundamental components of the nutrient cycling process (Melillo et al., 1982; Serrano et al., 2002). Anthocyanin concentration regulates the protective mechanisms of plants and provides important insight regarding response and adaptation to various environmental stresses 115  (e.g., drought, ultraviolet radiation, fungal infections) (Gitelson et al. 2001). While canopy water content also proved important for describing forest bird richness, the significance of carotenoid was additionally demonstrated, implying that these bioconstituents regulate and/or are influenced by forest bird richness. Detected differences in foliar water concentrations associated with open-country vs. forest habitats may be partly attributable to the time of year, because within this climatic zone most grasses have advanced to senescence by late July, whereas forests are typically not water or temperature stressed. This seasonal variation appears to help discriminate among guilds and predict richness based on observable bio-chemical differences. Of all hyperspectral derived bio-indicators pertaining specifically to aspects of ecosystem function, foliar water content was the most useful for differentiating between plots and describing bird species richness, demonstrating a consistent relationship.  6.4.2  Tree species diversity and bird species richness  While studies have demonstrated the role of broad-band VIs, the most common applications of remotely sensed data to faunal diversity studies involve combining expert knowledge of the habitat requirements of species with land-cover information derived from multispectral satellite imagery to create predictive maps of species distribution and density (e.g., Palmeirim, 1988; Morrison, 1997; Prins et al., 2005). Although successful in some instances, a key limitation remains a failure to represent micro-heterogeneity in land form which influences distribution for many species (Leyequien et al., 2007). Coarse spatial data can preclude identifying finer scale land-cover components associated with faunal species diversity. Thus, we asked if comparatively detailed and accurate tree 116  species information could enhance relationships between richness and fine scale aspects of land-cover configuration. We found support for this idea in the case of tree species richness in relation to open-country and forest bird species richness. Not surprisingly, these habitats differ in tree species richness. However, a detailed representation of this seemingly obvious difference serves well to characterize, quantify and exploit habitat discrepancies. Open-country areas exhibited typically lower average tree species richness (e.g., <3.0) with moderate variability (e.g., ±2.4). This variation was enough to permit a relationship with bird species richness, serving to characterize richness within this guild and further differentiate from forested plots, which exhibited higher average tree species diversity (e.g., ~5.0) with comparatively less variation (±1.8). Of all hyperspectral derived bio-indicators, aspects of tree species diversity were the most useful for differentiating between plots and describing bird species richness for all guilds, demonstrating the relatively unexplored and seemingly promising potential of finer scale aspects of land-cover heterogeneity in faunal diversity investigations.  6.4.3  Vegetation structure, bird species richness and LiDAR data  As an alternative to optical remotely sensed data (i.e., multispectral and/or hyperspectral imagery), strong relationships exhibited between vertical vegetation structure and bird species composition (Dunlavy, 1935; Morgan & Freedman, 1986; Buchanan et al., 1995; Villard et al., 1999; Siegel & DeSante, 2003) can be exploited through LiDAR. Different characteristics of vegetative structure have previously been demonstrated as important predictors for aspects of bird species habitat and diversity. For instance, information representing vegetation height has been both directly and indirectly correlated with  117  habitat suitability (e.g., North et al., 1999) and its importance towards biodiversity management has been verified (Bergen et al., 2009). However, few studies have employed the three-dimensional representation of vegetation offered by LiDAR for the explicit purpose of explaining avian diversity. The majority of these studies have relied solely on aspects of vegetation height and its variability, including Goetz et al. (2007), Clawges et al. (2008), Müller et al. (2009) and Müller et al. (2010), who all found these to be important explanatory variables. Within this context, distinct height ranges along continuous vertical profiles representing specific strata of canopy cover zones have also been found to be strongly correlated with avifaunal diversity (Willson, 1974; Hansen et al., 1995).  Our results further suggest that vegetation height and variability within  particular vegetation strata can help predict the occurrence of avian guilds. For instance, average midstory cover differentiated between and predicted richness in both opencountry and forest guilds. The ability of midstory cover to describe and differentiate between these guilds is likely attributable to a greater number of sampled forest guild species being directly associated with midstory habitat. Comparatively unique average midstory cover values accounted for a substantial fraction of variation in species richness within open-country and forest bird guilds. Similarly, average cover in the understory strata proved significant for describing forest bird richness, implying a specific association with habitat occurring ≤5m. Vegetation structural indices related to volume have also been proposed as potentially strong predictors of bird occurrence (Karr & Roth, 1971; Willson, 1974; Verner & Larson, 1989; Mills et al., 1991). Morgan and Freedman (1986) noted that volume is an effective surrogate for influential habitat characteristics such as successional stage and  118  age.  Along with maximum tree height and heterogeneity of tree height, within a  comparatively different and less complex vegetated system, Flaspohler et al. (2010) found LiDAR-derived vegetation volume to be amongst the strongest predictors of bird species richness (r² = 0.77, 0.78 and 0.78, respectively). We found some support for this with open-country birds, which are associated with habitat occupying a much smaller proportion of canopy volume in the euphotic strata as compared with forest birds. Variability within the euphotic strata permitted differentiation between guilds (Figure 6.1(A)) and facilitated describing richness for open-country birds.  6.4.4  Combining data  Combining information collected from LiDAR and multispectral sensors has been shown to improve the predictions of avifaunal diversity models (e.g., Goetz et al., 2007; Clawges et al., 2008). The synergistic benefits of combining LiDAR with hyperspectral data for studying diversity have also been demonstrated by Boelman et al. (2007). We found support for combining data types for models associated with the open-country guild, wherein the highest fraction of variance in species richness was accounted for based on hyperspectral and LiDAR-derived bio-indicators. In contrast, the simultaneous consideration of functional and structural variables did not improve models for generalist or forest guilds, for which as much variance was accounted for based on hyperspectralderived variables (i.e., generalist guild) or LiDAR-derived variables (i.e., forest guild) on their own. This implies that while aspects of function and structure both proved important for describing and differentiating between richness in all guilds, generalist and forest  119  habitat may be suitably described using either hyperspectral or LiDAR-derived variables on their own. Surprisingly, open-country habitat, which is most simply characterized and associated with birds that are comparatively easily observed, did not prove to be the most successfully described. For forest birds as much as 64.0% of variance in species richness was explained, whereas at best, 59.0% of variance in richness for open-country species was explained. As noted by Goetz et al. (2007), the most successfully described guild often involves species which are either most easily or most frequently observed, whereas the poorest is usually comprised of generalists whom by definition are expected to occur in a wide range of habitats. Our results support this, as forest birds were most frequently observed with the most variance explained and generalist species had the least amount of variance in richness explained (i.e., 32.0% at best).  120  6.5  Management Implications  This study identifies correlations between the richness of locally defined avian guilds and specific hyperspectral and/or LiDAR derived bio-indicators. Amongst hyperspectralderived variables, the importance of tree species diversity, and foliar water, carotenoid, lignin and anthocyanin content is clearly demonstrated. In addition, the significance of LiDAR-derived representations of cover strata (i.e., under, mid and overstory), heterogeneity in height, and mean volume properties is established. These relationships, which are affiliated with tangible aspects of ecosystem function and/or structure, can be exploited individually or in tandem to the benefit of disparate management goals in and around the Gulf Islands National Park Reserve. Numerous structural and compositional aspects govern avifaunal diversity (Flaspohler et al., 2010). As such, for certain guilds (i.e., open-country), the combined strength of both hyperspectral and LiDAR data is apparent. However, for other guilds, there is no added benefit in combining functional and structural attributes implying that hyperspectral (i.e., generalist guild) and/or LiDAR-derived (i.e., forest guild) variables are capable of characterization and differentiation on their own. Regardless, for all guilds, documented relationships help facilitate richness predictions throughout the extent of hyperspectral and/or LiDAR transects, supplying baseline information representing ecosystem condition from which measures of biodiversity and conservation value can be assessed. Periodic reassessment of richness estimates could help provide a mechanism for longterm monitoring wherein system change itself could be identified, investigated and quantified. While we make no claim that these relationships are direct and acknowledge  121  the potentially indirect nature of the cause and effect associated with correlations, we maintain that identifying specific variables which reflect the nature and condition of vegetation associated with preferred habitat types of native bird species is an extremely promising management tool.  6.6  Conclusion  This The design and implementation of effective conservation and management initiatives for highly valued ecosystems necessitates accurate and timely information representing faunal species composition (Nagendra et al., 2001; Turner et al., 2003). Using bio-indicators derived from airborne hyperspectral and LiDAR data, an increasing range of ecosystem characteristics pertaining to faunal habitat can be quantified across comparatively large areas in a detailed and accurate manner. Our results demonstrate that several sophisticated bio-indicators of vegetation condition proved useful for characterizing and differentiating between the richness values of locally defined avian guilds based on tangible aspects of ecosystem condition. Using faunal richness estimates derived from advanced remotely sensed data as surrogates to monitor long-term spatiotemporal dynamics could facilitate understanding, planning for and dealing with natural and anthropogenic changes. While these advanced geo-technologies do not offer a wholesale mechanism for describing faunal diversity, we suggest that the candidate pool of variables described in this study provides much promise for describing relationships with faunal diversity in an as of yet undetermined number of other environments. While our results provide a proof of concept, the full potential of these advanced data types as 122  tools for advancing methods which identify, characterize and facilitate conservation of high value systems remains unknown and as such warrants further investigation within a disparate range of environments.  123  7  EXTRAPOLATION OF TREE SPECIES MAPPING USING MEASURES OF HETEROGENEITY AND OBJECT-BASED CLASSIFICATION  7.1  Introduction  Identifying different tree species is critical for the effective management of forested ecosystems (Innes and Koch, 1996, Gong et al., 1997, Plourde et al., 2007, Dalponte et al., 2008, Voss and Sugumaran, 2008). Species-level data are utilized for many purposes, including inventory (Chubey et al., 2006), pest and invasive species mitigation (Everitt et al., 1996, Peterson, 2005, Morisette et al., 2006, White et al. 2006, and White et al., 2007), carbon sequestration calculations (van Aardt and Wynne, 2007), faunal habitat identification and characterization (Coops and Catling, 1997, Scarth et al., 1999), and biodiversity assessments (Nagendra, 2001, Turner et al., 2003). Conventionally, aerial photographs are interpreted to identify homogeneous patches of forest with similar attributes, containing a mixture of different species (Wulder et al., 2008a). Thus, aerial photograph interpretations often lack the spatial detail and accuracy required for applications requiring species-level information (e.g., locating, quantifying, and/or restoring rare and/or at-risk habitat) (Chapter 4, Chapter 5). Furthermore, because the interpretive process is extremely time and labor intensive, such inventories are not easily updated (Anderson et al., 1993, Leckie and Gillis, 1995, Gong et al., 1997, Gillis et al., 2005, Leckie et al., 2005, Evans et al., 2006, Lucas et al., 2008). Despite shortcomings and limitations, polygons delineated from aerial photographs remain a cornerstone of forest inventory (Wulder et al., 2008a). New methods are needed to provide managers  124  with more timely, cost-effective, consistent, detailed and accurate tree species information. Two technologies which permit characterizing tree species with decreased time and labor, and increased detail, accuracy, consistency and reproducibility, are fine spatial and spectral resolution airborne hyperspectral and light detection and ranging (LiDAR) remotely sensed data (Chapter 4, Chapter 5). Hyperspectral sensors acquire data in upwards of hundreds of extremely narrow spectral channels, providing an unprecedented tool for the detailed analysis of vegetation distribution, including differentiation of species based on subtle, bio-chemical properties (Okin et al. 2000, Ustin et al., 2004, Kumar et al., 2006, Smith, 2006). The successful generation of species-level maps from hyperspectral data has been demonstrated within tropical (e.g., Ustin et al., 2004, Carlson et al., 2007), sub-tropical (e.g., Lucas et al., 2008, Yang et al., 2009), and temperate (e.g., Boschetti et al., 2007, Chapter 4, Chapter 5) forest systems. In contrast, LiDAR sensors directly measure the vertical distribution of foliage, providing detailed information on vegetation height, cover, and structure (Lefsky et al., 2002, Wulder et al., 2008b). The ability of LiDAR to augment hyperspectral mapping endeavors has been verified by numerous studies (e.g., Geerling et al., 2007, Dalponte et al., 2008, Koetz et al., 2008, Chapter 4, Chapter 5). However, as with aerial photographs, advanced airborne data remain expensive to acquire and process (e.g., ~$5.00 USD per ha) (Wulder et al., 2008b). In addition, these data are normally collected in transects with coverage confined to  narrow swath  widths  (e.g.,  ~1km),  thus  limiting their applicability to  landscape/regional analysis (Chapter 5).  125  In contrast to these high (< 5 m) spatial resolution airborne/hyperspectral data sources, medium (5 – 30 m) spatial resolution multispectral satellite imagery (e.g., Landsat) are readily available at little to no cost and provide wall-to-wall landscape-level coverage (Franklin et al., 2003, Cohen and Goward, 2004, Woodcock et al., 2008). However, the primary disadvantage of Landsat imagery is that for most environs it is unable to provide detailed species-level information, due to both its limited spectral and spatial resolution. Thus, developing methodologies to extrapolate information derived from higher spatial resolution airborne sensors to the much broader spatial extent provided by satellite imagery has important ramifications for a variety of management goals (Vitousek et al., 1987, Huang and Asner, 2009, Chapter 4). However, effective and accurate linking of spatial data across multiple spatial scales, for the purposes of extrapolation, remains an incredibly vexing challenge. One approach for linking fine and moderate resolution data involves downscaling or coarsening the resolution of fine-grained imagery to correspond with coarser satellite imagery. Downscaling may involve the use of filters, such as a majority filter, which determines the dominant value within a predefined area or small grid. However, majority calculations are problematic as they risk inflating the presence of common values while under-representing rarer values (Thompson and Gergel, 2008). For the purposes of extrapolation over broad forested landscapes, such an impact is problematic as it can over-inflate the presence of dominant tree species while under-representing or potentially eliminating rarer species. Alternatively, measures of heterogeneity may be useful in degrading spatial resolution while maintaining important information content. Heterogeneity is the degree of spatial  126  variation exhibited within an ecosystem, and is widely appreciated to impact structure, function, and the distribution of biodiversity (Turner, 1989, Li and Reynolds, 1995, Huston, 1999, Hill and Smith 2005, Morgan and Gergel, 2010), and has proven relevant as an information source for restoring and conserving forested ecosystems (Sklenicka and Lhota, 2002, Lindenmayer et al., 2006, Morgan and Gergel, 2010). Heterogeneity measures are typically calculated at somewhat arbitrary spatial extents (Morgan and Gergel, 2010), such as within user defined windows, thus potentially limiting their applicability to forest management decisions made at varying scales. Therefore, heterogeneity measures would prove more applicable and representative if they corresponded with tangible, ecologically-relevant units (e.g., ecosystems, forest stands). Object-based methods assist with identifying ecologically-relevant units, as they emulate intuitive human interpretation of remotely sensed imagery (Castilla and Hay, 2008, Hay and Castilla, 2008). Through object-based analysis, segmentation initially partitions an image into spectrally and spatially similar multi-pixel regions, corresponding with definable scene elements (e.g., forest stands), which, instead of arbitrarily defined spatial entities (e.g., pixels), become the units of analysis (Woodcock and Harward, 1992, Blaschke and Strobl, 2001, Blaschke, 2003, Benz et al., 2004, Chubey et al., 2006, Castilla and Hay, 2008, Wulder et al., 2008a, Definiens, 2010). Image segments are then characterized according to their geometrical properties (e.g., size, shape) and/or their underlying statistical values (Chubey et al., 2006), which facilitates defining distinct classes to which all segments are assigned membership through object-based classification (Hay and Castilla, 2008). The utility of object-based methods for extrapolating fine spatial resolution yet spatially limited airborne metrics  127  across a broader area represented by Landsat data has been established, however, examples are scant, and are limited to extrapolating LiDAR-derived information (i.e., Wulder and Seeman, 2003). In this study we assess the use of object-based techniques to extrapolate tree species heterogeneity beyond the localized extent of non-contiguous hyperspectral/LiDAR flightlines to a broader area covered by Landsat imagery. First, to establish the superiority of coarsening through calculating heterogeneity, we compare the impact of coarsening tree species data derived from high spatial resolution imagery (airborne hyperspectral/LiDAR) using majority filtering versus three measures of heterogeneity. Secondly, we segmented a Landsat-5 TM image, providing wall-to-wall representation over hyperspectral/LiDAR flightlines, comprised of definable scene elements (e.g., forest patches) instead of image pixels. Thirdly, to establish the advantageousness of image segmentation, we compared representations of species heterogeneity at the scale of Landsat pixels versus variably sized image segments. We next pursued our primary goal: to develop and evaluate scaling relationships between species heterogeneity data and segmented Landsat data. To address this goal, we used regression trees to determine the strength of relationships between the statistical and geometric properties of Landsat segments (i.e., independent predictor variables) and their corresponding tree species heterogeneity values (i.e., dependent response variables). Then, using statistical relationships as rules defining distinct classes, we extrapolated species-level information to the extent of the British Columbian southern Gulf Islands (SGI) through object-based classification of the entire segmented Landsat scene. Lastly, we considered the  128  applicability of our results to local ongoing management goals and extension to other environs.  7.2  Methods  7.2.1  Study area  For a complete study area description, please consult section 2.2.1. Detailed, accurate, and comprehensive species-level information is currently unavailable for the forests of the SGI, but is increasingly required for a wide range of managerial tasks and initiatives (AXYS EC, 2004, Green, 2007, Chapter 4). Conventional ecological inventory information is available for the entire extent of the SGI in 1:10,000 (or coarser) scale polygons averaging 1.92 ha in size (derived from aerial photographs), however, it lacks the detail and accuracy vital to many managerial tasks (AXYS EC, 2004, Green, 2007, Chapter 4, Chapter 5). Estimated costs for acquisition, processing and interpretation of aerial photography were approximately $12.00 USD per ha, and required years to complete (Chapter 5).  7.2.2  Data  7.2.2.1 Hyperspectral/LiDAR-derived tree species data As described in Chapter 5, maps representing the distribution of 11 species had been produced (at a 2 m spatial resolution) through the classification of fused airborne  129  hyperspectral and LiDAR data, collected concurrently in mid-July, 2006 (Figure 7.1(a) provides an example of species data for a portion of the study area). Species maps provided detailed and accurate distribution information, with user’s and producer’s accuracies for most species ranging from >52-95.4 and >63-87.8%, respectively. In addition, estimated acquisition, processing and interpretation costs were approximately $6.00 USD per ha (Chapter 5). However, this more accurate, detailed and less costly species-level data was limited to 23 non-contiguous flightlines, ~1 km wide and of variable length, covering approximately ~2800 ha (Chapter 5). Therefore, the area of interest for this analysis is twofold: 1) the extent of airborne hyperspectral and LiDAR flightlines representing detailed tree species information, and 2) the extent of the BC SGI covered by Landsat-5 TM (Figure 7.2). For additional details on species maps, the reader should consult Chapter 5.  130  Figure 7.1. For a portion of the study area: (a) 2 m tree species distribution data derived from fused airborne hyperspectral/LiDAR data (Chapter 5) shown for and limited to the extent of three airborne transects. The background is area outside of the flightlines, represented by Landsat band 4 (NIR), (b) tree species richness calculated at a 30 m grain within and limited to the extent of airborne hyperspectral/LiDAR flightlines. The background is area outside of the flightlines, represented by Landsat band 4 (NIR), (c) Landsat-5 TM objects/segments falling within appropriate size thresholds and within the extent of airborne hyperspectral/LiDAR flightlines, and (d) 30 m tree species richness extrapolated beyond the extent of flightlines, based on rules defined through regression tree analysis and applied using object-based classification, wherein eight richness classes range from low to high.  131  7.2.2.2 Tree species heterogeneity To facilitate the extrapolation of species-level data confined to airborne flightline extents to the broader coverage of a multispectral satellite image, three measures of heterogeneity were applied to coarsen 2 m species data (within the extent of airborne hyperspectral and LiDAR flightlines (Figure 7.2)):  1. Tree species richness (R): the total number of species present within each window  2. Tree species diversity (D): calculated within each window using Simpson’s Index (Simpson 1949 (22)): D = 1 /Ʃpi2 Where pi represents the percentage cover of species i.  3. Tree species evenness (E): calculated within each window as: E = D/R  As heterogeneity calculations were based on 11 species, richness could range from 1-11. Similarly, diversity could range from 1-11; wherein the greater the value, the more species classes present and the more evenly distributed the classes. Diversity could only match richness if all species types were equally represented. Evenness could range from 1/R to 1, wherein a score of 1 occurred when all cover types are uniformly represented (Hill and Smith, 2005). Heterogeneity was calculated within increasingly coarser windows, ranging from 10-100 m.  132  Figure 7.2: The study area is two-fold, including: 1) the extent of hyperspectral/LiDAR flightlines shown in green/purple above, and 2) the British Columbian southern Gulf Islands, the extent of the figure as represented by Landsat-5 TM band 4 (near infrared) (30 km east/west by 35 km north/south, centered at latitude 48.76º and longitude -123.18º). 133  7.2.2.3 Tree species dominance To facilitate a multi-scale comparison of the impact of coarsening species data using filtering (i.e., majority) versus heterogeneity measures, and to confirm the superiority of heterogeneity as a coarsening technique, the majority (e.g., dominant) tree species was also determined within increasingly coarser windows, ranging from 10-100 m. Dominance was defined as the most common species within each window. As calculations were based on 11 possible species, the majority within each unit (i.e., pixel) at each scale was 1 of 11 possible species classes.  7.2.2.4 Landsat data To facilitate the extrapolation of tree species heterogeneity, a broader scale landscape representation (coarser resolution and broader spatial extent) was provided by a Landsat5 TM scene comprehensively covering the BC SGI (path: 47/row: 26) and encompassing the hyperspectral/LiDAR acquisition extent and timeframe (July, 2006). The image was delivered by MacDonald, Dettwiler and Associates (MDA) Ltd. and then precision georegistered, orthorectified, and resampled to a spatial resolution of 30 m, with an RMS error of 0.25. At-surface reflectance atmospheric correction was based upon the Cos(t) model (Chavez, 1996), which estimated the effects of absorption by atmospheric gases and Rayleigh scattering and removed systematic atmospheric haze. All Landsat bands were then masked to remove non-forest pixels from the analyses.  134  7.2.3  Segmentation of Landsat image  In order to extract ecological units from the Landsat imagery the image was segmented using eCognition Developer version 8.0.1 (Definiens, 2010). Objects or segments are formed by the creation of groups of similar homogenous pixels in a process called segmentation which  merges pixels using a bottom-up, pair-wise, region-growing  technique, which minimizes the heterogeneity within segments and maximizes their homogeneity (Wulder et al., 2004, Definiens, 2010). Once segmented, objects can be characterized based on their inherent spectral and geometric properties, as well as by their location and position relative to other objects. Based on segment-level attributes, rules can be developed which define distinct classes to which all segments can be assigned membership (Hay ad Castilla, 2008).  7.2.4  Selection of segment features (independent variables)  The characteristics of image segments (or objects) considered for use in subsequent regression analyses included 19 aspects of segment geometry and 11 Landsat-layer specific statistics (Table 7.1). This suite of potential independent predictor variables was extracted  from  all  Landsat  segments  completely  within  the  extent  of  hyperspectral/LiDAR flightlines and imported into Statistica version 7.0 (StatSoft, 2004) for analysis. Pearson-r correlation coefficients (p<0.05) were then calculated for all possible variable combinations. Based on coefficients of determination, an r2 threshold of 0.85 eliminated redundant variables. Where two or more variables were highly correlated, emphasis was placed on the simpler and/or more ecologically relevant variable (Morgan 135  and Gergel, 2010). Extremely small and large (outlier) segments were eliminated based on minimum and maximum segment sizes of 0.54 (i.e., 6 pixels) and 2.7 ha (30 pixels), respectively.  Layer summaries  Geometry/shape  Mean  area (meters)  standard deviation  area (pixels)  Skewness  border length  Minimum  length  maximum  length/width  mean inner border  volume  mean outer border  width  border contrast  asymmetry  contrast to neighbor pixels  border index  edge contrast to neighbor pixels  compactness  standard deviation to neighbor pixels  Density elliptic fit main direction radius of largest enclosed ellipse radius of smallest enclosed ellipse rectangular fit Roundness shape index  Table 7.1: Aspects of segment geometry and segment-level spectral (i.e., Landsat) features originally considered as independent predictor variables for building models to extrapolate fine tree species heterogeneity (i.e., richness, diversity, and evenness) data derived from airborne hyperspectral/LiDAR data to the broader extent of a Landsat scene.  7.2.5  Extraction of segment-level heterogeneity values (dependent variables)  Richness, diversity and evenness values were determined for all Landsat derived segments within the extent of flightlines, Segment-level heterogeneity values initially facilitated comparison with pixel-level heterogeneity values, which aimed to establish the  136  merit of image segmentation. In addition, segment-level heterogeneity values served as the response variables in regression tree analysis and subsequent extrapolation using object-based classification.  7.2.6  Model creation and validation  Regression tree analyses (Breiman et al., 1984, Steinberg and Colla, 1995) established which independent predictor variables exhibited relationships with and explained variance in species heterogeneity and further served to define rules for extrapolation. Regression trees analysis recursively partitions a dataset into increasingly homogenous subsets, determining which of a candidate pool of independent variables can be used to predict a response (dependent) variable (i.e., tree species heterogeneity measures), and how much variance each predictor variable accounts for. Resulting trees supply a series of Boolean then statements based on specific value thresholds of important predictor variables, and end in terminal nodes which define distinct classes, permitting a transparent, straightforward interpretation of results (Chubey et al., 2006, Bater and Coops, 2009). In addition to transparency, regression trees accommodate high dimensionality data sets, make no assumptions about input variables or their statistical distributions, and are typically robust to errors present in either independent or dependent variables (Breiman et al., 1998, Chubey et al., 2006). A regression tree was constructed to predict each measure of heterogeneity (i.e., richness, diversity, and evenness), wherein segment-level geometric properties and/or statistical summaries of Landsat data were the independent predictor variables. For each tree, rules defined the membership parameters for distinct classes of heterogeneity based on specific thresholds corresponding with 137  specific predictor variables. K-fold cross-validation (K=10) determined optimal tree size and accuracy for each tree (Venables and Ripley, 2002), wherein explained variance measured how well each tree fit the data.  7.2.7  Extrapolation of species heterogeneity through object-based classification  Using the rules generated through regression tree analysis all image segments were assigned memberships to richness, diversity, and evenness classes through the objectbased classifier resulting in three coverages representing each measure of heterogeneity (i.e., richness, diversity, and evenness) for the extent of the BC SGI. As class assignment was based entirely on defined rules, the accuracy of resulting maps reflected the amount of variance explained by each associated regression tree, and therefore, additional accuracy assessment was not conducted. Using this transparent rule-based object-level classification approach avoided salt-and-pepper effects common to traditional pixel-based approaches (Yu et al., 2006), permitting more realistic final products. In addition, using segments instead of pixels as the unit of analysis reduced computational time by orders of magnitude (Hay and Castilla, 2008).  7.3  Results  7.3.1  Tree species dominance  Increasingly coarser majority filters applied to the original 2 m species data (Figure 7.1(a)) (from the hyperspectral/LIDAR) changed the amount (%) of forested land  138  occupied by each species (Figure 7.3) in five key ways, 1) the overall amount of forested land occupied by a species continuously decreased at each coarser spatial resolution, the most common scenario, impacting five species (i.e., black cottonwood, grand fir, red alder, Western hemlock and bigleaf maple), 2) the overall amount occupied rose, but eventually decreased overall, as demonstrated by trembling aspen and Western redcedar, 3) the amount of forested land occupied continuously increased, as exhibited by arbutus, 4) the amount of the forested land occupied initially decreased, but eventually increased, as seen with Douglas-fir, or 5) the amount fluctuated, trending towards an increase, but ending up approximately the same (i.e., lodgepole pine and Garry oak). Increasing coarser spatial resolution typically resulted in an increase in the most dominant species (i.e., Douglas-fir), with decreases in rarer species (e.g., black cottonwood, trembling aspen, grand fir, Western hemlock, bigleaf maple). Coarsening to the spatial resolution of Landsat (i.e., 30 m) exaggerated the proportion of Douglas-fir, trembling aspen, Garry oak, lodgepole pine, arbutus and grand fir by 10, 25, 51, 69, 96 and 307 %, respectively, while many other species exhibited reductions of 5, 16, 38, 45 and 61 % (i.e., Western redcedar, red alder, bigleaf maple, black cottonwood, Western hemlock, respectively).  139  Figure 7.3: The effect of coarsening spatial resolution on the amount (%) of forested land occupied by each species.  7.3.2  Tree species heterogeneity (pixel-level)  Calculating species heterogeneity at increasingly coarser spatial resolutions resulted in markedly different representations of richness, diversity and evenness (Table 7.2). Minimum species richness remained 1 regardless of resolution, while the maximum, mean and standard deviation steadily increased as spatial resolution increased from 10 (7, 1.48 and 0.71 respectively) to 100 m (11, 4.68 and 2.1, respectively). Minimum diversity also exhibited no change across scales, and similar to richness, the mean and standard deviation steadily increased as spatial resolution increased from 10 (1.25 and 0.44) to 100 m (1.62 and 0.67). Although maximum diversity fluctuated it did not increase 140  consistently. While maximum species evenness remained consistent at all scales, the minimum and mean consistently decreased as spatial resolution increased from 10 (0.31 and 0.9) to 100 m (0.12 and 0.42).  Spatial Resolution  Heterogeneity  Minimum  Maximum  Mean  Standard deviation  10  Richness  1  7  1.48  0.71  Diversity  1  5.45  1.25  0.44  Evenness  0.31  1  0.9  0.16  20  30  40  50  60  70  80  90  100  Richness  1  9  2.04  1.08  Diversity  1  6.17  1.4  0.56  Evenness  0.21  1  0.77  0.22  Richness  1  10  2.5  1.33  Diversity  1  5.88  1.48  0.62  Evenness  0.19  1  0.67  0.25  Richness  1  9  2.94  1.5  Diversity  1  5.61  1.52  0.64  Evenness  0.16  1  0.6  0.24  Richness  1  10  3.33  1.64  Diversity  1  5.86  1.55  0.65  Evenness  0.14  1  0.55  0.24  Richness  1  10  3.67  1.76  Diversity  1  5.83  1.57  0.66  Evenness  0.13  1  0.51  0.24  Richness  1  10  3.96  1.86  Diversity  1  5.56  1.59  0.67  Evenness  0.13  1  0.48  0.24  Richness  1  11  4.21  1.95  Diversity  1  5.83  1.6  0.67  Evenness  0.13  1  0.45  0.23  Richness  1  10  4.45  2.02  Diversity  1  6.19  1.61  0.67  Evenness  0.12  1  0.44  0.23  Richness  1  11  4.68  2.1  Diversity  1  5.69  1.62  0.67  Evenness  0.12  1  0.42  0.23  Table 7.2: The effect of coarsening spatial resolution (i.e., pixel size in meters) on three measures of tree species heterogeneity (i.e., richness, diversity and evenness) derived from airborne hyperspectral/LIDAR data within and limited to the extent of flightlines. 141  7.3.3  Landsat segmentation  Segmentation resulted in 8,190 segments of which 1,818 were completely within the area of the hyperspectral/LiDAR flightlines (example segments shown for a portion of the study area in Figure 7.1(c)). Segment size ranged from 0.54-2.7 ha, averaging 1.23 ha (±0.54).  7.3.4  Tree species heterogeneity: comparison of pixel to segment-level  Comparing average pixel with segment-level statistics highlights the discrepancy in the range of predicted richness and diversity values (Table 7.3). Focusing on the spatial resolution of Landsat (i.e., 30 m), Figure 7.1 (b) provides an example of tree species heterogeneity (richness) calculated within the extent of hyperspectral/LiDAR flightlines for a portion of the study area. At this scale, no single segment contained all 11 species classes, with richness values ranging from 1-10 (Table 7.3, Figure 7.4 (a)), however, richness averaged 2.60 (± 1.33 (1 standard deviation)). Similarly, no single pixel contained species with equal or somewhat equal distribution, as diversity ranged from 1.00-5.90, averaging 1.48 (± 0.61) (Table 7.3, Figures 7.4 (b)). Species evenness ranged from 0.19-1.00, averaging 0.66 (± 0.24) (Table 7.3, Figures 7.4 (c)). For evenness, the maximum value (i.e., 1) occurred most often, representing pixels for which all species present were equally distributed. In contrast, at the segment-level, while minimum species richness remained 1, maximum richness was substantially more (i.e., 5.58 vs. 10.00) (Table 7.3, Figure 7.4 (a, d)). Similarly, while minimum species diversity values were 1 at both scales, maximum segment diversity was 1.45 more (i.e., 3.45 vs. 5.90)  142  (Table 7.3, Figure 7.4 (b, e). For species evenness, maximum values were the same at both scales, however, minimum values were typically higher for segments. Mean values for all three metrics were similar if not the same at both scales, whereas standard deviations were typically lower for segments (Table 7.3, Figure 7.4 (c, f)).  Pixel  Segment  Richness  Pixel  Segment  Diversity  Pixel  Segment  Evenness  Minimum  1  1  1  1  0.19  0.33  Maximum  10  5.58  5.9  3.45  1  1  Mean Standard deviation  2.6  2.66  1.48  1.48  0.66  0.64  1.33  0.9  0.61  0.4  0.24  0.12  Table 7.3: Comparison of descriptive statistics for tree species heterogeneity (i.e., richness, diversity and evenness) values calculated from and within the flightlines of airborne hyperspectral/LiDAR data, within 30 m image pixels versus variably sized image segments.  143  Figure 7.4: Comparison of frequency distribution for tree species heterogeneity values (i.e., richness, diversity and evenness) calculated from and within the flightlines of airborne hyperspectral/LiDAR data, within 30 m image pixels versus variably sized image segments. The distribution of richness values is represented by (a) 30 m pixels and (d) variably sized segments. Similarly, diversity and evenness are represented by (b) 30 m pixels and (e) variably sized segments, and (c) 30 m pixels and (f) variably sized segments, respectively.  7.3.5  Model definition and validation (regression tree analysis)  Based on pair-wise Pearson’s correlation coefficients, ecological significance, and preliminary relationship assessment (i.e., scatter-plots), 21 potential metrics representing the statistical and/or geometrical properties of Landsat segments were retained for  144  analysis as potential independent variables (Table 7.4). Of the potential predictor variables, regression tree analysis identified eight which partially explained variance within dependent response variables (i.e., measures of heterogeneity) (Tables 7.4). For species richness, eight distinct classes were identified based on the statistical properties of Landsat bands 4 (NIR) and 5 (SWIR), with generated decision rules accounting for 48.9% of variance. Similar to richness, for diversity, a final model with eight terminal nodes was generated using bands 4 and 5, accounting for 43.93% of variance. The importance of band 7 (SWIR) was also apparent. For species evenness, six terminal nodes defined rules which also indicated bands 4 and 5 were the important predictors, however, as compared with diversity and richness, much less variance was explained (i.e., 22%).  Wavelength (µm)  Minimum (Min)  Maximum (Max)  Mean  Standard deviation (Stdev)  band 1  0.45-0.52  x  band 2  0.52-0.60  x  x  band 3  0.63-0.69  x  x  band 4  0.76-0.90  xx  xx  band 5  1.55-1.75  xx  xx  xx  band 7  2.08-2.35  NDWI  band4-band5 band4+band5  xx x  xx x  x  x  x  x  Mean inner border (MIB)  xx x x  Table 7.4: Independent variables representing aspects of segment geometry and segment-level spectral (i.e., Landsat statistics) considered for and selected in regression tree analyses. Selected variables explained variance in tree species richness, diversity and/or evenness. An ‘x’ represents variables considered for but not selected in analyses, whereas an emboldened ‘xx’ represents selected variables. Greyed out cells represent variables which were not considered for analysis, based on pairwise Pearson’s correlation coefficients (p<0.05), a lack of ecological significance and/or preliminary relationship assessment.  145  Using richness as an example, Figure 7.5 provides an example of generated decision rules. The initial split was on average band 5 values of 0.09 (% reflectance/100), partitioning the image into segments with higher (>0.09) or lower (≤0.09) richness values. Segments with higher richness values were further partitioned based on minimum band 4 and 5 values, maximum band 4 values, and mean 5 band values. The highest richness classes were associated with minimum band 5 and 4 values of >0.09 and 0.23, respectively, isolating segments with the highest minimum reflectance thresholds for these spectral regions. In contrast, segments with lower richness values were further partitioned based on maximum values of 0.09 and 0.31 for bands 5 and 4, respectively. Segments with the least richness had maximum band 5 values ≤0.09, whereas segments with maximum 4 values >0.31 had slightly higher values.  146  Figure 7.5: Decision rules generated through regression tree analysis, wherein tree species richness was the dependent response variables, and segment-level spectral properties of Landsat data were selected as independent predictor variables. Generated rules identify spectral thresholds which define distinct richness classes, while explaining nearly 50% of variance.  7.3.6  Object-based classification (extrapolation)  Based on generated decision rules, three maps were produced representing comprehensive tree species richness, diversity and evenness. An example of extrapolated richness is shown for a portion of the study area in Figure 7.1 (d), whereas Figure 7.6 presents extrapolated richness for the extent of the BC SGI.  147  Figure 7.6: Tree species richness, derived from 2 m spatial resolution species distribution maps (from and with the flightlines of airborne hyperspectral/LiDAR data) extrapolated to the extent of the British Columbian Southern Gulf Islands. Richness was extrapolated based on relationships established with specific spectral properties of Landsat segments. Species richness has eight distinct classes, ranging from low to high values. The background image is Landsat band 4 (NIR) coverage.  148  7.4  Discussion  7.4.1  Comparison of coarsening techniques  Results confirmed that heterogeneity measures were a more appropriate technique for coarsening the spatial resolution of detailed airborne data. While heterogeneity measures cannot discriminate how much of a species exists within a predefined area, and increasingly coarser spatial resolutions can alter information content, their use involved less severe information alteration then majority filtering. This finding has important ramifications, given the common use of filters such as the majority for coarsening spatial environmental data. Therefore, given the established links of heterogeneity with ecological structure and function and its superiority as a coarsening method, its use is encouraged.  7.4.2  Comparison of pixel versus segment-level heterogeneity  There were notable differences in heterogeneity values at the scale of Landsat (i.e., 30 m pixels) versus the variable scale of segments. Owing to their multi-pixel nature, segments exhibited less range in richness, diversity and evenness values. Pixels with high values were typically reduced within larger segments, thus reducing their overall range. In addition, at the pixel-level, there was a markedly higher frequency of low richness and diversity values, and high/maximum evenness values. This discrepancy is partly explained by the dominance of Douglas-fir, which occupies the majority of the forested landscape, regardless of scale (i.e., 79.1% (at 2 m), 79.2% (at 30 m), 84% (at 100 m)). If only one species occupies a pixel, it is still assigned a maximum evenness score, whereas 149  pixels occupied by fewer species receive lower richness and diversity scores. While many segments were dominated by Douglas-fir, other species were typically also present. In addition very few segments were completely occupied by a single species, even Douglasfir. In aggregate, differences in values imply that information loss and/or manipulation does occur with consideration at the pixel versus segment scale, however, these changes are required to represent the landscape based on definable units (e.g., forest units) as opposed to arbitrary units (e.g., image pixels) (Hay and Castilla, 2008, Smith et al., 2008). In addition, as also indicated with cross-scale pixel-level analysis, the impacts of coarsening using heterogeneity measures are not as severe as they would be if employing standard filtering procedures (i.e., majority).  7.4.3  Regression tree results  Promisingly, regression trees provided transparent rules explaining approximately half the variance in richness and diversity. For all three heterogeneity measures, independent explanatory variables were consistently Landsat bands 4, 5 or 7, demonstrating the importance of vegetation’s reflective properties in certain NIR and SWIR spectral regions. In band 4 (i.e., NIR: 0.76-0.90 micrometers (µm)), reflectance is primarily regulated by internal leaf structure (Sinclair et al., 1971), whereas in bands 5 (i.e., SWIR: 1.55-1.75 µm) and 7 (SWIR: 2.08-2.35 µm), reflectance is much lower than in the NIR and dominated by the presence of moisture and various canopy biochemicals (e.g., protein, lignin, cellulose) (Curran, 1989, Elvidge, 1990, Kumar et al., 2006, van der Meer, 2006). Established associations between vegetation and spectral properties in the NIR and SWIR permit informed and meaningful ecological characterization of image 150  segments. Using richness as an example, segments with band 5 values >0.09 exhibited higher values, meaning they contained multiple tree species. Multi-specied segments are more prone to host coniferous and broadleaved species. Due to the comparatively higher levels of biomass associated with broadleaves, segments containing broadleaved species should exhibit higher SWIR reflectance values. Therefore, segments with higher SWIR values were more likely to be richer and contain broadleaved tree species. In addition to high SWIR values, highest richness classes were also associated with highest NIR values. Similar to the SWIR, healthy broadleaves exhibit higher reflectance than conifers in the NIR. Therefore, because segments with higher richness values are more likely to contain broadleaved species, they will typically exhibit higher SWIR and NIR reflectance. Exceptions to these associations could include, 1) segments containing trees exhibiting signs of stress and/or 2) a disproportionate dominance of multi-specied segments known to contain none or few broadleaved trees, however, previous studies (i.e., Chapter 4, Chapter 5) indicate these exceptions are rare in the SGI.  7.4.4  Management implications for extrapolated heterogeneity  7.4.4.1 Management implications in the SGI Regression rules generated on specific spectral properties of NIR and SWIR bands transparently facilitate the extrapolation of heterogeneity to the extent of the SGI through providing rules which assign membership into distinct classes for all image segments. For managers in the SGI, object classification results supply previously unavailable broad scale tree species heterogeneity information for the same extent as existent delineated  151  photographs. This information can aid a range of ongoing restoration and conservation tasks, and facilitates long term management goals (Chapter 4, Chapter 5).  7.4.4.2 Improvements over conventional methods While similar measures could be calculated from aerial photo attributes, as detailed in Chapter 4, our heterogeneity measures were derived from comparatively more detailed and accurate tree species distribution information (i.e., 2 m spatial resolution maps derived from airborne hyperspectral/LiDAR data). In addition, even after factoring in the time and labor associated with parameterization, image segmentation and classification require significantly less time and labor than conventional methods. Aerial photo interpretation also requires non-automatable manual labor, which relies on increasingly less available skilled analysts (Wulder et al., 2008a), and therefore, once established, final scaling parameters permit transparent, timely replication. In terms of costs, it has previously been demonstrated that the acquisition, processing and interpretation costs of hyperspectral and LiDAR data in the SGI are half the price of aerial photographs (Chapter 5). While this will not be the case everywhere, the costs of these technologies have been found comparable in numerous and variable environs (Wulder et al., 2008b). Within this monetary context, it is important to note that the advanced airborne data used in this study were collected in transects vs. the wall-to-wall acquisition of aerial photographs. However, relationships with Landsat segments facilitated low-cost extrapolation from the limited extent of detailed airborne measurements.  152  7.5  Conclusion  Image segmentation, regression tree analyses, and segment classification provided a transparent and effective means for extrapolating certain measures of tree species heterogeneity beyond the boundaries of non-contiguous hyperspectral and LiDAR flightlines to the wall-to-wall coverage provided by a Landsat scene. While heterogeneity measures alter the information content of finer scale species data, as opposed to filtering techniques, they do so in a less invasive and more meaningful manner. At the segment scale, calculating heterogeneity results in definable forest units containing tangible heterogeneity values, which as units of analyses, have pertinence to a variety of ongoing and/or potential management initiatives demanding species-level data. Image segmentation permits this comparatively superior landscape representation, and based on transparent relationships with spectral properties of Landsat data deduced through regression tree analyses, object-based classification permits extrapolation, resulting in unprecedented heterogeneity information for the extent of the SGI explaining nearly half the variance in tree species richness and diversity. In the SGI, the spectral properties of Landsat bands 4 (i.e., NIR: 0.76-0.90 µm), 5 (i.e., SWIR: 1.55-1.75 µm) and 7 (SWIR: 2.08-2.35 µm) were consistently selected as independent predictor variables, underscoring the importance of including NIR and SWIR channels in terrestrial forested ecosystem landscape-level analyses. Of the three measures of heterogeneity, richness performed best, followed closely by diversity. While evenness was predicted comparatively worse, future work involving additional independent variables may exhibit stronger relationships. Incorporating additional independent predictor variables may also explain more variance in richness and diversity. 153  In addition, generalizing the relationships exhibited between richness, diversity and properties of the NIR and SWIR remains untested, however, based on known associations between these spectral regions and vegetative properties, further evaluation in comparable and/or disparate environs is anticipated to yield similar results and the objectlevel analysis of tree species heterogeneity moves beyond the arbitrary pixel-scale into the realm of definable forest units, a scenario which most likely holds great promise for a wide range of environs.  154  CONCLUSIONS  Mapping and monitoring terrestrial ecosystems is critical for effective, sustainable forest management. While the delineation and interpretation of aerial photographs remains a foundation of ecological inventory, advanced remotely sensed data offer new approaches for accurately quantifying critical forest attributes. The objective of this thesis was to demonstrate and test how new generation remote sensing technologies can be integrated with field data to improve conventional forest species and structural information in the southern Gulf Islands. To meet this objective, 6 questions were posed in this thesis, the answers to which reveal the numerous advantages of using these innovative technologies within a terrestrial ecosystem mapping context. First, can ground-based spectroscopy distinguish between 11 dominant tree species? In Chapter 2, discriminant analyses of vegetation spectral curves measured using a groundbased full-range spectrometer elucidated specific spectral regions which significantly differentiated between all examined species (α level 0.05, overall cross-validation accuracies ≥98%, producers accuracies >85%). Most notably, spectral regions 501-550 nm (i.e., chlorophyll reflection) and 681-740 nm (i.e., red edge transition) exhibited consistent importance with respect to spectral discrimination, whereas wavelengths > 1794 nm had minimal to no influence. Results guided the selection of optimal airborne hyperspectral channels, thus scaling leaf-level observations to the canopy-scale. While there are notable differences in reflection at leaf versus canopy scales resulting from disparate influences, including non-photosynthetic vegetation, shadowing, understory presence, and/or atmospheric effects (among others), the results in Chapter 2 provide  155  insight for selecting specific spectral regions, thereby reducing data dimensionality and noise in airborne imaging spectrometer for applications (AISA) data-sets, as well as defining the wavelengths optimized for species classification used in subsequent research. The results of this chapter supported the findings of previous studies (e.g., van Aardt and Wynne, 2001, Clark et al., 2005), demonstrating the utility of feature selection methods employing discriminant analyses to reduce data dimensionality. In addition, Chapter 2 establishes a precedent for using ground-based spectrometry to spectrally describe and distinguish among ecologically important coastal Pacific Northwest tree species. In addition to the spectral differentiation of species, tree and stand structure are critical components of forested ecosystems. Light detection and ranging (LiDAR) sensors actively measure the vertical distribution of forest elements, providing a cutting-edge tool for characterizing vegetation structure (Lefsky et al., 2002, Wulder et al., 2008b). While a wide variety of metrics can be derived from LiDAR data, the three types most common to the literature include, 1) canopy height descriptors (CHDs), 2) height percentiles (HPs), and 3) canopy volume profiles (CVPs) (Lim et al., 2003, Wulder et al., 2008b). Question 2 asked which, if any, of these three suites of LiDAR metrics could differentiate among British Columbia’s terrestrial ecosystem mapping (TEM) initiative defined structural classes (herbaceous (HB), shrub/herbaceous (SH), pole/sapling (PS), young forest (YF), mature forest (MF), and old forest (OF)). Results described in Chapter 3 verified all three types of metrics could significantly differentiate (i.e., p≤0.01) between certain structural classes, however, the number and types of metrics able to distinguish between particular combinations decreased with stand age and complexity. Specifically,  156  HPs, CHDs, and CVPs could differentiate HB, SH and PS from all other structural types, and certain CHDs and CVPs could further distinguish among older, more complex classes. While older and comparatively complex structural classes were only differentiated using a few variables, encouragingly, all structural class combinations could be separated. In addition, the results of Chapter 3 are significant in that they identify explicit relationships between specific LiDAR variables and structural classes within the SGI, which permits enhancing structural detail within TEM units and/or predicting structural classes for the extent of LiDAR surveys. Furthermore, relationships between certain metrics and forest structure resulted in targeting specific metrics for inclusion as classification inputs in Chapters 4 and 5. Once spectral and stand structural stages were shown to be differentiable using leaf-level spectrometry and airborne LiDAR, question 3 (Chapter 4) asked if leaf-level spectral measurements could be scaled to airborne hyperspectral data to identify airborne channels optimal for species differentiation, and subsequently fuse optimal channels with targeted LiDAR metrics to predict the distribution of rare and/or at-risk Garry oak habitat. Chapter 4 accurately inventoried Garry oak distribution within an ~600 ha methodological test site (producer’s and user’s accuracies >81.5%), emphasizing previous findings that hyperspectral data effectively predicts tree species within temperate forest systems (e.g., Goodwin et al., 2005, Boschetti et al., 2007). Within the broader framework of the Parks Canada mandate to restore, conserve and maintain ecological integrity (Woodley, 1993, Parks Canada Agency, 1997), and with direct applicability to specific GINPR management goals and initiatives, Chapter 4 successfully located and quantified three types of Garry oak habitat (i.e., rocky bluff with woodland  157  patch, steep slope woodland, field with woodland patch (Green, 2007)) at a 2 m spatial resolution, significantly increasing detail and specificity compared with 1:10,000 scale (or coarser) TEM polygons mapped with a minimum unit of 0.04 ha (i.e., 400m2). Accurately characterizing Garry oak distribution filled a current knowledge gap, not only quantifying its occurrence within TEM units, but also locating previously unidentified habitat, directly aiding several primary management goals, including 1) restoring and 2) conserving degraded Garry oak habitats, while working towards 3) expanding conserved habitat and 4) reestablishing network connectivity through targeted land acquisition (i.e., >3 ha of previously unidentified private lands occupied by Garry oak were located), and 5) providing baseline contemporary reference information from which restoration activities can be assessed and revised as necessary (Hobbs and Harris, 2001, AXYS EX, 2004, Green, 2007, Hobbs, 2007). Question 4 sought to determine if optimal hyperspectral channels could be fused with targeted LiDAR metrics to predict the distribution of 11 tree species over the extent of airborne surveys. Building on the success of Chapter 4, Chapter 5 mapped the distribution of all 11 species over the entire 2800 ha extent of airborne surveys (yielding producer’s and user’s accuracies for most species ranging from >52-95.4 and >63-87.8%, respectively), supporting previous findings that hyperspectral data effectively predicts tree species distribution in temperate forest systems (e.g., Martin et al., 1998, Xiao et al., 2004). Results further indicated that including LiDAR metrics improved overall accuracies and many individual class accuracies. Pixel-level fusion of both height and volumetric canopy profile data increased producer’s (+5.1-11.6%) and user’s (+8.418.8%) accuracies for species characterized by distinct growth stages (e.g., black  158  cottonwood and lodgepole pine (height), arbutus and Western hemlock (volume)). Despite less pronounced differences in overall accuracies associated with including LiDAR data (e.g., +0.6-1.2%), McNemar’s tests confirmed overall increases were statistically significant (p<0.05). The results of Chapter 5 also directly augmented the inventory of Garry oak habitat presented in Chapter 4, providing detailed and accurate estimates for three habitat types for the extent of surveys. Chapter 5 also demonstrates the innovative application of employing a non-parametric support vector machine (SVM) classifier. Whereas conventional parametric classification approaches (e.g., maximum likelihood) are limited in their ability to classify high dimensional, multi-source data, SVMs have been shown to outperform conventional approaches (Dalponte et al., 2008, Koetz et al., 2008). Specific to hyperspectral data, as compared with conventional parametric classification algorithms and other non-parametric algorithms, (e.g., neural networks), SVMs have improved classification accuracies, computational time, and stability (Bruzzone and Carlin, 2006, Dalponte et al., 2009). Chapters 4 and 5 build upon the work of a limited number of studies employing SVMs for hyperspectral/LiDAR fusion (e.g., Dalponte et al. 2008). In addition, Chapters 4 and 5 are amongst a small number of studies involving airborne hyperspectral AISA sensors in south-western coastal Canada (i.e., Goodenough et al., 2008, Goodenough et al., 2009). Furthermore, the results of Chapter 4 and 5 are amongst the first to document successful, pixel-level fusion of hyperspectral/LiDAR data for vegetation type (i.e., Geerling et al., 2007, Koetz et al., 2008) and/or species mapping (i.e., Dalponte et al., 2008). While the utility of LiDAR-derived height information for increasing accuracy for species classes exhibiting similar spectral properties, yet different average heights, has previously been  159  demonstrated (i.e., Geerling et al., 2007, Koetz et al., 2007, Dalponte et al., 2008), Chapter 5 is innovative in that it verifies that three-dimensional LiDAR-derived volume increases species accuracies based on distinguishably unique architectural arrangements (e.g., arbutus; Western hemlock). The fifth question assessed if ecosystem function metrics (e.g., foliar lignin, water and/or carotenoid content) and tree species diversity derived from hyperspectral data and/or aspects of vegetation structure (e.g., mid and/or overstory cover; euphotic canopy volume) derived from LiDAR could describe the richness of, and differentiate among, avian guilds. Chapter 6 characterized the habitat of, and species richness for three distinct avifaunal guilds (i.e., open-country; generalist; forest). Previous studies have established the importance of LiDAR-derived height metrics as important explanatory variables for avian diversity (i.e., Goetz et al., 2007, Clawges et al., 2008, Müller et al., 2009, Müller et al., 2010). Furthermore, one study has demonstrated the importance of LiDAR-derived volume as an explanatory variable for avian diversity (i.e., Flaspohler et al. 2010). Chapter 6 supports these findings, identifying significant relationships (p<0.05) between avian richness and average mid and over-story cover (i.e., height), and variability in the euphotic strata (i.e., volume). In addition to LiDAR metrics, the results of Chapter 6 are innovative as they establish the importance of hyperspectral derived vegetation indices (i.e., estimates of foliar water and/or carotenoid content) for differentiating between avian guilds (p<0.01) and describing richness. A unique relationship between tree species diversity and open-country bird richness was also exposed. Lastly, Chapter 6 supports the findings of Goetz et al. (2007), Clawges et al. (2008) and Boelman et al. (2007), which purport that because both structural and compositional forest attributes govern avifaunal  160  diversity (Flaspohler et al. 2010), the highest fraction of variance in species richness models is accounted for through combining hyperspectral and LiDAR-derived bioindicators. The first five questions were directed towards the utilization of hyperspectral and LiDAR data within the extent of airborne surveys, thus limiting regional-scale management applications. The sixth and final question asked if detailed and accurate species information could be spatially extended beyond limited survey coverage to the extent of a multispectral, medium spatial resolution satellite image. Image segmentation, regression tree analysis, and object-based classification were successfully applied to predict measures of tree species heterogeneity beyond the extent of flightlines based on the coverage offered by, and relationships derived with, a segmented Landsat-5 TM scene (Chapter 7). Regression trees confirmed that significant segment-level relationships existed between the statistical properties of Landsat data bands 4 (i.e., NIR: 0.76-0.90 µm), 5 (i.e., SWIR: 1.55-1.75 µm), and 7 (SWIR: 2.08-2.35 µm) and each measure of heterogeneity, explaining upwards of approximately 50% of variance and showing the importance of vegetation’s reflective properties in NIR and SWIR. The characteristics of these spectral regions facilitated categorizing all image segments into distinct classes based on transparent, ecologically meaningful empirical relationships. In addition, using heterogeneity measures to coarsen tree species data was confirmed as superior to majority filtering, involving less severe information alteration and providing a firm ecological link with tangible forest units. Furthermore, the comparatively cheap acquisition and processing costs of Landsat data facilitated inexpensive extension and avoided the need for wall-to-wall airborne coverage. Chapter 7 is amongst just one other study (i.e.,  161  Wulder and Seeman 2003) to confirm the spatial object-based approaches for spatially extending the detailed yet limited coverage of advanced airborne transects (i.e., LiDAR data) to coarser satellite imagery. In addition, Chapter 7 is the first study to spatially extend airborne hyperspectral data. In answering the proposed questions, this thesis accomplished its objective of improving conventional terrestrial ecosystem mapping methods using a multi-scale combination of advanced remotely sensed data and field measurements. The unique outcomes of this thesis include the first spectral library of dominant tree species in the coastal Pacific northwest of Canada (Chapter 2), the first inventory in the SGI of LiDAR-metrics able to characterize and differentiate among structural classes (Chapter 3), vastly improved inventory data for rare and/or at-risk Garry oak habitat (Chapters 4, 5), markedly more detailed and accurate distribution information for dominant tree species derived using an innovative classification approach and newly developed LiDAR metrics as input (Chapters 4, 5), the first assessment in any environ of hyperspectral metrics for describing and differentiating faunal guilds based on aspects of diversity (Chapter 6) and, the first tree species heterogeneity predictions for the entire SGI, resulting from the first documented geospatial object-based classification incorporating airborne hyperspectral data and space-borne multispectral data (Chapter 7). The primary strength of these results is that they provide a state-of-the-art, step-by-step protocol for forest managers and ecologists to undertake detailed and accurate species and structural mapping of protected areas, while decreasing associated labor, time and subjectivity, and increasing the repeatability, at a cost comparable, if not less, than conventional aerial photography. In contrast, the challenges of using this approach 162  include the subjectivity and potential for labeling error associated with TEM plots’ structural attribute (Chapter 3), potential sources of SVM classification error associated with variable crown dimension (Chapters 4, 5), spectral and/or structural inseparability of certain species at the canopy-level (Chapters 4, 5), and/or geometric errors associated with airborne or GPS data (Chapters 4, 5). In addition, the success of predicting structural stage and/or avifaunal richness using additional, independent data sets is still required. Future research can help alleviate these challenges. For instance, forest canopy classes could be defined based on their species composition and their structural stage, potentially extending the utility of LiDAR derived metrics and reducing spectral confusion, leading to increased mapping accuracies. Increasing the amount of species-specific reference data collected, and/or optimizing the delineation of tree/tree-cluster in accordance with optimal GPS signal strength could also reduce species confusion, increasing producer’s and user’s accuracies. In addition, future work could exploit the relationships observed between certain LiDAR metrics and structural stages (Chapter 3) to predict the distribution of variation in forest structure throughout the extent of airborne surveys. Similarly, the relationships established between certain bio-indicators and bird species guilds could be used to predict avifaunal richness. Future work can also work towards strengthening forest structural and/or avifaunal richness models, through incorporating as-of-yet untested variables as input. The utility of spatially extending forest structural and avifaunal richness predictions based on image segmentation and object-based classification also remains worthy of investigation. If successful, such work would provide tree species, forest structure, and bird species richness information for the entire SGI, enabling regional-scale conservation value and biodiversity assessments.  163  Periodic reacquisition of airborne hyperspectral and LiDAR data and space-borne Landsat data would facilitate the updating these ecological attributes. With new imagery acquired, established methods provide an easily replicated approach. Field plots can be revisited to confirm and/or amend species specific reference data, ensuring reference data are as accurate and up-to-date as is possible. Spectral regions optimal for species differentiation could provide a recommended protocol for reducing the number of bands required to be acquired in hyperspectral data sets, and LiDAR metrics can be derived again, facilitating data fusion, resulting in up-to-date species maps from which changes in distribution (and their causes) can be assessed. The relationships established in Chapter 7 form the basis for spatially extending species, structural and/or avian diversity predictions using newly acquired, free (since 2008 (Woodcock 2008)) Landsat imagery. Given documented relationships between structural and functional aspects of forests and remotely sensed data, we suggest that the innovative methods described in this thesis are not limited to the SGI or the forests of the Pacific Northwest, and can be replicated, where targeted species/structural characteristics can be defined and differentiated based on hyperspectral-derived and/or LiDAR-derived metrics. Testing these methods in similar and disparate systems will reconfirm strengths, and further serve to identify areas in need of improvement. A robust application to a variety of environs will permit finetuning these methods to a degree that establishes them as operational tools employable for long-term mapping and monitoring.  164  Bibliography  Abramov, O. and McEwen, A. (2004). An evaluation of interpolation methods for Mars Orbitor Laser Altimeter (MOLA) data. International Journal of Remote Sensing 25,669−676. Ackermann, F. (1999). Airborne laser scanning—Present status and future expectations. ISPRS Journal of Photogrammetry and Remote Sensing 54, 64−67. Adams, W.W., Demmig-Adams, B., Logan, B.A., Barker, D.H. and Osmond, C.B. (1999). Rapid Changes in Xanthophyll Cycle Dependent Energy Dissipation and Photosystem II Efficiency in two Vines, Stephania japonica and Smilax australis, Growing in the Understory of an open Eucalyptus forest. Plant, Cell and Environment, 125-136. Abramov, O., and McEwen, A. (2004). An evaluation of interpolation methods for Mars Orbitor Laser Altimeter (MOLA) data. International Journal of Remote Sensing 25, 669–676. Agresti, A. (1996). An introduction to categorical data analysis. New York, N.Y.: Wiley. Akaike, H. (1973). Information theory and an extension of the maximum likelihood principle. In PETROV, B.N. and CASAKI, F. (Eds.). Second international symposium on information theory. Academiai Kiado, Budapest, Hungary, pp. 267-281. Anderson, G. L., Everitt, J. H., Richardson, A. J. and Escobar, D. E. (1993). Using satellite data to map false broomweed (Ericameria austrotexana) infestations on south Texas rangelands. Weed Technology 7, 865−871. Anderson, J. E., Plourde, L. C., Martin, M. E., Braswell, B. H., Smith, M., Dubayah, R. O., Hofton, M.A. and Blair, J.B. (2008). Integrating waveform lidar with hyperspectral imagery for inventory of a northern temperate forest. Remote Sensing of Environment 112, 1856−1870. Andrew, M.E. and Ustin, S.L. (2009). Habitat suitability modelling of an invasive plant with advanced remote sensing data. Diversity and Distributions 15, 627-640. Asner, G.P., Knapp, D.E., Kennedy-Bowdoin, T., Jones, M.O., Martin, R.E., Boardman, J. and Hughes, F.R. (2008). Invasive species detection in Hawaiian rainforests using airborne imaging spectroscopy and LiDAR. Remote Sensing of Environment 112, 1942-1955. Axelsson, P. (1999). Processing of laser scanner data—Algorithms and applications. ISPRS Journal of Photogrammetry and Remote Sensing 54, 138−147. AXYS Environmental Consulting Ltd. (2004). Gulf Islands National Park Reserve: Conservation assessment review. Gulf Islands National Park Reserve of Canada, Parks Canada, Sidney, BC. Bailey, S.A., Horner-Devine, M.C., Luck, G., Moore, L.A., Carney, K.M., Anderson, S., Betrus, C. and Fleishman, E. (2004). Primary productivity and species richness: relationships among functional guilds, residency groups and vagility classes at multiple spatial scales. Ecography, 27, 207–217. Bater, C.W. and Coops, N.C. (2009). Evaluating error associated with lidar-derived DEM interpolation. Computers and Geosciences 35, 289-300. Bergen, K.M, Goetz, S.J., Dubayah, R.O., Henebry, G.M., Hunsaker, C.T., Imhoff, M.L., Nelson, R.F., Parker, G.G. and Radeloff, V.C. (2009). Remote sensing of vegetation 3-D structure for  165  biodiversity and habitat: Review and implications for LiDAR and radar spaceborne missions. Journal of Geophysical Research 114, 1-13. Benediktsson, J. A., Swain, P. H., and Ersoy, O. K. (1990). Neural network approaches versus statisticalmethods in classification of multisource remote-sensing data. IEEE Transactions on Geoscience and Remote Sensing 28, 540−552. Benz, U.C., Hofmann,P., Willhauck, G., Lingenfelder, I. and Heynen, M. (2004). Multi-resolution, objectoriented fuzzy analysis of remote sensing data for GIS-ready information. ISPRS Journal of Photogrammetry Remote Sensing 58, 239–258. Blackburn, G. A. (1998). Spectral indices for estimating photosynthetic pigment concentrations: a test using senescent tree leaves, International Journal of Remote Sensing 19, 657-675. Blackburn, G.A. (2002). Remote sensing of forest pigments using airborne imaging spectrometer and LiDAR imagery. Remote Sensing of Environment 82, 311-321. Blaschke, T. (2003). Object-based contextual image classification built on Image Segmentation. In Proceedings of IEEE workshop on advances in techniques for analysis of remotely sensed data. Washington, DC. Blaschke, T. and J. Strobl. (2001). What’s wrong with pixels? Some recent developments interfacing remote sensing and GIS. GIS, 12-17. Biging, G.S. and Dobbertin, M. (1992). A comparison of distance dependent competition measures for height and basal area growth of individual conifer trees, Forest Science 38, 695-720. Boelman, N.T., Asner, G.P., Hart, P.J., and Martin, R.E. (2007). Multi-trophic invasion resistance in Hawaii: Bioacoustics, field surveys and airborne remote sensing. Ecological Applications 17, 2137-2144. Boschetti, M., Boschetti, L., Oliveri, S., Casati, L. and Canova, I. (2007). Tree species mapping with airborne hyper-spectral MIVIS data: the Ticino Park study case. International Journal of Remote Sensing 28, 1251−1261. Bradbury, R.B., Hill, R.A., Mason, D.C., Hinsley, S.A., Wilson, J.D., Balzter, H., Anderson, G.Q.A., Whittingham, M.J., Davenport, I.J. and Bellamy, P.E. (2005). Modeling relationships between birds and vegetation structure using airborne LiDAR data: a review with case studies from agricultural and woodland environments. Ibis 147, 443-452. Bradley, J. V. (1968). Distribution-free statistical tests.Englewood Cliffs, New Jersey: Prentice-Hall 388p. Breiman, L., Friedman, J., Olshen, R. and Stone, C. (1984). Classification and Regression Trees. Wadsworth, Pacific Grove, CA. pp. 358. Breiman, L., Friedman, J.H., Olshen, R.A. and Stone, C.J. (1998). Classification and regression trees. Chapman and Hall/CRC, Boca Raton, USA. British Columbia Ministry of Forests and British Columbia Ministry Of Environment. (1998). Field manual for describing terrestrial ecosystems. Victoria, BC. British Columbia Ecological Data Committee (BCEDC). (2000). Standard for Terrestrial Ecosystem Mapping (TEM) – digital data capture in British Columbia: Ecosystem Technical Standards and Database Manual, version 3.0. Terrestrial Ecosystem Task Force, Ecosystems Working Group. Victoria, British Columbia. 166  British Columbian Ministry Of Forests And Range. (2009). Biogeoclimactic Zones of British Columbia. [www.for.gov.bc.ca]. B.C. Conservation Data Centre. (2010). In BC Species and Ecosystems Explorer. B.C. Ministry of Environment Victoria, B.C. [http://a100.gov.bc.ca/pub/eswp/]. Bruzzone, L. and Carlin, L. (2006). A multilevel context-based system for classification of very high spatial resolution images. IEEE Transactions on Geoscience and Remote Sensing 44, 2587-2600. Buchanan, J.B., Irwin, L.L. and Mccutchen, E.L. )1995). Within-stand nest-site selection by spotted owls in the eastern Washington cascades. Journal of Wildlife Management 59, 301– 310. Carlson, K. M., Asner, G. P., Hughes, R. F., Ostertag, R. and Martin, R. E. (2007). Hyperspectral remote sensing of canopy biodiversity in Hawaiian lowland rainforests. Ecosystems 10, 526−549. Castilla, G. and Hay, G.J. (2008). Image objects and geographic objects. In Blaschke, T., Lang, S., and Hay, G. (Eds). Object-based image analysis: Spatial concepts for knowledge-driven remote sensing applications series. Springer-Verlag Berlin Chavez, P.S. (1996). Image-based atmospheric corrections: revisited and improved. Photogrammetric Engineering and Remote Sensing 62, 1025-1036. Chubey, M.S., Franklin, S.E. and Wulder, M.A. (2006). Object-based analysis of Ikonos-2 imagery for extraction of forest inventory parameters. Remote Sensing of Environment 72, 383-394. Clark, M., Roberts, D.A. and Clark, D.B. (2005). Hyperspectral discrimination of tropical rain forest tree species at leaf to crown scales. Remote Sensing of Environment 96, 375-398. Clawges, R., Vierling, K., Vierling, L. and Rowell, E. (2008). The use of airborne LiDAR to assess avian species diversity, density, and occurrence in a pine/aspen forest. Remote Sensing of Environment 112, 2064-2073. Cochrane, M.A., (2000). Using vegetation reflectance variability for species level classification of hyperspectral data. International Journal of Remote Sensing 21, 2075-2087. Cohen, W.B. and Goward, S.N. (2004). Landsat’s role in ecological applications of remote sensing. Bioscience 54, 535-545. Congalton, R. G. and Green, K. (1999). Assessing the accuracy of remotely sensed data: principles and practices. Lewis Publishers, Boca Raton, Florida. Coops, N.C. and Catling, P.C. (1997). Predicting the complexity of habitat in forests from airborne videography for wildlife management. International Journal of Remote Sensing 18, 2677-2682. Coops, N.C., Hilker, T., Wulder, M.A., St-Onge, B., Newnham, G., Siggins, A. and Trofymow, J.A. (2007). Estimating canopy structure of Douglas-fir forest stands from discrete-return LiDAR, Trees: Structure and Function 21, 295-310. Curran, P.J. (1989). Remote sensing of foliar chemistry. Remote Sensing of Environment 30, 271–278. Dalponte, M., Bruzzone, L. and Gianelle, D. (2008). Fusion of hyperspectral and LiDAR remote sensing data for classification of complex forest areas. IEEE Transactions on Geoscience and Remote Sensing 46, 1416−1427.  167  Dalponte, M., Bruzzone, L., Vescovo, L. and Gianelle, D. (2009). The role of spectral resolution and classifier complexity in the analysis of hyperspectral images of forest areas. Remote Sensing of Environment 113, 2345-2355. Datt, B. (1998). Remote sensing of chlorophyll a, chlorophyll b, chlorophyll aþb, and total cartenoid content in Eucalyptus leaves. Remote Sensing of Environment 66, 111–121. Definiens. (2010). Definiens Developer 7 User Guide. Munchen, Germany. Dennison, P. E. and Roberts, D. A. (2003). The effects of vegetation phenology on endmember selection and species mapping in southern California chaparral. Remote Sensing of Environment 87, 295−309. Drake, J. B., Dubayah, R. O., Knox, R. G., Clark, D. B. and Blair, J. B. (2002). Sensitivity of largefootprint LIDAR to canopy structure and biomass in a neotropical rainforest. Remote Sensing of Environment 81, 378−392. Dunlavy, J.C. (1935). Studies on the phyto-vertical distribution of birds. Auk 52, 425−431. Elvidge, C.D. (1990). Visible and near infrared reflectance characteristics of dry plant materials. International Journal of Remote Sensing 11, 1775-1795. Evans, D. L., Roberts, S. D. and Parker, R. C. (2006). LiDAR — A new tool for forest measurements? The Forestry Chronicle 82, 211−218. Everitt, J.H., Escobar D.E., Alaniz, M.A., Davis, M.R. and Richerson, J.V. (1996). Using spatial information technologies to map Chinese tamarisk (Tamarix chinensis) infestations. Weed Science 44, 194–201. Felde, G. W., Anderson, G.P., Cooley, T.W., Matthew, M.W., Adler- Golden, M.A., Berk, A. and Lee., J (2003). Analysis of Hyperion data with the FLAASH atmospheric correction algorithm. Proceedings of the Geoscience and Remote Sensing Symposium 1:90–92. Flaspohler, D. J., Giardina, C.P., Asner, G.P., Hart, P., Price, J., Lyons, C.K. and Castaneda, X. (2010). Long-term effects of fragmentation and fragment properties on bird species richness in Hawaiian forests. Biological Conservation 143, 280-288. Foley, S., Rivard, B., Sanchez-Azofeifa, G.A. and Calvo, J. (2006). Foliar spectral properties following leaf clipping and implications for handling techniques. Remote Sensing of Environment 103, 265–275. Foody, G. M. (2002). Status of land cover classification accuracy assessment. Remote Sensing of Environment 80, 185–201. Foody, G. M. (2004). Thematic map comparison: Evaluating the statistical significance of differences in classification accuracy. Photogrammetric Engineering and Remote Sensing 70, 627−633. Foody, G. M. and Mathur, A. (2004). A relative evaluation of multiclass image classification by support vector machines. IEEE Transactions on Geoscience and Remote Sensing 42, 1335−1343. Franklin, J., Phinn, S.R., Woodcock, C.E. and Rogan, J. (2003). Rationale and conceptual framework for classification approaches to assess forest resources and properties. Kluwer Academic Publishers. Freemark, K.E. and Merriam, H.G. (1986). Importance of area and habitat heterogeneity to bird assemblages in temperate forest fragments, Biological Conservation 36, 115-141.  168  Fuchs, M.A. (2001). Towards a Recovery Strategy for Garry oak and Associated Ecosystems in Canada: Ecological Assessment and Literature Review. Technical Report GBEI/EC-00-030. Environment Canada, Canadian Wildlife Service, Pacific and Yukon Region, Delta, British Columbia. Gao, B.C. (1996). NDWI-A normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sensing of Environment 58, 257-266. Gao, X., Huete, A.R., Ni, W. and Miura, T. (2000). Optical-biophysical relationships of vegetation spectra without background contamination. Remote Sensing of Environment 74, 609–620. Garry oak Ecosystems Recovery Team. (2003). Species at risk in Garry oak and associated ecosystems in British Columbia. Summary report. Garry oak ecosystems recovery team, Victoria, British Columbia. Gedalof, G., Pellatt, M. and Smith, D.J. (2006). From Prairie to Forest: Three Centuries of Environmental Change at Rocky Point, Vancouver Island, British Columbia. Northwest Science 80, 34-46. Geerling, G.W., Labrador-Garcia, M., Clevers, J.G.P.W., Ragas, A.M.J. and Smits, A.J.M. (2007). Classification of floodplain vegetation by data fusion of spectral (CASI) and LiDAR data. International Journal of Remote Sensing 28, 4263-4284. Geerling, G.W., Vreeken-Buus, M.J., Jesse, P., Ragas, A.M.J. and Smits, A.J.M. (2009). Mapping river floodplain ecotopes by segmentation of spectral (CASI_ and structural (LiDAR) remote sensing data. River Research and Applications 25, 795-813. Gillespie, T. W., Brock, J. and Wright, C. W. (2004). Prospects for quantifying structure, floristic composition and species richness of tropical forests. International Journal of Remote Sensing 25, 707−715. Gillis, M.D., Omule, A.Y. and Brierley, T. (2005). Monitoring Canada’s Forests: The National Forest Inventory. The Forestry Chronicle 81, 214-221. Gitelson, A.A., Merzlyak, M.N. and Chivkunova, O.B. (2001). Optical properties and non-destructive estimation of anthocyanin content in plant leaves. Photochemistry and Photobiology 74, 38-45. Gitelson, A.A., Zur, Y., Chivkunova, O.B. and Merzlyak, M.N. (2002). Assessing carotenoid content in plant leaves with reflectance spectroscopy. Photochemistry and Photobiology 75, 272-281. Goetz, A. F. H. (1992). Principles of narrow band spectrometry in the visible and IR: instruments and data analysis. In F. Toselli and J. Bodechtel (Eds.). Imaging spectroscopy: fundamentals and prospective applications. Kluwer Academic Publishers, Dordrecht, Netherlands. pp. 21-32. Goetz S., Steinberg, D., Dubayah, R. and Blair, B. (2007). Laser remote sensing of canopy habitat heterogeneity as a predictor of bird species richness in an eastern temperate forest. USA. Remote Sensing of Environment 108, 254-263. Golumbia, T. (2006). A history of species introductions in Gwaii Haanas and Gulf Islands National Park Reserves in British Columbia, Canada: Implication for park management. Transaction of the Western Section of the Wildlife Society, 42, 20-34. Goodwin, N., Turner, R. and Merton, R. (2005). Classifying Eucalyptus forests with high spatial and spectral resolution imagery: an investigation of individual species and vegetation communities, Australian Journal of Botany 53, 337-345.  169  Gong, P., Pu, R. and Zu, B. (1997). Conifer species recognition: an exploratory analysis of in situ hyperspectral data. Remote Sensing of Environment 62, 189−200. Goodenough, D.G., Niemann, K.O., Dyk, A., Hobart, G., Gordon, P. Loisel, M. and Chen, H. (2008). Comparison of AVIRIS and AISA Airborne Hyperspectral Sensing for Above-ground Forest Carbon Mapping. In Proceedings IEEE International Geoscience and Remote Sensing Symposium 2008, Boston, MA. pp. ii-129 - ii-132. Goodenough, D.G., Niemann, K.O., Quinn, G.S., Gordon, P., Gross, A.. Tian, H., Hobart, G., Hao, C., and Dyk, A. (2009). Comparison of AVIRIS and AISA for chemistry mapping. Geoscience and Remote Sensing Symposium 1, I-80-I-83. Goodwin, N., Turner, R. and Merton, R. (2005). Classifying Eucalyptus forests with high spatial and spectral resolution imagery: an investigation of individual species and vegetation communities. Australian Journal of Botany 53, 337-345. Gottschalk, T.K., Huettmann, F. and Ehlers, M. (2005). Thirty years of analyzing and modeling avian habitat relationships using satellite imagery data: a review. International Journal of Remote Sensing 12, 2631-2656. Green, R.N. (2007). Terrestrial ecosystem mapping of the southern Gulf Islands. Final report. B.A. Blackwell and Associates Ltd., North Vancouver, British Columbia. Grimm, N.B. (1995). Why link species and ecosystems? A perspective from ecosystem ecology. In Jones, C.G. and Lawton, J.H. (Eds.). Linking species and ecosystems. Chapman and Hall, New York, USA, pp. 5-15. Hall, R.K., Watkins, R.L., Heggem, D.T., Jones, K.B., Kaufmann, P.R., Moore, S.B. and Gregory, S.J. (2009). Quantifying structural physical habitat attributes using LiDAR and hyperspectral imagery. Environmental Monitoring and Assessment 159, 63-83. Hamilton, E. (1988). A system for the classification of seral ecosystems within the Biogeoclimatic Ecosystem Classification system, first approximation. BC Ministry of Forests and Lands. Victoria, BC. Research report RR87004-HQ. Hansen, A.J., Mccomb, W.C., Vega, R., Raphael, M.G. and Hunter, M. (1995). Bird habitat relationships in natural and managed forests in the West Cascades of Oregon. Ecological Applications 5, 555– 569.Hill, R. A. and Thomson, A. G. (2005). Mapping woodland species composition and structure using airborne spectral and LIDAR data. International Journal of Remote Sensing 26, 3763−3779. Hay, G.J., and Castilla, G. (2008). Geographic object-based image analysis (GEOBIA): a new name for a new discipline. In seriesBlaschke, T., Lang, S., and Hay, G. (Eds). Object-based image analysis: Spatial concepts for knowledge-driven remote sensing applications. Springer-Verlag Berlin. Hill, R.A. and Smith, G.N. (2005). Land cover heterogeneity in Great Britain as identified in Land Cover Map 2000. International Journal of Remote Sensing 26, 5467-5473. Hinsley, S.A., Hill, R.A., Bellamy, P.E., Harrison, N.M., Speakman, J.R., Wilson, A.K. and Ferns, P.N. (2008). Effects of structural and functional habitat gaps on breeding woodland birds: working harder for less. Landscape Ecology 23, 615-626. Hobbs, R. and Mooney, H.A. (1990). Remote Sensing of Biosphere Functioning: Concluding remarks. In: Mooney, H. A. and Hobbs, R.J., Eds., Ecological Studies 79: Remote Sensing of Biosphere Functioning.Springer-Verlag, New York, USA, pp. 305-306.  170  Hobbs, R. J. (2002). The ecological context: a landscape perspective. In M. R. Perrow and A. J. Davy (Eds.). Handbook of ecological restoration, vol. 1: principles of restoration. Cambridge University Press, United Kingdom. pp. 22-45. Hobbs, R.J. (2007). Setting Effective and Realistic Restoration Goals: Key Directions for Research. Restoration Ecology 15, 354-357. Hobbs, R. J. and Norton, D.A. (1996). Towards a conceptual framework for restoration ecology. Restoration Ecology 4, 93–110. Hobbs, R.J., and Harris, J.A. (2001). Restoration Ecology: Repairing the Earth’s Ecosystems in the New Millennium. Restoration Ecology 9, 239-246. Holmes, R.T. and Robinson, S.K. (1981). Tree species preferences of foraging insectivorous birds in a northern hardwoods forest. Oecologia 48, 31-35. Hsu, C., Chang, C. and Lin, C. (2010). A practical guide to support vector classification. Final Report. National Taiwan University, Taipei City, Taiwan: Department of Computer Science. Huang, C., Davis, L. S., and Townshend, J. R. (2002). An assessment of support vector machines for land cover classification. International Journal of Remote Sensing 23, 725−749. Huang, C. and Asner, G. P. (2009). Applications of remote sensing to alien invasive plant studies. Sensors 9, 4869−4889. Hudak, A.T., Lefsky, M.A., Cohen, W.B. and Beterretche, M. (2002). Integration of LiDAR and Landsat ETM+ data for estimating and mapping forest canopy height. Remote Sensing of Environment 82, 397-416. Hudak, A.T., Crookston, N.L., Evans, J.S., Falkowski, M.J., Smith, A.M.S., Gessler, P.E. and Morgan, P. (2006). Regression modeling and mapping of coniferous forest basal area and tree density from discrete-return LiDAR and multispectral satellite data. Canadian Journal of Remote Sensing 32, 126-138. Hurlbert, H. and Haskell, J.P. (2003). The effect of energy and seasonality on avian species richness and community composition. American Naturalist 161, 83–97. Huston, M.A. (1994). Biological Diversity: The Coexistence of Species on Changing Landscapes,.Cambridge University Press, Cambridge, UK, pp. 701. Huston, M.A .(1999). Local processes and regional patterns: appropriate scales for understanding variation in the diversity of plants and animals. Oikos 86, 393–401. Innes, J. L.. and Koch, B. (1998). Forest biodiversity and its assessment by remote sensing. Global Ecology and Biogeography Letters 7, 397−419. Jensen, J.R. (2007). Remote sensing of vegetation. In , J.R. Jenson (Ed.). Remote Sensing of the Environment. An Earth Resource Perspective. Upper Saddle River, NJ: Prentice Hall. Jewell, K., Arcese, P and Gergel, S.A. (2007). Robust predictions of species distribution: spatial habitat models for a brood parasite. Biological Conservation 140, 259-272. Jewell, K. and Arcese, P. (2008). Consequences of parasite invasion and land use on the spatial dynamics of host populations. Journal of Applied Ecology 45, 1180-1188.  171  Jones, T. G., Coops, N. C., and Sharma, T. (2010a). Employing ground-based spectroscopy for tree species differentiation in the Gulf Islands National Park Reserve. International Journal of Remote Sensing 31, 1121−1127. (Chapter 2) Jones, T.G., Coops, N.C., and Sharma, T. (in press). Assessing the utility of LiDAR to differentiate among vegetation structural classes. Remote Sensing Letters. (Chapter 3) Jones, T.G., Coops, N.C. & Sharma, T. (2011). Exploring the utility of hyperspectral imagery and LiDAR data for predicting Quercus garryana ecosystem distribution and aiding in habitat restoration. Restoration Ecology 19, 245-256. (Chapter 4) Jones, T.G., Coops, N.C. and Sharma, T. (2010b). Assessing the utility of airborne hyperspectral and LiDAR data for species distribution mapping in the coastal Pacific Northwest, Canada. Remote sensing of Environment 114, 2841-2852. Chapter 5 Jones, T.G., Arcese, P., Sharma, T., and Coops, N.C. (in revision). Describing avifaunal richness with functional and structural bio-indicators derived from advanced airborne remotely sensed data. Systematics and Biodiversity. (Chapter 6) Jones, T.G., Coops, N.C., Gergel, S. and Sharma, T. (under review). Extending tree species heterogeneity using object-based classification. Ecography. (Chapter 7) Jørgensen, A.F. and Nøhr, H. (1996). The use of satellite images for mapping of landscape and biological diversity in the Sahel. International Journal of Remote Sensing 17, 91–109. Karr, J.R. and Roth, R.R. (1971). Vegetation structure and avian diversity in several New World areas. American Naturalist 105, 423−435. Kerr J.T. and Ostrovsky, M. (2003). From space to species: ecological applications for remote sensing. TRENDS in Ecology and Evolution 18, 299-305.Kimmins, J.P. (1997). Biodiversity and its relationship to ecosystem health and integrity, Forest Chronicles 73, 229-232. Koetz, B., Sun, G., Morsdorf, F., Ranson, K.J., Kneubühler, M., Itten, K. and Allgöwer, B. (2007). Fusion of imaging spectrometer and LiDAR data over combined radiative transfer models for forest canopy characterization. Remote Sensing of Environment 106, 449–459. Koetz, B., Morsdorf, F., van der Linder, S., Curt, T. and Allgöwer, B. (2008). Multi-source land cover classification for forest fire management based on imaging spectrometry and LiDAR data. Forest Ecology and Management 256, 263-271. Kraus, K. and Pfeifer, N. (1999). Determination of terrain models in wooded areas with airborne laser scanner data. ISPRS Journal of Photogrammetry and Remote Sensing 53, 193−203. Kumar, L., Schmidt, K., Dury, S., and Skidmore, A. (2006). Imaging spectrometry in vegetation science. In: van der Meer, F.D., and de Jong, S.M., (Eds.) Imaging spectroscopy: Basic principles and prospective applications. Dordrecht, Springer Publishers, pp. 111-155. Lack, D. (1969). The numbers of bird species on islands. Bird Study 16, 193–209. Leckie, D. and M. Gillis. (1995). Forest inventory in Canada with an emphasis on map production. The Forestry Chronicle 71, 74-88. Leckie, D. G., Cloney, E., Jay, C. and Paradine, D. (2005). Automated mapping of stream features with high resolution multispectral imagery: an example of the capabilities. Photogrammetric Engineering and Remote Sensing, 71, 145−155. 172  Lefsky, M.A., Cohen, W.B., Hudak, A.T., Acker, S.A. and Ohmann, J.L. (1999a). Integration of LiDAR, Landsat ETM+ and forest inventory data for regions forest mapping, International Archives of Photogrammetry and Remote Sensing 32, 119-126. Lefsky, M.A., Cohen, W.B., Acker, S.A., Parker, G., Spies, T.A. and Harding, D. (1999b). LiDAR remote sensing of the canopy structure and biophysical properties of Douglas Fir-Western Hemlock Forests. Remote Sensing of Environment 70, 339−361. Lefsky, M.A., Cohen, W.B., Parker, G.G. and Harding, D.J. (2002). LiDAR remote sensing for ecosystem studies. Bioscience 52, 19-30. Leyequien, E., Verrelst, J., Slot, M., Schaepman-Strub, G., Heitköng, I.M.A. and Skidmore, A. (2007). Capturing the fugitive: Applying remote sensing to terrestrial animal distribution and diversity. International Journal of Applied Earth Observation and Geoinformation 9, 1-20. Li, H., and Reynolds, J.F. (1995). On definition and quantification of heterogeneity. Oikos 73, 280–284. Lindenmayer, D.B., Franklin, J.F. and Fischer, J. (2006).General management principles and a checklist of strategies to guide forest biodiversity conservation. Biological Conservation 131, 433–445. Lim, K., Treitz, P., Wulder, M., Benoit, S. and Flood, M. (2003). LiDAR remote sensing of forest structure. Progress in Physical Geography 27, 88-106. Lucas, R. M., Lee, A. C. and Bunting, P. J. (2008). Retrieving forest biomass through integration of CASI and LiDAR data. International Journal of Remote Sensing 29,1553−1577. Macarthur, R. H. (1972). Geographical ecology: patterns in the distribution of species. Princeton University Press, New Jersey, USA. Martin, M.E., Newman, S.D., Aber, J.D. and Congalton, R.G. (1998). Determining forest species composition using high spectral resolution remote sensing data. Remote Sensing of Environment 65, 249-254. Martin, T.G., Arcese, P and Scheerder, N. (2011). Browsing down our natural heritage: Deer impacts on vegetation structure and songbird communities across an island archipelago. Biological Conservation 144, 459-469. Mccullagh, P. and Nelder, J.A. (1989). Generalized Linear Models. Chapman and Hall, London, UK. Means, J. (1999). Design, capabilities and uses of large-footprint and small-footprint LiDAR systems. International Archives of the Photogrammetry and Remote Sensing 32, 201-207. Melillo, J.M., Aber, J.D. and Muratore, F.F. (1982). Nitrogen and lignin control of hardwood leaf litter decomposition dynamics. Ecology 63, 621-626. Meidinger, D. and J. Pojar. 1991. Ecosystems of British Columbia. BC Special Report Series No. 6. British Columbia Ministry of Forests, Victoria, British Columbia. Meinke, C. W., Knick, S.T. and Pyke, D.A. (2008). A spatial model to prioritize Sagebrush landscapes in the intermountain west (U.S.A.) for restoration. Restoration Ecology 17, 652–659. Melgani, F. and Bruzzone, L. (2004). Classification of hyperspectral remote sensing images with support vector machines. IEEE Transactions on Geoscience and Remote Sensing 42, 1778–1790.  173  Mills, G.S., Dunning Jr., J.B. and Bates, J.M. (1991). The relationship between breeding bird density and vegetation volume. Wilson Bulletin 103, 468−479. Ministry of Forests and Range. (2006). Biogeoclimactic zones of British Columbia 2006. [http://www.for.gov.bc.ca]. Mooney, H.A. and Hobbs, R.J. (1990). In Mooney, H.A. and Hobbs, R.J. (Eds.). Ecological Studies 79: Remote Sensing of Biosphere Functioning. Springer-Verlag, New York, USA, pp. 1-15. Morgan, K. and Freedman, B. (1986). Breeding bird communities in a hardwood forest succession in Nova Scotia. Canadian Field Naturalist 100, 506–519. Morgan, J.L. and Gergel, S.E. (2010). Quantifying historic landscape heterogeneity from aerial photographs using object-based analysis. Landscape Ecology 25, 985-998. Morisette, J.T., Jarnevich, C.S., Ullah, A., Cai, W., Pedelty, J.A., Gentle, J.E., Stohlgren, T.J. and Schnase, J.L. (2006). A taskmark habitat suitability map for the continental United States, Frontiers in Ecology and the Environment 4, 11-17. Morrison, M.L., Timosse, I.M. and With, K.A. (1987). Development and testing of linear regression models predicting bird-habitat relationships. Journal of Wildlife Management 51, 247-253. Morrison, R.I.G. (1997). The use of remote sensing to evaluate shorebird habitats and populations on Prince Charles Island, Foxe Basin, Canada. Arctic 50, 55–75. Müller, J., Moning, C., Bässler, C., Heurich, M. and Brandl, R. (2009). Using airborne laser scanning to model potential abundance and assemblages of forest passerines. Basic and Applied Ecology 10, 671-681. Müller, J., Stadler, J., and Brandl, R. (2010). Composition versus physiognomy of vegetation as predictors of bird assemblages: The role of LiDAR. Remote Sensing of Environment 114, 490-495. Mundt, J.T., Streutker, D.R. and Glenn, N.F. (2006). Mapping sagebrush distribution using fusion of hyperspectral and lidar classifications. Photogrammetric Engineering and Remote Sensing 72, 4754. Murthy, M.S.R., Giriraj, A. and Dutt, C.B.S. (2003). Geoinformatics for biodiversity assessment. Biology Letters 40, 75–100. Nagendra, H. (2001). Using remote sensing to assess biodiversity. International Journal of Remote Sensing 22, 249-254. North, M.P., Franklin, J.F., Carey, A.B., Forsman, E.D. and Hamer, T. (1999). Forest stand structure of the northern spotted owl’s foraging habitat. Forest Science 45, 520– 527. Noss, R.F. (1996). Ecosystems as conservation targets. Trends in Ecology and Evolution 11, 351. Okin, G.S., Roberts, D.A., Murray, B. and Okin, W.J. (2000). Practical limits on hyperspectral vegetation discrimination in arid and semiarid environments. Remote Sensing of Environment 77, 212-225. Oliver, C.D. and. Larson, B.C. (1990). Forest stand dynamics. McGraw-Hill, New York, NY. Olsson, H. and Næsset, E. (2004). Special issue–ScandLaser workshop in Umeå, Sweden, September 2–4, 2003—Preface. Scandinavian Journal of Forest Resources 19, 481−582.  174  Palacios-Orueta, A.P. and Ustin, S.L. (1996). Multivariate statistical classification of soil spectra. Remote Sensing of Environment 57, 108-118. Pal, M. and Mather, P. M. (2006). Some issues in the classification of DAIS hyperspectral data. International Journal of Remote Sensing 27, 2895−2916. Palmeirim, J.M. (1988). Automatic mapping of avian species habitat using satellite imagery. Oikos 52, 59– 68.Parks Canada Agency. (1997). State of the Parks Report. Internal Document. Ottawa, Ontario. Parker, G.G., Harmon, M.E., Lefsky, M.A., Chen, J.Q., Van Pelt, R., Weis, S.B., Thomas, S.C., Winner, W.E., Shaw, D.C. and Franklin, J.F. (2004). Three-dimensional structure of an old-growth Pseudotsuga-Tsuga canopy and its implications for radiation balance microclimate and gas exchange. Ecosystems 7, 440-453. Parks Canada Agency. (1997). State of the Parks Report. Internal Document. Ottawa, Ontario. Peck, K.M. (1989). Tree species preferences shown by foraging birds in forest plantations in northern England. Biological Conservation 48, 41-57. Peterson, E.B. (2005). Estimating cover of an invasive grass (Bromus tectorum) using tobit regression and phenology derived from two dates of Landsat ETM+ data. International Journal of Remote Sensing 26, 2491–507. Pfadenhauer, J. and Grootjans, A. (1999). Wetland restoration in Central Europe: aims and methods. Applied Vegetation Science 2:95–106. Sambridge, M., J. Braun, and H. McQueen. 1995. Geophysical parameterization and interpolation of irregular data using natural neighbours. Geophysical Journal International 122, 837–857. Pohl, C. and van Genderen, J. L. (1998). Multisensor image fusion in remote sensing: concepts, methods and applications. International Journal of Remote Sensing 19, 823−854. Plourde, L. C., Ollinger, S. V., Smith, M. and Martin, M. E. (2007). Estimating species abundance in a northern temperate forest using spectral mixture analysis. Photogrammetric Engineering and Remote Sensing 73, 829−840. Prins, E., Petersen, B.S., Aunins, A. and Priednieks, J. (2005). Using Landsat TM and field data to produce maps of predicted bird densities in Latvian farmland. International Journal of Remote Sensing 26, 1881-1891. Price, J.C. 1998. An approach for analysis of reflectance spectra. Remote Sensing of Environment 64, 316330. Resource Inventory Committee: Vegetation Inventory Working Group. (1995). The British Columbia Land Cover Classification: background and definitions, a working document. Victoria, BC. Reutebuch, S. E., Andersen, H. E. and McGaughey, R. J. (2005). Light detection and ranging (LIDAR): An emerging tool for multiple resource inventory. Journal of Forestry 103, 286−292. Rhoads, A.G., Hamburg, S.P, Fahey, T.J., Siccama, T.G. and Kobe, R. (2004). Comparing direct and indirect methods of assessing canopy structure in a northern hardwood forest. Canadian Journal of Forest Resources 34, 584-591. Richards, P.W. (1983). The three dimensional structure of tropical rain forest. In S.L. Sutton, T.C. Whitmore and A.C. Chadwick (Eds.). Tropical Rain Forest: Ecology and Management. Blackwell, Oxford. pp. 3-10. 175  Rock, B.N., Vogelmann, J.E., Williams, D.L., Vogelmann, A.F. and Hoshizaki, T. (1986). Remote sensing of forest damage. BioScience 36, 439-445. Rogan, J. and Chen, D. (2004). Remote sensing technology for mapping and monitoring land-cover and land-use change. Progress in Planning 61, 301-325. Roth, R.R. (1976). Spatial heterogeneity and bird species diversity. Ecology 57, 773-782. Sala, O.E., Chaplin Iii, F.S., Armesto, J.J., Berlow, E., Bloomfield, J., Dirzo, R., Huber-Sanwald, E., Huenneke, L.F., Jackson, R.B., Kinzig, A., Leemans, R., Lodge, D.M., Mooney, H.A., Oesterheld, M., Poff, N.L., Sykes, M.T., Walker, B.H. and Wall, D.H. (2000). Biodiversity-Global biodiversity scenarios for the year 2100. Science 287, 1770-1774. Scarth, P., Phinn, S.R., Held, A. and Mitchell, D. (1999). Mapping koala habitat and eucalypt trees: integration and scaling of field and airborne hyperspectral data. In Proceedings of the 10th Australasian remote sensing and Photogrammetry conference, 21-25 August 1999, Adelaide, Australia. CD-ROM. Seavy, N.E., Viers, J.H. and Wood, J.K. (2009). Riparian bird response to vegetation structure: a multiscale analysis using LiDAR measurements of canopy height. Ecological Applications 19, 1848-1857. Serrano, L., Penuelas, J. and Ustin, S.L. (2002). Remote sensing of nitrogen and lignin in Mediterranean vegetation from AVIRIS data: Decomposing biochemical from structural signals. Remote Sensing of Environment 81, 355-364. Seto, K.C., Fleishman, E., Fay, J.P. and Betrus, C.J. (2004). Linking spatial patterns of bird and butterfly species richness with Landsat TM derived NDVI. International Journal of Remote Sensing 25, 4309–4324. Shannon, C.E., (1948). A mathematical theory of communication. Bell System Technical Journal 27, 379– 423 and 623–656. Sibson, R. (1981). A brief description of natural neighbor interpolation. In V. Barnett (Ed.). Interpreting multivariate data. Wiley, Chichester, West Sussex. pp. 21-35. Siegel, R.B. and DESANTE, D.F. (2003). Bird communities in thinned versus unthinned Sierran mixed conifer stands. Wilson Bulletin 115, 155– 165. Simpson, E.H. (1949). Measurement of diversity. Nature 163, 688. Sims, D.A. and Gamon, J.A. (2002). Relationships between leaf pigment content and spectral reflectance across a wide range of species, leaf structures and developmental stages. Remote Sensing of Environment 81, 337-354. Sinclair, T.R., Hoffer, R.M. and Schreiber, M.M. (1971). Reflectance and internal structure of leaves from several crops during a growing season. Agronomy Journal 63, 864-868. Sklenicka, P. and Lhota, T. (2002). Landscape heterogeneity—a quantitative criterion for landscape reconstruction. Landscape Urban Plan 58,147–156. Smith, R.B. (2006). Introduction to hyperspectral imaging, TNTmips; MicroImages, Inc, 1-24. Smith, G.M. (2008). The development of integrated object-based analysis of EO data within UK national land cover products. In Blaschke, T., Lang, S., and Hay, G. (Eds). Object-based image analysis: Spatial concepts for knowledge-driven remote sensing applicationsseries. Springer-Verlag Berlin 176  Staudhammer, C.L. and LeMay, V.M. (2001). Introduction and evaluation of possible indices of stand structural diversity. Canadian Journal of Forest Research 31, 1105-1115. Steinberg, D. and Colla, P. (1995). CART: Tree-Structured Non-Parametric Data Analysis. Salford Systems, San Diego, CA, pp. 336. Stoms, D.M. and Estes, J.E. (1993). A remote sensing research agenda for mapping and monitoring biodiversity. International Journal of Remote Sensing 14 1839–1860. Technical Report GBEI/EC-00-030. Environment Canada. Canadian Wildlife Service, Pacific and Yukon Region, Delta, British Columbia. Tews, J., Brose, U., Grimm, V., Tielborger, K., Wichmann, M.C., Schwager, M. And Jeltsch, F. (2004). Animal species diversity driven by habitat heterogeneity/diversity: the importance of keystone structures. Journal of Biogeography 31, 79–92. Thompson, S.D. and Gergel, S.E. (2008). Conservation implications of mapping rare ecosystems using high spatial resolution imagery: recommendations for heterogeneous and fragmented landscapes. Landscape Ecology 23, 1023-1037. Tucker, C. J. (1977). Asymptotic nature of grass canopy spectral reflectance. Applied Optics 16, 1151– 1156. Turner, M.G. (1989). Landscape ecology: the effect of pattern on process. Annual Review of Ecology and Systematics 20, 171–197. Turner, W., Spector, S., Gardiner, N., Fladeland, M., Sterling, E. and Steininger, M. (2003). Remote sensing for biodiversity science and conservation. Trends in Ecology and Evolution 18, 306-314. Ustin, S.L., Roberts ,D.A., Gamon, J.A., Asner, G.P. and Green, R.O. (2004). Using Imaging Spectroscopy to Study Ecosystem Processes and Properties. BioScience 54, 523-534. van Aardt, J.A.N. and Wynne, R.H. (2001). Spectral seperability among six southern tree species. Photogrammetric Engineering & Remote Sensing 67, 1367-1375. van Aardt, J.A.N. and Wynne, R.H. (2007). Examining pine spectral seperability using hyperspectral data from airborne sensor: An extension of field-based results. International Journal of Remote Sensing 28, 431-436. van der Meer, F.D. (2006). Basic principles of spectrometry. In: van der Meer, F.D., and de Jong, S.M., (Eds.) Imaging spectroscopy: Basic principles and prospective applications. Dordrecht, Springer Publishers, pp. 3-16. Vapnik, V. N. (1998). Statistical learning theory. New York: Wiley. Vellend, M., Bjorkman, A.D. and Mcconchie, A. (2008). Environmentally biased fragmentation of oak savanna habitat on southeastern Vancouver Island, Canada. Biological Conservation 141, 25762584. Venables, W.N. and Ripley, B.D. (2002). Modern applied statistics with S., 4th edition. Springer-Verlag, New York, New York, USA. Verner, J. and Larson, T.A. (1989). Richness of breeding bird species in mixed conifer forests of the Sierra Nevada, California. Auk 106 447−463.  177  Vierling, K.T., Vierling, L.A., Gould, W.A., Martinuzzi, S. and Clawges, R.M. (2008). LiDAR: shedding new light on habitat characterization and modeling. Frontiers in Ecology and Environment 6, 9098. Villard, M. A., Trzcinski, M.K. and Merriam, G. (1999). Fragmentation effects on forest birds: Relative influence of woodland cover and configuration on landscape occupancy. Conservation Biology 13, 774 – 783. Vitousek, P. M., Walker, L. R., Whiteaker, L. D., Mueller-Dombois, D. and Matson, P. A. (1987). Biological invasion by Myrica faya alters ecosystem development in Hawaii. Science 238, 802−804. Voss, M. and Sugumaran, R. (2008). Seasonal effect on tree species classification in an urban environment using hyperspectral data, LiDAR, and an object-oriented approach. Sensors 8, 3020−3036. Wallace, L.W. and Gray, S.T. (2002). Feedbacks between organisms and ecosystem processes. In GERGEL, S.E. And TURNER, M.G. (Eds.). Learning Landscape Ecology: A practical guide to concepts and techniques.Springer Publishers, New York, USA, pp. 249-265. Wang, L., Sousa, W.P. and Gong, P. (2004). Integration of object-based and pixel-based classification for mapping mangroves with IKONOS imagery. International Journal of Remote Sensing 25, 56555668. Wehr, A. and Lohr, U. (1999). Airborne laser scanning-an introduction and overview. ISPRS Journal of Photogrammetry and Remote Sensing 54, 68–82Wu, T., Lin, C., and Weng, R. C. (2004). Probability estimates for multi-class classification by pairwise coupling. Journal of Machine Learning Research 5, 976−1005. Whisenant, S. G. (1999). Repairing damaged wildlands: a process-oriented, landscape-scale approach. Cambridge University Press. Williams, P.C. and Norris, K.H. (1987). Qualitative applications of near-infrared reflectance spectroscopy. In P.C. Williams and K.H. Norris (Eds.). Near-infrared Technology in the Agricultural and Food Industries. St Paul, MN, American Association. pp. 241–246. Willson, M.F. (1974). Avian community organization and habitat structure. Ecology 55, 1017−1029. Wilson, E.O. (1988). Biodiversity. National Academy Press, Washington, D.C., USA. Weetman, G.F., Ponozzo, E. and Marek, K. (1990). An assessment of opportunities for alternative silvicultural systems in the SBS, ICH, and ESSF biogeoclimatic zones of the Prince Rupert Forest Region. BC Ministry Of Forestry. Smithers, BC Contract Report. Woodcock, C.E. Allen, R., Anderson, M., Belward, A., Bindschadler, R., Cohen, W., Gao, F., Goward, S.N., Helder, D., Helmer,E., Nemani, R., Oreopoulos, L., Schott, J., Thenkabail, P.S., Vermote, E.F., Vogelmann, J., Wulder, M.A., and Wynne, R. (2008). Free Access to Landsat Imagery. Science 320, 1011. Woodley, S. (1993). Monitoring and measuring ecosystem integrity in Canadian National Parks. In Woodley, S.J., Francis, G. and Kay, J. (Eds.) Ecosystem integrity and the management of ecosystems. St. Lucie Press. pp. 1-14. White, P. S. and Walker, J.L. (1997). Approximating nature’s variation: selecting and using reference information in restoration ecology. Restoration Ecology 5, 338–349.  178  White, J., Wulder, M., and Grills, D. (2006). Detecting and mapping mountain pine beetle red attack damage with SPOT-5 10 m multispectral imagery. BC Journal of Ecosystems and Management 7, 105-118. White, J.C., Coops, N.C., Hilker, T., Wulder, M.A. and Carroll, A.L. (2007). Detecting mountain pine beetle red attack damage with EO-1 Hyperion moisture indices. International Journal of Remote Sensing 28, 2111-2121. Woodcock, C. and Harward, V.J. (1992). Nested-hierarchical scene models and image segmentation. Internatinal Journal of Remote Sensing 13, 3167–3187. Woodcock, C.E. Allen, R., Anderson, M., Belward, A., Bindschadler, R., Cohen, W., Gao, F., Goward, S.N., Helder, D., Helmer,E., Nemani, R., Oreopoulos, L., Schott, J., Thenkabail, P.S., Vermote, E.F., Vogelmann, J., Wulder, M.A. and Wynne, R. (2008). Free Access to Landsat Imagery. Science 320, 1011. Woodley, S. (1993). Monitoring and measuring ecosystem integrity in Canadian National Parks. In Woodley, S.J., Francis, G. and Kay, J. (Eds.) Ecosystem integrity and the management of ecosystems. St. Lucie Press. pp. 1-14. Wu, T., Lin, C. and Weng, R.C.. (2004). Probability estimates for multiclass classification by pairwise coupling. Journal of Machine Learning Research 5, 976–1005. Wulder, M.A. and Seemann, D. (2003). Forest inventory height update through the integration of LiDAR data with segmented Landsat imagery. Canadian Journal of Remote Sensing 29, 536-543. Wulder, M.A., Skakun, R., Kurz, W. and White, J.C. (2004). Estimating time since forest disturbance using segmented Landsat ETM+ imagery. Remote Sensing of Environment 93, 179-187. Wulder, M.A., Han, T., White, J.C., Sweda, T. and Tsuzuki, H.. (2007). Integrating profiling LiDAR with Landsat data for regional boreal forest canopy attribute estimation and change characterization. Remote Sensing of Environment 110, 123-137. Wulder, M.A., White, J.C., Hay, G.J., and Castilla, G. (2008a). Pixels to objects to information: Spatial context in forest characterization with remote sensing. In: Blaschke, T., Lang, S., and Hay, G., (Eds.) Object-based image analysis: Spatial concepts for knowledge-driven remote sensing application series. Berlin, Springer Publishers, pp. 345-364. Wulder, M.A., Bater, C.W., Coops, N.C., Hilker, T. and White, J.C. (2008b). The role of LiDAR in sustainable forest management. The Forestry Chronicle 84, 1-19. Xiao, Q., Ustin, S.L., and McPherson, E.G. (2004). Using AVIRIS data and multiple-masking techniques to map urban forest tree species. International Journal of Remote Sensing 25, 5637-5654. Yang, C., Everitt, J. H., Fletcher, R. S., Jensen, R. R., and Mausel, P. W. (2009). Evaluating AISA+hyperspectral imagery for mapping black mangrove along the south Texas Gulf Coast. Photogrammetric Engineering and Remote Sensing 75, 425−435. Yates, C.J., Hobbs, R.J. and Bell, R.W. (1994). Landscape-scale disturbances and regeneration of semi-arid woodlands of south-western Australia. Pacific Conservation Biology 1, 214-221. Yu, X., Hyyppä, J., Kaartinen, H. and Maltamo, M. (2004). Automatic detection of harvested trees and determination of forest growth using airborne laser scanning. Remote Sensing of Environment 90, 451-462.  179  APPENDIX A  Data collection/pre-processing Field program In July, 2007 a stratified random sampling design was implemented wherein suitable plot locations were initially stratified by island and filtered by land tenure. Within island strata plots were further stratified based on API-derived tree species distribution supplied by Parks Canada. Within overlapping strata plots were randomly selected, factoring in island size and aiming to relatively represent the proportion of the landscape estimated by Parks Canada to be occupied by each species. A total of 150 20 m radial plots were established on Saturna, North Pender, South Pender, Tumbo, D’arcy, Sidney and Portland islands. In each plot the perimeters of all trees/tree clusters observable on the AISA imagery and ≥5 m in height were delineated. The coordinates of all trees/tree clusters were recorded using post-differentially corrected Trimble GPS measurements, accurate within ±1.5 m. Within a subset of 26 plots, trees were climbed and branches collected using clippers attached to the end of a 6 m long pole. Branch selection was based on targeting unimpeded sunlit vegetation and samples were collected as high as operational constraints permitted. As recommended by Foley et al., (2006), broadleaves and coniferous needle clusters were removed and placed in sealed plastic freezer bags. Table 1 displays the allocation of samples per species.  180  Taxonomic group Broadleaf  Conifer  Species Black cottonwood Trembling aspen Red alder Bigleaf maple Garry oak Arbutus Grand fir Western redcedar Douglas-fir Western hemlock Lodgepole pine all species  Scientific name Populus balsamifera Populus tremuloides Alnus rubra Acer macrophylum Quercus garryana Arbutus menziesii Abies grandis Thuja plicata Pseudotsuga menziesii Tsuga heterophylla Pinus contorta  Samples 30 22 28 30 31 24 32 23 26 30 30  Total 165  141  306  Table 1: Allocation of samples per species.  In addition to the 150 plots, polygons composed purely of single species were selected from the API-derived inventory. Factoring in the same strata and in accordance with plot dimensions, segments of species-specific polygons were used to augment field-collected reference. As a result, 411 tree/tree clusters representing 11 species were located. For each species, two-thirds of trees/tree clusters were randomly selected as training data, with the remaining reference used to independently assess accuracy (Table 2). Taxonomic group Broadleaf  Conifer  Common name black cottonwood trembling aspen red alder bigleaf maple Garry oak arbutus grand fir Western redcedar Douglas-fir Western hemlock lodgepole pine All species  Scientific name Populus balsamifera Populus tremuloides Alnus rubra Acer macrophylum Quercus garryana Arbutus menziesii Abies grandis Thuja plicata Pseudotsuga menziesii Tsuga heterophylla Pinus contorta  Calibration 11 (513) 7 (121) 14 (2,570) 8 (331) 12 (418) 43 (668) 9 (96) 51 (1,025) 76 (8,801) 13 (105) 13 (256)  Validation 5 (283) 3 (83) 15 (1,210) 6 (186) 8 (225) 28 (570) 8 (58) 40 (766) 50 (4,092) 4 (70) 14 (138)  Total 16 (796) 10 (204) 29 (3,780) 14 (517) 20 (643) 71 (1,238) 17 (154) 91 (1,791) 126 (12,893) 17 (175) 27 (394) 411 (22,585)  Table 2: The quantity of reference tree/tree clusters and the number of associated pixels (in parentheses) per-species. Reference data were partitioned in to calibration and validation data.  181  Of the 411 tree/tree clusters, 200 representing nine species were used to initially test the classification methodology with emphasis on rare and/or at-risk Garry oak (Quercus garryana) habitat (Table 3). Taxonomic group Broadleaf  Species Scientific name Abbreviation Calibration Trembling aspen Populus tremuloides At 72 Red alder Alnus rubra Dr 2300 Bigleaf maple Acer macrophylum Mb 55 Garry oak Quercus garryana Qg 353 Arbutus Arbutus menziesii Ra 54 Conifer Western redcedar Thuja plicata Cw 475 Douglas-fir Pseudotsuga menziesii Fd 3520 Western hemlock Tsuga heterophylla Hw 68 Lodgepole pine Pinus contorta Pl 19 Table 3: Subset of species-specific reference data utilized to initially test the classification methodology with an emphasis on Garry oak (Quercus garryana).  Validation 45 963 31 213 32 312 2634 51 10  TEM plots From 2003-2006, >8,000 1:10,000 (or coarser) scale polygons were derived from fine spatial resolution (i.e., 1.6 m) aerial photography providing wall-to-wall coverage of the SGI. During April-July, 2006, 704 11.3 m radial field plots were established throughout the SGI to enhance polygon-level ecosystem characterizations. Plots completely within GINPR boundaries and LiDAR surveys were targeted for analysis, isolating 141 plots representing the full range of structural classes. Each TEM plot had a structural label and estimated tree species inventory. TEM defined structural stages include sparse/bryoid, herbaceous, shrub/herb, pole/sapling, young forest, mature forest and old forest (Table 4). Specific structural stage designations are made based on a variety of structural features and age criteria.  182  Table 4: Terrestrial Ecosystem Mapping (TEM) structural stage descriptions based on definitions provided by Hamilton (1988), Oliver and Larson (1990), Weetman et al. (1990), Resource Inventory: Vegetation Inventory Working Group (1995) and, BC Ministry of Forests and BC Ministry of Environment (1998).  Avian survey Bird data were collected within plots from early May to mid June, 2005-07, following methods described in detail and applied at landscape scales (e.g., Jewell and Arcese 2008). Briefly, 50m radial plots were established across a range of forested, opencountry and human-dominated habitats for the purpose of mapping species occurrence 183  and estimating species richness. At each visit to plots a trained observer recorded all species detected by sight or sound within a 10 min period inside the plot.  All  observations were made between sunrise and 11AM in the absence of audible rainfall or wind to minimize variation in detection due to time of day, weather conditions or season. During counts, observers moved slowly throughout the plot area to maximize the likelihood of detecting cryptic species. Suitable plot locations were identified for this study using vegetation cover information derived from aerial photographs provided by Parks Canada. 136 plots distributed across 11 islands were identified as: 1) completely within the confines of airborne transects, 2) free of water, intertidal and/or beach and rock and 3) almost entirely (i.e., > 95.0%) comprised of non-agricultural vegetation. The size of islands upon which plots were established varied from 12 to 3000 ha averaging 1170 ha. The combined size of all islands was ~12900 ha, 107 of which were represented by plots. Plots were visited twice on average, but up to 6 times in total. Richness is defined here as the cumulative number of species detected. Uneven sampling effort across plots was accommodated in statistical analyses by weighting estimates based on their reliability.  Ground-based hyperspectral data An Analytical Spectral Devices (ASD) full range (FR) spectrometer (Analytical Spectral Devices, Boulder, CO, USA) was used to measure the spectral reflectance of speciesspecific vegetative samples from 350–2500 nanometers (nm) in a controlled indoor setting. For each measurement, broadleaves or needle clusters were stacked six layers deep to approximate an infinite optical thickness, therefore, simulating canopy 184  reflectance and the maximum of near infrared (NIR) reflectance (Datt, 1998). To reduce noise and capture variability, each spectral curve was calculated by averaging 10 measurements, resulting in 306 curves (Table 5). The original sampling interval of the ASD FR is approximately 2 nm, however, during spectral-curve acquisition; measurements are automatically interpolated to a 1 nm sampling interval. To mimic the original sampling interval, baseline reflectances were obtained by averaging values using 2 nm wavelength increments and a 2 nm full width half maximum (fwhm). First and second derivatives were also calculated from the baseline data, resulting in three datasets. Wavelengths 1350–1416 and 1796–1970 nm, 350–429 and 2401–2500 nm, and 998– 1002 and 1798–1802 nm were removed as they represent water absorption regions (van Aardt and Wynne, 2001), sensor extremes, and sensor transitional zones, respectively (Table 6). Taxonomic group Broadleaf  Conifer  Species Black cottonwood Trembling aspen Red alder Bigleaf maple Garry oak Arbutus Grand fir Western redcedar Douglas-fir Western hemlock Lodgepole pine all species  Scientific name Populus balsamifera Populus tremuloides Alnus rubra Acer macrophylum Quercus garryana Arbutus menziesii Abies grandis Thuja plicata Pseudotsuga menziesii Tsuga heterophylla Pinus contorta  Samples 30 22 28 30 31 24 32 23 26 30 30  Total 165  141  306  Table 5: Number of spectral curves measured per taxonomic group and species. Wavelength region (nm) 350-429 998-1002 1350-1416 1796-1970 2401-2500  Reason for removal sensor extreme sensor transition water absorption water absorption sensor extreme  Table 6: Spectral regions removed from analysis and associated reasons for removal.  185  Airborne hyperspectral data Airborne hyperspectral data were collected in 23 transects (~2800 ha) covering 100 % of Sidney, D’arcy and Portland Islands and portions of North Pender, South Pender, Saturna, Tumbo and Mayne Islands (among others) in mid July, 2006, close to solar noon, by Terra Remote Sensing, Inc. (Sidney, BC, Canada), using an Airborne Imaging Spectrometer for Applications (AISA) Dual push-broom sensor (Spectral Imaging Ltd.) on a fixed-wing platform flown at an altitude of 1600 m yielding a spatial resolution of 2×2 m pixels. The AISA Dual combines the AISA Eagle and Hawk sensors in a dual sensor bracket mount, allowing for the simultaneous acquisition of hyperspectral imagery in the visible, VNIR, and SWIR portions of the EMS. The AISA Dual combines the spectral range of the AISA Eagle (i.e., 395-970 nm) and AISA Hawk (i.e., 970-2503 nm) to allow the collection of hyperspectral data simultaneously from 395-2503 nm over 492 spectral channels. The portion of the EMS abstracted by the AISA Eagle has a spectral resolution of 2.9 nm, whereas that abstracted by the Hawk has a spectral resolution of 8.5 nm. When flown at an approximate altitude of 1600 m, the AISA Dual yields a spatial resolution of 2 m (Table 7). Data were delivered georeferenced to Universal Transverse Mercator (UTM) zone 10 north, World Geodetic System (WGS) 1984 using a 2-m LiDAR derived digital elevation model (DEM). To assess and verify the accuracy of the co-registered, simultaneously collected hyperspectral/LiDAR data, 50 control points, randomly located along roads and/or the coast line, were selected on both the AISA imagery and a LiDAR canopy height model (CHM). A simple least squares adjustment indicated averaged error between the two was within 3 m, equal to 1.5 pixels. This level of co-registration was deemed suitable and well within the smallest average crown size.  186  Characteristic spectral range sampling interval full-width-at-half-maximum spectral resolution altitude spatial resolution  Eagle 395-970 (nm) 1.2 (nm) 1.2 (nm) 2.9 (nm) 1600 (m) 2 (m)  Hawk 970-2503 (nm) 6.3 (nm) 6.3 (nm) 8.5 (nm) 1600 (m) 2 (m)  Table 7: Airborne Imaging Spectrometer for Applications (AISA) Dual: Eagle and Hawk sensor characteristics.  Airborne LiDAR data Airborne LiDAR were collected concurrently with AISA data from the same platform for the same area using a TRSI Mark II two-return sensor onboard a fixed-wing platform. A mean flying height above the ground of 1600 m and a beam divergence of 0.05 mrad yielded a footprint diameter of 0.8 m. A swath width of approximately 980 m resulted from the flying height (i.e., 1600 m) and a maximum scan angle of 17° (Table 8). LiDAR point clouds were labeled using Terrascan v 4.006 (Terrasolid, Helsinki, Finland), which employs iterative algorithms combining filtering and thresholding methods (Axelsson, 1999, Kraus and Pfeifer, 1999) to separate the ground from objects (e.g., vegetation) and subsequently classify returns as either ground or non-ground. As described by Kraus and Pfeifer (1999), this involves: 1) generating a terrain surface to represent all points, equally weighted and forming a triangulated irregular network (TIN), 2) using the surface for averaging to determine the residuals of all z coordinates, and 3) using the residuals to adjust weights and establish thresholds to differentiate between terrain and vegetation. Categorized returns were georeferenced to Universal Transverse Mercator (UTM) zone 10 north (10N) and World Geodetic System (WGS) 84. Pulse frequency (50 kilohertz (kHz)), flight speed, and altitude were optimized to achieve return densities of 187  approximately 1.8 (ground) and 2.5 (non-ground)/m2. Based on comparisons with BC Government elevation benchmarks, average horizontal and vertical accuracies of 0.5 and 0.3 m, respectively, were exhibited. Characteristic wavelength center altitude beam divergence footprint diameter maximum scan angle approximate ground return density approximate non-ground return density  Mark II 1064 (nm) 1600 (m) 0.05 (mrad) 0.8 (m) 17 (°) 2 1.8/m 2 2.5/m  Table 8: Mark II small footprint discrete-return LiDAR sensor characteristics.  Pre-processing airborne hyperspectral data Effective utilization of remotely sensed data requires atmospheric correction, which minimizes atmospheric effects associated with particulate and molecular scattering and rescales raw radiance data to reflectance values (Felde et al., 2003, van der Meer et al., 2006). The Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) atmospheric correction code derived from the MODTRAN4 radiative transfer code was utilized to perform atmospheric correction for all 23 hyperspectral transects. FLAASH input parameters included the wavelength and FWHM sensor specifications, image acquisition information, and sensor altitude. Site specifications included geographic location, elevation, and atmospheric model. For this dataset, local visibility was determined to be good and a maritime atmospheric model was applied (Table 9). In addition, known water absorption bands (i.e., 1350-1416 and 1796-1970nm (PalaciosOrueta & Ustin, 1996, Price, 1998, van Aardt and Wynne, 2001)) and regions exhibiting  188  extreme levels of spectral noise (i.e., 395-429 and 2401-2500nm) were removed; reducing the AISA dataset to 453 channels (Table 10). Parameter wavelength full-width-at-half-maximum sensor altitude geographic location elevation atmospheric model local visibility  Setting 395-2500 (nm) 2.2-6.3 (nm) 1600 (m) latitude, longitude 0-430 (m) maritime good (clear)  Table 9: FLAASH parameters and settings.  Wavelength region (nm) 395-429 1350-1416 1796-1970 2401-2500  Reason for removal sensor extreme water absorption water absorption sensor extreme  Table 10: Spectral regions removed from analysis and associated reasons for removal.  Landsat data To facilitate the extrapolation of tree species heterogeneity, a broader scale landscape representation (coarser resolution and broader spatial extent) was provided by a Landsat5 TM scene comprehensively covering the BC SGI (path: 47/row: 26) and encompassing the hyperspectral/LiDAR acquisition extent and timeframe (July, 2006). The portions of the EMS abstracted by the Landsat 5 TM include the visible, NIR, SWIR and thermal. Specifically, seven TM bands abstract electromagnetic energy in the blue (band 1: 450520 nm); green (band 2: 520-600 nm); red (band 3: 630-690 nm); NIR (band 4: 760-900 nm); SWIR (band 5: 1550-1750 and band 7: 2080-2350 nm); and thermal (band 6: 10400-12500 nm) portions of the spectrum. The spaceborne nature of the Landsat 5  189  platform yields a spatial resolution of approximately 30 m (bands 1-5, 7) and 120 m (band 6). The image was delivered by MacDonald, Dettwiler and Associates (MDA) Ltd. and then precision georegistered, orthorectified and resampled to a spatial resolution of 30 m, with an RMS error of 0.25. At-surface reflectance atmospheric correction was based upon the Cos(t) model (Chavez, 1996), which estimated the effects of absorption by atmospheric gases and Rayleigh scattering and removed systematic atmospheric haze. All Landsat bands were then masked to remove non-forest pixels from the analyses. Figure 1 summarizes data acquisition and pre-processing.  190  Figure 1: Summarization of data acquisition and pre-processing.  191  APPENDIX B  Open-country/shrubland/grassland Scientific name Agelaius phoeniceus Callipepla californica  Common name red-winged Blackbird California Quail  Carpodacus mexicanus Cathartes aura  house Finch turkey Vulture  Circus cyaneus Cistothorus palustris  Northern Harrier marsh Wren  Dendroica aestiva  yellow Warbler  Empidonax traillii Geothlypis trichas  willow Flycatcher common Yellowthroat  Hirundo rustica  barn Swallow  Melospiza melodia  song Sparrow MacGillivray's Warbler house Sparrow savannah Sparrow fox Sparrow cliff Swallow purple Martin bank Swallow American Goldfinch chipping Sparrow  Oporomis tolmiei Passer domesticus Passerculus sandwichensis Passerella iliaca Petrochelidon pyrrhonota Progne subis Riparia riparia Spinus tristis Spizella passerina  Generalist Scientific name Accipiter cooperii Colaptes auratus Corvus caurinus Corvus corax Cyanocitta stelleri Junco hyemalis Molothrus ater Psaltriparus minimus Sturnus vulgaris Troglodytes aedon Turdus migratorius  Mature forest/woodland Scientific name  Common name  Cooper's Hawk Northern Flicker Northwestern Crow common Raven  Bombycilla cedrorum Bubo virginianus  cedar Waxwing great Horned Owl  Buteo jamaicensis Carduelis pinus  red-tailed Hawk pine Siskin  Stellar's Jay dark-eyed Junco brown-headed cowbird  Carpodacus purpureus Catharus guttatus  purple Finch hermit Thrush  Catharus ustulatus  Swainson's Thrush  bushtit European Starling  Certhia americana  brown Creeper  Contopus cooperi  olive-sided Flycatcher  house Wren  Contopus sordidulus  Western Wood Pewee  American Robin  Dendroica coronata  yellow-rumped Warbler  Dendroica nigrescens Dendroica townsendi Dryocopus pileatus Empidonax difficilis Empidonax hammondii Haliaeetus leucocephalus Ixoreus naevius  black-throated Gray Warbler Townsend's Warbler pileated Woodpecker Pacific-slope Flycatcher hammond's Flycatcher bald Eagle varied Thrush  Loxia curvirostra Patagioenas fasciata  red Crossbill band-tailed Pigeon  Common name  192  Open-country/shrubland/grassland Scientific name Spizella passerina Stelgidopteryx serripennis Tachycineta bicolor Tachycineta thalassina Zenaida macroura Zonotrichia atricapilla Zonotrichia leucophyrs  Common name chipping Sparrow Northern Roughwinged swallow tree Swallow violet-green Swallow mourning Dove golden-crowned Sparrow white-crowned Sparrow  Generalist Scientific name  Common name  Mature forest/woodland Scientific name Patagioenas fasciata  Common name band-tailed Pigeon  Pheucticus melanocephalus Picoides pubescens  black-headed Grosback downy Woodpecker  Picoides villosus Pipilo maculatus  hairy Woodpecker spotted Towhee  Piranga Iudoviciana  Western Tanager  Poecile rufescens Regulus calendula Regulus satrapa Selasphorus rufus Sitta canadensis Strix varia Thryomanes bewickii Troglodytes troglodytes Vermivora celata Vireo glivus Vireo olivaceus Wilsonia pusilla  chestnut-backed Chickadee ruby-crowned Kinglet golden-crowned Kinglet Rufous Hummingbird red-breasted Nuthatch barred Owl Bewick's Wren winter Wren orange-crowned Warbler warbling Vireo red-eyed Vireo Wilson's Warbler  Appendix C: Bird species stratified by guild.  193  

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}]}"
                            data-media="{[{embed.selectedMedia}]}"
                            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.24.1-0071720/manifest

Comment

Related Items