UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Using multispectral and hyperspectral satellite data for early detection of mountain pine beetle damage Sharma, Rajeev 2007

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

Item Metadata

Download

Media
831-ubc_2007-267936.pdf [ 19.29MB ]
Metadata
JSON: 831-1.0074966.json
JSON-LD: 831-1.0074966-ld.json
RDF/XML (Pretty): 831-1.0074966-rdf.xml
RDF/JSON: 831-1.0074966-rdf.json
Turtle: 831-1.0074966-turtle.txt
N-Triples: 831-1.0074966-rdf-ntriples.txt
Original Record: 831-1.0074966-source.json
Full Text
831-1.0074966-fulltext.txt
Citation
831-1.0074966.ris

Full Text

USING MULTISPECTRAL AND HYPERSPECTRAL SATELLITE DATA FOR EARLY DETECTION OF MOUNTAIN PINE BEETLE DAMAGE by RAJEEV SHARMA M.Sc. (Botany), Meerut University, 1975 M.Sc. (Forestry), The University of British Columbia, 2000 A THESIS SUBMITTED IN PARTIAL FULF ILMENT O F T H E R E Q U I R E M E N T S F O R T H E D E G R E E O F D O C T O R OF P H I L O S O P H Y in T H E F A C U L T Y OF G R A D U A T E STUDIES (Forestry) T H E UNIVERSITY O F BRITISH C O L U M B I A March 2007 © Rajeev Sharma, 2007 ABSTRACT Mountain pine beetle [MPB] [Dendroctonous ponderosae Hopk.) is the most serious pest of mature lodgepole pine (Pinus contorta) in western North Amer ica. Three key research issues important in developing satell i te-based methods for early M P B damage detection and mapping are examined in this thesis. Relevant questions relating to these issues are: i) is it possible to provide information on MPB-at tacked stands using satellite imagery at an earlier date than conventional methods; ii) is spectral variability in mature lodgepole pine stands significant enough to warrant consideration in M P B attack detection at a landscape level; and iii) are satel l i te-based hyperspectral bands useful in forest tree spec ies discrimination and early detection of MPB-at tacked stands. The first two questions were investigated using multispectral Landsat-7 ETM+ data; the third question was investigated using EO-1 Hyperion hyperspectral data. Using a multi-step deductive approach, MPB-at tacked stands were identified with an accuracy of 6 9 % using the Landsat imagery, approximately four months earlier than would be possible with conventional surveys. Significant spectral variability was found in mature stands of lodgepole pine, Douglas-fir (Pseudotsuga menziesii) and spruce (Picea spp.) at the landscape level. Among the three variables examined (stand age, site index and site ecology), site ecology ( B E C subzone/variants) had the largest influence on the spectral signatures of the three spec ies . Douglas-fir, lodgepole pine and spruce could be identified with an identification accuracy of 81.8%, 8 2 . 1 % and 78.9%, respectively, using a subset of nine narrow bands from the Hyperion sensor, mainly distributed in the 1500-1800 nm spectral region. Corresponding accuracies using Landsat data were 66 .1%, 74 .3% and 67.6%. Another set of nine spectral bands, optimized to identify M P B attack and distributed mainly in the 900-1100 nm spectral region, resulted in identification accuracies of 81 .7% and 80.2% for MPB-at tacked (mainly green-attack) and unattacked stands, respectively. The results of this thesis demonstrate that early detection of MPB-at tacked stands is possible using multispectral and hyperspectral data at a scale and resolution to be of practical use to the forest managers. S o m e of the results from this study have already been used operationally for planning the harvest of MPB-k i l led trees. ii TABLE OF CONTENTS Abstract ii Table of Contents iii List of Tables vii List of Figures x Acknowledgements xiii 1.0 INTRODUCTION 1 1.1 Research Needs 1 1.1.1 M P B Attack Detection 1 1.1.2 Host Spec ies Identification 2 1.1.3 Spectral Bands for Forest Spec ies and M P B Attack Detection 3 1.2 Background 4 1.2.1 Lodgepole Pine 4 1.2.2 Bark Beetles 5 1.2.3 Mountain Pine Beetle 5 1.2.4 H o s t - M P B Interactions 7 1.2.5 M P B Attack Symptoms 8 1.2.6 Economic Importance of M P B 8 1.3 M P B Attack Detection: Conventional Methods 9 1.4 Remote Sens ing-Based M P B Attack Detection 10 1.4.1 Ground Studies 11 1.4.2 Aerial Studies 12 1.4.3 Satellite Studies 13 1.4.4 Summary 15 1.5 Objectives 16 1.6 Thes is Structure 17 1.7 References 20 2.0 EARLY DETECTION AND MAPPING OF MOUNTAIN PINE BEETLE INFESTATIONS USING LANDSAT 7 ETM+ 26 2.1 Introduction 26 2.2 Study Area 29 2.2.1 Lodgepole Pine Distribution 29 2.2.2 M P B Status and Distribution 32 2.3 Materials and Methods 32 2.3.1 Data Analys is 33 2.3.2 Image Analys is 37 2.3.3 Verifying M P B Attack Detection 40 2.4 Resul ts 42 2.4.1 M P B Attack Map 42 2.4.2 Verification Based on Ground Surveys 42 2.4.3 Verification Based on Overview Surveys 45 2.4.4 Verification Based on Helicopter Surveys 47 2.4.5 Verification Based on the I K O N O S Image 47 2.5 Discuss ion 49 2.6 Conclus ions • 50 2.7 References 52 3.0 EFFECTS OF STAND AGE, BEC SUBZONE/VARIANTS AND SITE INDEX ON SPECTRAL VARIABILITY IN SELECTED CONIFER SPECIES 57 3.1 Introduction 57 3.2 Study Area 59 3.3 Materials and Methods 60 3.3.1 Remote Sens ing Data 60 3.3.2 Pre-Process ing of Satellite data 62 3.3.3 Select ion of Forest S tands 65 3.3.4 Sampl ing P lan 66 3.3.5 Data Analysis 73 3.4 Resul ts and Discussion 76 3.4.1 General Spectral Pattern 76 3.4.2 Effect of Stand A g e 81 3.4.3 Effect of B E C Subzone/Var iants 85 3.4.4 Effect of Site Index 88 3.4.5 Interactions Among the Categorical Var iables 89 3.4.6 Effect of Signature Extension 93 3.5 Summary and Conclus ions 96 3.6 References 9 8 iv 4.0 HYPERSPECTRAL BANDS FOR IDENTIFYING SELECTED TREE SPECIES AND MOUNTAIN PINE BEETLE ATTACKED STANDS 102 4.1 Introduction 102 4.2 Study Area 106 4.3 Data Used 106 4.4 Methods 109 4.4.1 Pre-Process ing the Hyperion Data 109 4.4.2 Creating a Georegistered Database 112 4.4.3 Generat ing Hyperspectral Signatures for Tree Spec ies 112 4.4.4 Generat ing Hyperspectral Signatures for MPB-At tacked Stands 113 4.4.5 Spectral Band Select ion 119 4.4.6 Evaluating Selected Spectral Bands : Tree Spec ies 122 4.4.7 Evaluating Selected Spectral Bands : MPB-At tacked Stands 125 4.5 Resul ts and Discussion 125 4.5.1 Spectral Bands for Spec ies 125 4.5.2 Spectral Bands for Identifying MPB-At tacked Stands 130 4.6 Conclus ions 133 4.7 References 136 5.0 CONCLUSIONS AND RECOMMENDATIONS 139 5.1 Approach 139 5.2 Findings 141 5.3 Recommendat ions 143 5.4 Overall Conclus ions 144 Appendix 2.1: Mountain pine beetle Red attack distribution (Red colour) in the Car iboo Forest Region (2000) 145 Appendix 2.2: Mountain pine beetle Red attack distribution Red colour) in the Car iboo Forest Region (2001) 146 Appendix 2.3: Sensor Characterist ics - Landsat-7 Enhanced Thematic Mapper and I K O N O S 147 Appendix 3.1: Equat ions for converting Digital number (DN) to at-satellite radiance and at-satellite-radiance to at-satellite-reflectance; and for estimate of Kappa 148 Appendix 3.2: Site Index distribution in the study area 149 v Appendix 3.3: Spec ies , age and B E C subzone/variant sample sets meeting the stand selection criteria 150 Appendix 3.4: Dataset 1: Pooled calibration and validation samples 151 Appendix 3.5: Genera l Linear Model : Douglas-fir 152 Appendix 3.6: Genera l Linear Model : Lodgepole pine 155 Appendix 3.7: Genera l Linear Model : Spruce 158 Appendix 3.8: Genera l Linear Model : Interactions 161 Appendix 4.1: Absorption features in visible, near-infrared wavebands that have been related to particular foliar chemical concentrations 164 Appendix 4.2: Step Wise Discriminant Analys is : Tree Spec ies (Case C) 165 Appendix 4.3: Step W i s e Discriminant Analys is : M P B Attack 168 Appendix 5.1: Mountain pine beetle update, 2006 172 vi LIST OF TABLES Table 1.1 : Timber harvested in British Columbia -1999/2000 ( B C M O F , 2000) 4 Table 1.2 : Major damaging bark beetles of British Columbia ( B C M O F , 1995) 6 Table 1.3 : Timing of life cycle events for bark beetles ( B C M O F , 1995) 6 Table 1.4 : Fol iage symptoms of successfu l bark beetle attack ( B C M O F , 1995) 6 Table 1.5: Data sources used for various study objectives 19 Table 2.1 : Summary of selected remote sensing studies for M P B attack detection 27 Table 2.2 : A rea (ha) under forest cover polygons with lodgepole pine (60 years and above) spec ies content 32 Table 2.3 : Partial error matrix based on third party independent field verification of 339 randomly selected MPB-at tacked stands identified from Landsat-7 ETM+(Augus t 2000) 44 Table 2.4 : Pests and d iseases confused with MPB-at tacked stand identified from Landsat data 45 Table 2.5 : Composi t ion of MPB-at tacked stands with respect to M P B attack stages 45 Table 3.1 : Environmental characteristics of B E C subzone /variants in the study area 61 Table 3.2 : A g e c lasses in the B C forest inventory 66 Table 3.3: Distribution of lodgepole pine, Douglas-fir, and spruce leading stands in the master dataset 67 Table 3.4 : Criteria for selecting datasets 69 Table 3.5 : Distribution of forest stands by B E C subzone/variant, age c lass, and site index for lodgepole pine, Douglas-fir and spruce in Calibration Dataset-1 69 Table 3.6 : Distribution of samples in Dataset-2 70 Table 3.7 : Distribution of samples in Dataset-3 73 Table 3.8 : Distribution of samples in Dataset-4 74 Table 3.9 : Stand level mean reflectance (%) and spectral range difference for lodgepole pine and Douglas-fir (all three age c lasses) and spruce stands (two age c lasses : A C 5 6 7 and AC89 ) 82 Table 3.10: Multivariate G L M Tests (Wilks' Lambda) for stand age c lass, B E C subzone / variants and site index for Douglas-fir, lodgepole pine and spruce 83 Table 3.11: Tests of between-subjects effects 84 Table 3.12: Multiple Compar isons: age c lass levels. Means followed by the same letter within bands and spec ies not significantly different (Tukey H S D Test, a < 0.05) 84 Table 3.13: Stand level mean (nine B E C subzone/variants) spectral reflectance (%) and spectral range difference for lodgepole pine, Douglas-fir and spruce 86 Table 3.14: Multiple Compar isons: B E C subzone/variant levels. Means followed by the same letter within bands and spec ies not significantly different (Tukey H S D Test, a <0.05) 87 Table 3.15: Stand level mean reflectance (%) and spectral range difference by site c lass for lodgepole pine, Douglas-fir and spruce stands 90 Table 3.16: Multiple Compar isons: site index levels. Means followed by the same letter within bands and spec ies not significantly different (Tukey H S D Test, a < 0.05) 91 Table 3.17: Multivariate tests (Wilks' Lambda): Interactions between species and age c lass, B E C subzone/variant and site index 92 Table 3.18: Band-wise signif icance for interactions between spec ies and age c lass, B E C subzone/variant and site index 92 Table 3.19: Relative effect s ize of categorical and independent variables 93 Table 3.20: Effect of age c lass on classification accuracy (%). The number of stands used in classification are given in parentheses 95 Table 3.21: Effect of site index on classification accuracy (%). The number of stands used in classif ication are given in parentheses 95 Table 3.22: Effect of B E C subzone/variant on classif ication accuracy (%). The number of stands used in classif ication are given in parentheses 95 Table 3.23: Effects of B E C subzone/variants: K H A T statistic 95 Table 3.24: Effect of B E C subzone/variants on spec ies classif ication: Z statistic 96 Table 4.1 : Selected hyperspectral studies for tree spec ies classification 103 Table 4.2 : Selected hyperspectral studies for detection of pests and d iseases 104 Table 4.3 : Data used in the study 108 Table 4.4 : Hyperion calibrated and non-calibrated bands ( U S G S EO-1 Users Guide, 2003) 108 Table 4.5 : Bands most affected by the atmosphere ( U S G S EO-1 Users Guide, 2003) 110 Table 4.6 : Bands recommended for use ( U S G S EO-1 Users Guide, 2003) and bands used in this study 110 Table 4.7: Details of training sites for the three tree spec ies 114 Table 4.8 : Spectral bands selected using different F entry/removal criteria in a step-wise discriminant analysis to discriminate among lodgepole pine, Douglas-fir and spruce 126 Table 4.9 : Spec ies classification accurac ies for different spectral band combinations 127 Table 4.10: Spec ies classif ication accuracy (%) using different Hyperion spectral bands and Landsat ETM+ data 128 Table 4.11: Hyperion spectral bands for M P B attack detection identified using step-wise Discriminant analysis 131 Table 4.12: S D A - b a s e d classification of MPB-at tacked and unattacked stands 132 Table 4.13: Training site based identification accuracy (%) of MPB-a t tacked and unattacked standsat the pixel level (n = 216) 132 ix LIST OF FIGURES Figure 1.1 : Location and relative extent of Test Areas A, B, and C 18 Figure 2.1 : Location of study area 30 Figure 2.2 : (A) Biogeocl imatic Ecosys tem Classif ication Subzone / Variants, (B) Lodgepole pine distribution, in the study area 31 Figure 2.3 : MPB-af fected areas in the Car iboo Forest Region during 1998 - 2002 33 Figure 2.4 : Red attack distribution in the study area in 2000 and 2001 (Source: B C M O F Forest Health Surveys). Infestation severity c lasses are based on percent of red attacked trees within the delineated polygons (Light: 1-10%; Moderate: 11-29%; Severe : >30%) 34 Figure 2.5 : Landsat-7 ETM+ (August, 2000) Band-4 image of study area (2340 km 2 ) . The I K O N O S image covers a small (100 km 2 ) area in south-eastern part of the study area. The partial extent of the IFPA (Innovative Forest Pract ices Agreement) is also shown 35 Figure 2.6 : Mountain pine beetle attacked lodgepole pine stands (red color) identified from Landsat-7 ETM+ data (August 2000). The M P B attacked stands occupy 3130 ha which is 1.3% of the total study area 43 Figure 2.7 : Distribution of randomly selected M P B attacked stands identified from Landsat-7 ETM+ (August 2000) for field verification 43 Figure 2.8 : The MPB-at tacked stands (identified from Landsat data) contained within a 300 m buffer zone around red-attacked stands identified from aerial overview surveys done in 2000 and 2001 46 Figure 2.9 : Existence of M P B attack outside stands with lodgepole pine content, identified from digital forest cover maps 47 Figure 2.10: Examples of Landsat-7 ETM+ (August 2000) based MPB-at tacked stands (yellow color, 30 m pixel) overlaid on I K O N O S normal colour imagery (acquired August 2001). Red attacked M P B attacked stands in 2001 (as seen on 4m I K O N O S image) were current attacked in 2000. The orthorectification rms error bounds ± 30 m are shown in magenta. (The portion of study area covered by the I K O N O S imagery is shown in Figure 2.5) 48 Figure 3.1 : Distribution of B E C subzone/variants in the study area 61 Figure 3.2 : Landsat mosaic of study area (July, 1999) (red: Band 5; green: Band 4; blue: Band 3) : 64 Figure 3.3 : Spatial distribution of forest stands (calibration dataset, 2590 samples) in different B E C subzone/variants in the study area (1: S B P S x c , 2: S B P S d c , 3:IDFdk4, 4: IDFdk3, 5: IDFxm, 6: S B P S m k , 7: S B S d w l , 8: S B S d w 2 , 9: ICHmk3; 10: other zones;11: data gap) 71 Figure 3.4 : Stand s ize distribution for lodgepole pine, Douglas-fir and spruce in Calibration Dataset-1 72 x Figure 3.5 : Proportion of different stand s ize c lasses for lodgepole pine, Douglas-fir and spruce in Calibration Dataset-1 72 Figure 3.6 : Spectral pattern of Douglas-fir, lodgepole pine and spruce stands as ordered by age c lass and B E C subzone/variant within spec ies (continued...) 77 Figure 3.6 : (continued) Spectral pattern of Douglas-fir, lodgepole pine and spruce stands as ordered by age c lass and B E C subzone/variant within spec ies 78 Figure 3.7 : Lateral canopy characteristics of lodgepole pine, Douglas-fir and spruce stands ( B C M O F , 2001) 80 Figure 3.8 : The average spectral reflectance pattern of Douglas-fir, lodgepole pine and spruce 80 Figure 3.9 : Effects of stand age on the spectral variability in Douglas-fir, lodgepole pine and spruce 82 Figure 3.10: Effects of B E C subzone/variants on the spectral variability in Douglas-fir, lodgepole pine and spruce 86 Figure 3.11: Effects of site index on the spectral variability in Douglas-fir, lodgepole pine and spruce 90 Figure 4.1 : Location of the study area within British Columbia 107 Figure 4.2 : Examples of bad pixels (B12-14) and striping (B183-185) in the Hyperion data for the study area 111 Figure 4.3 : Spectral reflectance pattern of unattacked Douglas-fir (red), lodgepole pine (blue) and spruce (green) stands (stand age > 60 years) 115 Figure 4.4 : Spectral reflectance pattern of Douglas-fir, lodgepole pine and spruce stands 116 Figure 4.5 : M P B red-attacked stands (red colored polygons) super imposed on the digital aerial photographs (part of Map sheet 93a002) 117 Figure 4.6 : Red-attacked trees as seen on the aerial photographs. A 30 x 30 m grid (Hyperion spatial resolution) is overlaid on the photograph 118 Figure 4.7 : Spectral reflectance pattern of unattacked (green) and M P B green-attacked (red) lodgepole pine stands (all training sites) 120 Figure 4.8 : Spectral variability due to age in Douglas-fir, lodgepole pine and spruce stands 123 Figure 4.9 : Spectral variability of Douglas-fir, lodgepole pine and spruce stands in different B E C subzone/variants 124 Figure 4.10: Spectral variability of Douglas-fir stands in low and high site index c lasses 125 Figure 4.11: Individual training site level classif ication accuracy in the combined data set using nine spectral bands 129 Figure 4.12: Mean spectral reflectance of unattacked (UA: 51 sites) and attacked (GA: 94 sites) lodgepole pine stands 131 xi Figure 4.13: Summary of steps in the selection of spectral band combinations suitable for conifer spec ies discrimination and for differentiating MPB-at tacked stands from unattacked stands 134 xii ACKNOWLEDGEMENTS I would like to thank and acknowledge the support and encouragement I have received from various people in my very rewarding graduate exper ience; without their help this thesis would not have been possible. I express my sincere gratitude to my supervisors, Dr. Peter A . Murtha and Dr. Peter Marshall for their support, guidance, and encouragement throughout this work. Dr. Murtha (Professor Emeritus, University of British Columbia) provided me the academic freedom at the beginning of my thesis work and also encouraged me to interact with people across academic, professional and industrial sectors, so essential for an applied research thesis. He always built my confidence and sustained faith in my technical abilities. I express my special gratitude to Dr. Peter Marshal l , my Academic Supervisor after Dr. Murtha's retirement, who not only guided me in the design of this study, and provided critical reading of this thesis, but was always there when needed providing advice, assurances and encouragement. He has been a continuous source of inspiration, support and guidance throughout my graduate programme. I am extremely thankful to my supervisory committee members- Dr. Donald Leckie, Dr. John McLean and Dr. Bill Bourgeois. This thesis has greatly benefited from their valuable inputs and suggest ions. I am indebted to Dr. Donald Leckie (Research Scientist, Canad ian Forest Serv ice, Pacif ic Forestry Centre, Victoria), my research supervisor from 2005, for in-depth reviews, critical insights, constructive criticism and helpful d iscussions into my research and support in the last years of my graduate programme. Dr. John McLean introduced me to the complexit ies of mountain pine beetle and lodgepole pine interactions and his seemingly innocuous questions and in-depth comments would often start a thought process in me which led to stimulating d iscussions. Appl ied research can not be completed without the active support and collaboration by an industrial partner. In Dr. Bill Bourgeois, I found a very enthusiastic advocate of applied research. He also helped me gain an insight into the user requirements in the field of mountain pine beetle damage assessment in particular, and in forestry applications in general . This work has greatly benefited from my associat ion with several M P B related initiatives and I thank Dr. Peter A . Murtha, Dr. Bill Bourgeois, Tyler Mitchell, G IS Co-coordinator, Lignum Ltd., Wil l iams Lake and Jeff Alexander, Planning Forester, Lignum Ltd., Wil l iams Lake, for giving me an opportunity to be a part of these initiatives. A lso , Mr. Tyler Mitchell and Mr. Jeff Alexander, not only provided the majority of the collateral and field data used in this thesis, but also organized field and aerial surveys to verify the research results. I owe them special thanks for their continuous support and interest in this research. Support received from Geneve Dagenais, Wil l iams Lake District Planner, Riverside Forest Products Ltd., Wil l iams Lake; Ma lcom Sutton, Planning Forester, Weldwood Forest Products Ltd.; Rick Welke, Planning Forester, West Fraser Ltd., Wil l iams Lake, xiii S e a n Donahue, Planning & Engineering, Smal l Bus iness Forest Enterprise Program, Wil l iams Lake Forest District, is a lso gratefully acknowledged. I also thank Leo Rankin , Forest Entomologist, Car iboo Regional Office, Wil l iams Lake and Guy Newsome, Forest Health Forester, Car iboo Regional Office, Wil l iams Lake for fruitful d iscussions and valuable suggest ions. Support provided by Mr. Jerry Maede l , G IS coordinator, U B C , is highly appreciated. Thanks are also due to Andrew Dyk (Pacif ic Forestry Centre, Victoria) for friendly and insightful d iscussion on hyperspectral remote sensing. Most importantly, I would like to acknowledge the huge moral support that I had from my family. Tara, my lifelong companion and wife of more than 25 years, has provided continuous support, care and encouragement and never let me lose the sight of my goal . My children, Rahul and R icha , have been amazingly supportive and understanding and often enjoyed role reversals and helped me concentrate on my work by sharing several responsibil it ies. I a lso acknowledge the great moral support I have received from my mother-in-law and brother-in-law. In the end I would like to thank my parents who made great sacri f ices in their lives to help me achieve my ambitions and goals in life. Though, they are not physically present in this world today, but I have always felt their presence around me. I dedicate this thesis to them. A part of my graduate research programme (2002-2003) was funded through the British Columbia Sc ience Counci l Graduate Engineering and Technology Scholarship. xiv 1.0 INTRODUCTION Lodgepole pine (Pinus contorta var. latifolia Engelm.) is the most ecological ly diverse of the commercial ly important coniferous tree spec ies in western North Amer ica (Schmidt and Alexander, 1984). Mountain pine beetle [MPB] (Dendroctonus ponderosae Hopk.) is its most destructive pest, which preferentially attacks trees of 80 years of age or more (Amman and Safranyik, 1984; Safranyik and Carrol l , 2006). Although M P B is endemic to mature lodgepole pine forests, under favorable climatic conditions and with abundant food supply, M P B populations can reach epidemic proportions resulting in large scale mortality of lodgepole pine. G iven the economic importance of lodgepole pine and consequently the large impact of M P B on timber supply with huge economic and social consequences, timely detection, mapping, and monitoring of M P B infestations is of paramount importance ( B C M O F , 2003a). Information on the location, s ize, attack stages (current or green attack, red attack, gray attack; explained in section 1.2.5) and severity levels of M P B infestations is needed for developing beetle management and mitigation strategies at local and provincial levels. Except for ground surveys, conventional survey methods (i.e. aerial sketch mapping, aerial photography ( B C M O F , 2001)) provide this information at least one year after a successful M P B attack. Both aerial sketch mapping and aerial photography are based on aerial detection of foliar color change from green to red (red attack stage), which generally takes place in the summer following M P B attack. However, foresters need information within a few months of successful M P B attack to plan harvest operations. From a beetle management perspective, it is essential that beetle infestations be detected during early stages of successfu l attack and when infestation s izes are smal l . After the initiation of M P B attack, all three attack stages can be found within a stand in subsequent years as the emerging beetles from dead trees attack healthy mature trees within that stand. This can result in a mixed spectral response from medium spatial resolution remote sensing data (e.g., from Landsat Thematic Mapper (TM), Enhanced Thematic Mapper (ETM+), 30 m spatial resolution, S P O T Multi-spectral Linear Array (MLA), 20 m spatial resolution). It may not be technically correct to label these pixels as red or green or gray attacked stands. Therefore, throughout this thesis the more general term " M P B attack" has been used. 1.1 RESEARCH NEEDS 1.1.1 MPB Attack Detection Because of the cycl ic nature of M P B epidemics in British Columbia (BC) and the economic importance of resulting losses, application of aerial and satell i te-based remote sensing techniques for 1 detection of MPB-at tacked stands has been an area of active research. A number of studies (e.g., Harris et al., 1978; R e n c z and Nemeth, 1985; Ahern, 1988; Sirois and Ahern , 1988; G imbarzevsky ef al., 1992; Franklin et al., 2003; Skakun et al., 2003; Bentz and Endreson, 2004; White et al., 2004) have demonstrated that large, red-attacked stands can be detected using medium (e.g., S P O T , Landsat T M / ETM+) and high resolution (e.g., IKONOS) satellite data. M P B red attack detection accuracy varied from under 25 percent to more than 70 percent depending upon the spatial resolution of the satellite data used. Discussions with practicing foresters revealed that their prime information need is to get M P B attack locations at a much earlier date than is currently available from red attack stage methods. Although aerial multispectral techniques have reached a level of maturity in early detection of M P B -attacked stands (Murtha, 1985; Murtha and Wiart, 1987; Roberts ef al., 2003; Roberts ef al., 2005), early detection of MPB-at tacked stands using satellite data is unproven. Satel l i te-based M P B attack detection is still at a research stage due to: i) the inferential nature of the information derived from medium resolution satellite data (Roller, 2000), and ii) inherent spectral variability in MPB-at tacked stands due to the co-existence of various attack stages, variable rates of infestations within stands, and the confounding influences of several d iseases and pests, as well as site ecology, on the spectral signatures. 1 .1 .2 Host Species Identification Lodgepole pine grows in even-aged pure stands, and in mixed stands in associat ion with other conifer spec ies such as Douglas-fir (Pseudotsuga menziesii), spruce (Picea spp.) and subalpine fir (Abies lasiocarpa). Separat ing lodgepole pine stands from other forest vegetation could reduce the confounding influence of other pests and d iseases on identification of MPB-at tacked stands. It would be more time and cost efficient to deal with only lodgepole pine stands compared to the entire forested area for multi-temporal mapping of M P B infestations and for monitoring changes in lodgepole pine cover resulting from harvesting and other disturbances. There is a need for forest species mapping at a much finer spatial resolution than is currently available from provincial forest inventory maps, where a forest stand can range in s ize from a few to tens of hectares. Stand age, site productivity, and ecological site factors may influence the reflectance spectra of tree species growing in different locations (Dadykin and Bedenko, 1960; Kalensky and Wi lson, 1975; Bartlett ef a/., 1988; Cihlar ef a/., 1991; Ramsey ef al., 1995). However, the effect of diversity on spectral response is either poorly understood or unknown (Cihlar ef al., 2003). A s lodgepole pine grows in 12 of the 14 Biogeocl imatic Ecosys tem Classif icat ion (BEC) zones in B C (Pojar et al., 1991), there is a need to study the spectral variability in lodgepole pine stands across these zones. 2 Understanding ecological ly- induced spectral variability in lodgepole pine stands would be useful in spectral stratification and establishment of permanent observation (ground truth) plots to generate area-specif ic spectral signatures in cover types of interest. 1.1.3 Spectral Bands for Forest Species and MPB Attack Detection Most satell i te-based M P B attack detection has been carried out using broad-band multispectral scanners such as Landsat M S S , Landsat T M , S P O T , Landsat-7 ETM+, and I K O N O S . Amongst other factors, the broad-band nature of the multispectral data used could also be responsible for the varying detection accuracy reported. Sensit ive features related to plant health and vigor (such as green peak, red edge) embedded within the spectral curve are interspersed with insensitive features (e.g., chlorophyll well, near-infrared plateau). A broad waveband can easi ly combine narrow sensitive and insensitive features masking the response of the sensit ive component of the spectral curve (Carter, 1994; Gitelson et al., 1996). Many researchers (e.g., Runesson , 1991; Davidson and Lataille, 2000; Heath, 2001; Smith ef al., 2002) have explored hyperspectral remote sensing data for M P B attack detection and have suggested suitable spectral regions. However, these studies were carried out using ground-based spectral measurements at the needle level or using aerial data at the tree level; the suitability of these spectral regions for M P B detection at the stand-level remains to be explored. A s well, there is a wide range in the spectral regions recommended for M P B attack detection. Therefore, the need to identify optimal spectral band combinations for early M P B attack detection still persists. The optimal bands could be used in designing an optimal sensor configuration, particularly for configurable sensors such as the Compact Airborne Spectrometer Imager (CASI) , in which the user has to make decis ions regarding the trade-off between the number, location, and widths of spectral bands (Warner and Shank, 1997). With the launch of the Earth-Observing 1 (EO-1) satellite, in November 2000, Hyperion hyperspectral data have become available. These data are acquired on a global scale, on a repetitive basis, previously unavailable with aircraft-based systems (Unger ef al., 2003). The Hyperion sensor is a pushbroom hyperspectral imaging spectrometer with 242 spectral channels covering visible, near-infrared, and shortwave infrared bands (400-2500 nm) with a 10 nm band interval (Datt et al., 2003). 3 1.2 BACKGROUND INFORMATION 1.2.1 Lodgepole Pine Lodgepole pine forests cover 6 million ha in the United States and about 20 million ha in C a n a d a (Lotan and Critchfield, 1990). In B C , lodgepole pine forests cover 14.9 million hectares ( B C M O F , 2003a). Lodgepole pine is the single largest contributor of any species to B C ' s annual timber harvest. In 1999/2000, lodgepole pine accounted for 29 .1% (23.19 million m 3) of the total provincial timber harvest (Table 1.1). In 2002-03, it accounted for 34 .2% (27.28 million m 3) of the total provincial timber harvest ( B C M O F , 2003b). Table 1.1: Timber harvested in British Columbia -1999/2000 ( B C M O F , 2000). Species Volume (Million m3) Percent Lodgepole pine 23.19 29.1 Spruce 12.53 15.7 Douglas-fir 10.88 13.6 Hemlock 11.21 14.1 Ba lsam 8.97 11.3 Cedar 7.03 8.8 Other spec ies 5.91 7.4 Total timber harvest 79.72 100 Lodgepole pine provides habitat for a variety of insect spec ies ranging from moth larvae that feed on seed and cones to bark beetles that cause widespread mortality of mature trees. Of the approximately 240 species of insects that feed on lodgepole pine, 35 are considered pests or potential pests (Amman and Safranyik, 1984). Bark beetles (Coleoptera: Scolytidae) pose serious threats to mature lodgepole pine. The major d iseases of lodgepole pine, in decreasing order of damage, are: dwarf mistletoe (Arceuthobium americanum Nutt.), stem rusts (Cronartium spp.), root d iseases (Armillaria spp.), cankers and foliage d isease (Van der Kamp and Hawksworth, 1984; McLean ef al., 2005). These pests and d iseases affect productivity through: i) killing lodgepole pine seedl ings leading to under-stocked stands); ii) reducing merchantability (through killing of terminal leaders resulting in deformed boles); iii) growth reduction (e.g., defoliators which prolong the time required for a tree to reach merchantable size); and iv) direct killing ( losses due to widespread killing of merchantable trees or reducing the number of merchantable trees to a point that the harvest of residual trees is unprofitable) (Amman and Safranyik, 1984). 4 1.2.2 Bark Beetles M P B , Douglas-fir beetle (Dendroctonus psudotsugae Hopk.; host spec ies - Douglas-fir), and spruce beetle (Dendroctonus rufipennis Kirby; host spec ies - spruce), are among the most important bark beetles in British Columbia (Table 1.2). These beetles form an integral part of the forest ecosystem and help maintain forest health through infesting branches, stumps, and stems of standing dead, severely weakened trees, or downed material. However, during extended favorable climatic conditions and abundant food supply, these spec ies periodically reach outbreak (epidemic) levels. During outbreaks, they kill large numbers of apparently healthy trees, over extensive areas. A general life cycle pattern and foliar color changes associated with successfu l bark beetle attack are given in Table 1.3 and Table 1.4, respectively. The M P B is the most destructive of the bark beetles as it attacks and kills live trees. 1.2.3 Mountain Pine Beetle M P B is a small (4.0-7.5 mm in length), aggressive bark beetle that can colonize healthy trees (Raffa et al., 1993). It prefers mature lodgepole pine trees which are older than 80 years, with diameters at breast height (DBHs) greater than 25 cm (Shrimpton and Thomson, 1983; Amman and Safranyik, 1984).The M P B also carries blue-staining fungi (Ophiostoma clavigerum and O. montium) (Nebeker et al., 1993). These fungi produce a pigment, melanin, which causes a blue to black discoloration of the sapwood (Kim ef al., 2005). The associat ion between the M P B and these fungi is mutualistic. The staining fungi lower the moisture content of the wood and may produce an environment more favorable for beetle broods and some of the fungal associates are potential nutrient sources for M P B larvae (Kim ef al., 2005). M P B complete their one-year long life cycle within a successful ly attacked tree. At the end of their life cycle, mature adult beetles emerge from the brood tree during July-August and find a suitable host to lay their eggs. The following summer, larvae pupate in late June and in July-August they mature into adult beetles, emerge, and the whole process starts again. M P B requires 833 degree-days above 5.6°C to complete its annual life cycle. Significant winter mortality occurs if temperatures beneath the bark drop to -40° C . During the last six to eight years B C has not experienced a significant cold event. S ince 1995 the minimum degree-day accumulat ions required to complete M P B life cycle have been consistently exceeded in central B C (Carroll and Linton, 2002). 5 Table 1.2: Major damaging bark beetles of British Columbia ( B C M O F , 1995). - Common name '^My ' Species name • «I#$'4'-- Insect code'i* Host species Mountain pine beetle Dendroctonus ponderosae Hopk. IBM PI, Py , Pw, P a Spruce beetle Dendroctonus rufipennis (Kirby) IBS S e , Sw , S s Douglas-fir beetle Dendroctonus pseudotsugae Hopk. IBD Fd Western Ba lsam bark beetle Dryocoetes confuses Swaine IBB BI PI: Lodgepole pine (Pinus contorta); Py : Ponderosa pine (Pinus Ponderosa); Pw: Western white pine (Pinus monticola); P a : Whitebark pine (Pinus albicaulis) S e : Engelmann spruce (Picea engelmanni); Sw: White spruce (Picea glauca); S s : Sitka spruce: (Picea sitchensis) Fd : Douglas-fir (Pseudotsuga menziesii) BI: Subalpine fir (Abies lasiocarpa) Table 1.3: Timing of life cycle events for bark beetles ( B C M O F , 1995). '"''•Factor ' ••'v''.' • IBM • N : '• IBS IBD :»'•' Main adult flight July through August May through June April through July Host preference Living tree Windfal l /s lash, or living tree Windfal l /s lash, or living tree Normal length of life cycle 1 year 2 yea rs 3 1 year Overwintering stage Larvae Larvae and adult" Larvae and adult [a] The length of the spruce beetle life cycle is highly dependent on the rate of heat accumulat ion. In warm summers, spruce beetle can complete its life cycle in 1 year. [b] Spruce beetle must overwinter as an adult before emerging to attack new host material. In the normal life cycle of two years, spruce beetle overwinters as a larva in the first year and as an adult in the second. Table 1.4: Fol iage symptoms of successful bark beetle attack ( B C M O F , 1995). IBM IBS IBD 1 -year post attack Bright red crown in spring following attack No color change in first spring following attack Bright red crown in spring following attack (may turn red in the year of attack) 2-year post attack Dull red remains on tree Genera l yellowing or graying of crowns after 24 to 30 months Crowns gray in spring 2 years after attack 3-year post attack Crowns gray, little foliage remains Gray crowns, needle drop Crowns gray, little foliage remains 6 1.2.4 Host-MPB Interactions Colonizat ion by M P B consists of four phases: dispersal, host select ion, mass attack, and establishment (i.e., elimination of host resistance followed by successfu l oviposition and fungal growth in the host substrate) (Wood, 1972; Raffa ef a/., 1993). Adult beetles emerge from the brood trees from mid-July through August. Emergence may occur over a period of several days to several weeks; however, it usually peaks over 2-3 days during this period, depending upon the temperature and other environmental condit ions. During dispersal flights, the pioneer female M P B locates a suitable host tree and initiates mass attack by releasing aggregation pheromones. It has been shown that the vigor and stress status of a tree play an important role in selection, with trees under greater stress being more vulnerable to infestations (Berryman, 1976; Mitchell ef al., 1983). However, at epidemic levels the beetles also attack trees of high vigor (Paine ef al., 1997). P ines are characterized by a well defined oleoresin duct system, which can mobil ize large quantities of oleoresin to pitch out attacking beetles (Christ iansen ef al., 1987). The preformed (constitutive) oleoresin system of pines is considered to be the primary defense against bark beetle attack. A s a second line of defense, after the beetle gains entry, trees respond with a hypersensit ive reaction. The tree t issue surrounding the entry point dies, and is f looded with tree resins and phenolics produced by the surrounding cells. These chemicals are toxic to bark beetles and inhibit fungal growth, and the tree t issues around the beetle and fungal attack sites are modified to form a barrier against the further spread of infection (Christ iansen ef al., 1987). However, both these tree defenses are overcome gradually through sustained M P B mass attack. Host resistance has a threshold (expressed in attacks per unit bark surface) that is related to host condition (Raffa and Berryman, 1983; Christ iansen ef al., 1987). If the beetle attack density is below the resistance threshold of the tree and the tree defenses have not been depleted, the M P B will continue to produce aggregation pheromones; the aggregation is terminated once host resistance has been exhausted. Stress reduces the ability of a tree to respond and lowers the number of beetles required to overcome the resistance (Berryman, 1972; Raffa and Berryman, 1983; Miller et al., 1986). After the M P B have successful ly overcome a tree's defensive system by mass attack, colonization begins. This involves two main components: i) the physical construction of egg galleries, and ii) introduction of blue-staining fungi into the phloem and xylem t issues (Nebeker ef al., 1993). The fungi penetrate and kill live t issue surrounding the bark beetle galleries and enhance host colonization by the beetles and cause rapid host death through: i) toxin production; ii) mycelial plugging of the tracheids; iii) release of gas bubbles into the tracheids disrupting water transportation; 7 iv) production of particles that block the pit openings; and v) reduction in stored food in the parenchyma cells and restricted water conduction by destroying the ray parenchyma cells that partially control water movement (Paine et al., 1997). 1.2.5 M P B Attack Symptoms The first signs of bark beetle attack are usually red boring dust in the bark crevices and/or at the base of attacked tree, as well as yellow pitch tubes at the point of entry on the bole. Foliar color changes follow as a result of bark beetle / fungal colonization. Newly attacked pines are termed green or current attack (GA) s ince the foliage is still visibly green. The foliage gradually dries, fades, and turns a yellow green over the fall and winter. Finally, in the summer following the attack, the lodgepole pine foliage turns red brown. This stage is called red attack (RA). At this stage beetles have left the dead tree and colonized other healthy trees. In the third year when the foliage has fallen off the tree and the branches are bare, the trees are referred to as gray attacked. The rate of color change varies with season , geographic location, and weather (Safranyik et al., 1974). In some cases only one side of a tree is successful ly attacked (strip attack). Such trees survive unless living portions are re-attacked in subsequent years ( B C M O F , 1995) After the color change is pronounced, infestations can be detected from the air. The warmer and dryer the season , and the greater the water stress level, the quicker the discoloration occurs. Fol iage symptoms of attacked lodgepole pine trees generally are not obvious until shortly before the mature adult beetles fly from the tree in the summer following attack. It is these visual and physiological changes in the canopies that can be used by aerial and satellite remote sensing techniques to detect M P B attack. 1.2.6 Economic Importance of M P B Endemic M P B populations are an integral part of lodgepole pine forests and help maintain stand health. An endemic beetle population is held in check by predators, limits to the food supply, natural host resistance (Goyer ef al., 1998) and lethal cold winters (temperatures <-40°C). At endemic levels the M P B survives in weakened or stressed trees. During extended favorable climatic conditions and abundant food supply, M P B populations periodically reach epidemic levels and cause large-scale mortality of mature lodgepole pine trees. Large sca le infestations impact forest management through short term requirements to "chase" beetle infestations, marketing of infested timber, and long-term impacts on sustained yield ( B C M O F , 2003c). They also affect aesthetic values, wildlife habitat, ecological success ion and watershed values. 8 The proportion of mature and over-mature lodgepole pine stands has increased three-fold in B C over the past century (from about 2.5 million hectares in 1910 to over 8 million hectares in 1990), largely due to fire suppression efforts. This created a vast supply of potential hosts for the buildup of large M P B populations under favorable weather conditions. There is currently more than 1 billion cubic meters of mature pine at risk of M P B infestation in the interior of the province ( B C M O F , 2003a). Large-scale M P B infestations have occurred with regular periodicity in B C . S ince the first recorded infestations in 1913 in the Okanagan and Merrit areas, major infestations have occurred in Kootenay National Park and the Chilcotin Plateau in the 1930s, on Vancouver Island during the 1940-1950s, near Takla and Babine lakes in the 1950s, and throughout much of the southern interior, Chilcotin Plateau and the Skeena and Nass River areas in the late 1970s and early 1980s. S ince 1994, M P B populations have grown steadily in the central interior of B C . Starting in 1999, infestations began to spread to new areas at a much faster rate. Infestations have spread from 0.3 million ha in 1999 to 5.7, 8.0, 8.9 and 10.1 million ha in 2000, 2001, 2002 and 2003, respectively. The cumulative volume of dead timber during this period was 6.0 million m 3 in (1999), 41.1 million m 3 (2000), 71.8 million m 3 (2001), 107.7 million m 3 (2002) and 173.5 million m 3 in 2003. The 173.5 million m 3 dead timber is equivalent to over two years of allowable annual cut (AAC) for the entire province and the value of timber products made from this volume is estimated to be just under $18 billion (COFI , 2003). 1.3 MPB ATTACK DETECTION: CONVENTIONAL METHODS Aerial sketch mapping, aerial photography or hel icopter-based G P S surveys and ground surveys ( B C M O F , 1995, 2001) are used operationally in B C for mapping MPB-at tacked stands. Aerial sketch mapping is used for mapping red-attacked stands on an annual basis by the B C M O F . Aerial flights are conducted during early July to early September and locations of M P B infestations are delineated on N T S (National Topographic Survey) map sheets. The outputs include maps showing locations of forest damage, severity levels, infestation size and reports or tables of statistical summaries. Aerial photography provides a permanent record of conditions over an area at a given point in time and is the most widely used remote sensing data source for bark beetle damage detection because of the high resolution of information. Hel icopter-based G P S surveys provide G P S co-ordinates of the centers of red-attacked stands with a positional accuracy of + 20 m. Possibi l i t ies of errors due to false positives in such surveys are extremely low; errors of omissions are not measured. The overall objective of detailed aerial surveys is to produce a map of locations with red-9 attacked trees so that ground crews may be directed expeditiously to areas where treatment is required. Ground surveys are conducted at a local level when pockets of discolored trees first appear in a stand to verify the causal agent (Unger, 1993). Ground surveys include walkthrough surveys (reconnaissance level surveys) and beetle probes (systematic surveys carried out in ± 5 m wide transect lines at an interval of 100 m). During beetle probes, information on several parameters such as the stage(s) of beetles under the bark (egg/larva/pupa/adult), the ratio of different stages and the percentage of various attack categories are collected ( B C M O F , 1995). Ground surveys can locate current attack, including strip attack, but they are expensive and cannot be used over large areas. Using aerial survey-based methods, information on the extent and severity of MPB-at tacked trees and stands is available only after 12-18 months following the attack. Even then, the information is only on the location and extent of red-attacked trees which are already dead and have lower timber value due to the blue staining of the dead wood. Foresters need this information within a few months after the attack to effectively plan harvest operations to combat the spread of M P B . Therefore, there is a need to reduce the time gap between the M P B attack and its detection. The need for alternate methods, suitable for early detection of M P B infestations, has long been recognized. 1.4 REMOTE SENSING-BASED MPB ATTACK DETECTION A number of attempts at detecting M P B attack have been made employing: i) different types of remote sensing data (e.g., aerial color-infrared photographs, aerial multispectral digital data, space-borne multispectral digital data); ii) different spatial resolutions (ranging from large scale aerial photographs to medium spatial resolution data available from space-borne sensors) ; iii) different analytical methods (visual interpretation and digital image analysis methods); and iv) pixel level to sub-pixel level image analysis procedures. These developments have been reviewed by Murtha (1978), Puritch (1981), Leckie ef al. (1983), Gimbarzevsky (1984), Wulder and Dymond (2004), Niemann and Visintini (2004), and Wulder ef al. (2006). Cho ice of remote sensing data and analysis techniques is largely governed by the spectral response of the feature of interest, its s ize, and its spatial distribution pattern. The application of remote sensing data in M P B attack detection is based on the premise that a successful attack induces stress in the attacked tree, which is manifested through changes in the spectral reflectance pattern in different spectral regions (during green attack), followed by foliar color change (during red attack), and lastly through spectral and morphological changes in the crown characteristics (gray attack). 10 The majority of the remote sensing studies on M P B attack detection have been carried out on known attacked sites at the: (i) needle level (Heller, 1968; Ahern , 1988; Runesson , 1991); (ii) tree level (Klein, 1973, 1982; Hobbs, 1983; Murtha, 1985; R e n c z and Nemeth, 1985; Murtha and Wiart, 1987; Kneppeck and Ahern, 1989; Murtha and Wiart, 1989a, b; Davidson and Lataille, 2000; Heath, 2001; Smith et al., 2002; Roberts et al., 2003); and (iii) stand level (Harris ef al., 1978; Brockhaus ef al., 1985; Sirois and A h e m , 1988; Sharma, 2000; Franklin ef al., 2003; Skakun ef al., 2003). These studies were aimed at identifying differences in the spectral response among healthy, green-attacked and red-attacked trees (Heller, 1968; Ahern, 1988; Runesson , 1991), early detection of green-attacked trees (Hobbs, 1983; Murtha and Wiart, 1987; Murtha and Wiart, 1989a,b; Davidson and Lataille, 2000; Heath, 2001; Smith ef al., 2002; Roberts ef al., 2003), and red attack detection (Harris ef a/., 1978; Brockhaus ef al., 1985; Sirois and Ahern , 1988; Franklin ef a/., 2003; Skakun ef al., 2003). Remote sensing data from a variety of sensors were used in the studies identified above. For example, red attack detection has been attempted using 80 m Landsa t -MSS (Harris ef al. 1978; Brockhaus ef al., 1985); panoramic color infrared photography from a U-2C airplane (Klein, 1982), and 30 m Landsat-7 ETM+ (Franklin ef al., 2003; Skakun ef al., 2003). Green attack detection has been attempted using aerial hyperspectral data (Davidson and Lataille, 2000; Heath, 2001; Smith ef al., 2002), and large scale aerial color infra-red imagery (Hobbs, 1983; Murtha and Wiart, 1987; Murtha and Wiart, 1989a,b; Roberts ef al., 2003). A variety of data analysis techniques have been used for the identification of red- and green-attacked trees, including clustering techniques on digitized aerial photos (Murtha and Wiart, 1987; Murtha and Wiart, 1989a,b), supervised classif ication (Franklin ef al., 2003; Roberts ef al., 2003), discriminant analysis (Heath, 2001; Skakun ef al., 2003), proprietary algorithms (Davidson and Lataille, 2000; Smith ef al., 2002), and visual interpretation methods (Harris ef al., 1978). 1.4.1 Ground Studies Heller (1968) studied the effect of M P B attack on the spectral reflectance of ponderosa pine (Pinus ponderosa, Laws.) trees. Spectral reflectance in the 400 to 2200 nm wavelengths from healthy, newly attacked (45 days after attack), and old attacked (discolored foliage) were measured. It was concluded that the greatest deviation between reflectance occurs between the old infested foliage and the healthy foliage in all the spectral regions. The newly infested foliage showed an increase in reflectance at 680 nm (red region) and a decrease in reflectance of 5 to 10 percent at 750 to 1200 nm. 11 Ahern (1988) studied the spectral reflectance properties of MPB-at tacked lodgepole pine trees to identify spectral bands (in the 360-1050 nm range) showing the earliest sign of attack. He identified four spectral regions (525-565 nm, 690-730 nm, 730-760 nm, and 760-1050 nm) as the most promising for detecting early effects of M P B attack. These spectral regions are the green peak, red edge red shift, NIR shoulder region and NIR plateau. Runesson (1991) had similar findings. Both studies consisted of a detailed ground survey examining the reflectance from individual needles from both unattacked and green-attacked lodgepole pine trees. The main difference between the studies was that Runesson found that the red edge for green-attacked lodgepole pine trees shifted to shorter wavelengths (a blue shift) instead of a shift to longer wavelengths (a red shift) as indicated in Ahern 's study. Runesson (1991) also found that the peak of the red edge for healthy trees had higher values (715.6 nm on average) than those of successful ly attacked trees (710.5 nm on average). 1.4.2 Aerial Studies Several studies have concentrated on the use of colour infrared film to monitor green-attacked lodgepole pine. Hobbs and Murtha (1984) successful ly detected green-attacked pine by measuring the densitometric changes on large scale colour infrared film. Murtha (1985) used visual analysis to detect green-attacked pine. His analysis examined the colour and the hue of the tree crown to determine whether the tree had been recently attacked. He identified different foliage age c lasses on a normal (healthy) tree which have different tones and hues. He noted that the crowns of green-attacked pine lose their variegated pattern. The variegated patterns in non-attacked trees are evident in remotely sensed data when the spatial resolution is fine enough to resolve different ages of needles along branches. However, large scale (e.g., 1:2000) colour infrared photographs are not operationally feasible over wide areas, due to the expense of obtaining and interpreting such photographs. Gimbarzevsky et al. (1992) used 1:56000, 1:19000, and 1:8000 scale normal color and color infrared photographs, aerial multispectral data (from the 11 band Daedalus / M D D A DS1260 system) and digital Landsat M S S to investigate the potential operational use of various remote sensing techniques for identifying MPB-k i l led forest stands, mapping their areal extent, and measuring damage intensity levels. The supporting data included aerial-sketch maps showing M P B red-attacked stands and damage severity, 70-mm color photography of 32 4-ha photo plots at a scale of 1:6000 and 90 stereo pairs at an average scale of 1:1000, and a series of 35-mm oblique color sl ides. This study was carried out in a 370-km 2 section in the G u n Lake area in south central B C . The authors concluded that airborne and satellite multispectral images were inferior to color aerial photography for outlining MPB-k i l led forest stands, s ince more area appeared to be infested by beetles on the Landsat M S S imagery than was actually the case . A lso , it was difficult to discriminate between 12 damage intensities on the M S S images. The color aerial photography was found to be the most efficient, practical and reliable remote sensing technique for delineation of M P B infested stands. The authors also recommended that Landsat T M and S P O T imagery (because of their higher spatial and spectral resolution) be a s s e s s e d in the future to determine their efficacy in mapping MPB-at tacked stands and in damage intensity mapping. The identification and counting of damaged trees was also influenced by the photograph sca les. While 1:1000 stereo-pairs permitted the recognition and counting of almost all trees, on the 1:6000 scale photographs, counts had to be adjusted with the aid of 0.25 ha ground-verified sub-plots. The ground sub-plot / photo-plot ratio averaged 1.19 for red-attacked trees, 3.23 for gray-attacked trees and 1.60 for the total number of counted trees. The count accuracy continued to decrease with smaller sca les. Heath (2001), using C A S I hyperspectral data (60 cm spatial resolution, 36 spectral bands), found that derived spectral reflectance curves (for green-attacked and unattacked populations) displayed greater variability within each population than between the populations. It was not possible to visually detect a difference between the populations. Using a discriminant analysis approach, 79 percent of the green-attacked trees and 68 percent of the unattacked trees were reported to be correctly classi f ied. However, classif ication accuracy decreased to 63 percent for green-attacked and 64 percent for red-attacked trees when the spatial resolution of C A S I data was reduced from 60 cm to 1.2 meters through resampling. However, similar accuracies could not be achieved at the pixel level. It was concluded that supervised classification methods (e.g., a maximum likelihood classifier) were unable to distinguish a difference between the populations due to the spectral variability within each population. Roberts ef al: (2003) stated that digitally converted multispectral aerial photography (blue, green, red and near-infrared spectral bands) performed best for early detection of current attack and mapping and monitoring of red attack. Using this approach, current attack could be reliably detected as early as mid-May to early June. For early detection of current attack, imagery at a 1:8000 sca le was found to be superior to that at 1:16,000; however, the latter was found to be sufficient for operational use. The authors concluded that multispectral and digitally converted aerial photographs can provide M P B detection and monitoring accurately and cost-effectively. 1.4.3 Satellite Studies The first attempt to map MPB- induced mortality in B C using satellite data (Landsat-1 M S S , 80 m spatial resolution) was made by Harris ef al. (1978). The results were not encouraging, largely because of the scattered nature of infestations and the coarse spatial resolution of Landsat M S S . An identification accuracy of only 2 5 % was achieved. Ahern and Archibald (1986), based on visual 13 analysis of Landsat T M false color composi tes, showed that gray-attacked areas were identifiable because of a distinct cyan color. Sirois and Ahern (1988) carried out a study to identify red-attacked lodgepole pine stands near Babine Lake, B C . Digital data from S P O T M L A and P L A (Panchromatic Linear Array, resolution 10 m) were used in the study. Different band combinations of digitally enhanced S P O T data for three test areas were visually interpreted to estimate the lower limit of red attack detectable. The authors concluded that the minimum red attack damage detectable with the S P O T satellite was approximately 1 to 2 ha in s ize with 80 to 100 percent of crowns red. This degree of mortality was found to be too high for M P B control programs where the requirement was to detect infestations of five or more trees. During a joint study between the Canada Center for Remote Sens ing ( C C R S ) and the British Columbia Ministry of Forests, R e n c z and Nemeth (1985) evaluated the capabil it ies of Landsat M S S and Thematic Mapper data (simulated from airborne multi-spectral scanner digital data) to detect M P B infestations in four test areas (each 5 x 8 km in size) in the Car iboo region of British Columbia. At one site the M P B red attack infestation was very small in size (few trees) and scattered, whereas at the remaining sites the red-attacked stands were larger than 1.5 ha. The 30 m simulated T M data provided better identification accuracy than the 80 m M S S data. R e n c z and Nemeth (1985) concluded that the "spatial and spectral resolution of Thematic Mapper data will permit insect damage, specifically red attack in lodgepole pine, to be monitored where areas of outbreak exceed 1.5 ha". It was suggested that the s ize of the outbreak is the major influencing factor in successful insect damage detection and that infestations must be larger than 3.0 ha for detection to be reliable. Sharma (2000) investigated identifying sub-pixel s ize MPB-at tacked stands using Tasse led C a p Transformations (brightness, greenness and wetness) generated from two Landsat-7 ETM+ scenes , one acquired in August 1999, and the other September 1999, in the Vanderhoof Forest District, B C . It was observed that Tasse led C a p indices for infestations of more than about 30 attacked trees per site varied in a relatively narrow range, compared to those from less than 30 attacked trees per site. It was also found that differences between mean brightness, greenness, and wetness of M P B attacked stands and healthy lodgepole pine stands were statistically significant. A linear relationship was observed between the number of attacked trees per pixel and identification accuracy. Franklin ef al. (2003) tested the feasibility of M P B red attack detection for infestation sites ranging from >10 to 50 red-attacked trees per site, in the Fort St. J a m e s Forest District, B C , using a maximum likelihood classifier on single-date Landsat-7 ETM+ data. To reduce training area variability, training areas were stratified by spec ies and age prior to signature generation using the 14 provincial forest inventory GIS database and logical decision rules (e.g., >60 years age and >40 percent lodgepole pine). The overall identification accuracy reported was 72.3%, with 73.3% and 71.1% for red-attacked and unattacked sites, respectively. As a continuation of this study, Skakun ef al. (2003) evaluated the sensitivity of a multi-temporal Landsat-7 ETM+-based Enhanced Wetness Difference Index (EWDI) to detect two levels (10- 29 trees and 30-50 trees) of red attack. The EWDI (Franklin ef al., 2001) was designed to improve the visual identification of canopy changes over time. The overall classification accuracy based on a one-year (1999-2000) as well as a two-year (1999-2001) wetness difference was nearly constant at 74%. Murtha ef al. (2000) reported a MPB identification accuracy of 72% for MPB attacked stands in a study area covered by 19 MOF Forest cover map sheets near Vanderhoof, British Columbia. Stand level MPB attack detection is adequate for planning purposes, and comparable with the existing MPB susceptibility and risk rating systems (Shore and Safranyik, 1992; Shore et al., 2000). However, for planning effective MPB control measures, forest managers require this information at a sub-stand level or small tree group level (pixel level). 1.4.4 Summary Aerial sketch mapping and aerial photography are the most commonly used tools to map MPB infestations at the red-attacked stage. MPB-attacked trees reach the red-attacked stage usually one year after the attack. From a harvest planning and bark beetle management perspective, it is essential that MPB infestations are detected at an earlier date. The use of aerial data has been influenced by the need to detect MPB attack at a tree level. Methodologies developed during these studies, with few exceptions, have not been tested over large areas and their potential applicability remains largely unexplored. In addition, such data have higher data acquisition and processing costs. Methodologies developed from satellite-based remote sensing data, if successful, may provide a cost-effective solution applicable to larger areas. Unlike aerial data where an individual tree can be seen, MPB detection using 30 m spatial resolution satellite data remains at the research and demonstration stage due to the inferential nature of the information. Its use is further complicated by the inherent spectral variability in MPB-attacked stands caused by the co-existence of various attack stages, variable rates of infestations, and the confounding influences of several diseases and pests. Moreover, the effects of ecological diversity on spectral response is either poorly understood or unknown (Cihlar ef al., 2003). It is necessary to study the effects of ecological variability on lodgepole pine's spectral signatures and its implications on the large area applicability of satellite-based MPB attack detection methods. 15 There are two major issues in satell i te-based M P B attack detection: i) mixed pixels due to the predominantly sub-pixel s ize of M P B infestations; and ii) lack of information on the nature of the spectral response from MPB-at tacked and unattacked stands of lodgepole pine at the landscape level. A forested landscape is characterized by the co-occurrence of a number of tree species, growing both as pure and mixed stands, with a range of ages, heights, density c lasses , site variability, and incidences of insects and pests. Similar complexit ies in identifying spruce budworm defoliated stands were reported by Leckie (1987). Discrimination of spectral response from each of these factors may not be feasible using broad-band spectral bands, such as those from Landsat, as plant reflectance is governed by a small number of physical parameters such as chlorophyll content, leaf structure and leaf water content (Leckie, 1987; Jacquemoud and Baret, 1990; Pr ice, 1994). 1.5 OBJECTIVES The overall goal of the research descr ibed in this thesis is to develop a satel l i te-based M P B early detection method that can be used for large area mapping and monitoring. Issues involved in satell i te-based early detection and mapping of MPB-at tacked mature lodgepole stands using 30 m spatial resolution Landsat-7 ETM+ and EO-1 Hyperion hyperspectral data are studied. The specif ic objectives are: 1. to detect M P B attack, specifically: i) identify MPB-at tacked stands using Landsat ETM+ data at an earlier date than conventional surveys; ii) a s s e s s the accuracy of Landsat ETM+ identification of MPB-at tacked stands, and iii) assess the effects of confounding factors on M P B attack detection; 2. to a s s e s s variability in the spectral signatures of tree spec ies (lodgepole pine, Douglas-fir and spruce) as a function of age, site-index, and B E C subzones/var iants using Landsat ETM+ data; and 3. to determine the key hyperspectral bands for conifer spec ies identification and M P B attack detection using EO-1 Hyperion hyperspectral data. The study was conducted in the former Car iboo Forest Reg ion 1 , in central British Columbia (Figure 1.1). Three test areas were selected, one for each of the principle objectives. A summary of the data used to meet these objectives is given in Table 1.5. 1 This area is now part of the Kamloops Forest Region of the B C Ministry of Forests and Range. 16 1.6 THESIS STRUCTURE The thesis is arranged in five chapters. Background information about lodgepole pine, M P B , and a review of pertinent remote sensing literature were presented in this chapter. Early M P B attack detection using Landsat ETM+ imagery is descr ibed in Chapter 2 (Objective 1). Chapter 3 provides information on the effects of ecological variability on the spectral signatures of certain conifer spec ies (Objective 2). This provides the basis for any large area remote sensing application (e.g. inventory, M P B attack detection). Results for this objective affect the data analysis approach, location, and amount of ground data required for M P B attack detection. The spectral regions which are best suited for identifying different conifer species and current MPB-at tacked stands are covered in Chapter 4 (Objective 3). Chapter 5 consists of overall conclusions and recommendat ions for further research. Chapters 2 to 4 are intended to be published as separate papers; therefore, some information is replicated among these chapters. 17 Figure 1.1: Location and relative extent of Test Areas A, B, and C Table 1.5: Data sources used for various study objectives. Objectives Early Detection of MPB Spectral Variability in conifer species. Hyperspectral band combinations r Conifer Species Early MPB Detection Study area sites Site A (234,000 ha) Site B (1.8 million ha) Site C (60,000 ha) Study area size (Ha) (covering stand age c lasses between 60 to >140 years; 3 site index c lasses and 9 B E C Subzone/variants) (42 samples) (145 samples) Primary remote sensing data Landsat-7 ETM+ (12 July 1999, 15 August 2000) Landsat-7 ETM+ (12 July 1999, 19 July 1999) EO-1 Hyperion (30 August 2002) EO-1 Hyperion (30 August 2002) Ancillary data Satell ite I K O N O S (13 August 2001) - - -Aerial • Helicopter overview survey (June 2001) Helicopter Survey (August 2002) Digital aerial photographs (August/September 2003) Aerial Color infrared photo Contracted but could not be acquired - - -Aerial Sketch mapping (BC M o F Health Survey) 2000, 2001 - - -TRIM maps Stream and Road network Stream and Road network Stream and Road network Stream and Road network Field data March 2001, June 2001 October 2001 August 2002 - March 2003 1.7 REFERENCES Ahern, F. J . , 1988. The effects of bark beetle stress on the foliar spectral reflectance of lodgepole pine, International Journal of Remote Sensing, 9:1451-1468. Ahern, F. J . , and P. D. Archibald, 1986. Thematic Mapper information about Canad ian Forests: early results from across the country, In: Proceedings of the 10th Canadian Symposium on Remote Sensing, Ottawa, pp 683-697. A m m a n , G . D., and L. Safranyik, 1984. Insects of lodgepole pine: Impact and control, In: Proceedings of the Symposium on Lodgepole pine: the species and its management, Spokane, Washington. Bartlett, D. S . , R. W. Hardisky, M. A . Johnson, V. K lemas, and J . M. Hartman, 1988. Continental scale variability in vegetation reflectance and its relationship to canopy morphology, International Journal of Remote Sensing, 9:1223-1241. B C M O F , 1995. Bark Beetle Management Guide Book, U R L : http://www.for.gov.bc.ca/tasb/legsregs/fpc/fpcguide/beetle/betletoc.htm, (last date accessed : 2 Feb 2 0 0 4 ) . B C M O F , 2001. Gener ic Forest Health Surveys Guidebook. 2nd E d . Forest Pract ices C o d e of British Columbia Guidebook, Forest Pract ices Branch, Ministry of Forests, Victoria, B C . B C M O F , 2000. Annual Report of the Ministry of Forests 1999/2000, http://www.for.gov.bc.ca/hfd/pubs/docs/mr/annual/ar 1999-00/, British Columbia Ministry of Forests, (last date accessed : 11 December 2006). B C M O F , 2001. Gener ic Forest Health Surveys Guidebook. 2nd E d . Forest Pract ices C o d e of British Columbia Guidebook, Forest Pract ices Branch, Ministry of Forests, Victoria, B C . B C M O F , 2003a. Timber supply and the mountain pine beetle infestation in British Columbia, Forest Analys is Branch, Ministry of Forests, Forest Analys is Branch, Victoria, B .C . B C M O F , 2003b. Annual Report of the Ministry of Forests 2002/03, http://www.for.gov.bc.ca/hfd/pubs/docs/mr/annual/ar 2002-03/tables/C03.pdf, British Columbia Ministry of Forests, (last date accessed : 11 December 2006). B C M O F , 2003c. Provincial Bark Beetle Management Technical Implementation Guidel ines, http://www.for.gov.bc.ca/hfp/forsite/fhdata/ProvBBStrategy.pdf, B C Ministry of Forests, (last date accessed : 5 April 2004). Bentz, B. J . , and D. Endreson, 2004. Evaluating satellite imagery for estimating mountain pine beetle-caused lodgepole pine mortality: current status, In: Proceedings of the Mountain pine beetle symposium, Oct 30-31, 2003, Kelowna, British Columbia, pp 154-163. Berryman, A . A. , 1972. Resis tance of conifers to invasion by bark beetle-fungus associat ions, Bioscience, 22:598-602. Berryman, A . A . , 1976. Theoretical explanation of mountain pine beetle dynamics in lodgepole pine forests, Environmental Entomology, 5:1225-1233. Brockhaus, J . A . , H. M. Cheshire, S . Khorram, and W. H. Klein, 1985. The use of Landsat M S S digital data to detect mortality caused by the mountain pine beetle, In: Proceedings of the Pecora 10, Remote sensing in forestry and range resource management, pp 220-227. 20 Carrol l , A . , and D. Linton, 2002. Managing mountain pine beetle populations in British Columbia, Forest Health and Biodiversity News, Canad ian Forest Serv ice, Spring 2002, Vol:6(1). Carter, G . A. , 1994. Ratios of leaf reflectance in narrow wavebands as indicators of plant stress, International Journal of Remote Sensing, 15:697-703. Christ iansen, E. , R. H. Waring, and A . A . Berryman, 1987. Res is tance of conifers to bark beetle attack: Searching for general relationships, Forest Ecology and Management, 22:89-106. Cihlar, J . , B. Guindon, J . Beaubien, R. Latifovic, D. Peddle, M. Wulder, R. Fernandes, and J . Kerr, 2003. From need to product: a methodology for completing a land cover map of C a n a d a with Landsat data, Canadian Journal of Remote Sensing, 29:171-186. Cihlar, J . , L. St-Laurent, and J . A . Dyer, 1991. Relation between the Normal ized Difference Vegetation Index and ecological variables, Remote Sensing of Environment, 35:279-298. C O F I , 2003. Mountain pine bark beetle update, http://www.mountainpinebeetle.com/article 2003 dec15.html, Counci l of Forest Industries, British Columbia, (last date accessed : 15 January 2004). Dadykin, V . P., and V. P. Bedenko, 1960. Concerning the geographic variability of optical properties in plant leaves, Forestry Laboratory, Academy of Sc iences U S S R , Translated from : Doklady Akad . N a u k S S S R Vol .130, pp 674-677. Datt, B., T. R. McVicar , T. G . Van Niel, D. L. B. Jupp, and J . S . Pear lman, 2003. Preprocessing EO-1 Hyperion hyperspectral data to support the application of agricultural indices, IEEE Transactions of Geoscience and Remote Sensing, 41:1246-1259. Davidson, D., and S . Lataille, 2000. Assessment of possible spectral indicators for green stage mountain pine beetle attack using airborne C A S I imagery at Tyee Lake. Prepared for the B .C . Ministry of Forests, I T R E S Research Ltd., Calgary, Alberta. Franklin, S . E., M. B. Lavigne, L. M. Moska l , M. A . Wulder, and T. M. McCaffrey, 2001. Interpretation of forest harvest conditions in New Brunswick using Landsat T M enhanced wetness difference imagery, Canadian Journal of Remote Sensing, 27 :118-128 . Franklin, S . E., M. A . Wulder, R. S . Skakun, and A. L. Carrol l , 2003. Mountain pine beetle red-attack forest damage classification using stratified Landsat T M data in British Columbia, C a n a d a , Photogrammetric Engineering & Remote Sensing, 69:283-288. Gimbarzevsky, P., 1984. Remote sensing in forest damage detection and appraisal - selected annotated bibliography, Information Report B C - X - 2 5 3 , Canad ian Forestry Service, Victoria. Gimbarzevsky, P., A . F. Dawson, and G . A . Van Sickle, 1992. Assessmen t of aerial photograph and multi-spectral scanner imagery for measuring mountain pine beetle damage. Pacif ic Forestry Centre, Information Report B C - X - 3 3 3 , Victoria. Gi te lson, A . A . , M. N. Merzlyak, and H. K. Lichtenthaler, 1996. Detection of red edge position and chlorophyll content by reflectance measurements near 700nm, Journal of Plant Physiology, 148:501-508. Goyer , R. A . , M. R. Wagner, and T. D. Schowalter, 1998. Current and proposed technologies for bark beetle management, Journal of Forestry, December:29-33. 21 Harris, J . W. E., A . F. Dawson, and D. Goodenough, 1978. Evaluation of Landsat data for forest pest detection and damage appraisal surveys in British Columbia. Pub. Rep . B C - X -182, Canad ian Forest Serv ice, Pacif ic Forest Research Centre, Victoria, B .C . Heath, J . , 2001. The detection of mountain pine beetle green attacked lodgepole pine using Compact Airborne Spectrographic Imager (CASI) data, M.Sc. dissertation, Faculty of Forestry, University of British Columbia, Vancouver, B.C. ,77 p. Heller, R. C , 1968. Previsual Detection of Ponderosa Pine Trees Dying from Bark Beetle Attack, In: Proceedings of the 5th Symposium on Remote Sensing of Environment, Wil low Run Lab, University of Michigan, Ann Arbor, Michigan, pp 387-434. Hobbs, A . J . , 1983. Effects of air photo scale on early detection of mountain pine beetle, M.Sc. dissertation, Faculty of Forestry, University of British Columbia, Vancouver , B .C. , 137 p. Hobbs, A . J . , and P. A . Murtha, 1984. Visual interpretation of four sca les of aerial photography for early detection of mountain pine beetle infestations, In: Proceedings of the RNRF symposium on the applications of remote sensing to resources management, Fal ls Church, Virginia, Amer ican Society of Photogrammetry, pp 433-444. Jacquemoud, S . , and F. Baret, 1990. P R O S P E C T : a model of leaf optical properties spectra, Remote Sensing of Environment, 34:75-91. Kalensky, Z. , and D. Wi lson, 1975. Spectral signatures of forest trees, In: Proceedings of the 3rd Canadian Symposium on Remote Sensing, Ottawa, Canad ian Remote Sens ing Society, pp 155-171. Kim, J . J . , E. A . Al len, L. M. Humble, and C . Breuil , 2005. Ophiostomatoid and basidiomycetous fungi associated with green, red, and grey lodgepole pines after mountain pine beetle (Dendroctonus ponderosae) infestation, Canadian Journal of Forest Research, 35:274-284. Klein, W. H., 1973. Beetle-kil led pine estimates, Photogrammetric Engineering, 39:385-388. Klein, W. H., 1982. Estimating bark beetle-killed lodgepole pine with high altitude panoramic photography, Photogrammetric Engineering & Remote Sensing, 48:733-737. Kneppeck, I. D., and F. J . Ahern, 1989. A compar ison of images from a pushbroom scanner with normal color aerial photography for detecting scattered recent conifer mortality, Photogrammetric Engineering and Remote Sensing, 55:333-337. Leckie, D. G . , 1987. Factors affecting the assessment of forest damage using airborne multispectral scanner data, Photogrammetric Engineering & Remote Sensing, 53(12): 1665-1674. Leckie, D. G . , D. P. Ostaff, and J . P. Peaker , 1983. Methodology for measuring losses due to forest insects and d iseases. Report of the Working Group of the Forest Insect and Disease Survey/Forestry Statistics and Systems Branch. Lotan, J . E., and W. B. Critchfield, 1990. Pinus contorta Dougl. ex Loud, Silvics of North America, Vol. 1 Conifers, Agriculture Handbook 654. Washington, DC, (R. M. Burns and B. H. Honkala, Editors), U S D A , pp. 302-315. M c L e a n , J . A. , B. J . V a n der Kamp, and A. L. Behennah, 2005. Forest Pest Management, Forestry Handbook for British Columbia, (S. B. Watts and L. Tol land, Editors), The Forestry Undergraduate Society, Faculty of Forestry, University of British Columbia, Vancouver . 22 Miller, R. H., H. S. Whitney, and A. A. Berryman, 1986. Effects of induced translocation stress and bark beetle attack (Dendroctonus ponderosae) on heat pulse velocity and the dynamic wound response of lodgepole pine (Pinus contorta var. latifolia), Canadian Journal of Botany, 64:2669-2674. Mitchell, R. G., R. H. Waring, and G. B. Pitman, 1983. Thinning lodgepole pine increases tree vigor and resistance to mountain pine beetle, Forest Science, 29:204-211. Murtha, P. A., 1978. Remote sensing and vegetation damage: A theory for detection and assessment, Photogrammetric Engineering & Remote Sensing, 44:1147-1158. Murtha, P. A., 1985. Interpretation of Large-scale Colour Infrared Photographs for Bark Beetle Incipient Attack Detection, In: Proceedings of the PECORA 10 Symposium, Falls Church, VA, USA, American Society for Photogrammetry and Remote Sensing, pp 209-219. Murtha, P. A., and R. Wiart, 1987. P. contorta-based digital image analysis for mountain pine beetle green attack: Preliminary results, Canadian Journal of Remote Sensing, 13:92-95. Murtha, P. A., and R. J . Wiart, 1989a. Cluster analysis of pine crown foliage patterns aid identification of mountain pine beetle current-attack, Photogrammetric Engineering & Remote Sensing, 55:83-96. Murtha, P. A., and R. J . Wiart, 1989b. PC-Based digital analysis of mountain pine beetle current-attacked and non-attacked lodgepole pine, Canadian Journal of Remote Sensing, 15:70-76. Murtha, P. A., Z. Bortolot, and J . Thurston, 2000. A Landsat TM spectral Unmixing mountain pine beetle attack fraction map in the Vanderhoof Forest District, British Columbia, In: Proceedings of the 22nd Canadian Symposium on Remote Sensing, Victoria, (Canadian Aeronautics and Space Institute, Ottawa, Ontario), unpaginated CD-ROM. Nebeker, T. E., J . D. Hodges, and C. A. Blanche, 1993. Host response to bark beetle and pathogen colonization, Beetle-pathogen interactions in conifer forests, (T. D. Schowalter and G. M. Filip, Editors), Academic Press, New York, pp. 157-173. Niemann, K. O., and F. Visintini, 2004. Assessment of potential for remote sensing detection of bark beetle-infested areas during green attack: a literature review, Mountain Pine Beetle Initiative Working Paper 2005-2. Natural Resources Canada, Canadian Forest Service, Victoria. Paine, T. D., K. F. Raffa, and T. C. Harrington, 1997. Interactions among scolytid bark beetles, their associated fungi, and live host conifers, Annual Review of Entomology, 42:179-206. Pojar, J . , D. Meidinger, and K. Klinka, 1991. Chapter 2: Concepts, Ecosystems of British Columbia, (D. Meidinger and J . Pojar, Editors), British Columbia Ministry of Forests, Victoria, B.C. Price, J . C , 1994. How unique are spectral signatures?, Remote Sensing of Environment, 49:181-186. Puritch, G. S., 1981. Nonvisual remote sensing of trees affected by stress: A review, Forestry Technical Report BC-X-30, Canadian Forestry Service, Victoria. Raffa, K. F., and A. A. Berryman, 1983. The role of host plant resistance in the colonization behavior and ecology of bark beetles (Coleoptera: Scolytidae), Ecological Monographs, 53:27-49. 23 Raffa, K. F., W. P. Thomas, and M. S . Scott, 1993. Strategies and mechanisms of host colonization by bark beetles, Beetle-pathogen Interactions in Conifer Forests, (T. D. Schowalter and G . M. Filip, Editors), Academic Press , New York, pp. 103-128. Ramsey , R. D., A . Falconer, and J . R. Jensen , 1995. The relationship between N O A A - A V H R R NDVI and Ecoregions in Utah, Remote Sensing of Environment, 53:188-198. Rencz , A . N., and S . Nemeth, 1985. Detection of mountain pine beetle infestation using Landsat and simulated thematic mapper data, Canadian Journal of Remote Sensing, 11:50-58. Roberts, A . , S . Dragicevic, J . Northrup, S. Wolf, L. Y . , and C . Coburn, 2003. Mountain pine beetle detection and monitoring: Remote sensing evaluations (Report for the B C Ministry of Forests), S imon Fraser University, Vancouver , B .C . Roberts, A. , J . Northrup, and R. Richard, 2005. Mountain pine beetle detection and monitoring: replication trials for early detection, In: Proceedings of the Third International Workshop on the Analysis of Multi-Temporal Remote Sensing Images, Biloxi, Mississ ippi , U S , (unpaginated C D - R O M . Roller, N., 2000. Intermediate multispectral satellite sensors, Journal of Forestry, 98:32-35. Runesson , U. T., 1991. Considerations for Early Remote Detection of Mountain Pine Beetle in Green-Foliaged Lodgepole Pine, P h D dissertation, Faculty of Forestry, University of British Columbia, Vancouver, B.C. ,237 p. Safranyik, L., and A . Carrol l , 2006. The biology and epidemiology of the mountain pine beetle in lodgepole pine forests, The Mountain Pine Beetle A Synthesis of Biology, Management, and Impacts on Lodgepole Pine, (L. Safranyik and B. Wi lson, Editors), Natural Resources Canada , Victoria, pp. 3-66. Safranyik, L., D. M. Shrimpton, and H. S . Whitney, 1974. Management of lodgepole pine to reduce losses from the mountain pine beetle, Forestry Technical Report 1, Government of Canada , Department of the Environment, Canad ian Forest Serv ice, Pacif ic Forest Research Centre, Victoria, B C . Schmidt, C . W., and R. R. Alexander, 1984. Strategies for managing lodgepole pine, In: Proceedings of the Symposium on Lodgepole pine: the species and its management, May 8-10, Spokane, Washington, pp 201-210. Sharma, R., 2000. Detection of mountain pine beetle infestations using Landsat TM Tasseled Cap Transformations, M .Sc dissertation, Faculty of Forestry, University of British Columbia, Vancouver , B.C.,57 p. Shrimpton, D. M., and A. J . Thomson, 1983. Growth characteristics of lodgepole pine associated with the start of mountain pine beetle outbreaks, Canadian Journal of Forest Research, 13:137-144. Sirois, J . , and F. J . Ahern, 1988. An investigation of S P O T H R V data for detecting recent mountain pine beetle mortality, Canadian Journal of Remote Sensing, 14:104-108. Skakun, R. S . , M. A . Wulder, and S. E. Franklin, 2003. Sensitivity of the Thematic Mapper Enhanced Wetness Difference Index (EWDI) to detect mountain pine beetle red attack damage, Remote Sensing of Environment, 86:433-443. 24 Smith, T., K. Whitehead, and A. Norquay, 2002. Mountain pine beetle green attack detection project -2001. Project Report, Earth Imaging Technologies Inc., Sa lmon Arms, B C . Unger, L , 1993. Mountain Pine Beetle, Forest Insect and Disease Survey, Forestry C a n a d a , Forest Pest Leaflet No.76. Unger, S . L., J . S . Pear lman, J . A . Mendenhal l , and D. Reuter, 2003. Overv iew of the Earth Observ ing One (EO-1) Miss ion, IEEE Transactions of Geoscience and Remote Sensing, 41:1149-1159. V a n der Kamp, B. J . , and F. G . Hawksworth, 1984. Damage and control of the major d iseases of lodgepole pine, In: Proceedings of the Lodgepole pine: the species and its management, Spokane, Washington, pp 125-131. Warner, T. A . , and M. C . Shank, 1997. Spatial auto-correlation analysis of hyperspectral imagery for feature selection, Remote Sensing of Environment, 60:58-70. White, J . C , M. A . Wulder, D. Brooks, R. Re ich , and R. D. Wheate, 2004. Mapping mountain pine beetle infestation with high spatial resolution satellite imagery, The Forestry Chronicle, 80:743-745. Wood , D. L., 1972. Select ion and colonization of ponderosa pine by bark beetles, In Proceedings of the Symposium of Royal Entomological Society, London, pp 110-117. Wulder, M. A . , and C. C. Dymond, 2004. Remote sensing in the survey of mountain pine beetle impacts: Rev iew and recommendat ions, MPBI Report, Canad ian Forest Serv ice, Natural Resources C a n a d a , Victoria. Wulder, M. A . , C . C . Dymond, J . C . White, D. G . Leckie, and A. L. Carrol l , 2006. Surveying mountain pine beetle damage of forests: A review of remote sensing opportunities, Forest Ecology and Management, 221:27-41. 25 2.0 EARLY DETECTION AND MAPPING OF MOUNTAIN PINE BEETLE INFESTATIONS USING LANDSAT-7 ETM+ 2.1 INTRODUCTION The most effective approach to reducing losses due to M P B is prompt detection and immediate suppression of small infestations as they develop. Suppress ion is the most aggressive of the M P B management strategies. It includes direct control actions (e.g., selective harvesting of the attacked trees) so that infestations do not get bigger, and are managed without large extra expenditures ( B C M O F , 1995). Information on MPB-at tacked stands at an early date is very useful in planning suppression activities. Conventional aerial sketch mapping and aerial photographic surveys are carried out at the red attack stage, which is one year after successfu l M P B attack. Therefore, information on the extent and severity of M P B infested trees/stands is available only after 12-15 months following an attack. There is a growing need for information on M P B infestations at an earlier date than is currently available from conventional survey methods. This information need, combined with the presence of a large pool of mature lodgepole pine susceptible to M P B attack in B C , has driven related remote sensing research. Many studies have been carried out to detect M P B infestations using remote sensing data acquired from ground-based, aerial, and satellite-borne sensors. Salient features of some of these studies are summarized in Table 2.1. A variety of data analysis techniques have been used for the identification of MPB-at tacked stands. These include visual interpretation, digital image analysis techniques such as supervised and unsupervised classif ications, spectral unmixing, statistical techniques such as discriminant analysis, and proprietary algorithms. The major focus of most studies has been M P B attack detection at the tree level. Therefore, high spatial resolution remote sensing data from airborne sensors have been the primary data source in the majority of studies. However, aerial data have high acquisition and processing costs. A lso, not all stakeholders involved in M P B management need attack information at this level of resolution. Information at the stand-level (group of trees) may suffice to direct crews on the ground for more detailed surveys required for harvest planning. Consequent ly, a number of attempts have been made using medium spatial resolution remote sensing data from satellites such as Landsat. Satel l i te-based remote sensing data are more cost-effective than airborne data. For example, at the onset of this study in 2001, Landsat-7 ETM+ was the most economical ($900 C A D / s c e n e , $0.03 C A D per km 2) source of 30 m spatial resolution, multispectral remote sensing data; digital aerial photography cost $0.21 C A D per ha. Note: A version of this chapter will be submitted for publication. (Suggested Journal : Canadian Journal of Forest Research) 26 Table 2.1: Summary of selected remote sensing studies for M P B attack detection. t : ™ r c f f i l ^ ; . r t : : f l R l S c a i ^ v - t ' i l l t u d y i f i i i ; • : U K - Se lS ip r "M. ' ^ M e t h o d 3 » \ ' ' ' ' I t ? ' . R p e r e n c e l i r . ~ Spectral reflectance pattern Needle level N A Ground based Spectrophotometer (Heller, 1968) R A and gray attack detection Tree level - Aerial photographs Visual interpretation (Klein, 1973) R A detection Group of trees - Landsat M S S Visual interpretation (Harris e ra / . , 1978) R A detection Tree level 300 miles* Aerial Color-IR photographs Visual interpretation (Klein, 1982) Early G A detection Tree level 32-66 trees Aerial colour-IR photographs Visual interpretation (Hobbs, 1983) R A detection Tree level 40 km* Landsat M S S and simulated TM Digital classification (Rencz and Nemeth, 1985) R A detection Tree level - MEIS Visual interpretation (Kneppeck and Ahern , 1988) R A detection Group of trees 2-3 ha S P O T Visual interpretation (Sirois and Ahern, 1988) Spectral reflectance pattern Needle level 6 trees Ground based Spectrophotometer (Ahern, 1988) G A detection Tree level 68 trees Aerial colour-IR photographs Digital techniques (Murtha and Wiart, 1989) R A and G A detection Needle level 75 trees Ground based Spectrophotometer (Runesson, 1991) G A detection Tree level 26 trees CASI Proprietary algorithm (Davidson and Lataille, 2000) " M P B attack" detection Group of trees 44 training and 568 validation sites Landsat-7 ETM+ Tasse led C a p Transformations (Sharma, 2000) " M P B attack" fraction Group of trees 2813 km* Landsat-5 T M Spectral unmixing (Murtha et al., 2000) G A detection Tree level 16-20 trees CASI Discriminant analysis (Heath, 2001) G A detection Tree level 10-31 G E R 3715 Proprietary algorithm (Smith e ra / . , 2002) R A and G A detection Tree level 614 trees Multispectral aerial photographs Digital classification (Roberts ef al., 2003) R A detection Group of trees 5070 km* Landsat-7 ETM+ Sup. Classif ication (Franklin ef al., 2003) R A detection Group of trees 100 training and 120 validation sites Landsat-7 ETM+ Discriminant analysis (Skakun ef al., 2003) R A : Red attack; G A : Green Attack Varying degrees of success in M P B red attack detection have been achieved using satellite data. Harris ef al. (1978), based on Landsat M S S , reported an identification accuracy of 25%. Franklin ef al. (2003), using a maximum likelihood classif ier on single date Landsat-7 ETM+ data, achieved an identification accuracy of 73 .3% and 71 .1% for red attack and unattacked stands, respectively. Skakun et al. (2003), using the Enhanced Wetness Difference Index, reported an overall accuracy of 74%. These studies demonstrated the potential of satel l i te-based remote sensing in detecting large concentrated groups of red-attacked trees. There are several major issues affecting the accuracy of satell i te-based mapping of M P B -attacked stands, including: i) the predominantly sub-pixel s ize and scattered distribution of M P B infestations; ii) the presence of different M P B attack stages within the same stand; iii) the presence of confounding d iseases and pests; and iv) the presence of mixed-species stands. In medium spatial resolution data, such as that obtained from the Landsat-7 ETM+ sensor with 30 m spatial resolution, the sub-pixel nature of M P B infestations and the presence of different attack stages within the same stand result in the formation of mixed pixels. The need for information on MPB-at tacked stands before the infestations reach the red attack stage persists and was identified as one of the high priority research area by B C Ministry of Forests ( B C M O F , 2005). A lso, the capabil it ies and limitations of satel l i te-based M P B attack assessment over large areas need to be assessed . This necessitates conducting prospective mapping of M P B -attacked stands. If the results of such mapping are satisfactory, it would be possible to provide information on MPB-at tacked stands at an earlier date than is available from conventional aerial survey methods. However, to achieve this, a number of studies in different areas representing diversity in M P B attack size and severity are needed. In this chapter, results obtained from a study carried out in central B C using Landsat-7 ETM+ data are presented. This study was conducted to map MPB-at tacked stands in 2000, before information on these infestations became available from conventional surveys. In 2000, M P B infestations in this area were scattered and still predominantly small in s ize (i.e., groups of only a few attacked trees). The objectives were: (1) to identify MPB-at tacked stands at an earlier date than conventional surveys; and (2) to characterize MPB-at tacked stands with respect to infestation s ize, attack stage and confounding factors such as other pests and d iseases. In this study, different attack stages were combined and called " M P B attack". A lso , the term "stand", which conventionally represents a polygon in the B C Ministry of Forests digital forest cover maps, has been used to represent a small group of trees, generally smaller in extent than a forest 28 inventory polygon. Early detection refers to M P B attack detection at a date before the commencement of conventional operational surveys. 2.2 STUDY AREA The study area (234,000 ha) is located in central B C near Wil l iams Lake (Figure 2.1). This area forms part of the Innovative Forestry Pract ices Agreement (IFPA) area of Lignum Ltd. 1 , a forest company. This area was selected because: i) it represented part of the outer edge of growing M P B infestations in the region at that time; ii) there were different forest cover types; iii) lodgepole pine grew under a variety of cl imatic/ecological conditions across the area; iv) there were both pure and mixed lodgepole pine stands; v) there were several other pest and d isease attacked stands; vi) it covered a portion of the operating areas of the Lignum Ltd., facilitating independent evaluation of the results; and vii) collateral data (e.g., digital forest cover maps, TR IM maps) for the area were available. The study area includes a wide range of vegetation types and growing conditions. Lodgepole pine, Douglas-fir, Engelmann spruce (Picea engelmannii Parry), and white spruce (Picea glauca), along with spruce c rosses 2 , are the common forest tree spec ies in the area. The area includes three Biogeocl imatic Ecosys tem Classif ication (BEC) subzones, the Interior Douglas-Fir Dry Cool Subzone (IDFdk), the Sub-Borea l Spruce Dry Warm Subzone (SBSdw) , and Sub-Borea l Pine and Spruce Moist Coo l Subzone ( S B P S m k ) . The distribution of the B E C subzones in the study area at the variant level is given in Figure 2.2a. These subzones differ in cl imax vegetation, temperature and rainfall characteristics (Klinka et al., 1991; Pojar ef al., 1991). In the S B P S zone, lodgepole pine is the dominant tree species, followed by spruce. Wetlands are common in this zone due to poor drainage patterns resulting from the subdued topography and fine textured soils (Steen and Coupe, 1997). Open to c losed forests of Douglas-fir, lodgepole pine, and spruce characterize the IDF zone. Spruce and lodgepole pine are the most common species present in the S B S zone. 2.2.1 Lodgepole Pine Distribution The study area is covered by 20 digital forest cover map sheets from the B C Ministry of Forests. These maps represent the forest inventory database in a geographic information system (GIS) environment. The inventory is used for management and planning purposes and gives the area, 1 Subsequent to the work described in this chapter, Lignum Ltd. was bought out by Riverside Forest Products, which was subsequently acquired by Tolko Industries Ltd. 2 The various spruce spec ies and their c rosses will simply be called "spruce" for the duration of this chapter. 29 British Columbia Figure 2.1: Location of study area. volume, and location of different forest species types at a stand level (Bonnor, 1982; Leckie and Gil l is, 1995). A detailed attribute table with information on various forest stand parameters (such as leading species, proportion of other species present, crown closure c lass, age c lass , and height class) is linked with the vector forest polygon coverage. Distribution of forest cover polygons with lodgepole pine (> 60 years age) in the study area is shown in Figure 2.2b. These polygons cover an area of 80,557 ha. The forest cover polygons have been categorized into five c lasses based on lodgepole pine content (Table 2.2). The majority of the lodgepole pine stands in the study area are mixed. Only about one-fourth are relatively pure or homogeneous; these are primarily concentrated in the S B P S zone. 30 F igu re 2.2: (A) Biogeocl imat ic Ecosys tem Classi f icat ion Subzone / Variants, (B) Lodgepole pine distribution, in the study area. Table 2.2: Area (ha) under forest cover polygons with lodgepole pine (60 years and above) spec ies content (Data Source: B C M O F Digital Forest Cover Maps) . B E C Subzones Lodgepole Pine species con ten t Total A r e a (Ha) Leading •ir-. species •„. 14:<20(%)|; ,21 -400ft) 41-60(%) 6j!h80(%JK & >80(%) IDFdk 7388 10017 2756 4 2883 23048 Douglas-fir S B P S m k 1414 2587 3701 471 12171 20344 Spruce and Lodgepole pine S B S d w 14225 8186 6873 773 7108 37165 Spruce Total 23027 20790 13330 1248 22162 80557 2.2.2 MPB Status and Distribution MPB-at tacked stands in the Car iboo Region increased from 32,275 ha in 1999, 38,473 ha in 2000 to 103,792 ha in 2001 (Figure 2.3). In 2000, the year of this study, the MPB-at tacked area in the Car iboo Forest Region was relatively small and the M P B infestations were small in s ize and scattered. Infestations rose sharply from 2001 onwards. The spatial distribution of the M P B infestations in the Car iboo Forest Region for 2000 and 2001, based on provincial health surveys, are given in Appendix 2.1 and Appendix 2.2, respectively. The study area was outside of the major concentration of M P B attack in the region during these years. The M P B distribution in the study area for 2000 and 2001, based on provincial aerial overview surveys, is shown in Figure 2.4. During 2000, the majority of the M P B infestations were at the current attack stage (i.e., the foliage of MPB-at tacked trees were still green and could not have been detected by the conventional aerial surveys). The majority of the M P B infestations recorded in 2001 were spot infestations, indicating that M P B attacked stands were small in size. Smal l infestations up to 50 trees are delineated as spot infestations on the overview survey maps ( B C M O F , 2001). 2.3 MATERIALS AND METHODS Multispectral Landsat-7 ETM+ data (path/row: 47/24; acquisition dates: 12 July 1999 and 15 August 2000; spatial resolution: 30 m) and I K O N O S multispectral data (acquisition date: 13 August 2001, spatial resolution: 4 m) were used in this study (Figure 2.5). The spectral and spatial resolutions of these sensors are given in Appendix 2.3. Lodgepole pine stands, attacked by M P B in July 2000, were at the current attack stage at the time of Landsat data acquisition in August 2000, unattacked in 1999 and at a red attack stage in 2001, when the I K O N O S data were acquired. 32 Increase in MPB affacted area across various forest regions in BC 800000 700000 600000 _ 500000 n X 400000 V < 300000 200000 100000 0 1998 1999 2000 2001 2002 Year Cariboo Forest Region —•—Vancouver Forest Region Kamloops Forest Region —K— Prince Rupert Forest Region Prince George Forest Region —•— Nelson Forest Region F igu re 2.3: MPB-af fected areas in the Car iboo Forest Region during 1998 - 2002 (Data source: B C M O F Annual Forest Health Surveys ( B C M O F , 2006)). 2.3.1 Data Analysis Data analysis consisted of three main components: i) identifying mature lodgepole pine stands; ii) identifying those lodgepole pine stands showing increased stress during the pre- and post-M P B attack phase; and iii) identifying MPB-at tacked lodgepole pine stands. Identifying Lodgepole Pine Stands A polygon in the vector coverage of the forest inventory database represents a forest stand. However, the spatial distribution of various species within in a forest cover polygon is not known. These inventory maps have been used previously to select forest stands with lodgepole pine content using a set of decision rules based on species content proportion and stand age (e.g., Murtha et al., 2000; Franklin ef al., 2003; Skakun ef al., 2003). However, this approach has certain limitations. For example, in the study area, forest cover polygons with lodgepole pine content cover 80,557 ha. In these polygons, the lodgepole pine content proportion varies from 20-80%. If the 33 MOUNTAIN PINE B E E T L E RED ATTACK F igu re 2.4: Red attack distribution in the study area in 2000 and 2001 (Source: B C M O F Forest Health Surveys). Infestation severity c lasses are based on percent of red attacked trees within the del ineated polygons (Light: 1-10%; Moderate: 11-29%; Severe : >30%). 122 WW 121°450"W F igu re 2.5: Landsat-7 ETM+ (August, 2000) Band-4 image of study area (2340 km 2 ) . The I K O N O S image covers a smal l (100 km 2 ) area in south-eastern part of the study area. The partial extent of the IFPA (Innovative Forest Pract ices Agreement) is a lso shown. content proportion threshold is kept too low (e.g., 20%), a large proportion of stands will get selected in the analysis set. O n the other hand, at a higher threshold (e.g., 40%) a significant proportion of stands containing lodgepole pine might be left out of analysis and, consequently, it may result in underestimation of M P B attack in the area. Landsat data (August 2000) were used to identify stands containing lodgepole pine in the study area. This had three advantages. First, Landsat images not only provide up-to-date information on forest types and extent of cover, but also on the distribution of lodgepole pine within forest cover polygons at a spatial resolution of 30 m, a much finer resolution than the forest inventory polygons, which can range anywhere between a few ha to tens of ha in s ize. Second , restricting analysis to only those stands / pixels that contain lodgepole pine may help minimize commiss ion errors due to other confounding pests and d iseases of forest vegetation. Third, it could be more time and cost efficient in 35 the future to deal with only lodgepole pine stands rather than the entire forested area for multi-temporal mapping of M P B infestations and for monitoring of changes in lodgepole pine cover resulting from harvesting and other disturbances. < Identifying MPB-Attacked Stands M P B introduces blue-staining fungi into the host tree. The growth of fungi in vascular t issues causes the disruption of normal plant water relations leading to severe water stress (DeAngel is et al., 1986; Nebeker et al., 1993). Most trees start exhibiting early signs of water stress three to four weeks after successful M P B attack (Amman, 1985). The prolonged water stress also results in loss of differentiation in needle age class (Murtha, 1978; Murtha and Wiart, 1987). These physiological changes lead to reduced vigor in MPB-at tacked lodgepole pine trees. Using pre- and pos t -MPB attack imagery, it may be possible to identify attacked stands through vegetation index-based change detection techniques. Lodgepole pine stands under a negative change mask are potential candidates where M P B attack might have taken place, hereafter cal led " M P B candidate lodgepole pine stands". Identifying such stands would further narrow the search for MPB-at tacked stands. Spectral unmixing processes provide information on the relative proportion of material of interest (end members) found within a pixel. The assumption behind linear mixture modeling is that pixel reflectance is the sum of reflectance for each cover type, weighted by their fractional presence within each pixel. The inputs to mixture models are endmember reflectance and multispectral images, and the output is a fraction image for each endmember giving its proportion in a pixel. Although spectral unmixing techniques were designed for use with hyperspectral data, Adams ef al. (1995) showed that they can work reasonably well with Landsat T M data when the number of endmembers is smal l . Spectral unmixing has been used for land use / cover mapping, spruce beetle (Dendroctonus rufipennis Kirby) attack detection in Utah, (Johnson ef al., 1997), jack pine budworm (Choristoneura pinus) defoliation in the Pine Barrens of Wiscons in (Radeloff ef al., 1999), MPB-at tack detection (Murtha ef al., 2000), and for detection of spec ies fraction (Huguenin ef al., 1997; Oki ef al., 2002). Due to the sub-pixel s ize of M P B infestations prevalent in 2000 in the study area, an ETM+ pixel may have varying proportions of different M P B attack stages as well as unattacked trees resulting in mixed pixels. In a heterogeneous landscape with mixed pixels, spectral unmixing provides a more realistic representation of the true nature of a cover type compared to the assignment of a single dominant c lass to every pixel. Therefore, spectral unmixing was considered to be better suited for identification of small MPB-at tacked stands than pixel-based algorithms such as a Maximum Likelihood Classif ier. 36 Verifying MPB-Attacked Stands Verification of MPB-at tacked stands was conducted to: i) to a s s e s s the identification accuracy; ii) to identify infestation size and M P B attack stages associated with the identified M P B -attacked stands; and iii) to identify other pests and d iseases confused with M P B attack. This focus was deliberate to meet the intended use of a M P B attack map, which is to get early information on the locations of MPB-at tacked stands where field crews could be directed for detailed surveys. With a cost of $700 per d i em 3 for a two-member field crew, the magnitude of commiss ion errors (false positives) have high cost implications in the operational use of the mapping results. In the context of accuracy assessment of M P B attack detection, it is very expensive to a s s e s s true omission errors using ground surveys due to the high costs involved in complete enumeration of attacked or unattacked trees on test sites covering the full range of M P B infestation s izes and severity levels present in an area. Therefore, estimates of omission errors, although important, are not made in conventional aerial surveys. It is assumed that the missed M P B infestations would be picked up while traversing between M P B hotspots during field surveys. However, a comparison of MPB-at tacked stands identified from remote sensing data with an aerial overview survey might provide some insight into the magnitude of omission errors. 2.3.2 Image Analysis The image analysis approach consisted of: i) pre-processing the satellite data for atmospheric correction, radiometric normalization and co-registration of the two-date Landsat data; ii) generating a lodgepole pine map of the study area from the Landsat ETM+ data acquired in August 2000; iii) identifying candidate lodgepole pine stands based on reduction in plant vigor using p re -MPB attack (July 1999) and pos t -MPB attack (August 2000) Landsat data; and iv) spectral unmixing of the August 2000 Landsat data under the candidate lodgepole pine stand mask to identify probable M P B -attacked stands. Data analysis was carried out using image analysis software (PCI Geomat ica V8.1) and GIS software (ArcView 3.1). The Landsat images were atmospherically corrected using the dark object subtraction method (Chavez, 1988). This technique is based on the premises that: i) atmospheric effects add a uniform offset to each image band; ii) all or nearly all near infra-red energy incident on clear deep water is absorbed; and iii) any reflectance over water represents path radiance. This method does not account for the multiplicative component of the atmospheric effects. Although this technique is one of the atmospheric correction methods, it is still widely used method (Price ef al, 1997; Song ef al., 2001) because of its simplicity and its requirement for little information beyond the image itself. 3 Alexander, J .A. , 2001, [Area Forester], Personal Communicat ion. 37 The Landsat (15 August 2000) and I K O N O S images were orthorectified using the nearest neighbor (NN) resampling algorithm using a digital elevation model (derived using 20 m contours) and B C ' s Terrain Resource Inventory Maps (TRIM). The July 1999 Landsat image was then registered to the orthorectified 2000 Landsat image. K-mean clustering was run on bands 1-5 and 7 of the Landsat (15 August 2000) digital data to successively separate vegetated areas from non-vegetated areas, forest vegetation from non-forest vegetation, and mature lodgepole pine-leading stands from other coniferous forest. At each of the three clustering stages, the resulting clusters were visually inspected and non-pine clusters were progressively excluded from the further iterative clustering. The B C Ministry of Forest 's digital forest cover maps were used as reference data for labeling various forest vegetation types during the classif ication. Local area foresters were involved in the digital analysis to identify the mature lodgepole pine leading stands. The distribution of Landsat-based mature lodgepole pine stands was visually compared with the forest cover polygons (on digital forest cover maps) containing lodgepole pine as one of the leading tree species. In addition, the identified lodgepole pine stands were also partially f ield-checked by local foresters. The lodgepole pine leading forest stands, thus identified, were found to be sufficiently accurate for this study. However, no quantitative estimate of identification accuracy was made. This process narrowed down the search for M P B attack from a total study area of 234,000 ha to 43,915 ha ( -19% of the total study area). The identified lodgepole pine stands were then analyzed using change detection techniques to identify stands which had shown increased stress between the dates of the two Landsat images. Tasse led cap greenness (Kauth and Thomas, 1976; Crist and Cicone, 1984) is an indicator of plant vigor and has been found to be strongly related to forest canopy changes (Coppin and Bauer, 1994). Therefore, greenness for the pre- and pos t -MPB attack phases was computed from the Landsat data for both dates. The tasseled cap transformation coefficients, developed for Landsat-7 ETM+ (Sharma, 2000), were used to generate the brightness, greenness and wetness image. These transformation coefficients are applicable to atmospherically corrected DN images. The tasseled cap transformation coefficients for Landsat-7 ETM+, applicable to at-satellite-reflectance images, were later developed by Huang ef al., (2002). A n image differencing technique (Singh, 1989) was used to generate a greenness difference image. The August 2000 greenness image was subtracted from the July 1999 greeness image). To account for variation in greenness due to differences in viewing geometry and environmental conditions on the image acquisition dates, pseudoinvariant features (Schott ef al., 1988) were identified in both Landsat images. The mean two-date greenness difference between these pseudoinvariant features was subtracted from the two date green difference image 38 and a normalized greenness difference image was generated. A number of empirical approaches for multi-date radiometric normalization have been proposed in the literature (Schott ef al., 1988; Joyce and O lsson , 1999; Peddle ef al., 2003). In this study the offset-subtraction approach was used. From the normalized difference image, a liberal change mask covering all the mature lodgepole leading stands, which had shown reduction in greenness in 2000, was generated. This was done to ensure that all lodgepole pine stands exhibiting reduced vigor were included in further analysis to identify MPB-at tacked stands. Mature lodgepole pine stands under this change mask were further analyzed using linear spectral unmixing. The main assumption in a linear mixing model is that a pixel consists of a small number of distinct endmembers (e.g., vegetation, soil , water), fewer than the number of spectral bands. The selection of endmembers is a critical component to successful application of mixture modeling. Basical ly, there are two ways to select endmembers: i) reflectance spectra measured in the field or laboratory (reference endmember); and ii) derivation of spectra directly from the multispectral image (image endmember). One of the main difficulties with the first method is that the reference signatures might not match the phenology at image acquisition and may need to be calibrated with respect to the image (Asner, 1998; Garc ia and Ustin, 2001). Image endmembers were used in this study, as the signatures derived from the image were applicable without calibration and represented the true conditions of the endmembers at the time of image acquisit ion. Spectral signatures from ETM+ bands 1-5 and band 7 for endmembers MPB-at tacked trees, unattacked trees, shadow, and soil were used in the analysis. Field data, acquired on the March 20, 2001 were used to generate training signatures for the MPB-at tacked and unattacked c lasses . The field sites were delineated as polygons and super imposed on the Landsat image. All the M P B -attacked training sites had predominantly green attacked trees (ranging from 22 to 88), along with some red attacked trees. Each pixel within a training site polygon was individually examined for its spectral response based on tasseled cap indices. Spectral responses from unattacked stands were also examined. Separabil i ty between attacked and unattacked stands was qualitatively assessed based on these indices. The values for brightness, greenness and wetness for MPB-unat tacked and attacked stands were 39.2 and 35.3, 26.9 and 20.7, and 54.7 and 53.4, respectively. This is similar to earlier observations (Sharma and Murtha, 2001) from a study in a nearby forest area in the Vanderhoof Forest District. Signatures for shadow and open areas were derived from the regions of topographic shadow in the image and homogeneous exposed areas in the harvested areas, respectively. Using the six-band spectral signatures of the four image endmembers, the Landsat data were spectrally unmixed and fractional abundance images for each endmember were generated. The 39 MPB-at tacked fraction image was exported to A R C / V I E W and a map showing MPB-at tacked lodgepole pine stands was prepared. 2.3.3 Verifying MPB Attack Identification The identified MPB-at tacked stands were verified through: i) field checking; ii) aerial helicopter surveys; iii) comparison with the provincial aerial overview surveys of 2000 and 2001; and iv) compar ison with an I K O N O S image from 2001. Field Surveys Of the 34,770 MPB-at tacked stands identified from the Landsat data, a random sample of 340 (-1%) was field checked in November 2001 by independent contractors, not associated with the data analysis. A spreadsheet with U T M coordinates of MPB-at tacked stands (i.e., pixels), identified from Landsat, was generated and given to the contractors along with a soft- and hardcopy of the M P B attack map. The guidance to the contractors was to navigate to the pixel U T M coordinates using G P S receivers and record their observations on the presence/absence of MPB-at tacked trees, infestation s ize, attack stage, and presence of other pathogens and d iseases within ± 30 m of the pixel centroid. This distance was set to accommodate errors in the geometric rectification of Landsat data. The MPB-at tacked stands with only red or grey attacked trees in 2001 were considered in estimating user's accuracy, as these would have been at a green or red attack stage in 2000 when the Landsat data were acquired. M P B - attacked stands with only green attacked trees were treated as commission error, as stands at a green attack stage in November 2001 would have been unattacked in 2000. From the end-users perspective, assessment of commission errors is more important because it is operationally expensive to chase false positives. A s well, the Landsat-based method is designed to capture M P B infestations earlier and if conventional surveys are conducted later some omissions may be captured or they will be detected in the next year using this method. Therefore, a decision was made to concentrate the limited available funds for surveys on commission errors only. However, some omission information was extracted from the overview and helicopter surveys. A lso , there were other studies (Alexander, 2003) testing this method that did look at omission error. Overview Surveys Aerial sketch mapping is used for mapping of red attacked stands on an annual basis by the B C M O F (Westfall, 2004). During these surveys, experienced observers visually spot affected forest stands, and record the extent and severity levels of M P B infestations. Aerial flights are conducted during early July to early September and locations and severity levels of M P B infestations are delineated on N T S map sheets (1:100,000 - 1:250,000 scale). Smal l infestations up to 50 red 40 attacked trees are marked as a dot on the maps and the damage level is classif ied as severe. For larger infestations, a polygon outlining the infested areas is drawn and damage levels are assigned based on the proportion of red attacked trees: light (1-10%), moderate (11-29%), and severe (>30%) ( B C M O F , 2000). For area estimation, a default value of 0.25 ha and 0.50 ha is applied to infestation s izes of 2-30 trees and 31-50 trees, respectively. The positional errors for delineation of infestations on the maps are not known. Apart from the inexact location of a site on the map, there are limitations due to the scale of the map and the digitizing process. A positional error of 1 mm at 1:100,000 and 1:250,000 sca les represent an error of 100 m and 250 m on ground, respectively. The standard for digitizing error tolerance is 100 m. Aerial overview surveys are highly cost effective and useful in rapid identification of areas of forests being attacked by various pests and d iseases ; however, they are subjective and observer dependent. Var ious studies (Harris and Dawson, 1979; Harris ef al., 1982; Gimbarzevsky ef al., 1992) have shown that defoliation estimates are frequently exaggerated during sketch mapping, while counts of bark beetle-killed trees are low when compared to aerial photographs or ground plots. The typical accuracy of sketch mapping of red-attacked stands is around 70 percent ( B C M O F , 2000). Therefore, aerial overview survey maps are generally not used as a reference in accuracy assessments . However, due to the lack of any other reference data, these maps were used in this study to compare the distribution pattern of MPB-at tacked stands identified from the Landsat data to get some indication of the levels of omission. The 2000 and 2001 overview surveys maps were combined to generate a composite 2000-2001 map because the MPB-at tacked stands on the Landsat-based map contained both red-attacked stands as well as the current-attacked stands of 2000. To account for delineation and digitizing errors, five buffer zones around M P B red attack locations at 100 m intervals were created on the composite overview survey map. The M P B attack map prepared from the Landsat data was super imposed on this composite overview map in a GIS environment. Using the spatial query functions in G IS , information on three parameters: i) number of remote sensing-based MPB-at tacked stands contained within the red-attacked stands on the overview map (identification accuracy); ii) number of remote sensing-based MPB-at tacked stands not contained within the red attack stands on overview map (commission error); and iii) number of red-attacked stands on the overview map with no remote sensing-based MPB-at tacked stands (omission error). Statistics on these three parameters were collected at each buffer level. Helicopter Survey A helicopter survey to verify MPB-at tacked stands identified from the Landsat data was carried out in the last week of June 2001. A normal colour printout of the Landsat imagery of the 41 study area, with the locations of the identified MPB-at tacked stands, was used. Large portions of the study area were overflown and area foresters verified the MPB-at tacked stands on this printout based on the presence or absence of M P B infestations as determined from the helicopter. Comparison with an IKONOS Image A n I K O N O S image covering a test area of 100 k m 2 was acquired on August 13, 2001. The M P B attack of 2000 was at the red-attacked stage in August 2001. Whereas large red attacked lodgepole pine stands could be seen on the normal color image, smal l infestations were difficult to identify. These observations were supported by the later studies by White ef al. (2004) who reported that average red attack detection accuracy using multispectral I K O N O S imagery was 7 1 % for low attack levels (1 to 5 % of trees red-attacked) and 9 2 % for a medium level attack (5 - 2 0 % of trees red-attacked). Therefore, the I K O N O S image could be used only as partial evidence to verify the M P B -attacked stands identified from the Landsat data. 2.4 RESULTS 2.4.1 MPB Attack Map The distribution of the M P B attack identified by remote sensing is shown in Figure 2.6. The potential infestations occupied over 3,130 ha (34770 x 0.09 ha), about 1.3% of the total study area. These observations conformed well to the field impressions of the area foresters in 2000, as well as with the M P B infestation extent identified through the annual survey of 2001 (COFI , 2003). Landsat-based M P B attack maps from this study were available to and used by the area foresters in June of 2001, while the overview survey maps were not available until early 2002. 2.4.2 Verification Based on Ground Surveys Ground survey data were collected at 339 randomly selected stands identified as M P B -attacked (Figure 2.7). The surveys indicated that 226 of these stands were positively associated with M P B attack (Table 2.3). The user's accuracy varied with M P B infestation size. Based on the number of MPB-at tacked trees, the 339 stands identified as MPB-at tacked were categorized into 4 c lasses : no attacked trees [Class 0] (28 stands); single attacked tree [Class 1] (98 stands); 2-25 attacked trees [Class 2] (149 stands); and >25 attacked tress [Class 3] (64 stands). The first three c lasses correspond to the M P B infestations mapping convention of the B C Ministry of Forests ( B C M O F , 2001). The identification accuracies were 62.2% (class 1), 70 .5% (class 2), and 93.7% (class 3). The mean user's accuracy was 67%. Only 26 .5% of the M P B infestations had >25 M P B attacked trees/site. This indicates that the majority of the M P B infestations were very small in extent in August 2000 when the image was acquired. 42 ' sir* % • - P • •* ••V* " ^ V • T mngm a t « • • a* • * 1 • • = • - _j F igu re 2.6: Mountain pine beetle attacked lodgepole pine stands (red F igu re 2.7: Distribution of randomly selected M P B attacked stands color) identified from Landsat-7 ETM+ data (August 2000). identified from Landsat-7 ETM+ (August 2000) for field The M P B attacked stands occupy 3130 ha which is 1.3% verification, of the total study area. A significant proportion of the commission error was due to confusion with other pests and d iseases. At 63 sites, stands affected by other pests and d iseases (e.g., spruce bark beetle, Douglas-fir beetle, Ips bark beetle, Armillaria root rot, and Atropellis canker) were identified as MPB-at tacked stands (Table 2.4). Spruce bark beetles were present at 42 of these sites, all of which were located in the S B S zone. This zone is the wettest of the zones considered in this study; spruce, the primary host of spruce bark beetle, is a moisture loving species and grows preferentially in areas of higher moisture. In this zone, lodgepole pine are usually found in mixtures with other conifers, especial ly spruce, and account for less than 20% of the forest stands. A s information on other pests and d iseases in an area is useful to forest managers for overall health management, the user's accuracy for combined pest and d isease affected stands, including M P B , was estimated as 85.25% ((226+63)/339). If the forest stands affected by other pests and d iseases were excluded from the 339 verified stands, the M P B attack identification accuracy would be 81.88% (226/(339-63)). The ground surveys also indicated that 50%, 2 9 % and 2 0 % of the 226 correctly identified MPB-at tacked stands were at the current attack, current plus red attack, and red attacked stage, respectively when the Landsat data were acquired in August 2000 (Table 2.5). The current-attacked and red-attacked stands in 2000 were at the red attack and gray attack stages, respectively, in 2001 when the field surveys were conducted. Table 2.3: Partial error matrix based on third party independent field verification of 339 randomly selected MPB-at tacked stands identified from Landsat-7 ETM+ (August 2000). MPB 4 infestation size Sample size . , i S^Landsat#fe based M P B 1 sites) MPB positive .y MPB False positives User's accuracy (%) Pests and Diseases Others* ? Class 1 98 61 32 5 62.2 C lass 2 149 105 27 17 70.5 C lass 3 64 60 4 - 93.7 C lass 0 28 - - 28 -Total 339 226 63 50 66.7 C lass 1 = Single tree; C lass 2 = 2-25 trees; C lass 3 = >25 trees; C lass 0 = No M P B attack * M P B current attacked stands in 2001; dead trees; trees with other physical damages 44 Table 2.4: Pests and d iseases confused with MPB-at tacked stand identified from Landsat data. •J\" ''-^'''1nsecttdiseasei!•',' ' ~-''7 "^:l^: ! umber of stands Spruce bark beetle 42 Douglas-Fir Beetle 2 Ips bark beetle 13 Armillaria root rot 3 Atropellis canker 3 Total 63 Table 2.5: Composi t ion of MPB-at tacked stands with respect to M P B attack stages. MPB attack Types :kiJAu?"M2000).';;,| Number of Stands Total Current 114 50.4% Current / red 66 29 .2% Red 46 20.4% Total 226 2.4.3 Verification Based on Overview Surveys According to the overview surveys taken in 2000 and 2001, red attack infestations were spread over 1.93% (4,508/234,000 ha) of the study area, compared to 1.32% (3,096/234,000 ha) estimated from the Landsat data. Of the 4,508 ha identified as red-attacked, 3,017 ha were under large infestations, delineated as polygons, and 1,491 ha were spot infestations. A s only a certain percent of the area in a polygon is comprised of recently killed trees, the actual difference between the two area estimates could be lower. A lso, some portion of the underestimation from the Landsat data could be due to omiss ions. The comparative assessment of MPB-at tacked stands identified from Landsat with the overview surveys showed that 30.5, 51.6, 65.0, 73.6 and 79.0% of these stands were contained within buffer zones of 100, 200, 300, 400, and 500 m, respectively, around the red-attacked stands identified from the overview surveys. The corresponding errors of omission were 49.5, 32.0, 22.9, 18.6 and 16.5%, respectively. There was a relatively sharp increase in the proportion of M P B -attacked stands in the 100 m to 300 m buffer zone. Figure 2.8 shows the distribution of MPB-at tacked stands, identified from Landsat, within a buffer zone of 300 m around the 2000 and 2001 red-attacked stands. This figure shows that both omissions and commiss ions were present, although most of the MPB-at tacked stands identified from the Landsat data were within the 300 m buffer zone. S o m e of the M P B infestations in the northwest portion of the study area could not be detected. This could be 45 due to the presence of mixed stands, as lodgepole pine grows mixed with Douglas-fir in this portion of the study area. It is also possible that the natural spectral variability in the MPB-at tacked stands might not have been fully captured, indicating the need for a more comprehensive set of spectral signatures representative of the natural spectral variability in lodgepole pine stands. Both the overview survey and the Landsat-based map showed that M P B infestations were present in forest stands not shown as having lodgepole pine in the forest cover maps (Figure 2.9). This indicates that there is a need for a more accurate spec ies map of the area. The current digital forest cover maps only partially met the information needs for determining lodgepole pine distribution. k F igu re 2.8: The MPB-at tacked stands (identified from Landsat data) contained within a 300 m buffer zone around red-attacked stands identified from aerial overview surveys done in 2000 and 2001. 46 F igu re 2.9: Existence of M P B attack outside stands with lodgepole pine content, identified from digital forest cover maps. 2.4.4 Verification Based on Helicopter Surveys During the helicopter survey it was observed that the trees were just fading and had still not turned fully red at many of the ground locations identified from the Landsat data as M P B infested. Plant stress detection g lasses (Rankin ef al., 2000) were used to help confirm M P B infestations or tree stress. This survey indicated a qualitatively estimated user's accuracy of 70%. This survey particularly looked at omiss ions. It was observed that there was little indication of stress in areas not indicated by the M P B attack map as infested. This survey was instrumental to foresters understanding and having confidence in the Landsat-based M P B infestation map and led them to use it operationally for that season . 2.4.5 Verification Based on the IKONOS Image A qualitative comparison of the Landsat-based M P B attack map with the I K O N O S multispectral data imaged one year later is shown in Figure 2.10. The pixels of various shades of red in the I K O N O S image represent MPB-at tacked stands at the red attack stage in 2001. The Landsat-47 F igu re 2 . 1 0 : Examples of Landsat-7 ETM+ (August 2000) based MPB-at tacked stands (yellow color, 30 m pixel) overlaid on I K O N O S normal colour imagery (acquired August 2001). Red attacked M P B attacked stands in 2001 (as seen on 4m I K O N O S image) were current attacked in 2000. The orthorectification rms error bounds ± 30 m are shown in magenta. (The portion of study area covered by the I K O N O S imagery is shown in Figure 2.5). based current attacked M P B stands (represented by yellow colored pixels) in 2000 were in the red attack stage at the time of I K O N O S image acquisition in 2001. The magenta line around these pixels indicates the spatial bounds of positional error (± 30m) that is equivalent to the rms error in Landsat orthorectification. MPB-at tacked lodgepole pine stands, as determined by the Landsat-based method, are seen as groups of red colored pixels in the I K O N O S image, indicating correspondence with areas that were MPB-at tacked in 2000. 2.5 D I S C U S S I O N The average identification accuracy of early M P B detection, at 67%, is similar to the satellite-data based accuracies reported in several studies. For example, Franklin et al. (2003) and Skakun et al. (2003) reported identification accuracies of 73 .3%, and 74% respectively, for red-attacked stands. However, it should be noted that the major focus of these studies was M P B attack detection at the red attack stage, whereas the current study aimed at the early detection of M P B attack. Another major distinction between this and the other studies was the extent of M P B infestations at the time of study. In this study, the area of M P B infestations was smal l , whereas the other studies were carried out when large contiguous areas were under M P B red attack. Because of the high cost involved in pre-harvest ground surveys, the magnitude of commiss ion errors has high cost implications for the end-user. Therefore, provincial guidelines stress having low commission errors in forest health surveys. Based on the field interactions with the end-users, a 7 0 % user's accuracy for early detection of MPB-at tacked stands was found to be satisfactory for guiding field crews to attacked stands. This led to the operational application of this approach to over 1.8 million ha and 600,000 ha in 2002 and 2003, respectively Compared to the costs of helicopter G P S surveys at $ 0.15 C A D / h a , 1:30,000 air photo surveys at $0.21 C A D / h a , I K O N O S multispectral data at $0.32 C A D / h a (including data and tasking costs), and field surveys at $11.00 C A D / ha (White ef al., 2005), remote sensing data from Landsat ETM+, at a cost of less than $0.03 C A D / k m 2 was the most cost effective data source. A s well , the intermediate products of the M P B attack detection process, such as a lodgepole pine spec ies map, can be integrated with the forest inventory database, providing additional information on the spec ies distribution within forest cover polygons. MPB-at tacked stands identified from Landsat data included current attacked stands. This was supported by observations at various stages of the study through comparison with: i) overview surveys in 2001 which indicated that the identified MPB-at tacked stands were at the current attack stage in 2000; ii) helicopter surveys of MPB-at tacked stands in June 2001 where some of the M P B -49 attacked stands were still green and had to be identified using stress detection g lasses; iii) I K O N O S multispectral data which was acquired one year after the M P B attack of 2000 had reached the red attack stage; and iv) detailed field verifications in November 2001 where 50 percent of the correctly identified MPB-at tacked stands were only at the current attack stage. Omiss ion errors were mainly due to misidentification of lodgepole pine, especial ly in mixed stands, natural spectral variability within stands, and the predominance of small s ized infestations. Commiss ion errors were mainly due to confusion with other pests and d iseases. These results indicate that there are benefits to having better inventory data (e.g., better spatial accuracy) and information on the natural spectral variability within lodgepole pine stands, especial ly if M P B attack detection is required at the landscape level. There were a number of applications and tests of this method between 2001 and 2003. The results of this research were used by local forest managers in planning and implementing harvest operations in the MPB-at tacked lodgepole pine stands in the study area. Secondly, the approach adopted in this study was tested in another study covering 1.8 million ha in the central Car iboo region. Using Landsat images acquired on 9 June 2002 and 12 July 2001, the M P B attack for 2001 was mapped within a month's time and field-verified in July/August of that year. The initial ground-based field verifications reported a M P B attack identification accuracy of 70%. A third trial of this approach was carried out in 2003 to map the 2002 M P B infestations in an area spread over 600,000 ha in the same region. Landsat images acquired in August 2002 along with previous images (August 2000) were used. A detailed accuracy assessment based on wall-to-wall ground truth was done on 12 20-ha test sites. The average M P B identification accuracy, commiss ion error and omission errors were estimated at 71.6%, 28.4% and 33.0% respectively (Alexander, 2003). The M P B attack map was made available to the area foresters in March 2003; approximately 8 to 9 months earlier than conventional surveys could have been available. 2.6 CONCLUSIONS Landsat satellite data acquired on July 12, 1999 and August 15, 2000 were used to map MPB-at tacked lodgepole pine stands in 234,000 ha in central B C . The M P B attack locations identified on the Landsat imagery were compared with third party independent field surveys, provincial aerial overview surveys, helicopter surveys, and an I K O N O S image. Detailed ground checks, at 339 randomly selected MPB-at tacked stands identified from the Landsat imagery, indicated that: i) the user's accuracy for early M P B attack identification was 67%; ii) 50 a significant proportion of the commission error was due to confusion with other pests and d iseases; iii) the identification accuracy for combined pest and disease-affected stands was 85.3%; and iv) of the 226 correctly identified M P B infestations, 50.4% were current attack, 29 .2% were a mixture of current and red attack and 20.4% were red-attacked lodgepole pine stands. Early satell i te-based M P B attack detection is useful in directing ground surveys. M P B attack detection is influenced by the inherent spectral variability in attacked stands due to the co-existence of various attack stages, variable rates of infestations within forest stands, confounding influences of several d iseases and pests, as well as site ecology. However, by combining appropriate remote sensing data analysis techniques, with an understanding of the biological and physiological interactions of the host lodgepole pine stands with M P B , satel l i te-based remote sensing data can be used for detection and mapping of M P B infestations about four months earlier than conventional surveys. Spectral variability among mature lodgepole pine stands across a landscape can influence the M P B attack detection. This is because lodgepole pine has a large ecological amplitude growing on a wide range of habitats characterized by different climatic and site condit ions. Therefore, it is necessary to assess the spectral variability in mature lodgepole pine stands. This research issue is explored in the next Chapter. 51 2.7 REFERENCES Adams, J . B., D. E. Sabo l , V . Kapos, R. A lmeida, D. A . Roberts, M. O. Smith, and A . R. Gi l lespie, 1995. Classif icat ion of multispectral images based on fractions of endmembers - application to land cover change in the Brazil ian A m a z o n , Remote Sensing of Environment, 52:137-154. Ahern, F. J . , 1988. The effects of bark beetle stress on the foliar spectral reflectance of lodgepole pine, International Journal of Remote Sensing, 9:1451-1468. Alexander, J . , 2003. Remote sensing summary Landsat 7 A n assessment of error of omission / commission and accuracy: Forest Investment Account Project#: 1017018, Lignum Ltd., Wil l iams Lake, B C . A m m a n , G . D., 1985. Dynamics of 1-year and 2-year life cycle populations of mountain pine beetle and related tree losses, In: Proceedings of the Mountain Pine Beetle Symposium, Smithers, Ministry of Forests, Victoria, B .C. , pp 37-50. Asner , G . P., 1998. Biophysical and biochemical sources of variability in canopy reflectance, Remote Sensing of Environment, 64:134-153. B C M O F , 1995. Bark Beetle Management Guide Book, U R L : http://www.for.qov.bc.ca/tasb/leqsreqs/fpc/fpcquide/beetle/betletoc.htm, (last date accessed : 2 Feb 2004). B C M O F , 2000. Forest health aerial overview survey standards for British Columbia, Forest Pract ices Branch, Ministry of Forests, Victoria, B C , U R L : http://srmwww.gov.bc.ca/risc/pubs/teveg/foresthealth/assets/aerial.pdf, (last date accessed : 12 March 2004). B C M O F , 2001. Gener ic Forest Health Surveys Guidebook. 2nd E d . Forest Pract ices C o d e of British Columbia Guidebook, Forest Pract ices Branch, Ministry of Forests, Victoria, B C . B C M O F , 2005. Mountain pine beetle stewardship research strategy, Research Branch, Ministry of Forests and Range, Victoria, B C . B C M O F , 2006, Forest health: Aerial overview surveys, Forest Pract ices Branch, Ministry of Forests and Range, B C , U R L : http://www.for.gov.bc.ca/hfp/health/overview/overview.htm (last date accessed : 28 November 2006) Bonnor, G . M., 1982. Forest inventories in Canada , Forestry Abstracts, 43:201-211. Chavez , P. S . , 1988. An improved dark-object subtraction technique for atmospheric scattering correction of multi-spectral data, Remote Sensing of Environment, 24:459-479. C O F I , 2003. Mountain pine bark beetle update, http://www.mountainpinebeetle.com/article 2003 dec15.html, Counci l of Forest Industries, British Columbia, (last date accessed : 15 January 2004). Coppin , P. R., and M. E. Bauer, 1994. Process ing of multitemporal Landsat T M imagery to optimize extraction of forest cover change features, IEEE Transactions of Geoscience and Remote Sensing, 34:918-927. Crist, E. P., and R. C . C icone, 1984. Application of the Tasse led C a p concept to simulated Thematic Mapper data, Photogrammetric Engineering & Remote Sensing, 50:343-352. 52 Davidson, D., and S . Lataille, 2000. Assessment of possible spectral indicators for green stage mountain pine beetle attack using airborne C A S I imagery at Tyee Lake. Prepared for the B .C . Ministry of Forests, I T R E S Research Ltd., Calgary, Alberta. DeAngel is , J . D., T. E. Nebeker, and J . D. Hodges, 1986. Influence of tree age and growth rate on the radial resin duct system in loblolly pine (Pinus taeda), Canadian Journal of Botany, 64:1046-1049. Franklin, S . E., M. A . Wulder, R. S . Skakun, and A. L. Carrol l , 2003. Mountain pine beetle red-attack forest damage classification using stratified Landsat T M data in British Columbia, C a n a d a , Photogrammetric Engineering & Remote Sensing, 69:283-288. Garc ia , M., and S . L. Ustin, 2001. Detection of interannual vegetation responses to climatic variability using AVIRIS data in a coastal Savanna in California, IEEE Transactions of Geoscience and Remote Sensing, 39:1480-1490. Gimbarzevsky, P., A . F. Dawson, and G . A. V a n Sick le, 1992. Assessmen t of aerial photograph and multi-spectral scanner imagery for measuring mountain pine beetle damage. Pacif ic Forestry Centre, Information Report B C - X - 3 3 3 , Victoria. Harris, J . W . E., and A . F. Dawson, 1979. Evaluation of aerial forest pest damage survey techniques in British Columbia, Pacif ic Forest Research Centre, Report B C - X - 1 9 8 , Victoria. Harris, J . W. E., A . F. Dawson, and R. G . Brown, 1982. Evaluation of mountain pine beetle damage using aerial photography, Flathead River, B C , 1980. Pacif ic Forest Research Centre, Report B C - X - 2 2 8 , Victoria. Harris, J . W. E., A . F. Dawson, and D. Goodenough, 1978. Evaluation of Landsat data for forest pest detection and damage appraisal surveys in British Columbia. Pub. Rep . B C -X -182, Canad ian Forest Serv ice, Pacif ic Forest Research Centre, Victoria, B .C. Heath, J . , 2001. The detection of mountain pine beetle green attacked lodgepole pine using Compact Airborne Spectrographic Imager (CASI) data, M.Sc. dissertation, Faculty of Forestry, University of British Columbia, Vancouver, B.C. ,77 p. Heller, R. C , 1968. Previsual Detection of Ponderosa Pine Trees Dying from Bark Beetle Attack, In: Proceedings of the 5th Symposium on Remote Sensing of Environment, Wil low Run Lab, University of Michigan, Ann Arbor, Michigan, pp 387-434. Hobbs, A . J . , 1983. Effects of air photo scale on early detection of mountain pine beetle, M .Sc . dissertation, University of British Columbia, Vancouver , B .C. , 137 p. Huang, C , B. Wyl ie, L. Y a n g , C . Homer, and G . , Zylstra, 2002. Derivation of a tasseled cap transformation based on Landsat 7 at-satellite-reflectance. International Journal of Remote Sensing, 23:1741-1748. Huguenin, R. L., M. A . Karaska, D. V . Blar icom, and J . R. Jensen , 1997. Subpixel classif ication of Bald Cypress and Tupelo G u m trees in thematic mapper imagery, Photogrammetric Engineering & Remote Sensing, 63:717-725. Johnson, J . , P. Greenfield, and A . S. Munson , 1997. An evaluation of the utility of sub-pixel analysis of thematic mapper imagery for the spruce beetle outbreak on the Mant i -LaSal National Forest. United States Forest Serv ice, Forest Health Technology Enterprise Team Publications. 53 Joyce , S . , and H. O lsson , 1999. Long-term forest monitoring with temporal-spectral trajectories from Landsat T M data, In: Proceedings of the IUFRO Conference on Remote Sensing and Forest Monitoring, Rogow, Po land, pp 68-81. Kauth, R. J . , and G . S. Thomas, 1976. The Tasse led C a p - a graphic description of the spectral -temporal development of agricultural crops as seen by Landsat, In: Proceedings of the Symposium on Machine Processing of Remotely Sensed data, 21 June-1 July 1976, Purdue University, West Lafayette, Indiana, pp 4B41-44B51 . Klein, W. H., 1973. Beetle-kil led pine estimates, Photogrammetric Engineering, 39:385-388. Klein, W. H., 1982. Estimating bark beetle-killed lodgepole pine with high altitude panoramic photography, Photogrammetric Engineering & Remote Sensing, 48:733-737. Kl inka, K., J . J . Pojar, and D. V . Meidinger, 1991. Revis ion of biogeoclimatic units of coastal British Columbia, Northwest Science, 65:32-47. Kneppeck, I. D., and F. J . Ahern, 1988. A comparison of images from a pushbroom scanner with normal color aerial photography for detecting scattered recent conifer mortality, Photogrammetric Engineering and Remote Sensing. Leckie, D. G . , and M. Gil l is, 1995. Forest inventory in C a n a d a with emphasis on map production, The Forestry Chronicle, 71:74-88. Murtha, P. A . , 1978. Remote sensing and vegetation damage: A theory for detection and assessment , Photogrammetric Engineering & Remote Sensing, 44:1147-1158. Murtha, P. A. , Z . Bortolot, and J . Thurston, 2000. A Landsat T M spectral Unmixing mountain pine beetle attack fraction map in the Vanderhoof Forest District, British Columbia, In: Proceedings of the 22nd Canadian Symposium on Remote Sensing, Victoria, (Canadian Aeronaut ics and Space Institute, Ottawa, Ontario), unpaginated C D - R O M . Murtha, P. A. , and R. Wiart; 1987. P. contorta-based digital image analysis for mountain pine beetle green attack: Preliminary results, Canadian Journal of Remote Sensing, 13:92-95. Murtha, P. A. , and R. J . Wiart, 1989. Cluster analysis of pine crown foliage patterns aid identification of mountain pine beetle current-attack, Photogrammetric Engineering & Remote Sensing, 55:83-96. Nebeker, T. E., J . D. Hodges, and C . A . Blanche, 1993. Host response to bark beetle and pathogen colonization, Beetle-pathogen interactions in conifer forests, (G. M. Filip, Editor), Academic Press , New York, pp. 157-173. Oki , K., H. Oguma, and M. Sugita, 2002. Subpixel classification of Alder trees using multitemporal Landsat thematic mapper imagery, Photogrammetric Engineering & Remote Sensing, 68:77-82. Peddle, D. R., P. M. Teillet, and M. A. Wulder, 2003. Radiometric image processing, Remote sensing of forest environments, (E. F. Steven, Editor), Kluwer Academic Publ ishers, Boston. Pojar, J . , D. Meidinger, and K. Kl inka, 1991. Chapter 2: Concepts , Ecosystems of British Columbia, (J. Pojar, Editor), British Columbia Ministry of Forests, Victoria, B .C. Pr ice, K .P. , L .E. , Stephen, M.D. Nell is, R.Y. Lee, and R. Boyce, 1997. Mapping land cover in a high 54 plains agro-ecosystem using a multi-date Landsat Thematic Mapper modeling approach. Transact ions of Kansas Academy of Sc iences , Vol .100, no. 1/2, pp 21-33. Radeloff, V . C , D. J . D.J . Mladenoff, and M. S . Boyce , 1999. Detecting jack pine budworm defoliation using spectral mixture analysis: separating effects from determinants, Remote Sensing of Environment, 60:156-169. Rankin, L., J . Heath, and P. A . Murtha, 2000. Eff icacy of N A S A plant stress g lasses for pine beetle detection, In: Proceedings of the 22nd Canadian Symposium on Remote Sensing, Victoria, (Aeronautics and Space Institute, Ottawa, Ontario), unpaginated C D - R O M . Rencz , A . N., and S . Nemeth, 1985. Detection of mountain pine beetle infestation using Landsat and simulated thematic mapper data, Canadian Journal of Remote Sensing, 11:50-58. Roberts, A . , S . Dragicevic, J . Northrup, S. Wolf, L. Y . , and C . Coburn, 2003. Mountain pine beetle detection and monitoring: Remote sensing evaluations (Report for the B C Ministry of Forests), S imon Fraser University, Vancouver , B .C . Runesson , U. T., 1991. Considerations for Early Remote Detection of Mountain Pine Beetle in Green-Foliaged Lodgepole Pine, P h D dissertation, Faculty of Forestry, University of British Columbia, Vancouver, B.C.,237 p. Schott, J . R., C. Savaggio, and W. J . Volghok, 1988. Radiometr ic scene normalization using pseudoinvariant features, Remote Sensing of Environment, 26:1-16. Sharma, R., 2000. Detection of mountain pine beetle infestations using Landsat TM Tasseled Cap Transformations, M .Sc dissertation, Faculty of Forestry, University of British Columbia, Vancouver , B.C.,57 p. Sharma, R., and P. A . Murtha, 2001. Application of Landsat T M tasseled cap transformations in detection of mountain pine beetle infestations. In: Proceedings of the 23rd Canadian Symposium on Remote Sensing, Quebec, (Canadian Aeronaut ics and S p a c e Institute, Ottawa, Ontario), unpaginated C D - R O M . S ingh, A . , 1989. Digital change detection techniques using remotely-sensed data, International Journal of Remote Sensing, 10:989-1003. Sirois, J . , and F. J . Ahern, 1988. An investigation of S P O T H R V data for detecting recent mountain pine beetle mortality, Canadian Journal of Remote Sensing, 14:104-108. Skakun, R. S . , M. A . Wulder, and S . E. Franklin, 2003. Sensitivity of the Thematic Mapper Enhanced Wetness Difference Index (EWDI) to detect mountain pine beetle red attack damage, Remote Sensing of Environment, 86:433-443. Smith, T., K. Whitehead, and A. Norquay, 2002. Mountain pine beetle green attack detection project -2001. Project Report, Earth Imaging Technologies Inc., Sa lmon Arm, B C . Song , C , C . E . Woodcock, K .C. Seto, M.P. Lenney, and S.A. Macomber , 2001. Classif ication and change detection using Landsat T M data: When and how to correct atmospheric effects. Remote Sensing of Environment, 75:230-244. Steen, O. A. , and R. A . Coupe, 1997. A field guide to forest site identification and interpretation for the Cariboo Forest Region: Part 1. Land Management Handbook Number 39, Ministry of Forests, British Columbia, Victoria. 55 Westfal l , J . , 2004. 2003 Summary of forest health conditions in B C , Forest Pract ices Branch, Ministry of Forests, Victoria, B C . White, J . C , M. A . Wulder, D. Brooks, R. Re ich , and R. D. Wheate, 2004. Mapping mountain pine beetle infestation with high spatial resolution satellite imagery, The Forestry Chronicle, 80:743-745. White, J . C , M. A . Wulder, D. Brooks, R. Re ich , and R. D. Wheate, 2005. Detection of red attack stage mountain pine beetle infestation with high spatial resolution satellite imagery, Remote Sensing of Environment, 96:340-351. 56 3.0 EFFECTS OF STAND AGE, BEC SUBZONE / VARIANTS AND SITE INDEX ON SPECTRAL VARIABILITY IN SELECTED CONIFER SPECIES 3.1 I N T R O D U C T I O N Spectral reflectance from forest stands varies widely in the spatial and temporal domains. Natural variations in spectral reflectance patterns for a given plant spec ies are associated with the spec ies phenotype and genotype, leaf morphology, leaf development, structural differences, leaf senescence, foliage age, and plant maturity. External environmental factors, including soil and site quality, and biotic (e.g., insects and disease) and abiotic factors (e.g., fire, moisture), also affect reflectance (Kalensky and Wi lson, 1975; Murtha et al., 1997). Spectral variability affects discrimination among various cover types and identifying spectral differences caused by pests and d iseases. In M P B damage assessment using remote sensing data, spectral variability may affect identifying lodgepole pine stands, as well as reliable detection of infestations within those stands. A s mentioned in Chapter 2, acquiring lodgepole pine maps of an area is one of the pre-requisites to any attempt to detect M P B attack. The resolution and accuracy (both identification and positional) of such maps strongly influence the reliability and accuracy of detecting and mapping M P B infestations. General ly, in satell i te-based M P B attack detection studies in B C (Murtha ef al., 2000; Franklin ef al., 2003b; Skakun ef al., 2003), B C Ministry of Forests digital forest cover maps have been used to identify stands containing lodgepole pine. In a stand, delineated as a polygon on these maps, the relative proportions of up to five forest spec ies are recorded (Leckie and Gill is, 1995). However, the spatial distribution of the different spec ies within the stand is not known. Therefore, a species- level map showing the distribution of various tree spec ies within a stand is desirable. A forest spec ies map has three advantages. First, such a map can provide information on forest cover types and their extent at a much finer spatial resolution (depending upon the type of remote sensing data used for mapping) than currently available from traditional forest cover maps. Second , if the data analysis is confined to a single tree spec ies alone, this could help in reducing the confounding influences of pest and d iseases of other spec ies on the detection accuracy of pests or d iseases of interest. Third, it would facilitate monitoring changes in the health, extent, and distribution of a single spec ies in a more efficient manner. A version of this chapter will be submitted for publication (Suggested Journal : Landscape Ecology) 57 Several research questions must be resolved if medium resolution satellite data (e.g., Landsat-7 ETM+) are to be used for tree spec ies mapping. I have chosen to study the effects of site ecology, as manifested through an ecological gradient, site index, and stand age on the spectral reflectance of three commercial ly important conifer spec ies: lodgepole pine, Douglas-fir, and spruce. Site ecology may influence natural spectral variability within forest stands and affect identification of host spec ies using remote sensing techniques. Site-index is a measure of site productivity and provides an indirect measure of tree vigor and health influencing host susceptibility. Stand age can influence spectral variability among forest stands and confuse spec ies classif ication. A g e also influences susceptibility of lodgepole pine trees to M P B . The specif ic research questions addressed are: 1. What is the variability in spectral reflectance patterns of Douglas-fir, lodgepole pine and spruce as a function of ecological site, site index and stand age? 2. Is the variability in spectral responses significant? 3. What is the effect of spectral signature extension on spec ies classif icat ions? In order to investigate these questions several factors must be considered. Certain phenomena, such as ecological variability, may not manifest in spectral variability at the local level, but might at the landscape level. Therefore, study areas should be large to cover the variations of interest. The complexit ies, expenses, and logistics involved in the study of large areas led Walker (2003) to recommend combining existing G IS databases and remote sensing data. The B C Ministry of Forests ' G IS-based digital forest cover maps were available for the study area. This database is in a G IS environment and consists of a vector map (forest cover map) layer showing the extent and distribution of forest stands as polygons with an associated attribute table. Stand attributes include forest spec ies types, age c lasses, height c lasses , and crown closure c lass. Because of the local variations found within most natural cover types at the canopy level, image analysis methods based on pixels are not suitable for studying ecological variability in spectral reflectance patterns. To study spectral variability at the landscape level, an analysis unit bigger than a pixel is required. Image segmentation has been suggested as one of the possible solutions to define the units of image analysis. Another method is to use land parcel boundaries to subdivide the images, aggregate raster data within the land parcel, and analyze it at the parcel level (Smith and Fuller, 2001; Dean and Smith, 2003). This study was carried out at the stand level over an area of - 1 9 0 0 0 k m 2 in central B C . The null hypothesis is that there are no significant differences in the mean stand level spectral reflectance (Bands 1 to 5, and Band 7 bands of Landsat-7 ETM+) of lodgepole pine, 58 Douglas-fir, and spruce as a function of variations in stand age, site ecology (at the B E C subzone/variant level) and site index. 3.2 STUDY AREA The study area was located in the central portion of the former Car iboo Forest Region, covering an area of about 19,000 km 2 . It was covered by 167 B C M O F digital forest cover map sheets. Lodgepole pine, Douglas-fir, white spruce, and crosses between white and Engelmann spruce are the most commercial ly important forest tree spec ies in the area. For operational purposes in B C , white and Engelmann spruce (and their crosses) are often treated as the same species, cal led interior spruce (Nigh ef al., 2004). These tree spec ies dominate the southern forests of B C . They comprise 7 3 % of the total volume harvested and about 90% of the total number of seedl ings planted in the B C interior in 2001 (Nigh ef al., 2004). Lodgepole pine is the most widespread tree in the province and can grow in diverse environments, from water-logged bogs to dry sandy soi ls. Although it is adapted to a wide range of sites, it grows best on deep, moist, well-drained loam. It generally grows in pure even-aged stands or in associat ion with Douglas-fir and spruce. Lodgepole pine is susceptible to M P B attacks, mistletoe, rusts, and root d isease. Douglas-fir is found throughout southern and central British Columbia. It prefers moist but well-drained sites and a humid atmosphere. White spruce grows best on well-drained, moist soil a long streams and around the border of swamps. It grows in pure stands, as well as in associat ion with black spruce (Picea mariana), balsam fir (Abies amabilis), white birch (Betula papyrifera), aspen (Populus tremuloides), Engelmann spruce, and subalpine fir (Abies lasiocarpa). It can grow in a variety of environments. Engelmann spruce is most common in southern and central British Columbia. It grows in pure and mixed stands and prefers deep, rich loamy soils with high moisture content (Hosie, 1975). The landscape in the study area is predominantly a plateau between two mountain systems, the Coast Mountains to the west and the Columbia Mountains to the east. The area is covered by the Fraser Plateau and in parts by the Fraser Bas in , Quesne l Highland, Nechako Plateau and Thompson Plateau. The Fraser Plateau is a level to gently rolling landscape with elevations ranging between 900-1500 m. It is underlain by basalt bedrock, which is covered by a mantle of medium to coarse textured glacial till in most areas. Western portions of the Fraser Plateau are among the driest parts of the region and are strongly affected by the Coas t Mountains rainshadow. The study area is covered by 9 B E C subzone/variants (Figure 3.1). These variants are: i) Interior Cedar-Hemlock moist cool subzone Horsefly Variant (ICHmk3); ii) Interior Douglas-fir dry cool 59 subzone Fraser Variant (IDFdk3); iii) Interior Douglas-fir dry cool subzone Chilcotin Variant (IDFdk4); iv) Interior Douglas-fir very dry mild subzone (IDFxm); v) Sub-Borea l P ine-Spruce dry cold subzone ( S B P S d c ) ; vi) Sub-Boreal P ine-Spruce moist cool subzone ( S B P S m k ) ; vii) Sub-Borea l P ine-Spruce very dry cold subzone ( S B P S x c ) ; viii) Sub-Borea l Spruce dry warm subzone Horsefly Variant ( S B S d w l ) ; and ix) Sub-Borea l Spruce dry warm subzone Blackwater Variant (SBSdw2) . These subzone/variants vary in soil , temperature, and precipitation characteristics (Meidinger and Pojar, 1991; Pojar et al., 1991). The precipitation and forest productivity is lower in the western part of the study area and gradually improves toward the eastern portion, whereas, temperature follows an opposite gradient (Table 3.1). 3.3 MATERIALS AND METHODS Multispectral remote sensing data from Landsat-7 ETM+ were the primary remote sensing data in this study. The data analysis consisted of: i) pre-processing the satellite data; ii) selecting forest stands from the digital forest inventory maps using a set of decision rules in a G IS environment; iii) generating spectral signatures for the selected forest stands; iv) generating an integrated database (i.e., integration of forest cover attribute database with spectral signatures); v) selecting samples (calibration and validation data sets) from the integrated database; and vi) statistical analysis of the multi-band spectral reflectance data. 3.3.1 Remote Sensing Data The study area is covered by two Landsat-7 ETM+ scenes (path/row 47/24, acquired on July 12, 1999; and path/row 48/24, acquired on July 19, 1999). The Landsat scene - path/row 48/24 covers approximately 90% of the study area. Of the eight spectral bands of ETM+, six were used in the analysis: Band 1 (blue-green, 450-520nm), Band 2 (green, 520-600nm); Band 3 (red, 630-690nm); Band 4 (near infrared (NIR), 760-900nm); Band 5 (short wave infrared (SWIR), 1550-1750nm); and Band 7 (SWIR), 2080-2350nm). The acquisition dates were selected because trees are at their maximum vegetative growth in the early summer season (June and July) facilitating better spectral discrimination. Incidentally, these were also the earliest cloud free images available for the study area since the launch of Landsat-7 in Apri l 1999. In 1999, M P B infestations in the study area were very low. Accord ing to the B C Ministry of Forest 's aerial overview surveys, less than one percent of the total area of the Car iboo Forest Region was infested by M P B , Douglas-fir beetle and spruce bark beetle in 1999 (http://www.for.gov.bc.ca/hfp/health/overview/1999table.htm; last visited on October 14, 2005). The overview survey maps indicated that these infestations were mainly outside the study area. Thus, 60 Figure 3.1: Distribution of B E C subzone/var iants in the study area (Source: http:/ /www.for.qov.bc.ca/hre/becweB/resources/maps/index.html#bqc mapping: last a c c e s s e d on March 8, 2006). Table 3.1: Environmental characterist ics of B E C subzone /variants in the study area. (Steen and Coupe , 1997). Biogeoclimatic Ecosystem Classification Subzones/Variants SBPSxc SBPSdc IDFxm IDFdk4 IDFdk3 SBPSmk SBSdw2 SBSdwl ICHmk3 A r e a (Km*) 10898 4327 2373 3994 8953 5658 2567 3443 1123 Elevat ion Range (m) 850-1300 900-1280 650-950 1050-1350 750-1200 950-1350 750-1200 750-1200 780-1250 M e a n annual precipitation (mm) 389 508 392 355 433 506 487 585 722 M e a n annual snowfall (cm) 179 178 145 138 231 No data 198 182 214 Mean annual temperature ( U C ) 1.7 1.9 4.0 3.8 3.3 3.2 3.4 3.7 4.2 Spatial distribution of Subzone / V E Ecological gradient M e a n annual increment riants within study area East ->West Medium (3.5-6.3 m J /ha/year) ^ H i q h 6.4 m J /ha/vear at the time of Landsat data acquisition in July 1999, the tree spec ies in the study area were largely unaffected by any major forest pests. 3.3.2 Pre-Processing of Satellite Data Spectral response from a target is affected not only by its characteristics, but also by other factors such as sun angle, viewing geometry, atmospheric effects, and topographic variations (Song ef al., 2001). Therefore, it is desirable to correct satellite data for these effects through various pre-processing steps. Spectral signatures can then be extracted for different forest stands for statistical analysis. The pre-processing steps included: i) converting the digital numbers (DN) to at-satellite-reflectance ii) atmospheric correction using a dark object subtraction method; and iii) orthorectification to a Universal Transverse Mercator (UTM) projection, datum N A D 8 3 and mosaicking. The raw digital numbers of a Landsat image are not only dependent on the reflectance characteristics of the specif ic scene, but also contain noise and are affected by the viewing geometry of the satellite, the angle of the sun's incoming radiation, and the design characteristic of the sensor. Use of at-satellite-reflectance, also referred to as "Top-of-Atmosphere Ref lectance", removes variations in solar illumination caused by cycl ical changes in the Earth-Sun distance and in solar zenith angle, as well as differences in irradiance in different spectral bands (Peddle ef al., 2003). This conversion accounts for the varying sun angle due to differences in latitude, season , and time of the day, and the variation in the distance between the Earth and Sun . Convers ions of the digital numbers to at-satellite radiance and at-satell ite-radiance to at-satellite-reflectance are done using equation 1 and 2, respectively (Appendix 3.1) (Markham and Barker, 1986). For each Landsat 7 ETM+ image the band-wise gain, bias and sun-elevation-angle are provided in the associated metadata file. The solar irradiance and tables for Earth-Sun distances are given in the Landsat 7 Sc ience Data Users Handbook (Irish, 2000). Different illumination and sun-object-viewer geometry due to slope also cause differences in spectral signatures. Effects of topographic variations on reflectance are not accounted for in the at-satellite-reflectance calculation. Therefore, corrections for topographic effects, especial ly in areas of high relief, are desirable (Walsh, 1980; Riano ef al., 2003). However, this study area lies largely on a plateau. The elevation difference in various B E C subzone/variants ranges between 350 m to 480 m (mean 405 m, median 450 m). Therefore, little influence of topography on the spectral reflectance of various tree spec ies was expected. 62 The electromagnetic radiation, reflected from ground targets, is modified by scattering and absorption by gases and aerosols while traveling through the atmosphere from the Earth's surface to a satellite borne sensor. Of the various methods available for atmospheric correction, dark object subtraction (Chavez, 1988) is one of the most widely used (Price ef al., 1997; Song ef al., 2001).This method is based on the premise that the contribution due to atmospheric haze will equal the values detected for a ground surface area having zero reflectance. Therefore, reduction or removal of atmospheric effects can be approached by subtracting the digital counts associated with the darkest pixels present in a scene (Chavez, 1988). I used this technique on both images. The satellite images, after atmospheric corrections, were geometrically rectified by transforming the images to U T M projection, NAD83 Datum with nearest neighbor (NN) resampling and were mosaicked (Figure 3.2). To cover the range of B E C subzone/variants in the study area, two images were needed. It should be noted that there were only seven days difference between the two image acquisit ions (the minimum possible difference between two adjacent Landsat satellite overpasses) , and atmospheric conditions were relatively similar at the time of both image acquisit ions. However, there might be small differences in the reflectance values for similar cover types between two images. Any differences between the images were examined to ensure that they would not affect results. Examinat ion of bands in the imagery indicated slight difference in the bright non-forest areas and negligible differences within the conifers. . The mosaicked image was further processed to generate a Normal ized Difference Vegetat ion Index (NDVI) image, and for extraction of forest edges by using an edge detection filter. The NDVI was computed as: NDVI = NIR - R e d / NIR + Red NDVI values range from -1 to 1. Vegetated areas have higher positive NDVI va lues because of higher reflectance in the near infrared region (Band 4) and low reflectance in the red spectral region (Band 3). In contrast clouds and water have negative NDVI values because of higher reflectance in Band 3 and low reflectance in Band 4. A s only forested areas were of interest in this study, non-forest areas were excluded from further analysis. Using NDVI thresholds, determined interactively, masks for non-forest areas, harvested areas, and other vegetation were generated. These masks were combined to create a composite mask for the image. In order to choose relatively homogeneous portions of a forest stand for generating training signatures, it was necessary to remove stand edge pixels. Stand edges may be influenced by a number of factors such as gradual change of cover types at their boundaries, possible partial harvesting, higher rates of environmental influences (e.g., wind throw) and natural disturbances 63 (disease and pests) and variations in the stand structure and composit ion. The spectral variation in the edge pixels, due to such effects, can affect the mean stand level reflectance. Therefore, forest edges were extracted using Sobel edge detection filter (PCI, 2003). A reas under edge pixels were also masked in the mosaicked image. 3.3.3 Selection of Forest Stands Spectral signatures of forest stands under different situations were needed. The reflectance mosaic image, with masked pixels (e.g., non-forest and edge pixels) labeled as null, was used for generating stand level spectral signatures for stands representing various spec ies types, ages and site index c lasses in the 9 B E C subzone/variants. Stands were selected using the forest inventory database of the B C Ministry of Forests. The study area was covered by 167 digital forest cover map sheets, each sheet comprising 155 km 2 . These maps were based on the Universal Transverse Mercator (UTM) coordinate system, North Amer ican Datum 1983 (NAD83). The entire study area falls under U T M Zone 10. In B C , tree spec ies composit ion in a stand is estimated as the percentage contribution of each commercial spec ies to the gross total volume of the stand, rounded to the nearest 10%. Up to five spec ies may be listed in decreasing order of contribution (Leckie and Gil l is, 1995). Twenty-seven commercial and 11 non-commercial tree spec ies are recognized in B C . The age of a stand is estimated as an average of the ages of the dominant and codominant trees in the stand. It is either estimated to the nearest year, or more commonly, ass igned to a age c lass (Leckie and Gill is, 1995). S ince the forest cover maps used were last updated in 1990, age c lass was projected to represent the 1999 forest conditions when the Landsat data were acquired. The B C forest inventory characterizes stands into 9 age c lasses (Table 3.2) (Leckie and Gill is, 1995). Trees in age c lasses 1-3 are not generally susceptible to M P B attack, whereas trees in age c lasses 5-9 are. Trees belonging to age c lass 4 are susceptible to M P B during epidemics (Shore and Safranyik, 1992). Site index represents the end-product of ecophysiological p rocesses subjected to the joint effect of local climate, edaphic factors such as soil nutrients, and the background genetic composit ion of the natural stand (Nigh et al., 2004). Site index provides a numeric description of site productivity, and varies for each spec ies within a biogeoclimatic unit. In B C , site index is defined as the height of a site tree at a breast height age of 50 years (www.for.qov.bc.ca/research/spwg/define.htm#sitein, April 7, 2003). A site tree is the largest diameter tree of the target spec ies on a 0.01 ha plot, if the growth of the tree is free of suppression, damage, insect and d isease attack, and silvicultural practices such as 65 fertilization. The height growth of the site tree reflects the potential productivity of the site: a high site index means that the trees are growing fast and the site is productive. Site index allows the productive potential among sites to be compared across a broad range of existing stand conditions (Nigh et al., 2004). Table 3.2: A g e c lasses in the B C forest inventory. rf&Fores%lnventory Age Class w * . Age-Range^ears) j , , 1 1-20 2 21-40 3 41-60 4 61-80 5 81-100 6 101-120 7 121-140 8 141-250 g >251 3.3.4 Sampling Plan Data Preparation All stands of at least 5 ha, which were at least 80% lodgepole pine, Douglas-fir or spruce, and at least 60 years of age, were extracted from the forest inventory database for the study area. Using the Geomat ica Focus module these extracted forest polygons were overlaid on the mosaic. Band-wise stand level mean spectral reflectance, mean NDVI, the total number of pixels, and the total number of data pixels in a forest polygon were recorded using the "Overlay" function of the analysis module. Due to the masking of non-vegetated areas and edge pixels in a polygon, the actual number of data pixels in a polygon were less than the total number of pixels in that polygon. Stands with less than 9 pixels were excluded from the dataset. V isual inspection of the selected forest stands on the Landsat images indicated that some of the forest stands had been recently harvested. This was because the forest cover maps were not updated for changes between the map date and actual forest cover status in 1999, when the satellite data were acquired. These stands were identified using a mean NDVI threshold decided interactively, and removed from the dataset. The final dataset, hereafter cal led the master dataset, consisted of 21264 stands (polygons) with a total area of 845,957 ha (Table 3.3). Each of these stands was assigned a B E C subzone/variant label, site index value, and a random number. The age c lasses in the master dataset ranged from 4 to 9. These age c lasses were grouped into three categories: A C 4 (61-80 years), A C 5 6 7 (81-140 years) and A C 8 9 (>141 years). This was broadly in accordance with the susceptibility and risk rating systems for M P B (Shore and Safranyik, 66 Table 3.3: Distribution of lodgepole pine, Douglas-fir, and spruce leading stands in the master dataset. Subzone Variants LODGEPOLE PINE DOUGLAS-FIR | SPRUCE* AC4 AC567 AC89 AC4 AC567 I AC89 i AC4 AC567 AC89 ICHmk3 7 54 8 2 51 30 3 0 24 IDFdk3 270 1150 279 171 894 1656 4 33 25 IDFdk4 274 1149 362 32 167 631 1 15 56 IDFxm 5 103 19 87 888 863 0 0 4 S B P S d c 563 1510 525 0 0 11 3 16 57 S B P S m k 165 945 206 1 4 62 2 23 66 S B P S x c 1184 3016 1616 0 4 101 1 15 118 S B S d w l 17 299 9 29 218 162 10 37 30 S B S d w 2 60 266 50 26 103 148 6 19 22 Sub-Total (Age class) 2545 8492 3074 348 2329 3664 30 158 402 Sub-total (Species) 14111 6341 812 Grand Total 21264 * Including Engelmann spruce, black spruce and white spruce 1992). Site index for the three spec ies ranged from 3 to 30 m (Appendix 3.2). This range was divided into three categories: low (<12), medium (12-16) and high (>17). Each forest polygon in the master dataset had the following attributes: leading spec ies (lodgepole pine or Douglas-fir or spruce); % species content, age c lass (AC4 / A C 5 6 7 / A C 8 9 ) ; site index (low / moderate / high); B E C subzone/variants ( ICHmk3 / IDFdk3 / IDFdk4 / IDFxm / S B P S d c / S B P S m k / S B P S x c / S B S d w l / S B S d w 2 ) ; polygon area (ha); number of forested (data) pixels, mean spectral reflectance for each of the six Landsat bands (band 1-5, band 7), mean NDVI and a random number. The stands with lodgepole pine (14,111 stands), Douglas-fir (6,341 stands) and spruce (812 stands) covered 514,037, 317,816 and 14,104 ha, respectively. Approximately 6 0 % of the lodgepole pine stands belonged to age c lass A C 5 6 7 , which is highly susceptible to M P B attack. Douglas-fir stands in age c lass A C 8 9 were dominant in the area. Spruce stands were almost equally distributed between age c lasses A C 5 6 7 and A C 8 9 . Lodgepole pine stands were distributed across all the B E C subzone/variants. Douglas-fir stands in age c lass A C 4 and A C 4 5 6 were absent from subzone/variant S B P S d c . A lso, spruce stands, age c lass A C 4 , were absent in the IDFxm subzone/variant. Sample Selection Independent sets of calibration and validation samples were selected from the master dataset using simple random sampling techniques. The desired sample size was set at 75 stands per sample 67 set (species x age x B E C subzone/variant). Whi le the minimum acceptable s ize was nine, Congalton and Green (1999) recommended using a sample s ize of at least 50 for classif ication accuracy assessment . A sample size of at least nine samples was thought to be minimum necessary to represent variability within a sample set. Ideally, from the combination of three species, three age c lasses and nine B E C subzone / variants, a total of 81 combinations would be available. However, only 55 of the combinations provided at least nine samples in each of the calibration and validation samples for that combination (Appendix 3.3). The signatures of each stand sampled in each of the 110 (55 calibration, 55 validation) sample sets were individually screened for outliers. Box plots of spectral reflectance in the various bands were used to identify the outliers (Lane, 1993). Stands identified as outliers were excluded from the dataset. Outliers could be caused by a variety of reasons, such as errors in the attribute labels or having mixed stands in a polygon. Outliers were distributed throughout the study area and were not concentrated in any particular portion of the study area or source image. There were a total of 2786 and 2797 samples in the initial calibration and validation datasets, respectively. After outlier removal, the numbers were reduced to 2590 and 2615 samples, respectively. Al l the samples, after outlier removal, were combined to prepare a pooled calibration and validation data set. The species, B E C subzone/variant and age c lass distribution of the selected stands (both the initial sample s ize, and sample s ize after outlier removal) is given in Appendix 3.4. Datasets Four datasets were generated from the calibration and validation samples. The selection criteria and the use of these four datasets are descr ibed in Table 3.4. The pooled calibration and validation samples comprised Dataset-1. This dataset was used to explore the spectral response of tree spec ies as a function of B E C subzone/variants, age c lass and site index, and to test the statistical separability of the spectral signatures of stands under different situations. The calibration dataset consisted of 1319, 929 and 342 samples for lodgepole pine, Douglas-fir and spruce, respectively (Table 3.5) and the validation dataset consisted of 1326, 945 and 344 samples for lodgepole pine, Douglas-fir and spruce, respectively. The distribution of calibration samples in Dataset-1 is shown in Figure 3.3. The calibration samples were also categorized into 11 10-ha interval s ize c lasses . Around 90 percent of the stands fell within the first five s ize c lasses (<10 to 50 ha) (Figure 3.4). Fifty-seven percent of the spruce stands were less than 10 ha in s ize; the percentages of lodgepole pine and Douglas-fir stands in this c lass were 38 and 34%, respectively (Figure 3.5). 68 Table 3.4: Criteria for selecting datasets. ' Sample Selection criteria Analysis Dataset-1* Optimum 75 (minimum 9) for each variable combination of species, age c lass and B E C subzone/variant Spectral response of tree spec ies as a function of B E C subzone/variants, age c lass and site index Effect of categorical variables and interactions on spectral reflectance using G L M techniques Spec ies level classification Dataset-2** Presence of all three spec ies in a B E C subzone/variant To study effect of site ecology ( B E C subzone/variants) on species classif ication Effect of extension of signatures from one B E C subzone/variant to another on spec ies discrimination Dataset-3** Presence of a minimum of two age c lasses for each of the three spec ies in a B E C subzone/variant To study effect of age c lass on spec ies classification Effect of extension of signatures of forest stands of one age c lass to another on spec ies discrimination Dataset-4** Presence of a minimum of two site index c lasses for each of the three spec ies in a B E C subzone/variant To study effect of site index on species classification Effect of extension of signatures of forest stands of one site index to another on spec ies discrimination * Pooled calibration and validation datasets constitute Dataset 1. " T h e s e datasets are subsets of Dataset-1 Table 3.5: Distribution of forest stands by B E C subzone/variant, age c lass, and site index for lodgepole pine, Douglas-fir and spruce in Calibration Dataset-1. B E C Subzone / Variants Species ? Total Lodgepole pine Douglas-fir Spruce ICHmk3 24 35 12 2590 IDFdk3 208 204 44 IDFdk4 213 156 40 IDFxm 49 187 -S B P S d c 216 - 41 S B P S m k 208 28 52 S B P S x c 222 49 72 S B S d w l 68 141 51 S B S d w 2 111 129 30 Total 1319 929 342 A g e C lass A C 4 379 153 -2590 A C 5 6 7 567 343 158 A C 8 9 373 433 184 Total 1319 929 342 Site Index High 254 247 59 2590 Medium 650 276 143 Low 415 406 140 Total 1319 929 342 69 Only those B E C subzones/var iants where at least nine samples were available for each of the three spec ies were included in Dataset-2. This criterion was met in seven B E C subzone/variants. The calibration, validation and extension sample distribution in each of these seven B E C subzone/variants is given in Table 3.6. Dataset-2 was used to study the effect of site ecology on spec ies classif ication, and to study the effect of extension of spectral signatures of spec ies generated from one B E C subzone/variant to classify stands of same species in another B E C subzone/variant. For each of the seven B E C subzone/variants, a spec ies classif ication model was developed and validated using the calibration and validation samples, respectively, from that B E C subzone/variant. Signature extensions were tested by applying the classif ication model to extension samples, generated by pooling all the validation samples from the seven B E C subzone/variants in the dataset. For example, for the ICHmk3 subzone/variant, there were 71 calibration, 69 validation, and 2120 extension samples (Table 3.6). Table 3.6: Distribution of samples in Dataset-2. Sample Type ff1 " : ' < " ' l r " " ' BEC Variants "1CHmk3- ' 'Ip!Fdk4''! -SBPSrrik SBPSxc SBSdwl SBSdw2 Calibration Douglas-fir 35 204 156 28 49 141 129 Lodgepole pine 24 208 213 208 222 68 111 Spruce 12 44 40 52 72 51 30 Sub-Total It 456 409 288 343 260 270 Validation Douglas-fir 34 213 161 25 51 140 130 Lodgepole pine 23 208 215 214 220 67 115 Spruce 12 47 44 53 70 50 28 Sub-Total 69 468 420 292 341 257 273 Extension Douglas-fir 754 Lodgepole pine 1062 Spruce 304 Sub-Total 2120 Only those cases where at least two age c lasses were present for each of the three spec ies in a B E C subzone/variant were included in Dataset-3. These conditions were met only in the IDFdk3, IDFdk4 and S B S d w 2 B E C subzone/variants. Datset-3 was used to study the effect of stand age on species level classif ication, and to study the effect of extension of spectral signatures of forest stands in one age c lass on the classif ication of forest stands in another age c lass. The number of calibration and validation samples were 872 (145 + 171 + 158 + 163 + 101 + 134) and 898 (154 + 174 + 173 + 160 + 103 + 134), respectively (Table 3.7). The extension samples in IDFdk3, IDFdk4 and S B S d w 2 were 328 (154+ 174), 333 (173 + 160) and 237 (103 + 134), respectively. 70 Figure 3 .3: Spatial distribution of forest stands (calibration dataset, 2590 samples) in different B E C subzone/variants in the study area (1: S B P S x c , 2: S B P S d c , 3:IDFdk4, 4:IDFdk3, 5: IDFxm, 6: S B P S m k , 7: S B S d w l , 8: S B S d w 2 , 9: ICHmk3; 10: other zones ; 11: data gap). 60 ^ 5 ° o o o o o o o o o o r - C N O ^ i - i n c o h - c o T -Area (ha) — • — L o d g e p o l e p i n e — • — D o u g l a s - f i r S p r u c e Figure 3.5: Proportion of different stand size c lasses for lodgepole pine, Douglas-f ir and spruce in Calibration Dataset-1. 72 Table 3.7: Distribution of samples in Dataset-3. Sample type BEC Variants IDFdk3 IDFdk4 SBSdw2 AC89 h AC89 i AC89 AC5674 ' Calibration Douglas-fir 66 67 68 73 68 49 Lodgepole pine 68 71 65 75 22 66 Spruce 11 33 25 15 11 19 Sub-Total 145 171 158 163 101 134 Validation Douglas-fir 72 69 73 72 71 48 Lodgepole pine 71 69 72 72 22 68 Spruce 11 36 28 16 10 18 Sub-Total 154 174 173 760 103 134 Extension Douglas-fir 141 145 119 Lodgepole pine 140 144 90 Spruce 47 44 28 Sub-Total 328 333 237 Only those cases where at least nine samples were present for each of the three spec ies in a B E C subzone/variant for at least two site index c lasses were included in Dataset-4. These conditions were met only in the IDFdk3, and S B P S x c B E C subzone/variants. Dataset-4 was used to study the effect of site index on spec ies classif ication, and to study the effect of extension of spectral signatures of stands in one site index c lass on the classification of forest stands in another site index c lass. The total numbers of calibration and validation samples were 742 and 750, respectively (Table 3.8). 3.3.5 Data Analysis The data analysis consisted of: i) visual inspection of the species-wise spectral reflectance patterns to a s s e s s possible influences of age c lass, site index and B E C subzone/variant; ii) statistical analysis of the spectral signatures of stands to test differences among spec ies growing in various B E C subzone/variants, and belonging to different age c lasses or site index c lasses , iii) statistical testing of the interaction between species and age c lass, spec ies and B E C subzone/variant and species and site index; iv) determining the "effect s ize" (Kotrlik and Wil l iams, 2003) of the above mentioned interactions to a s s e s s the meaningfulness of any statistical significant differences; and v) assess ing the effect of spectral variability due to age c lass, B E C subzone/variant and site index c lasses for species- level classif ication. Multivariate general linear model (GLM) and discriminant analysis (DA) were the two main statistical techniques used. 73 Table 3.8: Distribution of samples in Dataset-4. i Sample Type *: • --j - BEC Variants'"* :\ IDFdk3 SBPSxc Si-Low Si-Medium Sl-Low Si-Medium Calibration Douglas-fir 61 112 18 31 Lodgepole pine 56 131 199 23 Spruce 13 26 58 14 Sub-Total 130 269 275 68 Validation Douglas-fir 57 118 21 30 Lodgepole pine 58 134 182 38 Spruce 12 30 57 13 Sub-Total 127 282 260 81 Extension Douglas-fir 175 51 Lodgepole pine 192 220 Spruce 42 70 Sub-Total 409 341 Multivariate general linear model (GLM) techniques were used to: i) compare the spectral band means of groups formed by various levels of the categorical independent variables; ii) identify the relative influence of each of the variables on the spectral reflectance; and iii) identify the main and interaction effects of each of the variables on the multiple dependent variables. Multivariate G L M tests the differences in the centroid (vector) of means of the multiple dependent variables, for various categories of independent variables. The F test is used to test the null hypothesis that there is no difference in the means of the dependent variables for the different groups formed by categories of the categorical variables. Among the four leading tests of group differences (Hotelling, Wilks, Pi l lai-Bartlett, Roy 's greatest characteristics root) Wilks' Lambda is the most common when there are more than two groups formed by the independent variables (Garson, 2004). The smaller the lambda, the greater the differences among the groups. In the G L M analysis, the spec ies (3 levels), B E C subzone/variants (9 levels), site index (3 levels) and age c lass (3 levels)) formed the categorical variables and reflectance in the six Landsat bands (Bands 1-5, and 7) were the dependent variables. Tests of statistical signif icance reveal that results of an experiment are not simply due to chance. However, the practical signif icance of these statistically significant differences is also very important in applied research. Est imates of "effect s ize" (Kotrlik and Wil l iams, 2003) indicate how strongly two or more variables are related, or how large the difference between the groups is. Among various measures of effect s ize, Eta squared (q2) is one of the most commonly used measures. Eta 74 squared is defined as the proportion of the variance in the dependent measure accounted for by the independent variables. Theoretically, Eta squared can vary from 0 to 1. Va lues of Eta squared approximately correspond to the following "effect s ize" conventions: small (r)2=0.01), medium (n2=0.06), and large (n,2=0.14) (Kirk, 1996; Kotrlik and Wil l iams, 2003). Discriminant analysis (DA) was used to test the effects of spectral variability due to B E C subzone/variants, age c lass and site index on the spec ies classif ication and to test the effects of signature extension on spec ies discrimination. DA is a technique used to classify cases into groups using a discriminant prediction equation (Garson, 2004). In order to conduct DA, a training dataset is required. This dataset is used to determine the weighted linear combination (discriminant function) of the band variables that best descr ibe each group. Each observation is ass igned a probability of belonging to a given group based on the distance of its discriminant function from that of each c lass mean. The end output of DA is a classif ication table. The classif ication table gives the number of observations correctly classif ied and misclassif ied, where the rows are the observed groups and the columns the predicted groups. In case of 100% correct classif ication, all observations lie on the diagonal. In order to determine how well the discriminant functions perform, a set of observations other than those used for developing the discriminant functions should be used. The validation datasets were used for this purpose. Kappa analysis was used to determine if one error matrix was significantly different from another (e.g., whether the classification results from the calibration data and the validation data in a B E C subzone/variant differ from each other). The result of performing a Kappa analysis is a K -HAT statistic, which is a measure of the difference between the observed agreements between the reference and the classification (as reported by the diagonal entries in the error matrix) and the agreement that might be attained solely by chance (Congalton and Green, 1999). Landis and Koch (1977) characterized the ranges of K-HAT(-1 to 1) into three groups: a value greater than 0.80 represents strong agreement; a value between 0.40 and 0.80 represents moderate agreement; and a value below 0.40 represents poor agreement. Finally, a Z test was performed to determine if the two independent K-HAT values, and therefore the two error matrices, were significantly different. The test statistic for testing the signif icance of a single error matrix is expressed as (Congalton and Green , 1999): Z = K H A T , / S Q R T [Variance (KHAT,)] The test statistic for testing if two independent error matrices are significantly different is expressed as (Congalton and Green , 1999): 75 Z = IKHAT! - K H A T 2 | / S Q R T {[Variance (KHAT,)] + [Variance (KHAT 2)]} At a 9 5 % confidence level, the critical value of Z is 1.96. If the absolute value of the test Z statistic is greater than 1.96, the result is significant, and the classification is better than random (Congalton and Green, 1999). This was used to test the validity of the extension of the spectral signatures by classifying the validation dataset, which consisted of pooled samples, using the spectral signatures generated using samples from one B E C subzone/variant, age c lass, or site index c lass. The underlying assumption is that the classification results of the extension sample would not be different from the calibration and validation datasets if spectral signatures are not influenced by B E C subzone/variants, age c lass or site index. 3.4 RESULTS AND DISCUSSION 3.4.1 General Spectral Pattern The stand level spectral response of lodgepole pine, Douglas-fir and spruce in three age c lasses and nine B E C subzone/variants in the six Landsat spectral bands is shown in Figure 3.6. Stands containing a spec ies were ordered, firstly based on increasing age (AC4, A C 5 6 7 , AC89) , and secondly, within an age c lass along the ecological gradient represented by B E C subzone/variants ( S B P S x c , S B P S d c , IDFdk4, IDFdk3, IDFxm, S B P S m k , S B S d w l , S B S d w 2 , ICHmk3). Two major spectral trends can be observed: i) reflectance in Bands 1-3, 5 and 7 for all three spec ies decreased along the ecological gradient, indicating greater absorption (lacking in Band 4); and ii) lodgepole pine stands had more spectral variability than Douglas-fir or spruce stands. Ref lectance from vegetation is a function of plant optical properties, canopy biophysical attributes, viewing geometry, illumination conditions and background effects, and is influenced at the needle level, canopy level and stand level (Will iams, 1991). Solar radiation on the vegetation surface is either reflected, absorbed or transmitted. At the leaf scale, absorption is caused primarily by photosynthetic pigments (chlorophyll a and b) in the visible spectrum (Landsat Bands 1 and 3) (Gates et al., 1965; G a u s m a n and Al len, 1973). Plant reflectance in the near infrared (NIR) (Landsat Band 4) results primarily from multiple refractions occurring at the interface of hydrated cell walls with intercellular spaces , as a result of differing refraction indices (Gates, 1970; Wooley, 1971; Gausman , 1974). Throughout the short wave infrared (SWIR) range (Landsat Bands 5 and 7), leaf reflectance is approximately inversely proportional to the total water present in a leaf and is a function of both moisture content and thickness of the leaf (Gates et al., 1965; Gates, 1970; G a u s m a n , 1977). At the canopy level, reflectance is influenced by tree variables such as crown closure, height, foliage age, shapes and dimensions of crowns (Figure 3.7). For example, reflectance in the NIR increases with 76 CD O c s o CD CD or 2 W Douglas-fir B3 Lodgepole pine Spruce • F C X V O t • F D A 0 5 6 7 F D A C 3 9 > P L A 0 4 P L A 0 5 6 7 • P L A C 8 9 ^- S A 0 5 6 7 - S A C 8 9 jure 3.6: Spectral pattern of Douglas-fir, lodgepole pine and spruce stands as ordered by age c lass and B E C subzone/variant within spec ies (continued...). 7 7 B4 22 • F0A04 • FDA0567 FCAC89 PLAD4 i PLAC567 • PLAC89 • SAC567 - S A C 3 9 Figure 3.6: (continued) Spectral pattern of Douglas-fir, lodgepole pine and spruce stands as ordered by age c lass and B E C subzone/variant within species. 78 the number of layers of leaves in a canopy. At the stand level, canopy reflectance is further modified by variables such as site characteristics, understory and ground vegetation conditions, soil types, soil moisture, and stand density (Will iams, 1991). The stand level spectral response is also influenced by changes in canopy closure, tree-story leaf area index, spec ies composit ion, background reflectance (Nilson and Peterson, 1994) and shadowing effects within stands (Horler and Ahern , 1986; De Wulf et al., 1990; Ardo, 1992). Spectral reflectance in the visible region (Bands 1-3) was approximately the same for Douglas-fir and spruce stands (Figure 3.6 and 3.8). In the NIR region (Band 4), Douglas-fir stands had higher reflectance than spruce and lodgepole pine stands. This could be due to larger crown length and/or needle concentration in Douglas-fir crowns. The spectral variability in the S W I R spectral region (Bands 5 and 7) was largest for lodgepole pine stands and lowest for spruce. One of most distinguishing characteristics of lodgepole pine is its ability to thrive on a broad spectrum of soil types and soil moisture regimes. Spruce is a moisture loving spec ies and is mostly confined to moister areas. A s explained earlier (Table 3.1), the B E C subzone/variants vary in temperature, precipitation, and soil characteristics and represent an ecological gradient in the east-west direction within the study area. The S B P S x c subzone/variant in the western most part of the study area is the coldest and driest of the subzone/variants and the ICHmk3, in the eastern part of the area, is the warmest, wettest and most productive subzone/variant. There is a nearly linear relationship between the spectral reflectance in Bands 1-3 and Bands 5 and 7 and the ecological gradient of these subzone/variants. Lodgepole pine stands in the S B P S x c subzone/variant exhibited higher reflectance in the visible and SWIR spectral regions, whereas lodgepole pine stands in the ICHmk3 subzone/variant had the lowest reflectance. This reflectance pattern is indicative of higher photosynthetic absorption of incident radiation by lodgepole pine in the ICHmk3 subzone/variant. The lodgepole pine stands in this variant have higher canopy moisture. These observations are indirectly supported by observed rates of mean annual increment (MAI), which is a measure of forest productivity. The average MAI ranges from 3.5 m3/ha/year in the S B P S x c to 6.4 m3/ha/year in the ICHmk3 subzone/variant (MacKinnon et al. , 1991). In the NIR spectral region, no pattern in reflectance was seen across these subzone/variants. The spectral reflectance in NIR is primarily influenced by the internal structure in the leaf mesophyl l t issues, and foliage characteristics, such as number, arrangement and age variations in the needles (Gates, 1970; Wooley, 1971; G a u s m a n , 1974). 79 Lodgepole pine Douglas-fir Spruce Figure 3.7: Lateral canopy characteristics of lodgepole pine, Douglas-fir and spruce stands ( B C M O F , 2001). Figure 3.8: The average spectral reflectance pattern of Douglas-fir, lodgepole pine and spruce. The observed reflectance pattern indicates that variations in the climate and soils among the B E C subzone/variants studied have a distinct effect on the spectral reflectance of lodgepole pine, Douglas-fir and spruce stands. Irrespective of the magnitude of the scatter, the relative effects of site variability are consistent along the ecological gradient for each of the three species. 3.4.2 Effect of Stand Age The band-wise stand level mean spectral reflectance (%) and spectral range difference (i.e., difference between minimum and maximum reflectance) for Bands 1-5 and 7 for the three spec ies are given in Table 3.9. The mean spectral response of the three age c lasses for the three spec ies is shown in Figure 3.9. The number of stands in each age c lass are 379, 567, 373 (total 1319 stands); 153, 343, 433 (total 929); and 0, 158, 184 (total 342) for lodgepole pine, Douglas-fir, and spruce, respectively. Spectral reflectance varied within a narrow range for the different age c lasses . A g e c lasses had the largest influence on the spectral reflectance of lodgepole pine (Bands 1-3, 5, and 7), followed by Douglas-fir and spruce. There was no consistent trend in reflectance with an increase / decrease in stand age. In the NIR region (Band 4), Douglas-fir exhibited the largest spectral range difference with age c lass (1.43), compared to lodgepole pine (0.31) and spruce (0.02). Stands of different age c lasses are characterized by variability in tree s izes, crown morphology and canopy gaps. Therefore, it was expected that stands of the three spec ies in the three age c lasses would be spectrally different. However, no clear trends with age c lass were observed visually for any of the species. For lodgepole pine, the maximum difference in spectral reflectance in Bands 1-3 5, and 7 was between A C 5 6 7 and A C 8 9 . In Band 4, the maximum difference was between A C 4 and A C 5 6 7 . For Douglas-fir, Bands 1, 2 and 5 showed maximum difference between A C 4 and A C 5 6 7 , Bands 3 and 7 between A C 5 6 7 and A C 8 9 , and Band 4 between A C 4 and A C 8 9 . The multivariate G L M (Appendix 3.5-7) indicated that spectral differences among the different age c lasses were statistically significant (a = 0.05) for lodgepole pine (p= 0.000) and Douglas-fir stands (p= 0.000), but not for spruce (p= 0.103). The values of Wilk 's lambda for lodgepole pine, Douglas-fir and spruce were 0.957, 0.877 and 0.968 (Table 3.10), respectively, indicating a significant but weak difference in the spectral reflectance among the stand age c lasses . It is apparent from the analysis of variance (Table 3.11) that there was variation in the signif icance of spectral bands with respect to species. For lodgepole pine, Bands 1-4 and 7 showed significant differences; for Douglas-fir only Bands 4-5 and 7 showed significant differences. None of the six spectral bands showed significant differences in spectral reflectance for spruce. Table 3.12 shows which of the three age 81 Tab le 3.9: Stand level mean reflectance (%) and spectral range difference for lodgepole pine and Douglas-f ir (all three age c lasses) and spruce stands (two age c lasses : A C 5 6 7 and A C 8 9 ) . S p e c i e s B1 B2 B3 B4 B5 B7 Lodgepole pine 2.96 (0.15) 5.79 (0.23) 4.45 (0.41) 15.87 (1.43) 10.82 (1.24) 1.31 (0.03) Douglas-fir 2.77 (0.04) 5.48 (0.11) 3.73 (0.11) 17.55 (0.31) 8.16 (0.58) 1.21 (0.05) Spruce 2.64 (0.06) 5.33 (0.12) 3.61 (0.16) 15.85 (0.02) 7.76 (0.40) 1.20 (0.01) Douglas- f i r L o d g e p o l e pine S p r u c e 20 0 I — _ Z _ B1 B2 B3 B4 B5 B7 B1 B2 B3 B4 B5 B7 B1 B2 B3 B4 B5 B7 Landsat ETM+ Spectral Bands _ _ A C 4 _ ^ A C 5 6 7 — A C 8 9 Figure 3.9: Effects of stand age on the spectral variability in Douglas-fir, lodgepole pine and spruce. CO ro c lasses of lodgepole pine and Douglas-fir significantly differ from each other. The stand age c lass means followed by the same letter within bands and spec ies are not significantly different. Such compar isons were not made for spruce, as there were only two age c lasses of this spec ies in the dataset. Although spectral differences among the age c lasses of lodgepole pine and Douglas-fir were statistically significant, the effect s ize was small to very small (in the case of Douglas-fir r)2 =0.022 and in the case of lodgepole pine n,2 =0.002). A s spectral differences among different age c lasses in spruce stands were not statistically significant, effect s ize was not computed for spruce. Even though spectral differences were significant among the three age c lasses in lodgepole pine, the spectral variability due to stand age differences is generally too small to be of any practical signif icance in Landsat ETM+ based classif ications of these three species. Similar results were reported by Franklin et al. (2003a); they observed significant differences between two age c lasses (<100 and >100 years) for jack pine (Pinus banksiana Lamb) and white spruce. However, they found the relationship to be weak. Horler and Ahern (1986) demonstrated that jack pine and black spruce (Picea mariana (Mill.) B S P ) stands aged 61 years or older did not have any spectral differences of consequence among age c lasses. Table 3.10: Multivariate G L M Tests (Wilks' Lambda) for stand age c lass, B E C subzone / variants and site index for Douglas-fir, lodgepole pine and spruce. ?/> "Effect-*? fSpecies- :'"VilLie: - Hypotliesis? df Error df f Sig. • Squared (n2) A g e c lass (3 levels) Fd 0.877 10.33 12 1824.00 0.000 0.0220 PI 0.957 4.82 12 2602.00 0.000 0.0024 Sx 0.968 1.78 6 326.00 0.103 -B E C Variants (9 levels) Fd 0.339 26.47 42 4281.11 0.000 0.1190 PI 0.461 22.77 48 6405.53 0.000 0.1210 Sx 0.276 11.54 42 1532.53 0.000 0.1188 Site Index (3 levels) Fd 0.918 6.68 12 1824.00 0.000 0.0203 PI 0.872 15.32 12 2602.00 0.000 0.0279 Sx 0.896 3.07 12 652.00 0.000 0.0151 (r) 2:0.010=small effect s ize, 0.059=medium effect s ize, 0.138=large effect s ize (Kotrlik and Wil l iams, 2003). 83 Table 3.11: Tests of between-subjects effects. Species* ' B1 wy B2:'-*is • ;-?B3 • • •WB5 : . f ' B7f!-: Sig. Sig. Sig. Sig. • f Sig. n 2 -\ Sig. n 2 A g e c lass Fd .318 0.0015 .180 0.0019 .056 0.0031 .000 0.0498 .018 0.0052 .002 0.0073 PI .020 0.0030 .014 0.0029 .010 0.0027 .015 0.0052 .107 0.0017 .024 0.0026 S x .532 0.0006 .134 0.0033 .209 0.0017 .378 0.0022 .579 0.0006 .776 0.0001 B E C Variant Fd .000 0.1993 .000 0.1865 .000 0.2053 .000 0.0553 .000 0.1523 .000 0.1661 PI .000 0.1247 .000 0.1298 .000 0.1352 .000 0.1386 .000 0.1147 .000 0.1222 S x .000 0.3125 .000 0.2800 .000 0.3875 .105 0.0342 .000 0.1949 .000 0.2891 Site index Fd .000 0.0227 .000 0.0292 .000 0.0251 .000 0.0161 .000 0.0224 .000 0.0248 PI .000 0.0328 .000 0.0340 .000 0.0382 .036 0.0041 .000 0.0325 .000 0.0340 Sx .042 0.0104 .000 0.0230 .001 0.0152 .093 0.0136 .013 0.0168 .018 0.0120 Bold indicates signif icance at a = .05. Table 3.12: Multiple Compar isons: age c lass levels. Means followed by the same letter within bands and species not significantly different (Tukey H S D Test, a < 0.05). jSPECIES; BI B2 B3 B4 B5 B7 Mean.: |,Level..; 1-iMean' [ Level : Mean Level Mean Level Mean "Level! Mean ILevel •-. PI 2.88a A C 5 6 7 5.68a A C 5 6 7 4.22a A C 5 6 7 15.74a A C 4 10.14a A C 5 6 7 1.28a A C 5 6 7 2.93b A C 4 5.78b A C 4 4.50b A C 4 15.80a A C 8 9 10.94b A C 4 1.31b A C 4 3.03c A C 8 9 5.90c A C 8 9 4.63c A C 8 9 16.06b A C 5 6 7 11.38c A C 8 9 1.33c A C 8 9 Fd 2.75a A C 5 6 7 5.43a A C 5 6 7 3.66a A C 5 6 7 16.88a A C 8 9 7.80a A C 5 6 7 1.19a A C 5 6 7 2.77a A C 8 9 5.48a,b A C 8 9 3.77b A C 4 17.46b A C 5 6 7 8.30b A C 8 9 1.21b A C 4 2.79a A C 4 5.54b A C 4 3.77b A C 8 9 18.32c A C 4 8.38b A C 4 1.22b A C 8 9 3.4.3 Effect of BEC Subzone/Variants The band-wise stand level mean spectral reflectance (%) and spectral range difference (i.e., difference between minimum and maximum reflectance) for Bands 1-5 and 7 for the three spec ies are given in Table 3.13. Figure 3.10 shows the relative variations in the mean stand level spectral reflectance of lodgepole pine, Douglas-fir, and spruce in the 9 B E C subzone/variants. Data were not available for Douglas-fir in S B P S d c and spruce in IDFxm. Visual analysis of the spectral reflectance pattern revealed that the largest heterogeneity was in the spectral reflectance of lodgepole pine, followed by Douglas-fir and spruce. A lso , lodgepole pine exhibited the largest spectral reflectance range in each of the six spectral bands (Table 3.13). In general, the minima of the spectral range for each of the spec ies are in ICHmk3 and S B P S m k and the maxima are in S B P S x c and S B P S d c . ICHmk3 and S B P S m k are neighboring subzone/variants in the eastern portion of the study area and S B P S x c and S B P S d c are neighboring subzone/variants in the western portion. The multivariate G L M for the different B E C subzone/variants indicated that their impact on spectral reflectance was statistically significant (a = 0.05) for each of the three spec ies (Table 3.10). Wilk 's lambda was 0.461 (F=22.768, p=.000), 0.339 (F=26.468, p = 000) and 0.276 (F=11.538, p =.000), for lodgepole pine, Douglas-fir, and spruce, respectively. These values were much lower than those for stand age indicating that the spectral differences among lodgepole pine, Douglas-fir and spruce stands in the B E C subzone/variants were relatively large. This is also supported by the large effect s izes of 0.121, 0.119, and 0.119 for lodgepole pine, Douglas-fir, and spruce, respectively. Unlike stand age, there were statistically significant differences at rx=0.05 in the stand level means in each of the six spectral bands for the three conifer species, except for band 4 in spruce (Table 3.11). The B E C subzone/variant level spectral differences in the three spec ies in each of the 6 Landsat ETM+ spectral bands are shown in Table 3.14. For example, in Band 1, the lodgepole pine stands in B E C subzone/variant ICHmk3 are significantly different from the stands in S B S d w 2 . However, lodgepole pine stands growing in the S B S d w l , S B S d w 2 and S B P S m k are spectrally not significantly different in this band. A lso, the lodgepole pine stands growing in the IDFxm are spectrally similar to those growing in IDFdk4 subzone/variant. Except for the B4 band in the case of Douglas-fir (r\2= 0.055, medium effect size) and spruce (r)2= 0.034, small to medium effect size), the effect s izes were greater than 0.12 (approaching a large effect size) for the three spec ies (Table 3.11). The B E C subzone/var iants had more influence on the spectral reflectance in the visible spectral region (Bands 1-3) than in the moisture sensitive S W I R 85 Table 3.13: Stand level mean (nine B E C subzone/variants) spectral reflectance (%) and spectral range difference for lodgepole pine, Douglas-f ir and spruce. S p e c i e s B1 B 2 B 3 B 4 B 5 B 7 Lodgepole pine 2.85 (0.94) 5.63 (1.27) 4.18 (1.86) 15.93 (2.67) 9.96 (6.05) 1.28 (0.23) Douglas-fir 2.73 (0.59) 5.42 (0.92) 3 .69(1.27) 17.06 (2.48) 7.94 (4.16) 1.20 (0.19) Spruce 2.61 (0.46) 5.29 (0.68) 3.56 (0.90) 15.81 (0.96) 7.62 (2.53) 1.19 (0.09) Douglas-fir Lodgepole pine Spruce 0 J B1 B2 B3 B4 B5 B7 B1 B2 B3 B4 B5 B7 B1 B2 B3 B4 B5 B7 Landsat ETM+ Spectral Bands I C H m k 3 — — I D F d k 3 IDFdk4 — — I D F x m ^ — S B P S d c S B P S m k — S B P S x c ^ — S B S d w l — — S B S d w : Figure 3.10: Effects of B E C subzone/variants on the spectral variability in Douglas-fir, lodgepole pine and spruce. Table 3.14: Multiple Compar isons: B E C subzone/variant levels. Means followed by the same letter within bands and species not significantly different (Tukey H S D Test, a < 0.05). ^SPECIES" W B 1 . - ^ " • "•f:.'B2' ' : ' B 3 i F ' Mean L e v e l Mean Level ; Mean Level Mean Level Mean™ Level 'Z Mean Level PI 2.35a ICHmk3 4.95a ICHmk3 3.29a ICHmk3 14.84a ICHmk3 6.86a ICHmk3 1.16a ICHmk3 2.57b S B S d w l 5.19b S B S d w l 3.44a,b S B S d w l 15.11a,b S B P S m k 7.65a,b S B S d w l 1.18a,b S B S d w l 2.58b S B P S m k 5.22b S B S d w 2 3.58b S B S d w 2 15.53b,c S B P S d c 8.06b S B S d w 2 1.20b S B S d w 2 2.58c S B S d w 2 5.34b S B P S m k 3.86c S B P S m k 15.63b,c,d S B S d w 2 9.22c S B P S m k 1.25c S B P S m k 2.86c IDFdk3 5.67c IDFdk3 4.19d IDFdk3 16.01c,d,e S B P S x c 10.12d IDFdk3 1.28d IDFdk3 3.01c,d S B P S d c 5.89d S B P S d c 4.43e IDFxm 16.06c,d,e S B S d w l 10.84d IDFxm 1.31d IDFxm 3.12d,e IDFxm 5.97d IDFxm 4.74f S B P S d c 16.15d,e IDFdk3 11.72e S B P S d c 1.34e S B P S d c 3.27e,f IDFdk4 6.21e IDFdk4 4.95f,g IDFdk4 16.50e IDFdk4 12.25e,f IDFdk4 1.37e,f IDFdk4 3.29f S B P S x c 6.22e S B P S x c 5.16g S B P S x c 17.51f IDFxm 12.91f S B P S x c 1.39f S B P S x c Fd 2.44a S B P S m k 5.06a S B P S m k 3.23a ICHmk3 15.68a S B P S m k 5.82a ICHmk3 1.12a ICHmk3 2.58b ICHmk3 5.12a S B S d w l 3.27a S B S d w l 16.73b S B S d w 2 6.56a,b S B S d w l 1.14a,b S B S d w l 2.58b S B S d w 2 5.13a ICHmk3 3.33a S B P S m k 16.79b ICHmk3 6.93b S B S d w 2 1.16b S B S d w 2 2.60b S B S d w l 5.15a S B S d w 2 3.35a S B S d w 2 16.98b S B S d w l 6.96b S B P S m k 1.16b S B P S m k 2.70b IDFdk3 5.43b IDFdk3 3.65b IDFdk3 17.15b,c IDFxm 8.11c IDFdk3 1.20c IDFdk3 2.92c IDFdk4 5.67c IDFxm 4.02c IDFxm 17.17b,c S B P S x c 8.95d IDFxm 1.25d IDFxm 3.00c IDFxm 5.85d IDFdk4 4.17c IDFdk4 17.85c,d IDFdk3 9.47d IDFdk4 1.26d IDFdk4 3.03c S B P S x c 5.98d S B P S x c 4.51d S B P S x c 18.16d IDFdk4 10.73e S B P S x c 1.31e S B P S x c S x 2.37a ICHmk3 4.96a ICHmk3 3.16a ICHmk3 15.37a IDFdk3 6.36a ICHmk3 1.15a ICHmk3 2.40a,b S B P S m k 5.07a,b S B S d w l 3.21a,b S B P S m k 15.51a S B P S d c 6.72a,b S B S d w l 1.15a S B S d w l 2.52b,c S B S d w l 5.08a,b S B P S m k 3.24a,b S B S d w l 15.57a ICHmk3 7.00a,b,c S B P S m k 1.16a,b S B P S m k 2.56c S B S d w 2 5.18b S B S d w 2 3.40b,c S B S d w 2 15.67a S B P S m k 7.55b,c,d IDFdk3 1.18b,c S B S d w 2 2.63c IDFdk3 5.25b IDFdk3 3.54c IDFdk3 15.68a IDFdk4 7.67c,d S B S d w 2 1.19c,d IDFdk3 2.77d IDFdk4 5.50c IDFdk4 3.82d IDFdk4 16.10a S B P S x c 8.08d,e IDFdk4 1.21d,e IDFdk4 2.82d S B P S d c 5.62c S B P S d c 4.02e S B P S x c 16.26a S B S d w l 8.67e S B P S x c 1.23e,f S B P S x c 2.83d S B P S x c 5.64c S B P S x c 4.06e S B P S d c 16.33a S B S d w 2 8.89e S B P S d c 1.24f S B P S d c OO spectral region (Bands 5 and 7) in case of Douglas-fir and spruce. In contrast, the B E C subzone/variants had nearly equal influence in the visible, NIR and S W I R spectral regions for lodgepole pine. This indicates that variations in the site and canopy moisture, in addition to the optical plant characteristics, are important factors in introducing variability to the spectral response of lodgepole pine stands. The subzone/variants accounted for more of the variance in reflectance than age c lass for all three species. Ramsey et al. (1995) demonstrated the effect of ecological variability on spectral reflectance of vegetation at the ecoregion-level in Utah (USA) using N O A A A V H R R data (~1km spatial resolution). Soi ls, vegetation, climate, geology, and physiography, are relatively homogeneous within an ecoregion. My study demonstrated the influence of ecological variability at the spec ies level at a much finer spatial scale. The statistically significant differences in the spectral reflectance of lodgepole pine, Douglas-fir and spruce stands due to ecological variability were large in effect s ize and consequently should be considered in landscape level Landsat ETM+ forestry applications. In the case of satel l i te-based M P B attack detection, field surveys aimed at collecting calibration and validation data should be designed to account for ecology-induced spectral variability across stands. 3.4.4 Effect of Site Index Site index represents the end-product of ecophysiological p rocesses subjected to the joint effect of local climate, edaphic factors such as soil nutrients, and the background genetic composit ion of the natural stand and reflects the potential productivity of the site (Nigh ef al., 2004). A high site index means that the trees are growing fast and the site is productive. Site index allows the comparison of productive potential among sites across a broad range of existing stand conditions. It has been observed that tree growth is related to the intercepted radiation (Kaufmann and Watkins, 1990). Therefore, it was expected that stands within the three site index categories would have different spectral responses. The relative variations in the spectral reflectance of the three spec ies in the three site index c lasses are shown in Figure 3.11. The band-wise mean spectral reflectance of the three site index c lasses for lodgepole pine, Douglas-fir and spruce stands and the spectral range difference is shown in Table 3.15. A visual inspection indicates that spectral differences among the three site index c lasses are largest for lodgepole pine, followed by Douglas-fir and then spruce. Lodgepole pine also exhibited the highest degree of variability in the spectral reflectance range in five of the six spectral bands (Bands 1-3, 5 and 7). Douglas-fir exhibited the largest difference (1.01) in the NIR (Band 4) compared to lodgepole pine (0.58) and spruce (0.60). 88 There was consistency in the spectral reflectance pattern with respect to the site index c lasses. Spectral reflectance increased from high to low site index c lass in the six bands for the three species. Trees in the high site index c lass have higher b iomass for a given age, indicating higher photosynthetic activity compared to those in the low site index c lass. Therefore, it was expected that the reflectance in bands 1 and 3 would be lower for higher site index stands than for low site index stands due to higher absorption of the incident radiation by the photosynthetic pigments. The multivariate G L M indicated that spectral reflectance differences due to site index c lass variations were statistically significant (a = 0.05) for all three species (Table 3.10). Wilk 's lambda for lodgepole pine equaled 0.872, (F = 15.315; p = .000); for Douglas-fir, Wilk 's lambda equaled 0.918 (F = 6.676, p = .000), and for spruce Wilk's lambda equaled 0.896 (F = 3.069, p = .000). These values of Wilk 's lambda fall between those for stand age and B E C subzone/variants, indicating that mean spectral differences due to variations in site index were higher than stand age but lower than B E C subzone/variants. There were statistically significant differences in the mean stand level spectral reflectance in all spectral bands for the three species, except for the NIR region in spruce. The band-wise spectral differences in the high, medium and low site index c lasses of lodgepole pine, Douglas-fir and spruce stands are given in Table 3.16. The effect s ize of site index was small to medium for Douglas-fir (n,2=0.020), lodgepole pine (n2=0.028) and spruce (n.2=0.015) and site index explained less than 4 percent of the variability in the spectral reflectance of the six spectral bands for all three species (Table 3.11). The effect s ize for Band 4 was the smallest for each of the species. In the case of Douglas-fir, the effect s ize in all bands other than Band 4 was approximately the same. A similar pattern was observed for lodgepole pine in each spectral band, except that the effect s ize was slightly smaller in Band 4. The effect s ize was lowest for spruce in all the spectral bands except Band 2, where it was similar to that of Douglas-fir. These results indicate that site c lasses should be considered in the design for field data collection. 3.4.5 Interactions Among the Categorical Variables In the previous section, the effects of stand age, B E C subzone/variants and site index on spectral reflectance were studied separately for each species. However, at a landscape level all three tree spec ies coexist and an overlap among spectral characteristics of these spec ies is possible due to the influence of stand age, B E C subzone/variants and site index. Therefore, it was necessary to study the interactions of spec ies with the effects of stand age, B E C subzone/variants and site index. 89 Tab le 3.15: Stand level mean reflectance (%) and spectral range difference by site c lass for lodgepole pine, Douglas-fir and spruce stands. Species B1 B2 B3 B4 B5 B7 Lodgepole pine 2.84 (0.67) 5.63 (0.91) 4.21 (1.39) 15.80 (0 58) 10.11 (4.08) 1.28 ( 0.16) Douglas-fir 2.77 (0.34) 5.46 (0.64) 3.72 (0.81) 17.38 (1.01) 8.08 (2.74) 1.21 (0.11) Spruce 2.62 (0.22) 5.29 (0.45) 3.55 (0.25) 15.91 (0.60) 7.59 (1.51) 1.19 (0.06) F igu re 3.11: Effects of site index on the spectral variability in Douglas-fir, lodgepole pine and spruce. Table 3.16: Multiple Compar isons: site index levels. Means followed by the same letter within bands and species not significantly different (Tukey H S D Test, a < 0.05). S P E C I E S B1 B2 B3 B4 ' B5 B7 Mean. Level Mean Level Mean Level Mean Level Mean Level Mean Level PI 2.52a High 5.19a High 3.52a High 15.56a High 8.16a High 1.20a High 2.81b Medium 5.60b Medium 4.20b Medium 15.71a Medium 9.93b Medium 1.28b Medium 3.19c Low 6.10c Low 4.91c Low 16.14c Low 12.23c Low 1.36c Low Fd 2.60a High 5.16a High 3.32a High 17.02a Medium 6.69a High 1.15a High 2.75b Medium 5.44b Medium 3.71b Medium 17.08a High 8.12b Medium 1.21b Medium 2.95c Low 5.78c Low 4.13c Low 18.03b Low 9.43c Low 1.26c Low S x 2.51a High 5.08a High 3.27a High 15.59a Medium 6.87a High 1.16a High 2.59b Medium 5.25b Medium 3.52b Medium 15.97a,b Low 7.53b Medium 1.19b Medium 2.75c Low 5.53c Low 3.86c Low 16.18b High 8.38c Low 1.22c Low The multivariate G L M (Appendix 3.8) indicated that spectral reflectance differences due to interactions between species and each of age c lass, B E C subzone/variants, and site index c lass were statistically significant (a = 0.05) (Table 3.17). This Wilk's lambda for spec ies and age c lass equaled 0.921 (F = 7.058; p = .000); for spec ies and B E C subzone/variant it equaled 0.383 (F = 20.175, p = .000); for spec ies and site index it equaled 0.870 (F = 10.058, p = .000). The trends in the value of Wilk 's lambda for these interactions were similar to those observed for individual spec ies. Differences in the mean spectral reflectance of all three spec ies were largest for B E C subzone/variants, followed by site index, and then by stand age. At a = 0.05, there were statistically significant differences between the group means of spectral reflectance in each of the six spectral bands for these three interactions (Table 3.18). Interactions between species and B E C subzone/variant accounted for 9% of the variance in spectral reflectance (r)2= 0.092), whereas interactions between species and site index accounted for only 2 % (n2= 0.02). Interactions between species and stand age had no influence on the spectral reflectance. This again supports the earlier observation that B E C subzone/variants were the single largest influence on the spectral reflectance of the three species. Table 3.17: Multivariate tests (Wilks' Lambda): Interactions between spec ies and age c lass, B E C subzone/variant and site index. if:'- Interactions " :" ; t"' Value ? / . • - p ^ - - - . Spec ies * A g e c lass 0.921 7.058 0.000 Spec ies * B E C subzone/variant 0.383 20.175 0.000 Spec ies * site index 0.870 10.058 0.000 Table 3.18: Band-wise signif icance for interactions between species and age c lass, B E C subzone/variant and site index. Interactions B1 B2 B3 B4 B5 B7 Spec ies * age c lass .033 .013 .004 .000 .025 .001 Spec ies * B E C variant .000 .000 .000 .000 .000 .000 Spec ies * site index .000 .000 .000 .000 .000 .000 Significant at a = .05 With respect to the relative effect s ize of individual spectral bands on spec ies and stand age c lass interactions (Table 3.19), only the NIR band (Band 4) had a small to medium effect s ize (n.2= 0.021). Spectral bands in the visible range had the largest influence for spec ies and subzone/variants followed by bands in the SWIR and NIR spectral region. In the case of site index, the spectral bands 92 in the visible and SWIR region had approximately equal influences on the spectral reflectance; however, spectral reflectance in the near infrared region had none to only a small effect. The effect s ize analysis (Table 3.19) indicated that interactions between spec ies and B E C subzone/variants were most important, followed by interactions between species and site index. Even though group means of individual spectral bands in all three interactions were statistically different, the spec ies by age c lass interaction was of little practical importance. Table 3.19: Relative effect s ize of categorical and independent variables. ; • ,-'^Effect- ' * Dependent variables '•".a"-.'. .'• , -^Effect,Sjze\ Eta squared (n/) Interpretation S p e c i e s * A g e C l a s s B 1 0 . 0 0 2 N o n e B 2 0 . 0 0 2 N o n e B 3 0 . 0 0 2 N o n e B 4 0 . 021 S m a l l t o m e d i u m B 5 0 . 0 0 2 N o n e B 7 0 . 0 0 3 N o n e Overall 0.007 None S p e c i e s * B E C v a r i a n t s B 1 0 . 141 L a r g e B 2 0 . 1 3 9 M e d i u m to l a r g e B 3 0 . 1 3 0 M e d i u m to l a r g e B 4 0 . 0 6 9 M e d i u m to l a r g e B 5 0 . 0 9 6 M e d i u m to l a r g e B 7 0 . 1 0 5 M e d i u m to l a r g e Overall 0.092 Medium to large S p e c i e s * s i t e i n d e x B 1 0 . 0 2 7 S m a l l t o m e d i u m B 2 0 . 0 2 8 S m a l l t o m e d i u m B 3 0 . 0 2 6 S m a l l t o m e d i u m B 4 0 . 0 0 9 N o n e B 5 0 . 021 S m a l l t o m e d i u m B 7 0 . 0 2 2 S m a l l t o m e d i u m Overall 0.019 Small to medium 3.4.6 Effect of Signature Extension The spectral signatures of the species generated from a B E C subzone/variant should result in higher classification accuracy when applied to stands within same B E C subzone/variant since the B E C variants exerted a large influence (n,2= 0.092). Similarly, stand age should have little impact on classification accuracy since age c lass exerted little influence (r]2= 0.007). Finally, site index variability 93 should have a small to moderate influence on the classification accuracy (r\2= 0.019). These expectations were tested using Dataset-2 ( B E C subzone/variants), Dataset-3 (age class) and Dataset-4 (site index class). The applicability of signature extensions was tested for age c lasses A C 5 6 7 and A C 8 9 in three B E C subzone/variants (IDFdk3, IDFdk4 and SBSdw2) using Dataset 3. W h e n the discriminant functions, developed from the calibration samples in age c lass A C 5 6 7 , B E C subzone/variant IDFdk3, were used to classify the three spec ies in the validation samples, the overall classification accuracies for the calibration and validation samples were 79.5 and 67.6%, respectively. When the same discriminant functions were used to classify stands from both age c lasses (i.e., A C 5 6 7 and AC89) , the overall classification accuracy was 66.6% (Table 3.20). This indicates that the spectral signatures of lodgepole pine, Douglas-fir and spruce stands of age c lass A C 5 6 7 can also be used to classify stands of age c lass A C 8 9 . This is not surprising because, as d iscussed earlier, the spectral differences due to stand age variations were very smal l . For example, the overall classification accuracy for age c lass A C 8 9 , B E C subzone/variant IDFdk4 was 79.8%, 69 .5% and 71 .3% for the calibration, validation and extension datasets, respectively. However, when spectral signatures generated from stands in age c lass A C 5 6 7 were used to classify stands of both age c lasses in B E C subzone/variant IDFdk4, differences in the overall classif ication accuracies were larger (classification accuracy: 88.4, 84.3, and 72 .2% for calibration, validation and extension samples, respectively). A reverse trend occurred in the case of forest stands in B E C subzone/variant S B S d w 2 , age c lass A C 8 9 . Similarly, the classification accuracies for the calibration, validation and extension samples in both B E C subzone/variant IDFdk3 (site index class: medium and low) and S B P S x c (site index c lass: low) were comparable. For example, the classification accuracies for the calibration, validation and extension samples for B E C subzone/variant IDFdk3, site index c lass: medium were 66.84%, 72 .73% and 72 .83% (Table 3.21), respectively. However, there was a definite trend with respect to overall classif ication accuracy in the case of B E C subzone/variants. This is reflected by the decrease in the overall classification accuracy (Table 3.22) and associated Kappa statistic (Table 3.23) of the extension samples, compared to the calibration and validation samples, and comparison of the z statistic of the calibration and validation, and calibration and extension classif ications (Table 3.24). 94 Table 3.20: Effect of age c lass on classification accuracy (%). The number of stands used in classification are given in parentheses. S a m p l e type . • i ^ i f v r - • • » ^ : v - B E C W a r l a n t s ^ ' Z . ' • . 1 '' r IDFdk3 IDFdk4 SBSdw2 AC89 AC567 AC89 AC567 AC89 AC567 Calibration 79.5 (145) 76.1 (171) 79.8(158) 88.4(163) 81.1 (101) 73.1 (134) Validation 67.6 (154) 71.9 (174) 69.5 (173) 84.3 (160) 57.1 (103) 71.1 (134) Extension* 66.6 (328) 76.1 (328) 71.3 (333) 72.2 (333) 68.9 (237) 67.5 (237) *Sum of validation samples in a B E C variant. Table 3.21: Effect of site index on classification accuracy (%). The number of stands used in classification are given in parentheses. Sample Type BEC Variants IDFdkS SBPSxc Si-Low Sl-Medium Si-Low Sl-Medium Calibration 89.9(130) 66.8 (269) 77.5 (275) 81.5 (68) Validation 72.7 (127) 72.7 (282) 80.5 (260) 64.8 (81) Extension* 69.1 (409) 72.8 (409) 74.7 (341) 69.0 (341) *Sum of validation samples in a B E C variant. Table 3.22: Effect of B E C subzone/variant on classification accuracy (%). The number of stands used in classification are given in parentheses. Sample ; T VP e BEC Variants ICHmk3 IDFdk3 IDFdk4 SBPSmk SBPSxc SBSdwl SBSdw2 Calibration 70.3(71) 74.2 (456) 82.6(409) 72.3 (288) 73.5 (343) 70.6 (260) 71.4 (270) Validation 65.25(69) 71.6 (468) 74.4 (420) 75.1 (292) 76.9(341) 71.9(257) 72.8 (270) Extension* 53.4 (2120) 62.0 (2120) 59.7 (2120) 61.6 (2120) 51.0 (2120) 50.9(2120) 60.00 (2120) *Sum of validation samples in all the seven B E C variants; Table 3.23: Effects of B E C subzone/variants: K H A T statistic. Sample Type BEC Variants ICHmk3 IDFdk3 IDFdk4 SBPSmk SBPSxc SBSdwl SBSdw2 Calibration 0.65 0.58 0.72 0.57 0.64 0.58 0.59 Validation 0.50 0.52 0.61 0.52 0.63 0.60 0.54 Extension 0.35 0.43 0.37 0.43 0.21 0.34 0.45 (KHAT > 0.80 — strong agreement; K H A T 0.40 to 0.80 — moderate agreement; K H A T < 0.40 — poor agreement (Congalton and Green, 1999)). 95 Table 3.24: Effect of B E C subzone/variants on spec ies classif ication: Z statistic. BEC Variant Z Star ZD/ff Calibration (a) Validation (b) Extension (c) : a and b a and c ICHmk3 8.89 8.26 43.07 0.09 3.964* IDFdk3 21.53 21.65 44.20 1.35 4.385* IDFdk4 21.67 20.10 48.20 2.43* 10.640* S B P S m k 15.17 14.84 43.72 0.82 3.087* S B P S x c 17.93 17.95 56.68 0.20 11.31* S B S d w l 15.40 15.43 43.40 0.34 5.360* S B S d w 2 16.44 16.34 40.23 0.90 3.278* Significant at a = 0.05 In each of the seven B E C subzone/variants in Dataset-2, the overall classif ication accuracy for stand identification in the validation samples was closer to that achieved in the calibration samples. However, the overall classification accuracy for the extension samples was consistently lower. For example, in B E C subzone/variant ICHmk3, the classif ication accurac ies for the calibration, validation and extension datasets were 70 .3% (K-HAT: 0.65), 65 .2% (K-HAT: 0.50) and 53.4% (K-HAT: 0.35), respectively. Similarly, for B E C subzone/variant S B P S x c , the classif ication accuracies for the calibration, validation and extension samples, were 73 .5% (K-HAT: 0.64), 76 .9% (K-HAT: 0.63) and 51.0% (K-HAT: 0.21), respectively. The Kappa coefficient for the extension dataset was much lower, indicating that the spectral extension resulted in a higher proportion of misclassif ications. Similar trends in the Kappa statistic were observed for each of the seven B E C subzone/variants. The z statistic for the classif ication differences indicated that there were significant differences between the error matrices of the calibration sample and extension sample, but no significant difference between the calibration and validation samples for the same B E C subzone/variant (except for IDFdk4) (Table 3.24). This implies that classif ications were similar when spectral signatures generated from a B E C subzone/variant were applied to classify tree spec ies within the same subzone/variant. However, if spectral signatures, generated in local ized areas, were applied over larger areas with ecological heterogeneity, the classif ications were different. This provides further evidence that spectral signatures are area dependent and that B E C subzone/variants have a significant influence on the spectral reflectances of these three tree species. 3.5 SUMMARY AND CONCLUSIONS A total of 5205 randomly selected forest stands were used to study the effects of three levels of stand age, three levels of site index and nine B E C subzone/variants on the spectral reflectance pattern of lodgepole pine, Douglas-fir and spruce stands in central British Columbia. The stand-level 96 means of six ETM+ bands (Bands 1-5 and 7) were analyzed to a s s e s s the spectral behavior and differences under these different conditions. Lodgepole pine stands exhibited the highest degree of variation in spectral response pattern, followed by Douglas-fir and then spruce. No definite trend in spectral reflectance with increasing age c lass was evident in any of the three tree species. Although there were statistically significant differences in the mean stand level reflectance due to age c lasses in each species, the effect s ize was small . Therefore, spectral variations due to stand age were of little practical signif icance. Compared to age c lass, the magnitude of spectral variation due to site index in each of the three spec ies was relatively large; a definite trend in the spectral reflectance pattern was observed for each of the three tree spec ies with increasing site index. Stands with low site index values had a lower absorption of incident radiation in the blue and red chlorophyll absorption bands, as well as in the SWIR infrared region. There were also statistically significant differences in the mean stand level spectral reflectance of the three site index c lasses among the species. Ecological variability among the nine B E C subzone/variants examined had the largest influence on the spectral reflectance of lodgepole pine stands, followed by Douglas-fir and spruce. The absorption of incident radiation by a spec ies increased in the visible spectral region (Bands 1 and 2) and the S W I R spectral region (Bands 5 and 7) along an ecological gradient for each species. At the B E C subzone/variant level, spectral reflectance in the visible region played a major role in explaining spectral variability. This could be due to the combined influence of vegetation and site characteristics in the study areas. The western part of the study area is character ized by low productivity rates compared to those in the eastern part, which are among the most productive zones in this region. Similar to stand age and site index, there were statistically significant differences in the stand level spectral means of a species in the B E C subzone/variants. However, unlike age and site index c lasses, the effect s ize was moderate to large, and therefore, of much higher practical signif icance. This study demonstrated that spectral signatures of coniferous tree species, as detected by remote sensing data, are sensit ive to site ecology, site productivity and age differences, in that order. The spectral signatures of the three spec ies I studied varied along climatic and edaphic gradients that were characterized by the B E C subzone/variants. In landscape-level forestry applications of satellite remote sensing, spectral variability due to B E C subzone/variant should be accounted for in selecting calibration and validation sites. Extension of signatures across B E C subzone/variants will lower the identification accuracy of lodgepole pine stands and thereby lower the identification accuracy of M P B -attacked stands. 97 3.6 REFERENCES Ardo, J . , 1992. Vo lume quantification of coniferous forest compartments using spectral radiance recorded by Landsat Thematic Mapper. International Journal of Remote Sensing, 13:1779-1786. B C M O F , 2001. Tree Book: Learning to Recogn ize Trees of British Columbia, http://www.for.qov.bc.ca/hfd/librarv/documents/treebook/treebook.pdf, B C Ministry of Forests, (last date accessed : 18 September 2002). Chavez , P. S. , 1988. A n improved dark-object subtraction technique for atmospheric scattering correction of multi-spectral data, Remote Sensing of Environment, 24:459-479. Congal ton, R. G . , and K. Green , 1999. Assess ing the accuracy of remote sensing data: principles and practices. Lewis Publ ishers, New York. Dean, A . M., and G . M. Smith, 2003. A n evaluation of per-parcel land cover mapping using maximum likelihood c lass probabilities, International Journal of Remote Sensing, 24:2905-2920. De Wulf, R.R., R .E . Goossens , B .P . De Roover, and F C . Borry, 1990. Extraction of forest stand parameters from panchromatic and multispectral S P O T - 1 data. International Journal of Remote Sensing, 11:1571-1588. Franklin, S . E., R. J . Hall, L. Smith, and G . R. Gerylo, 2003a. Discrimination of conifer height, age and crown closure c lasses using Landsat-5 T M imagery in the Canad ian Northwest Territories, International Journal of Remote Sensing, 24:1823-1834. Franklin, S . E., M. A . Wulder, R. S. Skakun, and A. L. Carrol l , 2003b. Mountain pine beetle red-attack forest damage classification using stratified Landsat T M data in British Columbia, Canada , Photogrammetric Engineering & Remote Sensing, 69:283-288. Garson , G . D., 2004. Topics in Multivariate Analys is , http://www2.chass.ncsu.edu/qarson/pa765/statnote.htm, (last date accessed : 20 March 2004). Gates , D. M., 1970. Physical and physiological properties of plants, Remote Sens ing with Specia l Reference to Agriculture and Forestry, National Academy of Sc ience, Washington, D.C. Gates , D. M., H. J . Keegan, J . C . Schleter, and V . R. Weidner, 1965. Spectral properties of plants, Applied Optics, 4:11-20. G a u s m a n , H., and W. Al len, 1973. Optical parameters of leaves of 30 plant species, Plant Physiology, 52:57-62. G a u s m a n , H. W., 1974. Leaf reflectance of near-infrared, Photogrammetric Engineering, 40:183-191. G a u s m a n , H. W., 1977. Ref lectance of leaf components, Remote Sensing of Environment, 6:1-9. 98 Horler, D. N. H., and F. J . Ahern , 1986. Forestry information content of Thematic Mapper data, International Journal of Remote Sensing, 7:405-428. Hosie, R. C , 1975. Native trees of Canada . Canad ian Forestry Serv ice, Department of the Environment, Ottawa, 380 pp. Irish, R.R. , 2000. Landsat 7 science data user's handbook. Report 430-15-01-003-0, National Aeronaut ics and Space Administration, U R L : http:/ / landsathandbook.qsfc.nasa.qov/handbook/handbook htmls/chapterl 1/chapter11 .html (last date accessed : 15 February 2006) Kalensky, Z. , and D. Wi lson, 1975. Spectral signatures of forest trees, In: Proceedings of the 3rd Canadian Symposium on Remote Sensing, Ottawa, Canad ian Remote Sens ing Society, pp 155-171. Kaufmann, M. R., and R. K. Watkins, 1990. Character ist ics of high- and low-vigor lodgepole pine trees in old-growth stands, Tree Physiology, 7:239-246. Kirk, R. E., 1996. Practical signif icance: A concept whose time has come, Educational and Pychological Measurement, 56:746-759. Kotrlik, J . W., and H. A . Wil l iams, 2003. The incorporation of effect s ize in information technology, learning and performance research, Information Technology, Learning, and Performance Journal, 21:1-7. Landis, J . , and G . Koch, 1977. The measurement of observer agreement for categorical data, Biometrics, 33:159-174. Lane, D. M., 1993. Box Plot, http://davidmlane.com/hyperstat/A37797.html, (last date accessed : December 12, 2003). Leckie, D. G . , and M. Gill is, 1995. Forest inventory in Canada with emphasis on map production, The Forestry Chronicle, 71:74-88. MacKinnon, A. , W. L. Harper, S . Chatwin, and B. M. Wikeem, 1991. Chapter 4: Resource Va lues, Ecosystems of British Columbia, (J. Pojar, Editor), B C Ministry of Forests, Victoria. Markham, B. L., and J . L. Barker, 1986. Landsat M S S and T M post-calibration dynamic ranges, exoatmospheric reflectances and at-satellite temperatures. E O S A T Landsat Technical Notes. Meidinger, D., and J . Pojar, 1991. Ecosys tems of British Columbia, B C Ministry of Forests, Victoria. Murtha, P. A. , Z. Bortolot, and J . Thurston, 2000. A Landsat T M spectral Unmixing mountain pine beetle attack fraction map in the Vanderhoof Forest District, British Columbia, In: Proceedings of the 22nd Canadian Symposium on Remote Sensing, Victoria, (Canadian Aeronaut ics and Space Institute, Ottawa, Ontario), unpaginated C D - R O M . 99 Murtha, P. A . , D. W. Deering, C . E. Olson, and G . A . Bracher, 1997. Vegetat ion, Manual of Photographic Interpretation, (W. R. Phi l ipson, Editor), Amer ican Society of Photogrammetry and Remote Sens ing. Nigh, G . D., C . C. Y ing, and H. Qian, 2004. Cl imate and productivity of major conifer spec ies in the Interior British Columbia, Canada , Forest Science, 50:659-671. Ni lson, T., and U. Peterson, 1994. A g e dependence of forest reflectance: Analys is of main driving factors, Remote Sensing of Environment, 48:319-331. PCI , 2003. PCI Geomat ica, Richmond Hill, Ontario, Peddle, D. R., P. M. Teillet, and M. A. Wulder, 2003. Radiometric image processing, Remote sensing of forest environments, (E. F. Steven, Editor), Kluwer Academic Publ ishers, Boston. Pojar, J . , D. Meidinger, and K. Kl inka, 1991. Chapter 2: Concepts , Ecosystems of British Columbia, (J. Pojar, Editor), British Columbia Ministry of Forests, Victoria. Price, K. P., L. E. Stephen, M. D. Nellis, R. Y . Lee, and R. Boyce, 1997. Mapping land cover in a High Plains agro-ecosystem using a multi-date Landsat Thematic Mapper modell ing approach, Transactions of Kansas Academy of Sciences, 100:21-33. Ramsey , R. D., A . Falconer, and J . R. Jensen , 1995. The relationship between N O A A - A V H R R NDVI and Ecoregions in Utah, Remote Sensing of Environment, 53:188-198. Riano, D., E. Chuvieco, J . Sa las , and I. Aguado, 2003. Assessmen t of different topographic corrections in Landsat -TM data for mapping vegetation types, IEEE Transactions of Geoscience and Remote Sensing, 41:1056-1061. Shore, T. L., and L. Safranyik, 1992. Susceptibil ity and risk rating system for mountain pine beetle in lodgepole pine stands. Canad ian Forestry Service, Pacif ic Forestry Center, Information Report B C - X - 3 3 6 . Skakun, R. S. , M. A . Wulder, and S. E. Franklin, 2003. Sensitivity of the Themat ic Mapper Enhanced Wetness Difference Index (EWDI) to detect mountain pine beetle red attack damage, Remote Sensing of Environment, 86:433-443. Smith, G . M., and R. M. Fuller, 2001. A n integrated approach to land cover classif ication: an example in the Island of Jersey, International Journal of Remote Sensing, 22:3123-3142. Song , C , C. E. Woodcock, K. C . Seto, M. P. Lenney, and S . A . Macomber, 2001. Classif icat ion and change detection using Landsat T M data: W h e n and how to correct atmospheric effects?, Remote Sensing of Environment, 75:230-244. Steen, O. A. , and R. A . Coupe, 1997. A field guide to forest site identification and interpretation for the Car iboo Forest Region: Part 1. Land Management Handbook Number 39. Ministry of Forests, British Columbia, Victoria. 100 Walker, R., 2003. Evaluating the performance of Spatially Explicit Models , Photogrammetric Engineering & Remote Sensing, 69:1271-1278. Wa lsh , S. J . , 1980. Coniferous tree spec ies mapping using Landsat data, Remote Sensing of Environment, 9:11 -26. Wil l iams, D. L., 1991. A comparison of spectral reflectance properties at the needle, branch, and canopy level for selected conifer spec ies, Remote Sensing of Environment, 35:79-93. Wooley, J . T., 1971. Ref lectance and transmittance of light by leaves, Plant Physiology, 47:656-662. 101 4.0 HYPERSPECTRAL BANDS FOR IDENTIFYING SELECTED TREE SPECIES AND MOUNTAIN PINE BEETLE ATTACKED STANDS 4.1 INTRODUCTION In this chapter, satell i te-based hyperspectral data are examined for tree spec ies identification and M P B green attack detection. The difference between multispectral and hyperspectral imagery is the detail of the spectral signature of an object. Most multispectral sensors take one measurement in a wide portion of each major wavelength band, such as visible blue, near infrared, etc. Hyperspectral sensors, on the other hand, measure energy in numerous narrow units of each band. A s a result of higher spectral resolution, the hyperspectral signatures are more detailed and contain specif ic information about vegetation. Often diagnostic spectral features are contained within narrow spectral regions (Appendix 4.1) and get masked within the relatively coarse bandwidth of multispectral sensors (Carter, 1994; Gitelson ef al., 1996). Existing methods for mapping spec ies distribution include field surveys and photographic interpretation. Digital remote sensing data have also been used to identify broad categories of forest cover, mainly using broadband sensors (e.g., Thematic Mapper (Landsat 4 & 5), Enhanced Thematic Mapper Plus (Landsat 7); Multi-spectral Linear Array (SPOT) , Linear Imaging Self Scanner (IRS) and Multispectral Sensor ( IKONOS) (Martin ef al., 1998). Using these sensors a number of studies have been conducted to classify forest type at a detailed resolution with varying success (Franklin, 1994; White ef al., 1995). Therefore, there is a need to evaluate data from new sensors and to develop methods for identifying tree species. Broad-band multispectral imagery from space-borne sensors (e.g. Landsat T M / ETM+) is the most common data source for satell i te-based M P B attack detection. Due to better spectral resolution, hence likely better discrimination, hyperspectral data have been investigated in many aerial and ground-based studies on tree spec ies identification (Table 4.1) and in M P B attack detection (Table 4.2). Note: A version of this chapter will be submitted for publication. (Suggested Journal: International Journal of Remote Sensing) 102 Table 4 .1 : Selected hyperspectral studies for tree species classification. X tree Species ' Sensor' Spectral regions of interest (nm) Reference ' ." . -Sugar Pine (Pinus lambertiana), Ponderosa Pine (Pinus ponderosa), White Fir (Abies concolor), Douglas-fir (Pseudotsuga menziesii), Incense Cedar (Calocedrus decurrens), Giant Sequo ia (Sequoiadendron giganteum), and California Black Oak (Quercus kelogii) P S D 1000 (aerial) 462-505; 700-744; bands in visible region (333-700) better than bands in NIR region; Yel low edge (590-641) least important (Gong etal., 1997) R e d Maple (Acer rubrum), Red Oak (Quercus rubra), White P ine (Pinus strobes), Red Pine (Pinus resinosa), Norway Spruce (Picea abies), and Hemlock (Tsuga canadensis) AVIRIS* 627, 750, 783, 822, 1641, 1660, 2140, 2280, 2290 (Martin etal., 1998) 25 conifer and broadleaf tree spec ies P S D 2 0 0 0 (Ground-based; 210 mm to 1050 nm) 427.29, 476.24, 516.89, 550.00, 597.62, 630.43, 670.37, 710.13, 728.14, 742.52, 760.46, 835.39 (Tung etal., 1999) Loblolly P ine (Pinus taeda), Virginia Pine (Pinus viginiana), Shortleaf Pine (Pinus echinata) Spectroradiometer ( F R Fieldspec™ 354, 404, 421, 435, 490, 712, 1463, 1771, 2460 (Van Aardt and Wynne, 2001) F ive tree spec ies Hyperion Band 8-57 (427.55-925.85) Bands 79-224 (932.72-2395.53) (Total 157 bands) PC1:Ear l y mid infrared (EMIR: 1300-1900); followed by FNIR: 1100-1300; P C 2 : Far mid infrared (FMIR: 1900-2350) and red bands (600-700) (Thenkabail ef al., 2004) *AVIRIS: Airborne Visible/Infrared Imaging Spectrometer Table 4 .2: Selected hyperspectral studies for detection of pests and d iseases. y S p e c t r » v - range (nfrf)^ Feature of interest »|S;Spectral regions of intereftT(hm) : i * v Referenced Bark beetles M P B 400-2200 Green attack 680, 750-1200 (Heller, 1968) M P B 360-1050 Green attack 525-565, 690-730, 730-760; 760-1050 (Ahern, 1988) M P B 427-977 Green attack 539.5, 645.4, 706.4; 829.4 (Heath, 2001) Southern pine beetle 675,698,840 Damage detection 675 ;698 (Carter etai., 1998) Southern pine beetle Green attack 698 (Carter and Knapp, 2001) Spruce budworm 430-881 Defoliation levels 445, 590, 665 (Leckie etai, 1992) Other insects Brown Planthoppers 350-1800 Severity levels 755; 890 (Yang and Cheng , 2001) Diseases Mildew 400-2500 Spectral change 405-418,428-738, 1390-1595, 1743-1805, 1826-2508 (Carter, 1993) Dutch elm d isease 350-2500 D isease levels Early Late Rapid increase in green and red, decrease in NIR Decrease in green and NIR, increase in SWIR (Wilson etai, 1998) Abiotic factors Progress ive dehydration (0-6 hrs) 350-3000 Spectral change Increase in reflectance at all wavelengths. Highest change in SWIR region when Equivalent water thickness of the leaf reduced by 58%. (Aldakheel and Danson, 1997) Var ious biotic and abiotic stress agents 400-2500 Competit ion S e n e s c e n c e Dehydration 403-409, 525-650, 686-728 498-715 506-519, 571-708, 1,119-2,508 (Carter, 1993) Branch cutting 306-1138 T ime lag (72hrs) Hardwood Softwoods Increase in NIR reflectance (-5%) 520,700, 740 (Richardson and Berlyn, 2002) Hyperspectral data have been used to assess the spectral characteristics of various bark beetle attacked stands (Heller, 1968; Ahern, 1988; Runesson , 1991; Carter etal., 1998; Heath, 2001, to identify different defoliation levels (Leckie et al., 1992), and to identify spectral regions suitable for infestations of various pest and d iseases (Carter, 1993; Yang and Cheng , 2001). Severa l researchers (e.g., Ahern, 1988; Runesson , 1991; Heath, 2001) have explored hyperspectral remote sensing data for early M P B attack detection and suggested suitable spectral regions. These studies were carried out using ground-based spectral measurements at the needle level or using aerial data at the tree level. Although these studies have proven valuable in understanding the biophysical and physiological characteristics of trees, both attacked and unattacked, the results from ground or aerial studies cannot be linearly scaled up to understand spectral characteristics at stand level. A s well, these studies were conducted on smal l , homogenous samples; therefore, results from these studies may not be replicable over large forested areas characterized by a wide range of heterogeneity induced by different age c lasses , site characteristics and stand attributes. There is a need to conduct studies over forest stands representing the spectral variability occurring in a natural forested landscape. A lso, there is wide range in the recommended spectral regions identified as being suitable for M P B attack detection (Table 4.2). This could be due to a number of reasons including sensor and data characteristics, and spectral and spatial variability in MPB-at tacked and unattacked trees. A lso , airborne hyperspectral data are costly and have limited spatial and temporal coverage. Satel l i te-based hyperspectral sensors with medium spatial resolution provide data over large areas with good temporal coverage; however, because of their 30 m spatial resolution, they are appropriate only at the stand level. Therefore, there is a need for identifying an optimum spectral band combination for M P B attack detection at the stand level. The November 2000 launch of Hyperion, the first space-based hyperspectral sensor on N A S A ' s EO-1 satellite, provided capability for repeated acquisition of hyperspectral data over extended areas, redefining the "frontier" for land cover mapping and characterization (Pear lman ef al., 2003). Optimal band selection is a key factor in creating practical, accurate predictive models for a given remote sensing application using hyperspectral data. A proper subset of bands can contain the same information, with less noise, than the complete set of bands. This can lead to both an increase in accuracy and a decrease in computational complexity. Another application of band selection is in designing an optimal sensor configuration. An effective method of doing this repeatedly is particularly important with the advent of configurable sensors such as the Compact Airborne Spectrometer Imager (CASI), in which the user has to make decisions regarding the trade-off between the number, location, and widths of spectral bands (Warner and Shank, 1997). The problem then becomes: how does one determine which bands to use? 105 The objectives of this study are: 1) to identify a subset of narrow spectral bands that optimally differentiate conifer spec ies types (Douglas-fir, lodgepole pine, spruce); and 2) to identify a subset of narrow spectral bands that optimally differentiate MPB-at tacked lodgepole pine stands from unattacked stands. 4.2 STUDY AREA The study area, centered at 52.14° N and 121.78° E, is in the former Car iboo Forest Region in B C (now part of the Kamloops Forest Region) and covers an area of 60,000 ha (Figure 4.1). This area is covered by several Biogeocl imatic Ecological Classif icat ion (BEC) subzone/variants (Meidinger and Pojar, 1991): Sub-Borea l Spruce dry warm subzone Horsefly Variant ( S B S d w l ) , Sub -Boreal Spruce dry warm subzone Blackwater Variant (SBSdw2) , Sub-Borea l P ine-Spruce moist cool subzone ( S B P S m k ) and (Interior Douglas-fir dry cool subzone Fraser Variant (IDFdk3). These subzone/variants differ in climatic, soil and vegetation characteristics. Lodgepole pine, Douglas-fir, white spruce, and Engelmann spruce, plus white and Engelmann spruce crosses are the most commercial ly important forest tree spec ies in the area. In this study all spruce types were treated as a single c lass. The forested landscape is characterized by a mosaic of stands of several age c lasses . Even though the M P B infestation in this area were on the rise in 2002 when the Hyperion data were acquired, large contiguous patches of M P B infestations were absent. Smal l patches of MPB-at tacked trees were scattered in mature lodgepole pine stands. 4.3 DATA USED Satel l i te-based hyperspectral data, acquired from the Hyperion sensor, were the main remote sensing data used in this study. Digital forest cover maps and Landsat-7 ETM+ imagery were used to aid in identifying Douglas-fir, lodgepole pine and spruce stands in the area covered by the Hyperion image and in generating training signatures for these forest cover types. B C Ministry of Forest 's aerial photography-based M P B maps, available for a portion of the study area, were used to select M P B -attacked polygons. The orthorectified aerial photographs were used to further identify MPB-at tacked trees or groups of trees within these polygons. The red-attacked trees in the aerial photographs, acquired in August / September 2003, were at the green attack stage in August 2002 at the time of Hyperion data acquisit ion. TR IM (Terrain Resource Information Mapping) maps were used for georectification of the Hyperion and Landsat scenes. These base maps (1:20,000 scale) are produced by the B C government. A summary of the data used in this study is given in Table 4.3. 106 F igu re 4 .1 : Location of the study area within British Co lumbia . Table 4.3: Data used in the study. Primary Data: Hyperion data Acquisit ion Date: August 30, 2002 Collateral data: 73: X • Landsat-7 ETM+ Acquisit ion Date: 12 July, 1999 Ortho Aerial Photographs Acquired during August / September 2003 M P B red attack Map Based on the aerial photographs of 2003 Digital Forest Cover maps TRIM Map sheets B E C Map The Hyperion sensor is a hyperspectral imaging spectrometer with 242 spectral channels covering visible, near-infrared, and shortwave infrared bands (400-2500 nm) with 10 nm bands and 30 m spatial resolution. Each image contains data for a 7.65 km wide (cross-track) by 185 km long (along-track) region. Hyperion has two spectrometers, one VNIR (Visible Near Infra Red) spectrometer (400-1000 nm) and one S W I R (Short Wave Infra Red) spectrometer (900-2500 nm). The overlap with from 900-1000 nm allows cross calibration between the two spectrometers (Pear lman ef al., 2003). The VNIR spectrometer uses a 70 by 256 pixel array; the S W I R spectrometer has a 172 x 256 pixel array. Hyperion operates in a "pushbroom" mode, which provides higher dwell time on a specif ic point on the earth, thus resulting in an improved signal-to-noise (S/N) ratio (Pear lman ef a/., 2003). The U S G S E R O S Data Center (EDC) has supplied Hyperion data products with Level 1R processing since December 2001. These Level 1R products contain 196 radiometrically calibrated spectral bands and 44 non-calibrated bands (Table 4.4). Due to decreased sensitivity of detectors in the spectral regions covered by non-calibrated bands, the spectral value in these bands is set to zero by E D C . The digital values are 16-bit radiances and are stored as a 16-bit s igned integer. The radiometrically corrected radiance images are rescaled for DN (digital number) output. A scal ing factor of 40 is applied to spectral bands 1-70, and a factor of 80 to spectral bands 71-242. Table 4.4: Hyperion calibrated and non-calibrated bands ( U S G S EO-1 Users Guide, 2003). VNIR Channels Band # Wavelength (nm) ^Status ' ' "' 1 - 7 3 5 6 - 4 1 7 n m Not calibrated 8 - 5 5 426 - 895 nm Calibrated 5 6 - 5 7 9 1 3 - 9 2 6 nm Calibrated (overlaps with S W I R 77-78) 5 8 - 7 0 936 - 1058 nm Not calibrated SWiRfCKa^ rihels^  71 - 7 6 852 - 902 nm Not calibrated 7 7 - 7 8 9 1 2 - 9 2 3 nm Calibrated (overlaps with VNIR 56-57) 79 - 224 933 - 2396 nm Calibrated 225 - 242 2406 - 2578 nm Not calibrated 108 4.4 METHODS Data analysis consisted of: i) pre-processing the Hyperion data; ii) creating a geo-registered database in a GIS environment; iii) identifying training sites for the three conifer spec ies; iv) identifying training sites for MPB-at tacked (green attack) and unattacked lodgepole pine stands; v) generating hyperspectral signatures for the three tree species, MPB-at tacked stands, and unattacked lodgepole pine stands; vi) identifying an appropriate spectral subset for spec ies identification and early M P B attack detection; and vii) validating the selected spectral subsets. 4.4.1 Pre-Processing the Hyperion Data Pre-processing consisted of identifying usable bands from the original 242 bands and performing atmospheric correction. Hyperion scenes contain 196 calibrated bands, after excluding the 44 non-calibrated and 2 overlapping bands. However, due to atmospheric absorption and data quality issues, not all 196 bands are usable. There are three major water vapor absorption areas: 1350-1480 nm; 1800-1970 nm; and > 2480 nm (Table 4.5). A s well, there are some lesser, but still significant, water absorption areas between 930- 960 nm and 1115-1150 nm. There are also areas where absorption due to ozone occurs, including a major band at 760 nm, and smaller bands at 690 nm and 1260 nm. Carbon dioxide impacts on bands in the 1950-2050 nm and 1600-1610 nm regions. Two small water absorption features flank the ozone line at 760 nm as well. A s a consequence of these absorption areas, the U S G S , EO-1 User 's Guide (2003) identified 159 spectral bands (Table 4.6) which could be used in Hyperion data based studies. Data quality issues with the remaining 159 bands include the presence of bad pixels (due to dead detectors), streaking / striping (due to uneven detector response), and the presence of anomalous grey values. If a detector has a slightly modified or unbalanced response compared to that of its neighbor or from its normal conditions, the result is a vertical "stripe" or "streak" in the corresponding band of the image data. Streaking or striping is observed in the along-track direction. "Smi le" refers to an across-track wavelength shift from the nominal center wavelength. The observed shift in the VNIR spectral region ranges between 2.6 - 3.5 nm. In the S W I R spectral region, shifts are less than 1 nm and are not significant for forestry applications (Goodenough et al., 2003). Whereas bad pixels and striping are visually apparent (Figure 4.2) in a band, anomalous grey values and spectral smile are detectable only through digital analysis of data. 109 Table 4 .5 : Bands most affected by the atmosphere ( U S G S EO-1 Users Guide, 2003). ;Waiyelfengths,(nri)Sj ' '• C h a n n e l s . ':<1Q„,' i . "•". ' .-.^Comment. •^ 15;-760 41 Oxygen "notch" 930-960 79-81 Water Vapor 1115-1150 98-101 Water Vapor 1350-1480 121-133 Water Vapor 1800-1970 165-182 Water Vapor Table 4.6: Bands recommended for use ( U S G S EO-1 Users Guide, 2003) and bands used in this study. Wavelengths (nm) Channels Name Bands used 446-750 1 0 - 4 0 (31) V N I F M 12-27, 29-40 770-926 42 - 57 (6) VNIR_2 42-53 902-920 76-78 (3) Overlap S W I R 970-1105 82-97 (16) SWIR_1 85-95 1160-1340 102-120(19) S W I R _ 2 101-115, 117-118 1490-1790 134-164 (31) S W I R _ 3 135-163 1980-2406 183-225 (43) SWIR_4 -Total number of bands 159 98 Var ious techniques have been used to reduce the effect of bad pixels or striping. One of the most commonly used techniques involves calculating the mean and var iance of the whole image. The radiance values of the pixels with striping are then set to values representing the mean and var iance for the entire image. This technique has worked well for images of flat uniform terrain such as desert scenes. However, this approach, when applied over large agricultural f ields, resulted in altered spectral characteristics within the image (Cudahy et al., 2002). A second approach involves the pixel values of defective detectors being replaced by the mean and var iance from the neighboring pixels (Datt et al., 2003). Methods for striping corrections are primarily interpolation based; the computed spectral values may be different and result in significant departures from the original spectral values. A s the M P B attack in the study area is local ized, any interpolation from the adjoining pixels which may be unattacked could modify the spectral values of pixels of features of interest. Each spectral band was visually inspected to identify those bands with excess ive striping and/or bad pixels. Visual inspection of these 159 spectral bands revealed the presence of excess ive striping in the bands beyond 1780 nm in the S W I R region. This region is characterized by a low signal-to-noise ratio (Pear lman et al., 2003). Excess ive striping was a lso visible in some of the spectral bands in the VNIR region. These bands were excluded from further analysis. 110 The final dataset consisted of 98 spectral bands (Table 4.6). These spectral bands covered a spectral regime of 457 to 1780 nm. Anomalous grey values (both negative and positive) in each of these 98 spectral band were identified using the image histogram and were masked. It is not possible to completely remove the effect of smile (Jupp ef al., 2002; Goodenough ef al., 2003; C S I R O , 2004). Therefore, smile correction was not attempted. The effect of not correcting smile is that there is a small change in the wavelength being detected across the image. However, this was not expected to affect the results. Three methods of atmospheric correction were evaluated: i) the F L A A S H (Fast Line-of-Sight Atmospher ic Analys is of Spectral Hypercubes) atmospheric correction modules of ENVI image analysis software (Research Systems Inc. Boulder, C O ) , ii) dark object subtraction using clear water as a dark object, and iii) dark object subtraction using the minimum grey value in a band as the dark object. A comparison of the spectral difference or gradient among various cover types on images generated using these three methods and the raw band-wise images were made. The original spectral gradient between the various cover types were modified in the images atmospherical ly corrected using the first method. This could have an impact on between-class discriminability. Images generated using the dark-object subtraction method using clear water as the dark object a lso resulted in a modified grey value gradient between the two cover types. A s well, water pixels in the Hyperion image showed one of the largest heterogeneities observed for a cover type in the data used. Dark object subtraction using the minimum grey value in a band as the dark object retained a spectral gradient similar to that observed in the raw image. Consequent ly, I used this method for atmospheric correction in this study. 4.4.2 Creating a Geo-Registered Database The raw Hyperion image, after atmospheric correction, was geometrically registered using the B C TRIM maps. Twenty-six, uniformly distributed, ground-control-points ( G C P s ) were used to geo-rectify the image (UTM zone 10, Nad83, N N resampling). The root mean square error of these G C P s after correction was <1 pixel. This image was used for generating spectral training signatures for the three tree spec ies and for MPB-at tacked and unattacked lodgepole pine stands. The remote sensing imagery, provincial M P B survey maps, and forest inventory maps were assembled in a GIS environment. 4.4.3 Generating Hyperspectral Signatures for Tree Species Mature stands of lodgepole pine, Douglas-fir, and spruce (age c lass 4 and above) were of interest in this study. This is because mature lodgepole pine stands are the preferred hosts of M P B . To identify training sites for the three species, B C M O F forest inventory maps, Landsat ETM+ (July 112 1999), Hyperion data (August 2002), and digital aerial photographs (August/September 2003) were used. Firstly, using a set of decision rules (i.e., age > 60 years, spec ies content: > 80%), Douglas-fir, lodgepole pine, and spruce stands were identified from the forest inventory maps. Secondly, these stands (polygons) were overlaid on the 1999 Landsat false colour composite (B5: red, B4: green, B 3 : blue) and training sites were delineated within homogeneous portions of the stands. These training sites were then examined using Hyperion data for any change due to harvesting between 1999 and 2002 and for the presence of bad pixels, and were modified if required. Thirdly, to ensure that the species training sites for lodgepole pine did not include MPB-at tacked trees, the lodgepole pine stands were examined for the presence of red-attacked or gray-attacked trees using the digital aerial photographs, acquired in 2003. These red-attacked and gray-attacked trees would have been at the green attack and red attack stage, respectively, in 2002. Any lodgepole pine training sites with identified MPB-at tacked trees were removed from the training sites for spec ies identification. Initially 46 training sites were identified and training statistics (mean, standard deviation and number of pixels) for the six Landsat ETM+ bands (B1-5 and B7) and 98 Hyperion bands (Table 4.6) were generated. The band-wise training statistics were examined for any anomaly which might have been caused by the existence of anomalous values in the Hyperion image. Based on this evaluation, 42 stands (14 Douglas-fir, 16 lodgepole pine, and 12 spruce) were selected for use in the analysis. These training sites covered 170 ha and represented a part of the natural heterogeneity found in the landscape by covering a range of age c lasses , site indices and site conditions (Table 4.7). A spectral subset identified from such heterogeneous training sites should have wide application in forest spec ies classif ication. The hyperspectral reflectance pattern of all the training sites is shown in Figure 4.3. The hyperspectral reflectance pattern of individual tree spec ies is shown in Figure 4.4. 4.4.4 Generating Hyperspectral Signatures for MPB-Attacked Stands Training sites for MPB-at tacked (green attack) and unattacked stands were identified using the M P B red attack map of 2003 (map sheet 93a002) and the digital aerial photographs (acquired in August /September 2003; 0.50 m spatial resolution) from which the M P B red attack map was derived. The lodgepole pine stands at the red attack stage in 2003 were at the green attack stage in August 2002 at the time of the Hyperion data acquisit ion. Firstly, the M P B red attack map was used to identify stands (polygons) containing red-attacked lodgepole pine trees (Figure 4.5). There were large MPB-at tacked polygons marked on the red attack map, but the distribution of red-attacked trees within these polygons was patchy (Figure 4.6). Secondly, digital aerial photographs were visually interpreted to delineate the red-attacked and unattacked trees within these stands. The red-attacked lodgepole pine trees had red crowns at the time of acquisition of the aerial photographs; therefore, it 113 Tab le 4.7: Details of training sites for the three tree spec ies . graining Site Size (Pixels) :; f* Age Class & Site Index •;- : BECVAR A FD1 187 4 High S B S d w l F D 2 34 4 Medium IDFdk3 F D 3 133 4 High IDFdk3 FD4 31 5 Medium S B S d w l F D 5 48 5 High IDFdk3 FD6 20 5 Low IDFdk3 F D 7 21 6 High S B S d w l F D 8 54 6 High S B S d w l F D 9 43 7 Medium IDFdk3 FD10 44 8 Medium S B P S m k FD11 71 8 Medium IDFdk3 FD12 142 8 Medium IDFdk3 F D 1 3 51 8 Medium IDFdk3 FD14 53 8 Medium IDFdk3 PL1 96 4 Medium S B S d w 2 P L 2 17 4 Medium S B S d w 2 P L 3 11 4 Medium S B P S m k P L 4 13 4 Medium S B P S m k P L 5 21 4 Medium IDFdk3 P L 6 34 4 Medium IDFdk3 P L 7 26 5 High S B S d w l P L 8 21 5 High S B S d w l P L 9 15 5 High S B S d w 2 PL10 74 5 High S B S d w 2 PL11 59 5 High S B P S m k P L 1 2 55 5 High S B P S m k P L 1 3 25 5 High IDFdk3 P L 1 4 37 5 High IDFdk3 P L 1 5 19 5 High IDFdk3 PL16 40 6 Medium IDFdk3 S1 16 4 High S B S d w l S 2 27 4 High S B S d w l S 3 43 4 Low S B S d w 2 S4 19 4 Medium S B P S m k S 5 43 4 Medium S B P S m k S 6 53 4 High IDFdk3 S 7 31 5 Medium S B S d w 2 S 8 25 5 Medium S B S d w 2 S 9 48 5 Medium S B P S m k S 1 0 50 5 Medium S B P S m k S11 17 5 Medium IDFdk3 S12 17 5 Medium IDFdk3 (Douglas-fir: 932 pixels; lodgepole pine: 563 pixels; Spruce: 389 pixels; Total: 1884 pixels) 114 3000 2500 400 600 800 1000 1200 1400 1600 1800 Wavelength (nm) F igu re 4.3: Spectral reflectance pattern of unattacked Douglas-fir (red), lodgepole pine (blue) and spruce (green) stands (stand age > 60 years). 2500 2000 a E 1500 £ 1000 • 500 0 400 600 800 1000 1200 Wavelength (nm) 1400 1600 1800 FD1 FD2 FD3 FD4 FD5 FD6 FD7 FD8 FD9 FD11 FD10 FD12 FD13 FD14 Douglas-fir 2500 2000 0! E — 1500 1000 ' r o b 500 0 I 400 600 800 1000 1200 Wavelength (nm) • PL1 -PL9 -PL2 • PL10 •PL3 PL11 •PL4 • PL12 •PL5 • PL13 1400 -PL6 • FL14 1600 • PL7 • PL15 1800 • FL8 PL16 Lodgepole pine 2500 r 2000 a a E 1500 3 C "3 1000 m b 500 0 400 600 800 1000 1200 Wavelength (nm) 1400 1600 1800 S1 S2 S3 S4 — — S5 S6 S7 S8 S9 S10 S11 S12 Spruce F igu re 4.4: Spectral reflectance pattern of Douglas-fir, lodgepole pine and spruce stands. 116 F i g u r e 4.5: M P B red-attacked stands (red colored polygons) superimposed on the digital aerial photographs (part of Map sheet 93a002). F igu re 4.6: Red-attacked trees as seen on the aerial photographs. A 30 x 30 m grid (Hyperion spatial resolution) is overlaid on the photograph. was easy to delineate the red-attacked trees. However, it should be noted that lodgepole pine crown foliage can also turn red because of needle d iseases like pine needle cast (Lophodermella concolor).Trees infected with this d isease have reddish foliage in May and June, turning straw-colored by July (Hunt, 1995). A lso , in certain cases the previous year's red foliage is retained on attacked trees. Such foliage is characterized by a dull red color on normal color photographs. Such trees crowns were excluded from the analysis. The red-attacked trees of 2002 were almost at the gray attack stage in 2003. At the gray attack stage, trees are essentially devoid of foliage. It is easy to visually identify such trees on a photograph. Such trees were also excluded from the analysis. Using these clues, effort was taken to ensure the MPB-at tacked training sites contained mainly green-attacked trees. The proportion of red-attacked trees (green attack in 2002) in a pixel in training sites varied from 20 to 40%. Separate MPB-at tacked and unattacked vector layers were generated. These two layers were super imposed on the georegistered Hyperion image and band-wise training statistics were generated for each training site. 118 The band-wise training statistics for each site were examined to identify contaminated signatures, if any, due to the presence of bad pixels or anomalous grey values and for presence of broad-leaved tree spec ies (e.g., trembling aspen - Populus tremuloides). This process resulted in identifying 94 MPB-at tacked and 51 unattacked training sites. All the training sites were within the S B P S m k subzone/variant. The spectral reflectance of the MPB-at tacked and unattacked training sites are shown in Figure 4.7. 4.4.5 Spectral Band Selection Stepwise discriminant analysis (SDA) , employing the forward method, was used to identify spectral bands important for discriminating Douglas-fir, lodgepole pine, and spruce stands, and discriminating M P B green-attacked stands from unattacked stands. This approach has been used by several researchers for subset band selection in hyperspectral data-based studies (e.g., Tung ef al., 1999; Heath, 2001; Van Aardt and Wynne, 2001; Thenkabai l etai., 2004; Cheng etai., 2005). S D A may be used to determine which weighted linear combination of spectral variables (i.e., spectral bands) best discriminates between two or more groups (e.g., species). The method identifies variables which maximize differences between groups. The default method is to select the variable, at each step, that minimizes the overall Wilks' lambda. S D A computes n-1 discriminant functions, with n being the number of groups. Using these functions, discriminant scores are calculated for each observation in the dataset and used to classify them into the different groups. The forward method is generally used when few variables are to be extracted from a very large set of independent variables. Under the forward method, all variables at each step are evaluated to determine which one will contribute most to the discrimination between groups. The variable with the largest F-to-enter statistic is entered into the initial model, provided that this is greater than the threshold value for F-to-enter. When there are no variables left to enter whose F-to-enter statistics are above the threshold, the model is checked to see whether the F-to-remove statistic of any variables added previously have fallen below the F-to-remove threshold. If so, those variables are removed from the model and the stepwise procedure continues. It finally stops when no further variables are either entered or removed from the model. Model parameters that may be examined include Wilk's Lambda (an index of the discriminating power ranging between 0 and 1, the lower the value the higher the discriminating power), eigenvalues (a measure of the variance in the dependent variable for each function), and canonical correlation (a measure of associat ion between the groups formed by the dependent variable and the given discriminant function) (Tabachnick and Fidell, 2001). The predictive validity of the model can be assessed using several methods, including self-classification of the calibration data, cross validation using the leave-one-out method, and classifying 119 F igu re 4.7: Spectral reflectance pattern of unattacked (green) and M P B green-attacked (red) lodgepole pine stands (all training sites). an independent validation dataset. Classif icat ion accuracies based on self-classif ication are generally higher than the other methods. Cross validation using the "leave-one-out" method involves classifying each observation into a group according to the classification functions computed from all the data except the case being classif ied. This process is run till each observation is individually classif ied. The third method involves randomly partitioning all the observations into training and validation samples. Typically, the training set consists of 70% of the data and the validation set consists of remaining 30% data points. This method is preferred to the leave-one-out classification if enough data are available in the initial sample. There are limitations to S D A . For example, when the number of independent variables is large compared to the number of observations in the dataset or if there is multicollinearity (linear dependence) among the variables, as is normal in hyperspectral data, the stepwise algorithm ends up with a large number of variables in the final model , especial ly at low F-to-enter and F-to-remove thresholds. Multicollinearity exists when two or more independent variables are "highly" correlated to one another. Ideally, independent variables should be independent of each other, a situation difficult to meet with hyperspectral data. If multicollinearity exists, the selection of variables is optimized for the current sample, but could vary considerably for a new sample drawn from the same data set. S ince the intent of using S D A - b a s e d bands is generally to predict group membership in future datasets, this is a serious drawback. Therefore, critical evaluation of selected spectral bands and model validation is very important. Pars imony is an important goal of S D A . This means using the fewest spectral bands needed for accurate classif ication, although not necessari ly the smallest set of classif ication functions. Fewer spectral bands mean lower costs of data collection and processing, and easier interpretation. It is especial ly pertinent in the case of hyperspectral remote sensing data, character ized by hundreds of narrow spectral bands. Huberty (1994) recommends having a goal of only 8 to 10 response variables in the final model. In this study a forward stepwise procedure was used to identify spectral band subsets suitable for discrimination among the three tree spec ies and to identify MPB-at tacked stands from unattacked stands. The selected spectral bands were entered into a canonical discriminant function analysis to quantify the predictability of group membership (i.e., spec ies type and MPB-at tacked versus unattacked stands). Equal classification probabilities were assigned for each group. A s there were only 42 observations in the spec ies data set, the leave-one-out classif ication method was used for validation. The M P B data set (145 observations) was split into training and test datasets. 121 For discriminating among Douglas-fir, lodgepole pine and spruce stands, three sets of spectral bands were identified using three sets of critical p values (Cases A (.05/. 10), B (.05/.06), C (.10/.11)) to guide entry and removal of spectral bands in the model. Spectral bands for discriminating MPB-at tacked stands from unattacked stands were identified using the default p values (entry/removal: .05/.10). Use of default values resulted in 100% discrimination between the M P B -attacked and unattacked lodgepole pine stands. 4.4.6 Evaluating Selected Spectral Bands: Tree Species The efficiency of the selected spectral bands for spec ies discrimination was evaluated using self classification and cross-validation using the leave-one-out method. Individual spec ies accuracy and overall accuracy were used as the criteria for assess ing the effectiveness of the selected spectral bands. Performance of the selected spectral bands for species classif ication was also evaluated at a pixel level through supervised classification using a maximum likelihood classifier. Each of the 42 training sites was used as a separate c lass in the classif ication. These training sites represented natural heterogeneity manifested through variations in stand age (Figure 4.8), biogeoclimatic zones (Figure 4.9) and site index (Figure 4.10). Using the spectral signatures of these training sites, pixels, under the spec ies mask (1884 pixels) generated by combining all the training sites, were classif ied. Classif ication results were organized in a confusion matrix to assess the average overall spec ies classification accuracy, as well as the classification accuracy for individual spec ies. Four sets of classif ications were conducted and their accuracies computed. The first three sets were based on the spectral bands identified in C a s e s A , B and C and the fourth set was generated using all the spectral bands identified in the first three sets. The training site classification accuracies were based on self-classification of the training sites from which the training signatures were generated. Therefore, the classification accuracies may not be representative of the accuracies applicable over the entire study area (Lil lesand and Kiefer, 2000). However, these accuracies were used here only to assess the relative performance of the different spectral band combinations. In order to compare the selected Hyperion spectral bands with Landsat ETM+ bands for spec ies classif ication, training signatures for the three spec ies were also generated from the six Landsat ETM+ bands (B1-5 and 7). The classification results were tabulated in a confusion matrix to assess the overall and individual spec ies classif ication accuracies. 122 2500 2000 ai A E 1500 3 z I 1000 oi 5 500 -J 400 2500 2000 i-a E 1500 3 S 1000 01 5 500 400 2500 2000 1500 5 1000 5 500 400 600 600 800 1000 1200 Wavelength (nm) 1400 1600 1800 -AC4 • -AC7 • • AC8 5 800 1000 1200 Wavelength (nm) 1400 1600 1800 -AC4 • -AC6 600 800 1000 1200 Wavelength (nm) 1400 1600 1800 -AC4 • -AC5 Douglas-fir B E C Subzone/Var ia nt: IDFdk3 Site index: Medium Lodgepole pine B E C Subzone/Var ia nt: IDFdk3 Site index: Medium Spruce B E C Subzone/Var ia nt: S B P S m k Site index: Med ium Figure 4.8: Spectral variability due to age in Douglas-fir, lodgepole pine and spruce stands. 123 2500 2000 E 1500 5 1000 a 5 500 400 2500 2000 E 1500 3 Z I 1000 5 500 400 2500 400 600 800 1000 1200 Wavelength (nm) 1400 1600 - IDFdk3 • - S B S d w 1 600 800 • IDFdk3 • 1000 1200 Wavelength (nm) • S B P S m k 1400 S B S d w 2 1600 600 800 1000 1200 Wavelength (nm) 1400 1600 • IDFdk3 • • S B P S m k S B S d w 2 1800 1800 1800 Douglas-fir Age c lass: 4 Site index: High Lodgepole pine Age c lass: 5 Site index: High Spruce Age c lass: 5 Site index: Medium Figure 4.9: Spectral variability of Douglas-fir, lodgepole pine and spruce stands in different B E C subzone/variants. 124 2500 2000 E 1500 3 S 1000 • 500 400 600 800 1000 1200 Wawelength (nm) 1400 •High • Low 1600 1800 Douglas-fir B E C Subzone/Var iant : IDFdk3 A g e c lass: 5 F igure 4.10: Spectral variability of Douglas-fir stands in low and high site index c lasses . 4.4.7 Evaluating Selected Spectral Bands: MPB-Attacked Stands The selected spectral bands were evaluated at both the stand and pixel level. For the pixel level classif ication, all 94 training sites in the MPB-at tacked c lass were combined to generate a single MPB-at tack c lass. Similarly, all 51 training sites in the unattacked c lass were combined into a single unattacked c lass. Analyz ing each site individually was not appropriate as the 145 sites were smal l , ranging from one to few pixels in size. Superv ised classification using a maximum likelihood classif ier was run with two c lasses (MPB-at tacked (green attack) and unattacked) under the mask created by merging all the 145 training sites (216 pixels). The resulting confusion matrix was used to assess the separability of the MPB-a t tacked pixels from the unattacked pixels. 4.5 RESULTS AND DISCUSSION 4.5.1 Spectral Bands for Species The sets of spectral bands for discriminating among Douglas-fir, lodgepole pine and spruce stands for C a s e s A , B, C are given in Table 4.8. Output from the S D A for C a s e C is given in Appendix 4.2. Identical results were achieved for cases A and B, where six spectral bands were selected. In C a s e C , eight spectral bands were selected; five of these matched with C a s e s A and B. The spectral bands in each subset were well distributed over the VNIR_1 (446-750nm), S W I R _ 2 (1160-1340nm) and S W I R _ 3 (1490-1790nm) spectral regions. However, there was a notable lack of any spectral band in the VNIR_2 (770-926nm) and SWIR_1 (970-1105nm) spectral regions. Plant reflectance in the VNIR_1 spectral region is influenced by chlorophyll a and chlorophyll b, in the S W I R _ 2 region it is 125 influenced by water, starch, cel lulose and lignin, and in the S W I R _ 3 region it is influenced by water, lignin, starch, sugar, cel lulose, protein, nitrogen (Curran, 1989). However, at this stage it is not possible to specifically d iagnose and relate these plant characteristics with a tree spec ies of interest. The spectral bands identified in the VNIR_1 and SWIR_2 spectral regions were consistent across the cases . In C a s e C , three additional spectral bands (B140, B143 and B161) were selected in the S W I R _ 3 spectral region. C a s e C had the lowest (highest p value) threshold values for variables to enter and remove from the model . General ly, at lower threshold values a larger number of variables are retained in a discriminant model . For example, in a trial run at threshold p values of 0.15 and 0.16 for entry and removal, respectively, 17 spectral bands were identified for the s a m e set of 42 observations. Table 4.8: Spectral bands selected using different F entry/removal criteria in a step-wise discriminant analysis to discriminate among lodgepole pine, Douglas-fir and spruce. Spectral regions Bands' Central •, Wavelength (nm) Case A , Case B , Case C VNIR 1 B15 498.04 * * B21 559.09 * * * SWIR 2 B104 1184.86 * * * B117 1316.05 * * * SWIR 3 B138 1527.92 * * * B140 1548.02 - - * B143 1568.21 - -B148 1628.81 * * -B161 1759.89 - -Spectral regions similar to that of the bands identified in the VNIR_1 region in this study have been reported to be significant in spec ies discrimination by several researchers (e.g., Gong ef al., 1997; Martin ef al., 1998; Tung ef al., 1999; Van Aardt and Wynne, 2001). The notable lack of spectral bands in the NIR region reconfirmed the findings reported in Chapter 3 that the spectral differences among the three spec ies were not statistically significant in the NIR region. Gong ef al. (1997) reported that bands in the visible region are better than bands in the NIR region. However, Martin ef al. (1998) also found spectral bands in the NIR region (750 nm, 783 nm, 822 nm) useful in spec ies discrimination. The identified spectral subsets in this study did not contain any band in the red region (600-700 nm). Gong ef al. (1997), Van Aardt and Wynne (2001) and Thenkabai l ef al. (2004) also did not identify the red band. However, these observations are at 126 variance with Tung ef al. (1999) who identified bands centered at 630.43 and 670.37 nm as important. The model validation results are shown in Table 4.9 for the three cases . In the calibration dataset, the discriminant functions developed from the selected bands had moderate overall spec ies classification accuracy at 85.7% (Case A and B), and 88 .1% in C a s e C . Spec ies specif ic classif ication accuracy was constant at 81 .3% and 100% for lodgepole pine and spruce, respectively, in all three cases . For Douglas-fir, accuracy increased from 78.6% (Case A and B) to 85 .7% in C a s e C . Both overall and spec ies specif ic identification accuracies were consistently lower using cross-val idation. The overall identification accuracies for the three spec ies were 73.8% (Case A and B) and 81 .0% for C a s e C . The classification accuracies for lodgepole pine stands were the lowest among the three species. The highest accuracies were for spruce stands. Tab le 4.9: Spec ies classif ication accuracies for different spectral band combinations. Ca l i b ra t i on Douglas-fir Lodgepole pine Spruce Overall C a s e A 78.6 81.3 100 85.7 C a s e B 78.6 81.3 100 85.7 C a s e C 85.7 81.3 100 88.1 C r o s s Va l i da t i on Douglas- Lodgepole Spruce Overal l fir pine 78.6 56.3 91.7 73.8 78.6 56.3 91.7 73.8 85.7 68.8 91.7 81.0 The classification results were also examined on an individual test site basis for each of the three cases to check for patterns. Lodgepole pine stands were mainly confused with Douglas-fir stands. For example, the same two Douglas-fir sites were consistently confused with lodgepole pine in all cases . For lodgepole pine, five sites were confused with Douglas-fir and three with spruce. For spruce, two sites were confused with Douglas-fir and one site with lodgepole pine. These sites were distributed in various stand age and site index c lasses in several B E C subzone/variants across the study area. These results are from the classif ication of mean stand level spectral values derived from the training sites, generally ranging from 20-30 pixels in s ize. The results of classifying at a pixel level were also tested. Individual spec ies and overall species classification accuracies achieved using four sets of Hyperion spectral bands and six Landsat ETM+ bands are presented in Table 4.10. In C a s e s A and B, the spectral bands were identical. A fourth set (9 bands) was generated by combining the bands identified in C a s e s A, B and C . These accuracies were computed using the same set of 127 training sites. The training site accuracy for each of the 42 training sites for the combined set is shown in Figure 4.11. Table 4.10: Spec ies classif ication accuracy (%) using different Hyperion spectral bands and Landsat ETM+ data. Douglas-fir Lodgepole pine Spruce Overal l Accuracy (%) Combined Set ABC 9 bands) Douglas-fir 81.76 11.59 6.65 81.261 Lodgepole pine 10.66 82.06 7.28 Spruce 9.00 12.08 78.92 " Case C (8 bands) Douglas-fir . 79.18 13.63 7.19 79.252 Lodgepole pine 11.90 A 8153 6.57 Spruce 10.28 13.62 76.09 Case B (6 bands) Douglas-fir 78.65 13.84 7.51 78.233 Lodgepole pine 13.50 78.86 7.64 Spruce 10.03 13.62 76.35 Case A (6 bands) Douglas-fir 78.65 13.84 7.51 78.233 Lodgepole pine 13.50 78.86 7.64 Spruce 10.03 13.62 76.35 Douglas-fir I 66.09 22.53 11.37 68.844 Lodgepole pine 9.59 74.25;, 16.16 Spruce | 11.57 20.82 67.61 Number of pixels used in validation: Douglas-fir: 932; Lodgepole pine: 563; Spruce: 389 (Total: 1884) 1 Using 9 Hyperion bands B15, B21, B104, B117, B138, B140, B143, B148, B161 (Combined A , B, C) 2 Using 8 Hyperion bands B15, B21, B104, B117, B138, B140, B143, B161 (Case C) 3 Using 6 Hyperion bands B15, B21, B104, B117, B138, B148 (Case A and B) 4 Using 6 Landsat ETM+ bands B 1-5 and B7 The classification accuracies for the training sites of each spec ies were examined to a s s e s s species identification patterns in the study area. For example, when a cover type is classif ied using multiple c lasses derived from individual training sites, it is normal to find crossovers among the c lasses. However, as long as such crossovers are primarily among the same cover types, it is acceptable because the main objective of having multiple c lasses based on individual training sites is to capture the heterogeneity in that cover type. In such cases , mean c lass accuracy is a better 128 100 I 8 0 t 70 3 60 £ 50 o f 40 u S 30 </> "> on ra 20 0 • • • • • • • • • • • • • • • • • • • • • • • • • • • • • I I I I I I I I I I I I I r - C S I r t ^ l O t O N a O a i O T - C N C O ^ - T - C N J C O ^ L T J t O h - Q O C J l O ^ - C N l r O ^ r k O C D ^ C N J C O - ^ - L n c O r ^ - Q O C T J O T - C N l U_LLL1_L1_LL 0 - 0_ 0_ Q_ 0_ CL Q_ Training sites • Douglas-fir • Lodgepole pine • Spruce F igu re : 4 .11 : Individual training site level classif ication accuracy in the combined data set using nine spectral bands. indicator of classification accuracy. Therefore, the species- level mean classification accuracies were derived by using the sum of correctly classif ied pixels across all the training sites for a given species. For example, in C a s e s A and B, 733 pixels were correctly classif ied out of 932 Douglas-fir pixels in the 14 training sites, leading to a mean accuracy of 78.7%. The overall classification accuracies were 78.2%, 79 .3% and 81 .3% for C a s e s A and B (6 bands), C (8 bands), and the combined set (9 bands), respectively. These comparative accuracies provide the relative efficiency of different spectral band combinations in spec ies discrimination and should not be confused with regular accuracy assessment . Goodenough et al. (2002) reported a training site accuracy of 88.4% and test data accuracy of 84 .2% for classification of some forest types (red alder {Alnus rubra), western hemlock (Tsuga heterophylla), lodgepole pine and western red cedar (Thuja plicata)) and general land use / cover c lasses using Hyperion data. By increasing the number of bands from 6 to 9 there was a marginal gain of approximately 3 percent in classification accuracy. This indicates that the first six Hyperion spectral bands distributed over the VNIR-1 , SWIR_2 and S W I R _ 3 spectral regions provided the majority of the discrimination power for the three species. Select ion of the additional spectral bands in the S D A was largely due to multicollinearity in the hyperspectral data. 129 To test how well the selected Hyperion spectral bands discriminated among the three tree spec ies compared to the Landsat bands, the same training area pixels (1884 pixels) were classif ied using the six Landsat ETM+ spectral bands. The overall spec ies level classif ication accuracy was higher for Hyperion than that for Landsat (68.8%). A lso , individual species level classification accuracies were higher for the Hyperion bands than for the Landsat bands. The classification accuracies for Douglas-fir, lodgepole pine and spruce were 81.8%, 8 2 . 1 % and 78.9%, respectively, for the Hyperion combined set of bands, and 66 .1%, 74 .3% and 67.6%, respectively, for the Landsat bands (Table 4.10). In a comparative study on forest spec ies discrimination, Goodenough ef al. (2002) also observed a better performance for Hyperion data than Landsat ETM+ data. For example, for Landsat ETM+ and Hyperion, the training site accuracies they found for lodgepole pine were 38.0% and 84.2%, respectively. The overall training si te-based and test data-based accuracies for all their 17 c lasses were 88.2% and 84.2%, respectively, using the Hyperion bands and 67.6% and 61 .3%, respectively, for the Landsat ETM+ bands. 4.5.2 Spectral Bands for Identifying MPB-Attacked Stands Using 145 training sites (94 MPB-at tacked and 51 unattacked), a subset of 9 Hyperion spectral bands (Table 4.11) was identified through S D A (Appendix 4.3). The selected bands were distributed over the V N I F M (446-750 nm), SWIR_1 (970-1105 nm), S W I R _ 2 (1160-1340 nm) and S W I R _ 3 (1490-1790 nm) spectral regions. However, the major concentration of the selected spectral bands was in the SWIR_1 region. From examining specif ic bands/spectral regions on their own, it is apparent that MPB-at tacked and unattacked lodgepole pine stands were most different in the 750-1040 nm spectral range. The differences ranged from about 20 to 200 D N , with the attacked stands having lower reflectance (Figure 4.12). Validation of the spectral bands suitable for separating MPB-at tacked stands from unattacked stands was carried out using cross validation using the ' leave-one-out" procedure and by using an independent validation dataset. A calibration dataset consisting of 102 samples (70%; MPB-at tacked: 66; unattacked: 36) was used in model development and cross validation. The validation data set consisted of the remaining 43 samples (30%; 28 MPB-at tacked and 15 unattacked sites). The M P B identification results are presented in Table 4.12. In each case , all the MPB-at tacked and unattacked stands were correctly classif ied. 130 Table 4.11: Hyperion spectral bands for M P B attack detection identified using step-wise discriminant analysis. Spectral Region Selected Bands Wavelength (nm) VNIR_1 B14 487.87 SWIR_1 B85 993.17 B87 1013.30 B89 1033.49 B90 1043.59 B93 1073.89 B95 1094.09 S W I R _ 2 B103 1174.77 S W I R _ 3 B136 1507.73 2500 2000 z Q 1500 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700 1800 Wavelength (nm) UA GA Figure 4.12: Mean spectral reflectance of unattacked (UA: 51 sites) and attacked (GA: 94 sites) lodgepole pine stands. A total of 216 pixels were classif ied using the maximum likelihood classif ier. The 120 pixels of the MPB-at tacked c lass were spread over 94 sites and the 96 pixels of the MPB-unat tacked c lass came from 51 sites. Ninety-eight pixels of the MPB-at tacked c lass and 77 pixels of the M P B -unattacked c lass were correctly classif ied, resulting in an identification accuracy of 81 .7% and 80.2%, respectively. The overall identification accuracy was 81.0% (Table 4.13). 131 Table 4.12: S D A - b a s e d classification of MPB-at tacked and unattacked stands. .iiClassiflcationlResulfsIr: - '•*•*# Predicted Group Membership Total Class MPB-Attacked Unattacked Training Calibration MPB-At tacked 66 - 66 data Unattacked - 36 36 (70%) Cross-val idat ion (leave- MPB-At tacked 66 - 66 one-out method) Unattacked - 36 36 Test data (30%) MPB-At tacked 28 - 28 Unattacked - 15 15 Table 4.13: Training site based identification accuracy (%) of MPB-at tacked and unattacked stands at the pixel level (n = 216). MPB-Attacked MPB-Unattacked ['Overall Accuracy (%) MPB-At tacked 81.67 18.33 81.02 Unattacked 19.79 80.21 Kappa Coefficient: 0.62 Using the identified spectral subset of Hyperion bands, MPB-at tacked stands could be identified with moderate accuracy. However, the training site accuracies are indicative only, not definitive. Therefore, more validation studies in different areas need to be conducted. The identified spectral bands, although distributed well over the 400-1800 nm spectral range, were mainly concentrated in the NIR plateau region (750-1200 nm). This spectral region was reported by Heller (1968) as a region of interest in the early detection of MPB-at tacked stands. Ahern (1988) reported differences between the foliar reflectance of attacked and unattacked trees to be highly significant throughout the NIR plateau. Ahern also stated that "bark beetle stress has the most pronounced effect in the pigment region of current foliage and in the cell structure region of previous foliage" of green-attacked trees. These inferences were drawn from a needle level study. It is not possible to make this distinction at the stand level because of a number of factors, including the presence of the foliage of various age c lasses . The NIR plateau region has also been found to be one of the spectral regions influenced by defoliators. For example, Leckie et al. (1988) reported progressively reduced spectral reflectance in this region with increasing levels of defoliation by the spruce budworm (Choristoneura fumiferana (Clem.)) in balsam fir (Abies balsamea (L.)) trees. Lawrence (2003) reported better discriminability of Douglas-fir beetle early attack from unattacked Douglas-fir stands in the NIR plateau region, especial ly at 1007 nm and 1069 nm. Douglas-fir stands at the early attack stage had lower spectral reflectances than unattacked stands. 132 Spectral reflectance in the NIR region is influenced by leaf internal cell structure, which in turn, is strongly affected by the leaf water status; starch, protein and oil are the other foliar constituents affecting the spectral reflectance (Curran, 1989). Spectral reflectance in this region is caused by multiple reflections and scattering of light in the spongy mesophyl l structure (Knipling, 1970; Colwel l , 1974). Leaf reflectance changes as water deficit stress develops in all tree spec ies (Knipling, 1970; Wooley, 1973; Gausman , 1974). M P B introduces blue stain fungi into a successful ly attacked tree. Water stress is caused by the movement of this fungi through the phloem and xylem t issues, disrupting water transportation by plugging trachieds in the xylem t issue, releasing gas bubbles in the water conducting t issue, and destroying ray parenchyma cells that partially control water movement (Nebeker et al., 1993). It is uncertain how soon the physiological changes triggered by successfu l M P B attack manifest in the spectral reflectance. General ly, attacked trees, except for visible symptoms on the bole, show no visible change until the following spring. However, it has been reported that "most successful ly attacked trees stop conducting moisture after about three or four weeks" (Les Safranyik in ( B C M O F , 1985), page 48). In a recent review on remote sensing of M P B (Wulder ef al., 2006), it is reported that effects due to water stress in lodgepole pine foliage can be detected within 45-90 days of attack. The MPB-at tacked stands used in this study were mainly at the green attack stage in August 2002 when the Hyperion data were acquired. The Hyperion data acquisit ion was about 6 weeks after the annual beetle flights that year. 4.6 CONCLUSIONS In this study, weighted linear combinations of hyperspectral bands suitable for discriminating Douglas-fir, lodgepole pine, and spruce stands, and band combinations suitable for discriminating green-attacked stands from unattacked stands were identified from an original set of 161 spectral bands in the 400-800 nm spectral region (Figure 4.13). For species identification, the most important bands were: i) V N I R _ 1 : B15 (498.04 nm), B21 (559.09 nm); ii) S W I R _ 2 : B104 (1184.86 nm), B117 (1316.05 nm); and iii) S W I R _ 3 : B138 (1527.92 nm), B140 (1548.02 nm), B143 (1568.21 nm), B148 (1628.81 nm) and B161 (1759.89 nm). This subset provided higher overall spec ies identification accuracy, as well as much better discrimination among the three species, than broadband Landsat ETM+ data. However, it is necessary to validate this spectral band set in various other forest areas before using it operationally to generate spec ies level maps for MPB-at tack detection. 133 400 600 )00 1000 1200 W a v e l e n g t h (nm) 1 400 1 600 1 800 - Cal ibrated - Non-Cal ib ra ted «• Atmospher ic Absorp t ion » R e c o m m e n d e d bands =• B a n d s used in study * Bands- fo r -spec ies t Bands - fo r -MPB Figure 4.13: Summary of steps in the selection of spectral band combinations suitable for conifer species discrimination and for differentiating MPB-attacked stands from unattacked stands. For discriminating MPB-attacked stands from unattacked stands, the most important spectral bands were: i) VNIR_1: B14 (487.87 nm); ii) SWIR_1: B85 (993.17 nm), B87 (1013.30 nm), B89 (1033.49 nm), B90 (1043.59 nm), B93 (1073.89), B95 (1094.09); iii) SWIR_2: B103 (1174.77); and iv) SWIR_3: B136 (1507.73). The most important spectral region for discrimination was 970-1105 nm covering the NIR plateau, whereas for species discrimination the major concentration of selected spectral bands was in the SWIR_3 spectral region (1490-1790 nm). These findings emphasize that narrow well placed spectral bands are necessary in the MPB attack detection process. There are three benefits from this study. Firstly, the spatial resolution of the Hyperion is 30 m, the same as that of multispectral Landsat satellite. This should make the results more widely applicable. Most of the previous hyperspectral studies for species identification or MPB attack detection were based on spectra measured either from leaves/needles or in-situ measurements of trees. Although these studies have been valuable in understanding the biophysical and physiological characteristics of attacked and unattacked trees, the results from these studies can not be linearly scaled up to understand spectral characteristics at stand level. Secondly, the insights gained through this study will be useful in analyzing and using hyperspectral data available from existing or proposed space-borne sensors with a similar spatial resolution. Finally, information on the optimum spectral regions will help users to make decisions regarding the trade-off between the number, location, and 134 widths of spectral bands for data acquisit ion from configurable sensors such as the Compact Airborne Spectrometer Imager (CASI). 135 4.7 REFERENCES Ahern, F. J., 1988. The effects of bark beetle stress on the foliar spectral reflectance of lodgepole pine, International Journal of Remote Sensing, 9:1451-1468. Aldakheel, Y. Y., and F. M. Danson, 1997. Spectral reflectance of dehydrating leaves: measurements and modeling, International Journal of Remote Sensing, 18:3683-3690. BCMOF, 1985. Mountain pine beetle symposium, Ministry of Forests, B.C., Smithers, B.C. Carter, G. A., 1993. Responses of leaf spectral reflectance to plant stress, American Journal of Botany, 80:239-243. Carter, G. A., 1994. Ratios of leaf reflectance in narrow wavebands as indicators of plant stress, International Journal of Remote Sensing, 15:697-703. Carter, G. A., and A. K. Knapp, 2001. Leaf optical properties in higher plants: linking spectral characteristics to stress and chlorophyll concentration, American Journal of Botany, 88:677-684. Carter, G. A., M. R. Seal, and T. Haley, 1998. Airborne detection of southern pine beetle damage using key spectral bands, Canadian Journal of Forest Research, 28:1040-1045. Cheng, X., Y. Tao, Y. R. Chen, and X. Chen, 2005. Feature extraction and band selection methods for hyperspectral imagery applied for identifying for defects. In: Proceedings of The International Society For Optical Engineering-Spie Conference, pp 5996:5996. Colwell, J. E., 1974. Vegetation canopy reflectance, Remote Sensing of Environment, 3:175-183. CSIRO, 2004. Evaluation of the EO-1 Hyperion Hyperspectral instrument & its applications at Australian validation sites 2001 -2003 (Collected papers written by the Australian members of the NASA Science Validation Team), CSIRO Earth Observation Centre. Cudahy, T. J., A. P. Rodger, P. S. Barry, P. Mason, M. Quigley, M. Folkman, and J. S. Pearlman, 2002. Assessment of the stability of the Hyperion SWIR module for hyperspectral mineral mapping using multi-date images from Mount Fitton, Australia, In: Proceedings of the IGARSS, Vol. 6, pp 3504-3506. Curran, P. J., 1989. Remote sensing of foliar chemistry, Remote Sensing of Environment, 30:271-278. Datt, B., T. R. McVicar, T. G. Van Niel, D. L. B. Jupp, and J. S. Pearlman, 2003. Preprocessing EO-1 Hyperion hyperspectral data to support the application of agricultural indices, IEEE Transactions of Geoscience and Remote Sensing, 41:1246-1259. Franklin, S. E., 1994. Discrimination of subalpine forest species and canopy density using CASI, SPOT PLA and Landsat TM data, Photogrammetric Engineering & Remote Sensing, 60:1233-1241. Gausman, H. W., 1974. Leaf reflectance of near-infrared, Photogrammetric Engineering, 40:183-191. Gitelson, A. A., M. N. Merzlyak, and H. K. Lichtenthaler, 1996. Detection of red edge position and chlorophyll content by reflectance measurements near 700nm, Journal of Plant Physiology, 148:501-508. 136 Gong , P., R. P u , and B. Y u , 1997. Conifer spec ies recognition: an exploratory analysis of in situ hyperspectral data, Remote Sensing of Environment, 62:189-200. Goodenough, D., A . S . Bhogal , A . Dyk, A . Hollinger, Z . Mah , K. O. Niemann, J . Pear lman, H. C h e n , T. Han, J . Love, and S . McDona ld , 2002. Monitoring forests with Hyperion and ALI , In Proceedings of the Geoscience and Remote Sensing Symposium, IGARSS '02, I E E E , pp 882-885. Goodenough, D. G . , A . Dyk, K. O. Niemann, J . S . Pear lman, H. C h e n , T. Han, M. Murdoch, and C . West, 2003. Process ing Hyperion and ALI for forest classif ication, IEEE Transactions of Geoscience and Remote Sensing, 41:1321-1331. Heath, J . , 2001. The detection of mountain pine beetle green attacked lodgepole pine using Compact Airborne Spectrographic Imager (CASI) data, M .Sc . dissertation, Faculty of Forestry, University of British Columbia, Vancouver, B.C. ,77 p. Heller, R. C , 1968. Previsual Detection of Ponderosa Pine Trees Dying from Bark Beetle Attack, In: Proceedings of the 5th Symposium on Remote Sensing of Environment, Wil low Run Lab, University of Michigan, Ann Arbor, Michigan, pp 387-434. Huberty, C . J . , 1994. Applied discriminant analysis, Wi ley and Sons , New York, p. Hunt, R. S . , 1995. P ine needle casts and blights in the Pacif ic region, Forest Pest Leaf. No. 43. , Canad ian Forestry Serv ice, Pacif ic Forestry Centre, Victoria, B .C . Jupp, D. L. B., B. Datt, T. R. McVicar , T. G . Van Niel , J . S . Pear lman, J . Lovel l , and E. A . King, 2002. Improving the analysis of Hyperion red edge index from an agricultural area, In: Proceedings ofSPIE, 4898:78-92. Knipling, E. B., 1970. Physical and physiological basis for the reflectance of visible and near-infrared radiation from vegetation, Remote Sensing of Environment, 1:155-159. Lawrence, R., 2003. Early detection of Douglas-fir beetle infestation with subcanopy resolution hyperspectral imagery, Western Journal of Applied Forestry, 18:1-5. Leckie, D. G . , P. M. Teillet, D. P. Ostaff, and G . Fedosejevs, 1988. Sensor band selection for detecting current defoliation caused by the Spruce Budworm, Remote Sensing of Environment, 26:31-50. Leckie, D. G . , X . Y u a n , D. P. Ostaff, H. P iene, and D. A . M a c L e a n , 1992. Analys is of high resolution multispectral M E I S imagery for spruce budworm damage assessment on a single tree basis, Remote Sensing of Environment, 40:125-136. Li l lesand, T. M., and R. W. Kiefer, 2000. Remote sensing and image interpretation, John Wiley & Sons , New York, 724 p. Martin, M. E., S . D. Newman, J . D. Aber, and R. G . Congal ton, 1998. Determining forest spec ies composit ion using high spectral resolution remote sensing data, Remote Sensing of Environment, 65:249-254. Meidinger, D., and J . Pojar, 1991. Ecosystems of British Columbia, B C Ministry of Forests, Victoria, . 137 Nebeker, T. E., J . D. Hodges, and C . A . Blanche, 1993. Host response to bark beetle and pathogen colonization, Beetle-pathogen interactions in conifer forests, (T. D. Schowalter and G . M. Filip, Editors), Academic Press , New York, pp. 157-173. Pear lman, J . S . , P. S . Barry, C . Sega l , J . Shepansk i , D. B ieso, and S. Carman , 2003. Hyperion, a space-based imaging spectrometer, IEEE Transactions of Geoscience and Remote Sensing, 41:1160-1173. Richardson, A . D., and G . P. Berlyn, 2002. Changes in foliar spectral reflectance and chlorophyll f luorescence of four temperate spec ies following branch cutting, Tree Physiology, 22:499-506. Runesson , U. T., 1991. Considerations for Early Remote Detection of Mountain Pine Beetle in Green-Foliaged Lodgepole Pine, P h D dissertation, Faculty of Forestry, University of British Columbia, Vancouver, B.C. ,237 p. Tabachnick, B. G . , and L. S . Fidell, 2001. Using multivariate statistics, 4th edition, Al lyn and Bacon , Boston. Thenkabai l , P. S. , E. A . Enc lona, E. A . Ashton, C . Legg, and M. J . D. Dieu, 2004. Hyperion, I K O N O S , ALI, and ETM+ sensors in the study of Afr ican rainforests, Remote Sensing of Environment, 90:23-43. Tung, F., F. Y . M a , and W. L. S iu , 1999. Band selection using hyperspectral data of subtropical trees, In: Proceedings of the 20th Asian Conference on Remote Sensing, Hongkong, Ch ina. U S G S , 2003. EO-1 Data User Guide, U R L : http://eo1 .usgs.gov/userGuide/hyp prod.html, U S G S , (last date accessed : 26 February 2006). V a n Aardt, J . , and R. H. Wynne, 2001. Spectral separability among six southern tree species, Photogrammetric Engineering & Remote Sensing, 67:1367-1375. Warner, T. A . , and M. C . Shank, 1997. Spatial auto-correlation analysis of hyperspectral imagery for feature selection, Remote Sensing of Environment, 60:58-70. White, J . D., G . C . Kroh, and J . E. Pinder, 1995. Forest mapping at Lassen Volcanic National Park, Cali fornia, using Landsat T M data and a geographical information system, Photogrammetric Engineering & Remote Sensing, 61:299-305. Wi lson, B. A. , J . E. Luther, and T. D. T. Stuart, 1998. Spectral reflectance characteristics of Dutch Elm d isease, Canadian Journal of Remote Sensing, 24:200-205. Wooley, J . T., 1973. Change of leaf dimensions and air volume with change in water content, Plant Physiology, 41:815-816. Wulder, M. A. , C . C . Dymond, J . C . White, D. G . Leckie, and A. L. Carrol l , 2006. Surveying mountain pine beetle damage of forests: A review of remote sensing opportunities, Forest Ecology and Management, 221:27-41. Yang , C . M., and C . H. Cheng , 2001. Spectral characteristics of rice plants infested by brown planthoppers, Proc. Natl. Sci. Counc. ROC(B), 25:180-186. 138 5.0 CONCLUSIONS AND RECOMMENDATIONS 5.1 APPROACH Lodgepole pine is one of the most important commercial tree spec ies in western North Amer ica. Mature lodgepole pine trees are the preferred hosts of M P B , which is its most destructive pest. After finding a suitable host tree and successful ly overcoming its resistance through mass attack, M P B colonize the attacked tree and start their l ifecycle. M P B also carry blue staining fungi, which grows within the conducting t issues of the attacked tree, disrupting its internal water and food transportation. The combined actions of the growing beetle and fungi kill the attacked tree. Although M P B is normally endemic in the lodgepole pine forests, under favorable climate and food conditions its population reaches epidemic proportions and causes extensive mortality in mature trees. The foliage symptoms of the attacked tree are used to characterize the M P B attack stages. During the green/current attack stage, foliage of the attacked trees is still green. At this stage visual discrimination between MPB-at tacked and unattacked trees is difficult based on the foliage color alone; M P B attack is identifiable only through ground surveys. The foliage of the attacked trees turns red in the summer following the beetle attack (red attack). This facilitates visually identifying M P B -attacked trees from the air. Conventional aerial surveys to identify and map M P B infestations are conducted at this stage. Through these surveys, M P B infestations are categorized into several severity levels such as trace (<1 percent), light (1-10 percent), moderate (11-30 percent), severe (31-50 percent) and very severe (> 50 percent), depending upon the percent of red-attack trees. M P B infestations are small and scattered in nature at most severity levels. At the red-attacked stage, high resolution, normal colour, digital or analogue aerial photography is effective for M P B attack detection and mapping. However, for planning the harvest of beetle-killed trees and taking measures to control beetle populations more effectively, foresters need information on M P B infestations over large areas, as early as possible after successful M P B attacks. Remote sensing-based M P B attack detection efforts have primarily focused on the green- and red-attacked stages. The literature on M P B attack detection has demonstrated the utility of remote sensing in M P B red attack detection at various sca les . The success of M P B attack detection at the green-attack stage has been mixed and warranted more research, especial ly using medium spatial resolution multi-spectral and hyperspectral remote sensing data. 139 The work presented in this thesis is applications-oriented, examining three key research issues important in developing methods for M P B damage assessment using satel l i te-based remote sensing data. My first research question was whether it was possible to provide good location information on M P B - attacked stands at an earlier date than conventional methods based on visual identification of red-attacked trees. It was hypothesized that it should be possible to identify M P B attack before the red-attacked stage using spectral characteristics associated with the changing physiological conditions within the MPB-at tacked trees. Being able to identify stands containing significant numbers of mature lodgepole pine is an important initial component of M P B attack detection. I expected that lodgepole pine might behave differently spectrally across a landscape because of its large ecological amplitude. However, the magnitude of the spectral variability was not known. My second research question was whether spectral variability in mature lodgepole pine trees was significant enough to warrant consideration in designing M P B attack detection across a broad landscape. Lastly, I expected that narrow band hyperspectral data may contain more diagnostic features to help early detection of MPB-at tacked stands than multi-spectral data. Existing research in this area was inconclusive. A lso , medium resolution hyperspectral data were hitherto unavailable for research. My third research question was whether hyperspectral bands' from medium resolution satellite imagery were useful in tree spec ies identification and early detection of MPB-at tacked stands. The first two research questions were examined using multispectral remote sensing data from Landsat ETM+ and the third question was addressed using hyperspectral data from the Hyperion sensor onboard the EO-1 satellite. Both sources provide data at a spatial resolution of 30 m. Landsat 7 ETM+ provides very cost-effective multi-temporal remote sensing data. EO-1 Hyperion is the first sensor to provide hyperspectral data from a space-borne platform. These data were acquired between 1999 and 2002 and covered a part of the Wil l iams Lake Forest District in central British Columbia. In this area during this period, M P B infestations were on the rise. However, prior to 2002 the area was largely free of large contiguous M P B infestations. Most of the infestations were characterized by small patches of MPB-at tacked trees scattered within mature lodgepole pine stands. 140 5.2 FINDINGS The results of the studies included in this thesis indicate that it is possible to provide good information on MPB-at tacked stands at an earlier date than conventional methods using a composite approach, combining host spec ies identification, change detection, and sub-pixel level image analysis algorithms. The spectral differences between MPB-at tacked and unattacked lodgepole pine stands could be detected and mapped with 69% identification accuracy. The mapping results could be provided four months earlier than conventional surveys. The overall identification accuracy for all forest health factors (i.e., including damaging agents other than M P B ) was 84%. The multi-step approach developed for this study is not restricted to any specif ic remote sensing data source, pest or d isease, or area; thus it should have wide applicability. This finding is important as the B C Ministry of Forest identified early detection of M P B infestations as a research area of high importance. Using the method developed in this study, three test studies on the early detection of M P B -attacked stands were conducted between 2001 and 2004 in areas ranging from 234,000 ha to 1,800,000 ha in s ize. The results from these test studies were utilized operationally by local forest managers in planning and implementing harvest operations in MPB-k i l led lodgepole stands in the study area. In one of the ground surveys associated with the studies, it was observed that more than 5 0 % of the identified MPB-at tacked sites contained only new M P B attack, giving forest managers a head-start in harvest planning and increasing the efficacy of preventative harvest a imed at reducing the spread of M P B . Significant spectral variability in mature lodgepole pine stands, as well as in stands with Douglas-fir and spruce, was documented and the effects of three variables (stand age: age c lasses 61-80, 81-140 and >140 years; site c lass: low, medium, high; site ecology: 9 B E C subzones/variants) on stand-level spectral reflectance were examined. Of the three variables, site ecology showed the greatest influence on the spectral signatures of the three species. This finding has relevance in operational remote sensing-based M P B attack detection because lodgepole pine grows in diverse habitats covering a wide range of ecological conditions. In this study, I used the B E C system because it is the standard for B C . However, other ecological classif ications in areas outside B C could provide a suitable criterion for stratifying forested areas prior to landscape level mapping of M P B infestations. Accurate information on the spatial distribution and extent of a host tree species is essential for reliable mapping of pest or d iseases infestations. The existing B C provincial forest inventory does not provide information on the detailed spatial distribution of various tree species, although it does provide an estimate of spec ies composit ion at the stand (polygon) level. Therefore, mapping of the detailed spatial distribution of mature lodgepole pine using satellite data was one of the essential 141 components of the methodology for M P B attack detection developed in this study. Knowledge about the causes and magnitude of spectral variability will help in better identifying and mapping the spatial distribution of lodgepole pine in large study areas. This study also demonstrated the utility of hyperspectral data for tree spec ies identification. Better spec ies identification accuracies were achieved using a subset of nine narrow spectral bands (498 nm, 559 nm, 1184 nm, 1316 nm, 1527 nm, 1548 nm, 1568 nm, 1628 nm, and 1759 nm) compared to broadband (Landsat) multispectral data. The distribution of these spectral bands indicates the signif icance of the S W I R spectral region in tree spec ies discrimination. Identification accuracies using the Hyperion data for Douglas-fir, lodgepole pine and spruce were 81.8%, 82 .1% and 78.9%, respectively. The corresponding accuracies from Landsat ETM+ were 66 .1%, 74 .3% and 67.6%, respectively. The hyperspectral data were also useful in detecting M P B infestations when the attacked trees were still predominantly at the green attack stage. A subset of nine narrow spectral bands (487 nm, 993 nm, 1013 nm, 1033 nm, 1043 nm, 1073, 1094 nm, 1174 nm, and 1507 nm) was found useful in discriminating attacked from non-attacked stands. The identification accuracy for MPB-at tacked and unattacked stands was 81 .7% and 80.2%, respectively. The identified spectral bands were mainly concentrated within a small (990-1100 nm) spectral region, signifying its importance in early M P B attack detection. The results from this study can also be useful in designing future studies using hyperspectral data acquired from other spaceborne hyperspectral sensors, such as H E R O (Hyperspectral and Environmental Resource Observer; spatial resolution 30m, spectral range 400-2500 nm), proposed by Canada . The spectral regions identified as significant in this study could be used for guiding new acquisition of hyperspectral data from programmable airborne hyperspectral sensors. Although the primary focus of this thesis was early detection of MPB-at tacked lodgepole pine stands, all bark beetles of the genus Dendroctonus (e.g., Douglas-fir beetle, spruce bark beetle) behave in a similar fashion. Throughout this thesis, Douglas-fir and spruce stands were included in the analysis along with lodgepole pine stands. Therefore, methods and results appl icable to M P B , theoretically, can a lso be extrapolated to damage from these two other beetles. These beetles can also be major pests and periodically attain epidemic proportions causing extensive damage and mortality in mature stands of the respective host trees. 142 5 . 3 RECOMMENDATIONS At endemic and incipient epidemic population levels of M P B , infestations are smal l , scattered and light in severity. A review of the 2005 B C provincial aerial overview surveys indicates that trace, light, moderate, severe and very severe severity levels of M P B covered 26.4%, 26.4%, 24 .1%, 13.8% and 9.3% area, respectively (Appendix 5.1). This distribution pattern indicates that the majority of M P B infestations are still sub-pixel in s ize even during the peak of a M P B epidemic. At the Landsat and Hyperion spatial resolutions of 30m, these sub-pixel s ized M P B infestations are represented by mixed pixels in an image. Such mixed pixels contain varying proportions of unattacked as well as attacked lodgepole pine trees at various attack stages. In this study, the generic term " M P B attack" was used to label MPB-at tacked stands identified using satellite data. It is desirable not to label M P B infestations identified using satellite data as green attack, red attack, or grey attack. These three attack stages have specif ic entomological definitions and are mainly appl icable at a tree level. Use of entomological definitions to characterize a mixed stand, which is what is actually identified using medium spatial resolution satellite data, is erroneous and potentially misleading to end users. Accuracy assessment of M P B attack for this study included collecting field data and comparing these data with results from the remote sensing techniques. Field work is time consuming and governed by a variety of factors such as access , timing of the survey relative to the M P B attack symptoms, and field crew experience. There is a need for appropriate protocols to assess the accuracy of M P B attack identified from early detection studies. Specif ic guidelines on the sampling strategy, sample s ize, and sampling unit s ize (the minimum number of trees to be used), applicable for operational trials are needed. Wall-to-wall field surveys over large areas to a s s e s s identification accuracy are not a practical solution. Additional studies are also needed to test the effectiveness of the key hyperspectral bands identified in this study for discriminating tree spec ies types and M P B -attacked stands in a wider range of forests. In this study and elsewhere, it was observed that stands that were successful ly attacked by M P B had lower vigor than unattacked stands even before the beetle attack. This needs further investigation. Understanding the characteristics of host trees which affect their susceptibility to bark beetle attack under changing environmental conditions may improve our ability to detect attack early. The greatest chal lenge in the study of interactions among bark beetles and pathogens, host conifers, and the environment is to integrate the roles of all the elements effectively, leading to a dynamic index reflecting the changing ground situations. Multi-temporal vigor assessment of mature lodgepole pine might help in developing such an index. Spectral vegetation indices reflect the integrated effects of phenology, tree conditions and the local environment. Vegetation index trajectories generated using multi-temporal pre- and pos t -MPB attack images might provide clues on host susceptibility. 143 5.4 OVERALL CONCLUSIONS The results of this thesis demonstrate that early detection of MPB-at tacked stands using multispectral and hyperspectral data is possible at a scale and resolution to be of practical use to forest managers. Spectral variability in mature lodgepole stands should be considered in landscape-level M P B attack detection. B E C subzone/variants or a similar ecological unit provide a suitable stratification criterion. For tree spec ies identification, hyperspectral data provided better identification accuracies than multispectral data at the same spatial resolution. 144 Appendix 2.1 Appendix 2.2 Mountain pine beetle Red attack distribution (Red colour) in the Cariboo Forest Region (2001) Appendix 2.3 Sensor Characteristics Landsat-7 Enhanced Thematic Mapper (ETM+) (Launched: April 1999) Band Number Spectral range Spatial Resolution Spectral Region 1 4 5 0 - 5 1 5 nm 30 m Blue 2 5 2 5 - 6 0 5 nm 30 m Green 3 630 - 690 nm 30 m Red 4 750 - 900 nm 30 m Near Infrared (NIR) 5 1 5 5 0 - 1750 nm 30 m Short Wave Infrared (SWIR) 6 1 0 . 4 0 - 1 2 . 5 0 urn 60 m Thermal Infrared (TIR) 7 2090 - 2350 nm 30 m Short Wave Infrared (SWIR) 8 520 - 900 nm 1 5 m Panchromat ic (Source: Landsat 7, Sc ience data Users handbook (http:/ / landsathandbook.gsfc.nasa.gov/handbook/handbook htmls/chapter3/chapter3.html, Last accessed : Dec 15, 2006). IKONOS (Launched: September 1999) Band Number Spectral range Spatial Resolution Spectral Region 1 4 4 4 - 5 1 6 nm 4 m Blue 2 506 - 595 nm 4 m Green 3 631 - 6 9 7 nm 4 m Red 4 757 - 852 nm 4 m Near Infrared (NIR) 5 525 - 928 nm 1 m Panchromat ic (Source: S p a c e Imaging http://www.spaceimaging.com/products/ikon6s/spectral.htm, last accessed : Dec 15, 2006). 147 Appendix 3.1 Equations Digital number (DN) to at-satellite radiance and at-satell ite-radiance to at-satellite-reflectance conversions are done using equation 1 and 2 respectively (Markham and Barker, 1986). Lx = Gain* . DNA + BiasA (1) Pi = 7t.Lx .tf/ESUNi. sin (0) (2) Where: L = At satellite radiance p = At-satellite reflectance A = ETM+ band number d = Earth-Sun distance in astronomical unit ESUN = Mean Solar exoatmospheric irradiance G = Sun elevation angle, provided in the scene header file For each Landsat7 ETM+ image the gains, biases, and sun elevation are provided in the header file of the image. The Solar irradiance E S U N * and the tables for Earth-Distance are given in LandsatT Sc ience Data Users Handbook (Irish, 2000). The estimate of Kappa is given by: K H A T = (P0-Pc)/(1-Pc) Where P0 is the 'True agreement' and Pc is the 'Chance agreement' (Congalton and Green , 1999) K H A T statistic is computed as: K= / V 2 - £ ( * „ - x + , . ) i=l where: r = number of rows in the error matrix; Xu = the number of observations in row /' and column /' (on the major diagonal); x,+ = total of observations in row /'; x+i = total of observations in column /; and N = total number of observations included in matrix. 148 Appendix 3.2 Site Index distribution in the study area o o CL 4500 4000 3500 3000 2500 2000 1500 1000 500 0 Lodgepole pine 1 • I  i i n i i . i 13 15 17 19 21 23 25 27 29 S i t e I n d e x 1600 » c 1400 o 0 ) 1200 >. o 1000 0. 800 o i - 600 CB S2 400 E 200 z Douglas-fir 13 15 17 19 21 23 25 27 29 S i t e I n d e x Spruce Spp. 9 11 13 15 17 19 21 23 25 27 29 S i t e I n d e x 149 Appendix 3.3 Spec ies , age and B E C subzone/variant sample sets meeting the stand selection criteria (e.g., stand size: >=5 ha, spec ies content: >=80%, stand age >=61 year, minimum number of forest pixels in a stand: 9) BEC subzone/variants Douglas-fir Lodgepole pine "Spruce species AC4 AC567 AC89 AC4 AC567 AC89 AC4 AC567 AC89 ICHmk3 * * IDFdk3 * * * * * IDFdk4 * * * * * * * * IDFxm * * * * S B P S d c * * * * * S B P S m k * * * * * * S B P S x c * * * * * S B S d w l * * * * * S B S d w 2 * * * * * * * * 150 Appendix 3.4 Dataset 1: Pooled calibration and validation samples EC Variants Species Age Class AC4 AC567 AC89 Cal. Val. Cal. Val. Cal. Val. a b a b a b a b a b a b ICHmk3 F D - - - - 25 22 26 23 15 13 15 11 PL - - - - 27 24 27 23 - - - -S - - - - - - - - 12 12 12 12 IDFdk3 F D 75 71 75 72 75 67 75 69 75 66 72 72 P L 75 69 75 68 75 71 75 69 75 68 75 71 S - - - - 38 33 39 36 12 11 13 11 IDFdk4 F D 15 15 17 16 75 73 75 72 75 68 75 73 P L 75 73 75 71 75 75 75 72 75 65 75 72 S - - - - 17 15 17 16 28 25 28 28 IDFxm F D 43 43 44 42 75 70 75 75 75 74 75 74 P L - - - - 51 49 52 51 - - - -S - - - - - - - - - - - -S B P S d c FD - - - - - - - - - - - -PL 75 71 75 72 75 74 75 75 75 71 75 66 S - - - - 16 16 16 15 28 25 29 25 S B P S m k F D - - - - - - - - 30 28 32 25 P L 75 69 75 71 75 67 75 69 75 72 75 74 S - - - - 27 23 27 25 33 29 33 28 S B P S x c F D - - - - - - - - 50 49 51 51 P L 75 74 75 72 75 73 75 74 75 75 75 74 S - - - - 16 15 17 14 59 57 59 56 S B S d w l F D 14 12 15 14 75 . 62 75 63 75 67 75 63 PL - - - - 75 68 75 67 - - - -S - - - - 40 37 40 37 15 14 15 13 S B S d w 2 F D 13 12 12 11 51 49 52 48 74 68 74 71 P L 30 23 30 25 75 66 75 68 25 22 25 22 S - - - - 21 19 22 18 11 11 11 10 Total 565 532 568 534 1154 1068 1160 1079 1067 990 1069 1002 Total samples for calibration 532 + 1068 + 990 = 2590 Total samples for validation 534+1079+1002 = 2615 a: Initial sample; b: Sample after outlier r emova l , Ca l . ; Calibration; Va l . : Validation Total population: Calibration (dataset a: 2786, dataset b: 2590, difference: 7%) Total population: Validation (dataset: a 2797, dataset b: 2615, difference: 6.5%) 151 Appendix 3.5 General Linear Model: Douglas-fir Between-Subjects Factors N A G E _ C L A S S AC4 153 AC567 343 AC89 433 B E C VAR ICHmk3 35 IDFdk3 204 IDFdk4 156 IDFxm 187 SBPSmk 28 SBPSxc 49 SBSdwl 141 SBSdw2 129 SITE INDEX High 247 Low 276 Medium 406 Multivariate Tests(c) Effect Value F Hypothesis df Error df Sig. Intercept Pillai's Trace 1.000 1133251.714(a) 6.000 912.000 .000 Wilks' Lambda .000 1133251.714(a) 6.000 912.000 .000 Hotelling's Trace 7455.603 1133251.714(a) 6.000 912.000 .000 Roy's Largest Root 7455.603 1133251.714(a) 6.000 912.000 .000 A G E C L A S S Pillai's Trace .125 10.109 12.000 1826.000 .000 Wilks' Lambda .877 10.336(a) 12.000 1824.000 .000 Hotelling's Trace .139 10.562 12.000 1822.000 .000 Roy's Largest Root .127 19.384(b) 6.000 913.000 .000 B E C VAR Pillai's Trace .906 23.285 42.000 5502.000 .000 Wilks' Lambda .339 26.468 42.000 4281.111 .000 Hotelling's Trace 1.322 28.657 42.000 5462.000 .000 Roy's Largest Root .688 90.166(b) 7.000 917.000 .000 SITEJNDEX Pillai's Trace .083 6.591 12.000 1826.000 .000 Wilks' Lambda .918 6.676(a) 12.000 1824.000 .000 Hotelling's Trace .089 6.760 12.000 1822.000 .000 Roy's Largest Root .080 12.183(b) 6.000 913.000 .000 a Exact statistic b The statistic is an upper bound on F that yields a lower bound on the signif icance level, c Design: l n t e r cep t+AGE_CLASS+BEC V A R + S I T E INDEX 152 T e s t s of B e t w e e n - S u b j e c t s E f fec ts S o u r c e Dependen t Va r iab le T y p e III S u m of S q u a r e s df M e a n S q u a r e F S i g . C o r r e c t e d M o d e l B a n d 1 33.006(a) 11 3.001 55.049 .000 B a n d 2 87.839(b) 11 7.985 77.029 .000 B a n d 3 147.669(c) 11 13.424 84.728 .000 B a n d 4 534.238(d) 11 48.567 20.206 .000 B a n d 5 1597.196(e) 11 145.200 58.481 .000 B a n d 7 2 . 5 3 5 ( f ) 11 .230 69.910 .000 Intercept B a n d 1 3156.524 1 3156.524 57911.246 .000 B a n d 2 12522.845 1 12522.845 120799.490 .000 B a n d 3 5779.299 1 5779.299 36475.900 .000 B a n d 4 127604.810 1 127604.810 53090.093 .000 B a n d 5 26928.538 1 26928.538 10845.889 .000 B a n d 7 608.323 1 608.323 184507.991 .000 A G E _ C L A S S B a n d 1 .125 2 .062 1.146 .318 B a n d 2 .357 2 .178 1.720 .180 B a n d 3 :917 2 .459 2.894 .056 B a n d 4 136.470 2 68.235 28.389 .000 B a n d 5 19.987 2 9.993 4.025 .018 B a n d 7 .041 2 .020 6.151 .002 B E C V A R B a n d 1 16.537 7 2.362 43.342 .000 B a n d 2 34.114 7 4.873 47.011 .000 B a n d 3 60.153 7 8.593 54.236 .000 B a n d 4 152.091 7 21.727 9.040 .000 B a n d 5 589.990 7 84.284 33.947 .000 B a n d 7 .923 7 .132 40.014 .000 S ITE I N D E X B a n d 1 1.883 2 .941 17.270 .000 B a n d 2 5.337 2 2.668 25.740 .000 B a n d 3 7.365 2 3.682 23.242 .000 B a n d 4 44.060 2 22.030 9.166 .000 B a n d 5 86.654 2 43.327 17.451 .000 B a n d 7 .138 2 .069 20.873 .000 E r ro r B a n d 1 49.982 917 .055 B a n d 2 95.062 917 .104 B a n d 3 145.291 917 .158 B a n d 4 2204.057 917 2.404 B a n d 5 2276.758 917 2.483 B a n d 7 3.023 917 .003 Tota l B a n d 1 7203.967 929 B a n d 2 27984.121 929 B a n d 3 13225.804 929 B a n d 4 281828.767 929 B a n d 5 65235.868 929 B a n d 7 1359.278 929 C o r r e c t e d Tota l B a n d 1 82.988 928 B a n d 2 182.901 928 B a n d 3 292.960 928 B a n d 4 2738.295 928 B a n d 5 3873.954 928 153 B a n d 7 5.559 I 928 a R Squared = .398 (Adjusted R Squared = .390) b R Squared = .480 (Adjusted R Squared = .474) c R Squared = .504 (Adjusted R Squared = .498) d R Squared = .195 (Adjusted R Squared = .185) e R Squared = .412 (Adjusted R Squared = .405) f R Squared = .456 (Adjusted R Squared = .450) 154 Appendix 3 . 6 General Linear Model: Lodgepole pine Between-Subjects Factors N AGE CLASS AC4 379 AC567 567 AC89 373 BEC VAR ICHmk3 24 IDFdk3 208 IDFdk4 213 IDFxm 49 SBPSdc 216 SBPSmk 208 SBPSxc 222 SBSdwl 68 SBSdw2 111 SITE INDEX High 254 Low 650 Medium 415 Multivariate Tests(c) Effect Value F Hypothesis df Error df Sig. Intercept Pillai's Trace 1.000 1790467.597(a) 6.000 1301.000 .000 Wilks' Lambda .000 1790467.597(a) 6.000 1301.000 .000 Hotelling's Trace 8257.345 1790467.597(a) 6.000 1301.000 .000 Roy's Largest Root 8257.345 1790467.597(a) 6.000 1301.000 .000 AGE CLASS Pillai's Trace .043 4.820 12.000 2604.000 .000 Wilks' Lambda .957 4.823(a) 12.000 2602.000 .000 Hotelling's Trace .045 4.826 12.000 2600.000 .000 Roy's Largest Root .030 6.513(b) 6.000 1302.000 .000 BEC VAR Pillai's Trace .674 20.645 48.000 7836.000 .000 Wilks' Lambda .461 22.768 48.000 6405.528 .000 Hotelling's Trace .902 24.418 48.000 7796.000 .000 Roy's Largest Root .460 75.062(b) 8.000 1306.000 .000 SITEJNDEX Pillai's Trace .130 15.130 12.000 2604.000 .000 Wilks' Lambda .872 15.315(a) 12.000 2602.000 .000 Hotelling's Trace .143 15.500 12.000 2600.000 .000 Roy's Largest Root .116 25.088(b) 6.000 1302.000 .000 a Exact statistic b The statistic is an upper bound on F that yields a lower bound on the significance level, c Design: lntercept+AGE_CLASS+BEC_VAR+SITEJNDEX 155 Tests of Between-Subjects Effects Source Dependent Variable Type III Sum of Squares df Mean Square F Sig. Corrected Model Band 1 123.154(a) 12 10.263 108.427 .000 Band 2 220.893(b) 12 18.408 133.909 .000 Band 3 492.570(c) 12 41.047 170.290 .000 Band 4 435.759(d) 12 36.313 26.845 .000 Band 5 4444.642(e) 12 370.387 113.762 .000 Band 7 6.886(f) 12 .574 133.091 .000 Intercept Band 1 4946.113 1 4946.113 52255.988 .000 Band 2 19399.195 1 19399.195 141121.708 .000 Band 3 10739.168 1 10739.168 44552.508 .000 Band 4 154803.462 1 154803.462 114440.601 .000 Band 5 60593.739 1 60593.739 18611.018 .000 Band 7 1000.801 1 1000.801 232132.792 .000 AGE CLASS Band 1 .744 2 .372 3.930 .020 Band 2 1.175 2 .587 4.273 .014 Band 3 2.207 2 1.103 4.578 .010 Band 4 11.352 2 5.676 4.196 .015 Band 5 14.552 2 7.276 2.235 .107 Band 7 .032 2 .016 3.723 .024 BEC VAR Band 1 30.778 8 3.847 40.646 .000 Band 2 51.964 8 6.496 47.253 .000 Band 3 109.169 8 13.646 56.612 .000 Band 4 305.268 8 38.158 28.209 .000 Band 5 997.727 8 124.716 38.306 .000 Band 7 1.529 8 .191 44.337 .000 SITE INDEX Band 1 8.096 2 4.048 42.765 .000 Band 2 13.622 2 6.811 49.546 .000 Band 3 30.830 2 15.415 63.950 .000 Band 4 8.992 2 4.496 3.324 .036 Band 5 282.614 2 141.307 43.402 .000 Band 7 .426 2 .213 49.372 .000 Error Band 1 123.615 1306 .095 Band 2 179.528 1306 .137 Band 3 314.805 1306 .241 Band 4 1766.622 1306 1.353 Band 5 4252.074 1306 3.256 Band 7 5.631 1306 .004 Total Band 1 11638.800 1319 Band 2 44308.448 1319 Band 3 26535.680 1319 Band 4 335497.882 1319 Band 5 160298.957 1319 Band 7 2260.046 1319 Corrected Total Band 1 246.769 1318 Band 2 400.421 1318 Band 3 807.375 1318 Band 4 2202.381 1318 Band 5 8696.716 1318 156 B a n d 7 12.516 1318 a R Squared = .499 (Adjusted R Squared = .494) b R Squared = .552 (Adjusted R Squared = .548) c R Squared = .610 (Adjusted R Squared = .607) d R Squared = .198 (Adjusted R Squared = .190) e R Squared = .511 (Adjusted R Squared = .507) f R Squared = 550 (Adjusted R Squared = 546) 157 Appendix 3 . 7 General Linear Model: Spruce Between-Subjects Factors N AGE CLASS AC567 158 AC89 184 BEC VAR ICHmk3 12 IDFdk3 44 IDFdk4 40 SBPSdc 41 SBPSmk 52 SBPSxc 72 SBSdwl 51 SBSdw2 30 SITE INDEX High 59 Low 143 Medium 140 Multivariate Tests(c) Effect Value F Hypothesis df Error df Sig. Intercept Pillai's Trace 1.000 1108039.501(a) 6.000 326.000 .000 Wilks' Lambda .000 1108039.501(a) 6.000 326.000 .000 Hotelling's Trace 20393.365 1108039.501(a) 6.000 326.000 .000 Roy's Largest Root 20393.365 1108039.501(a) 6.000 326.000 .000 AGE CLASS Pillai's Trace .032 1.776(a) 6.000 326.000 .103 Wilks' Lambda .968 1.776(a) 6.000 326.000 .103 Hotelling's Trace .033 1.776(a) 6.000 326.000 .103 Roy's Largest Root .033 1.776(a) 6.000 326.000 .103 BEC VAR Pillai's Trace .959 8.994 42.000 1986.000 .000 Wilks' Lambda .276 11.538 42.000 1532.528 .000 Hotelling's Trace 1.871 14.445 42.000 1946.000 .000 Roy's Largest Root 1.448 68.451(b) 7.000 331.000 .000 SITEJNDEX Pillai's Trace .106 3.052 12.000 654.000 .000 Wilks' Lambda .896 3.069(a) 12.000 652.000 .000 Hotelling's Trace .114 3.087 12.000 650.000 .000 Roy's Largest Root .089 4.865(b) 6.000 327.000 .000 a Exact statistic b The statistic is an upper bound on F that yields a lower bound on the signif icance level, c Design: l n t e r cep t+AGE_CLASS+BEC V A R + S I T E J N D E X 158 Tests of Between-Subjects Effects Source Dependent Variable Type III Sum of Squares df Mean Square F Sig. Corrected Model bandl 9.659(a) 10 .966 28.578 .000 Band 2 21.931(b) 10 2.193 35.803 .000 Band 3 42.431(c) 10 4.243 60.214 .000 Band 4 51.989(d) 10 5.199 1.960 .037 Band 5 238.234(e) 10 23.823 18.800 .000 band7 .405(f) 10 .040 34.354 .000 Intercept bandl 1541.739 1 1541.739 45614.088 .000 Band 2 6301.141 1 6301.141 102869.915 .000 Band 3 2848.985 1 2848.985 40430.037 .000 Band 4 56723.660 1 56723.660 21382.306 .000 Band 5 13024.315 1 13024.315 10278.091 .000 band7 319.390 1 319.390 271214.328 .000 AGE CLASS bandl .013 1 .013 .391 .532 Band 2 .138 1 .138 2.260 .134 Band 3 .112 1 .112 1.585 .209 Band 4 2.064 1 2.064 .778 .378 Band 5 .391 1 .391 .309 .579 band7 9.53E-005 1 9.53E-005 .081 .776 BEC VAR bandl 6.514 7 .931 27.534 .000 Band 2 11.816 7 1.688 27.558 .000 Band 3 25.481 7 3.640 51.657 .000 Band 4 31.836 7 4.548 1.714 .105 Band 5 128.212 7 18.316 14.454 .000 band7 .230 7 .033 27.857 .000 SITE INDEX bandl .217 2 .108 3.206 .042 Band 2 .971 2 .485 7.922 .000 Band 3 1.001 2 .501 7.106 .001 Band 4 12.686 2 6.343 2.391 .093 Band 5 11.063 2 5.531 4.365 .013 band7 .010 2 .005 4.057 .018 Error bandl 11.188 331 .034 Band 2 20.275 331 .061 Band 3 23.325 331 .070 Band 4 878.087 331 2.653 Band 5 419.441 331 1.267 band7 .390 331 .001 Total bandl 2414.334 342 Band 2 9784.398 342 Band 3 4541.743 342 Band 4 86832.005 342 Band 5 21321.682 342 band7 490.423 342 Corrected Total bandl 20.847 341 Band 2 42.206 341 Band 3 65.756 341 Band 4 930.076 341 Band 5 657.674 341 159 band7 .794 341 a R Squared = .463 (Adjusted R Squared = .447) b R Squared = .520 (Adjusted R Squared = .505) c R Squared = .645 (Adjusted R Squared = .635) d R Squared = .056 (Adjusted R Squared = .027) e R Squared = .362 (Adjusted R Squared = .343) f R Squared = 509 (Adjusted R Squared = .494) 160 Appendix 3 . 8 General Linear Model: Interactions Between-Subjects Factors N SPP FD 929 PL 1319 SP 342 AC AC4 532 AC567 1068 AC89 990 BECVAR ICHmk3 71 IDFdk3 456 IDFdk4 409 IDFxm 236 SBPSdc 257 SBPSmk 288 SBPSxc 343 SBSdwl 260 SBSdw2 270 SI High 560 Low 1069 Medium 961 Multivariate Tests(d) Effect Value F Hypothesis df Error df Sig. Intercept Pillai's Trace 1.000 3073483.351(b) 6.000 2549.000 .000 Wilks' Lambda .000 3073483.351(b) 6.000 2549.000 .000 Hotelling's Trace 7234.563 3073483.351(b) 6.000 2549.000 .000 Roy's Largest Root 7234.563 3073483.351(b) 6.000 2549.000 .000 SPP * AC Pillai's Trace .081 6.982 30.000 12765.000 .000 Wilks' Lambda .921 7.058 30.000 10198.000 .000 Hotelling's Trace .084 7.110 30.000 12737.000 .000 Roy's Largest Root .052 21.927(c) 6.000 2553.000 .000 SPP* BECVAR Pillai's Trace .837 18.829 132.000 15324.000 .000 Wilks' Lambda .383 20.175 132.000 14832.549 .000 Hotelling's Trace 1.117 21.558 132.000 15284.000 .000 Roy's Largest Root .544 63.134(c) 22.000 2554.000 .000 SPP * SI Pillai's Trace .135 9.826 36.000 15324.000 .000 Wilks' Lambda .870 10.058 36.000 11196.204 .000 Hotelling's Trace .144 10.220 36.000 15284.000 .000 161 R o y ' s La rges t .095 40.538(c) 6.000 2554.000 .000 R o o t a Computed using alpha = .05 b Exact statistic c The statistic is an upper bound on F that yields a lower bound on the signif icance level. d Design: Intercept+SPP * A C + S P P * B E C V A R + S P P * SI T e s t s of B e t w e e n - S u b j e c t s E f fec ts S o u r c e Dependen t Type III S u m of df M e a n F S i g . Va r i ab le S q u a r e s S q u a r e C o r r e c t e d B a n d 1 196.383(b) 35 5.611 77.552 .000 M o d e l B a n d 2 407.759(c) 35 11.650 100.910 .000 B a n d 3 1017.673(d) 35 29.076 153.616 .000 B a n d 4 2268.819(e) 35 64.823 34.145 .000 B a n d 5 10991.534(f) 35 314.044 115.434 .000 B a n d 7 16.467(g) 35 .470 132.869 .000 Intercept B a n d 1 8551.767 1 8551.767 118198.052 .000 B a n d 2 34065.895 1 34065.895 295064.511 .000 B a n d 3 16749.592 1 16749.592 88491.185 .000 B a n d 4 307218.473 1 307218.473 161821.754 .000 B a n d 5 83871.425 1 83871.425 30828.901 .000 B a n d 7 1712.517 1 1712.517 483623.719 .000 S P P * A C B a n d 1 .882 5 .176 2.438 .033 B a n d 2 1.670 5 .334 2.893 .013 B a n d 3 3.236 5 .647 3.419 .004 B a n d 4 149.885 5 29.977 15.790 .000 B a n d 5 34.931 5 6.986 2.568 .025 B a n d 7 .073 5 .015 4.109 .001 S P P * B a n d 1 53.829 22 2.447 33.818 .000 B E C V A R B a n d 2 97.895 22 4.450 38.542 .000 B a n d 3 194.803 22 8.855 46.781 .000 B a n d 4 489.195 22 22.236 11.712 .000 B a n d 5 1715.929 22 77.997 28.670 .000 B a n d 7 2.682 22 .122 34.432 .000 S P P * SI B a n d 1 10.195 6 1.699 23.485 .000 B a n d 2 19.929 6 3.321 28.769 .000 B a n d 3 39.196 6 6.533 34.513 .000 B a n d 4 65.738 6 10.956 5.771 .000 B a n d 5 380.331 6 63.389 23.300 .000 B a n d 7 .573 6 .095 26.966 .000 E r ro r B a n d 1 184.785 2554 .072 B a n d 2 294.865 2554 .115 B a n d 3 483.421 2554 .189 B a n d 4 4848.767 2554 1.898 B a n d 5 6948.273 2554 2.721 B a n d 7 9.044 2554 .004 Tota l B a n d 1 21257.102 2590 B a n d 2 82076.967 2590 B a n d 3 44303.227 2590 | B a n d 4 704158.654 2590 162 Band 5 246856.507 2590 Band 7 4109.747 2590 Corrected Total Band 1 381.168 2589 Band 2 702.624 2589 Band 3 1501.093 2589 Band 4 7117.586 2589 Band 5 17939.806 2589 Band 7 25.511 2589 a Computed using alpha = .05 b R Squared = .515 (Adjusted R Squared = .509) c R Squared = .580 (Adjusted R Squared = .575) d R Squared = .678 (Adjusted R Squared = .674) e R Squared = .319 (Adjusted R Squared = .309) f R Squared = .613 (Adjusted R Squared = .607) g R Squared = .645 (Adjusted R Squared = .641) 163 Appendix 4.1 Absorption features in visible, near-infrared wavebands that have been related to particular foliar chemical concentrations. (Adapted from Curran (1989)) Wavelength (pm) Chemical(s) Remote sensing considerations 0.43 Chlorophyll a Atmospher ic scattering 0.46 Chlorophyll b 0.64 Chlorophyll b 0.66 Chlorophyll a 0.91 Protein 0.93 Oil 0.97 Water, starch 0.99 Starch 1.02 Protein 1.04 Oil 1.12 Lignin 1.20 Water, cel lulose, starch, lignin 1.40 Water 1.42 Lignin 1.45 Starch, sugar, lignin, water Atmospher ic absorption 1.49 Cel lu lose, sugar I I I I I I I I I I Rapid decrease in signal-to-noise ratios of sensors I I I I I I I I I I I I I I 1.51 Protein, nitrogen 1.53 Starch 1.54 Starch, cel lulose 1.58 Starch, sugar 1.69 Lignin, starch, protein, nitrogen 1.78 Cellulose, sugar, starch 1.82 Cel lu lose 1.90 Starch 1.94 Water, lignin, protein, nitrogen, starch, cel lulose 1.96 Sugar, starch 1.98 Protein 2.00 Starch 2.06 Protein, nitrogen 2.08 Sugar, starch 2.10 Starch, cel lulose 2.13 Protein 2.18 Protein, nitrogen 2.24 Protein 2.25 Starch 2.27 Cel lu lose, sugar, starch 2.28 Starch, cel lulose 2.30 Protein, nitrogen 2.31 Oil 2.32 Starch 2.34 Cel lu lose 2.35 Cel lu lose, protein, nitrogen Chemica ls in bold have a wavelength of stronger absorption 164 Appendix 4.2 Step Wise Discriminant Analysis: Tree Species (Case C) Analys is C a s e Process ing Summary Unweighted C a s e s N Percent Val id 42 100.0 Exc luded Missing or out-of-range group codes 0 .0 At least one missing discriminating variable 0 .0 Both missing or out-of-range group codes and at least one missing discriminating variable 0 .0 Total 0 .0 Total 42 100.0 Stepwise Statistics Var iables Entered/Removed(a,b,c,d) Step Entered Removed Wilks' Lambda Statistic df1 df2 df3 Exact F Statistic df1 df2 Sig. 1 B117 .764 1 2 39.000 6.028 2 39.000 .005 2 B21 .512 2 2 39.000 7.558 4 76.000 .000 3 B15 .371 3 2 39.000 7.909 6 74.000 .000 4 B148 .271 4 2 39.000 8.297 8 72.000 .000 5 B104 .222 5 2 39.000 7.850 10 70.000 .000 6 B138 .182 6 2 39.000 7.624 12 68.000 .000 7 B143 .152 7 2 39.000 7.387 14 66.000 .000 8 B140 .117 8 2 39.000 7.678 16 64.000 .000 9 B148 .133 7 2 39.000 8.211 14 66.000 .000 10 B161 .111 8 2 39.000 8.006 16 64.000 .000 At each step, the variable that minimizes the overall Wilks' Lambda is entered. a Max imum number of steps is 194. b Maximum signif icance of F to enter is . 10. c Minimum signif icance of F to remove is .11. d F level, tolerance, or VIN insufficient for further computation. Summary of Canonical Discriminant Functions Eigenvalues Function Eigenvalue % of Variance Cumulative % Canonical Correlation 1 3.886(a) 82.2 82.2 .892 2 .844(a) 17.8 100.0 .676 a First 2 canonical discriminant functions were used in the analysis. 165 Wilks' Lambda Test of Function(s) Wilks' Lambda Chi-square df Sig. 1 through 2 .111 78.034 16 .000 2 .542 21.715 7 .003 Standardized Canonica l Discriminant Function Coefficients Function 1 2 B15 2.099 .014 B21 -4.120 -1.249 B104 1.415 2.338 B117 .621 -3.442 B138 -1.539 -5.095 B140 -4.112 .584 B143 8.740 1.061 B161 -2.564 5.923 Canonica l Discriminant Function Coefficients Function 1 2 B15 .083 .001 B21 -.132 -.040 B104 .020 .033 B117 .010 -.057 B138 -.035 -.116 B140 -.080 .011 B143 .186 .023 B161 -.087 .201 (Constant) -1.507 9.896 Unstandardized coefficients Funct ions at Group Centroids Species Function 1 2 Douglas-fir 1.975 -.849 Lodgepole pine .379 1.114 Spruce -2.809 -.496 Unstandardized canonical discriminant functions evaluated at group means Classification Statistics Classif ication Process ing Summary Processed 42 Excluded Miss ing or out-of-range group codes 0 At least one missing discriminating variable 0 Used in Output 42 166 Prior Probabilit ies for Groups Species Prior Cases Used in Analysis Unweighted Weighted Douglas-fir .333 14 14.000 Lodgepole pine .333 16 16.000 Spruce .333 12 12.000 Total 1.000 42 42.000 Classif ication Function Coefficients Species Douglas-fir Lodgepole pine Spruce B15 .615 .483 .216 B21 -.002 .131 .616 B104 .005 .038 -.078 B117 .542 .414 .473 B138 1.149 .978 1.276 B140 -1.162 -1.013 -.777 B143 -.468 -.720 -1.350 B161 -.042 .492 .446 (Constant) -185.016 -161.569 -176.071 Fisher's linear discriminant functions Classif icat ion Results(b.c) Species Predicted Group Membership Total 1 2 3 Original Count Douglas-fir 12 2 0 14 Lodgepole pine 2 13 1 16 Spruce 0 0 12 12 % Douglas-fir 85.7 14.3 .0 100.0 Lodgepole pine 12.5 81.3 6.3 100.0 Spruce .0 .0 100.0 100.0 Cross-validated(a) Count Douglas-fir 12 2 0 14 Lodgepole pine 3 11 2 16 Spruce 0 1 11 12 % Douglas-fir 85.7 14.3 .0 100.0 Lodgepole pine 18.8 68.8 12.5 100.0 Spruce .0 8.3 91.7 100.0 a C ross validation is done only for those c a s e s in the analysis. In cross validation, each case is classif ied by the functions derived from all cases other than that case. b 8 8 . 1 % of original qrouped cases correctly classif ied. c 81.0% of cross-validated grouped cases correctly classif ied. 167 Appendix 4.3 Step Wise Discriminant Analysis: MPB Attack Analys is C a s e Process ing Summary Unweighted Cases N Percent Valid 102 70.3 Excluded Missinq or out-of-range group codes 0 .0 At least one missing discriminating variable 0 .0 Both missing or out-of-range group codes and at least one missing discriminating variable 0 .0 Unselected 43 29.7 Total 43 29.7 Total 145 100.0 Stepwise Statistics Variables Entered/Removed(a,b,c,d) Step Entered Removed Wilks' Lambda Statistic df1 df2 df3 Exact F Statistic df1 df2 Sig. 1 B87 .564 1 1 100.000 77.149 1 100.000 .000 2 B94 .088 2 1 100.000 514.705 2 99.000 .000 3 B85 .062 3 1 100.000 497.138 3 98.000 .000 4 B95 .051 4 1 100.000 455.338 4 97.000 .000 5 B103 .045 5 1 100.000 411.507 5 96.000 .000 6 B161 .037 6 1 100.000 413.679 6 95.000 .000 7 B89 .034 7 1 100.000 385.883 7 94.000 .000 8 B90 .017 8 1 100.000 690.037 8 93.000 .000 9 B93 .016 9 1 100.000 643.386 9 92.000 .000 10 B94 .016 8 1 100.000 730.235 8 93.000 .000 11 B136 .015 9 1 100.000 684.607 9 92.000 .000 12 B161 .015 8 1 100.000 777.900 8 93.000 .000 13 B14 .014 9 1 100.000 717.641 9 92.000 .000 At each step, the variable that minimizes the overall Wilks' Lambda is entered. a Max imum number of steps is 194. b Maximum signif icance of F to enter is .05. c Minimum signif icance of F to remove is .10. d F level, tolerance, or VIN insufficient for further computation. 168 Wilks' Lambda Step Number of Variables Lambda df1 df2 df3 Exact F Statistic df1 df2 Sig. 1 1 .564 1 1 100 77.149 1 100.000 .000 2 2 .088 2 1 100 514.705 2 99.000 .000 3 3 .062 3 1 100 497.138 3 98.000 .000 4 4 .051 4 1 100 455.338 4 97.000 .000 5 5 .045 5 1 100 411.507 5 96.000 .000 6 6 .037 6 1 100 413.679 6 95.000 .000 7 7 .034 7 1 100 385.883 7 94.000 .000 8 8 .017 8 1 100 690.037 8 93.000 .000 9 9 .016 9 1 100 643.386 9 92.000 .000 10 8 .016 8 1 100 730.235 8 93.000 .000 11 9 .015 9 1 100 684.607 9 92.000 .000 12 8 .015 8 1 100 777.900 8 93.000 .000 13 9 .014 9 1 100 717.641 9 92.000 .000 Summary of Canonical Discriminant Functions Eigenvalues Function Eigenvalue % of Variance Cumulative % Canonical Correlation 1 70.204(a) 100.0 100.0 .993 a First 1 canonical discriminant functions were used in the analysis. Wilks' Lambda Test of Function(s) Wilks' Lambda Chi-square df Sig. 1 .014 407.360 9 .000 Standardized Canonica l Discriminant Function Coefficients Function 1 B14 - .255 B85 1.001 B87 2.742 B89 7.501 B90 -6.697 B93 -1.967 B95 -1.561 B103 -1.180 B136 .575 169 Canonica l Discriminant Function Coefficients Function 1 B14 -.022 B85 .006 B87 .019 B89 .052 B90 -.047 B93 -.016 B95 -.014 B103 -.017 B136 .034 (Constant) 12.629 Unstandardized coefficients Funct ions at Group Centroids Code V Function 1 1 -6.127 2 11.233 Unstandardized canonical discriminant functions evaluated at group means Classification Statistics Classif icat ion Process ing Summary Processed 145 Excluded Missing or out-of-range group codes 0 At least one missing discriminating variable 0 Used in Output 145 Prior Probabil it ies for Groups Code_V Prior Cases Used in Analysis Unweighted Weighted 1 .500 66 66.000 2 .500 36 36.000 Total 1.000 102 102.000 Classif icat ion Function Coefficients Code V 1 2 B14 1.619 1.233 B85 -.233 -.124 B87 -.295 .040 B89 -.931 -.031 B90 .945 .133 B93 .378 .107 B95 .220 -.017 B103 .217 -.074 B136 -.329 .258 (Constant) -391.805 -216.881 Fisher 's linear discriminant functions Classif icat ion Results(b,c,d) Code_V Predicted Group Membership Total 1 2 Cases Selected Original Count 1 66 0 66 2 0 36 36 % 1 100.0 .0 100.0 2 .0 100.0 100.0 Cross-validated(a) Count 1 66 0 66 2 0 36 36 % 1 100.0 .0 100.0 2 .0 100.0 100.0 Cases Not Selected Original Count 1 28 0 28 2 0 15 15 % 1 100.0 .0 100.0 2 .0 100.0 100.0 a C ross validation is done only for those cases in the analysis. In cross validation, each case is classif ied by the functions derived from all cases other than that case. b 100.0% of selected original grouped cases correctly classif ied. c 100.0% of unselected original grouped c a s e s correctly classif ied. d 100.0% of selected cross-val idated grouped cases correctly classif ied. 171 Mountain pine beetle update, 2006 BRITISH C O L U M B I A U P D A T E For Immediate Release Ministry of Forests and Range March 2006" MOUNTAIN PINE B E E T L E A F F E C T S 8.7 MILLION H E C T A R E S 2005 aerial overview surveys show the mountain pins beetle infestation proceeding according to Ministry o f Forests arid Range forecasts, with about &,7 million hectares of B . C . forests ui the red-attack stage. This is up from just over .seven oii i i ion hectajgs Uie previous year. The red-attack stage occurs in the year following the initial attack (green-attack). The beetles have left the tree and the needles have turned red, indicating that the tree Is dead because the beetles, and the fungus (hay carry, have cut the tree off from its supply of water and nutrients. In subsequent years tlie needles fell off the trees and only the branches remain (grey^attack). The ministry relies on aeriaJ overview surveys for monitoring infestation levels and assessing the amount of damage caused by the beetle. These surveys are non-cumulative and only record red-attack for the year. Affected pine in the green-attack or grey-attack stages is not included as part of the survey. These annual surveys are part of B.C. ' s Mountain Pine Beetle Action Plan. The survey data provides the provincial government and industry with the Information to make the necessary decisions in managing the Infestation. O f 1he total area affected i n 2005 by the mountain pine beetle: • 2.J mill ion hectares showed trace amounts of red-attack (less than one per cent of the trees killed in the past year) • 2.3 mill ion hectares sustained light amounts of" red-attack (one to 10 per cent o f the trees killed in the past year) • 2.1 mill ion hectares sustained moderate mortality ( i 1 to }0 percent of the trees killed in the past year) • 1.2 mill ion hectares experienced severe levels o f red-attack (31 to 50 per cent of the killed in tlie past year) • 744,000 hectares saw very severe signs o f Infestation (more than 50 percent of the trees killed in the past year) Using the annual aerial overview surveys, the Ministry of Forests and Range and the Council of Forest Industries are able to determine the amount of timber affected by the mountain pine beetle. A s of Fall 2005, it's estimated that the mountain pine beetle infestation has now affected more than 400 mil l ion cubic metres of timber, up from about 243 million cubic metres at this lime last year. This volume survey Is cumulative and records timber in all three stages of attack - green, ted and grey - as well as infested pine that has already been harvested and processed. See the attached backgrounder for a regional and district, breakdown of the total area affected by the mountain pine beetle in 2005. Contact: Max Cleeveley Communications Director Ministry of Forests and Range 250 3S7-S486 Visit the Province's website at www.gov.bc.ca for online information and services. 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0074966/manifest

Comment

Related Items