The remote sensing of forest canopy gaps in a selectively logged submontane tropical forest reserve in Kenya Colbert Mutiso Jackson A thesis submitted to the Faculty of Science, at the University of Witwatersrand, in fulfilment of the academic requirements for the degree of Doctor of Philosophy in Geography and Environmental Sciences November 2022 Johannesburg South Africa ii Abstract Forests constitute 31% (about 4 billion ha) of the land area of the earth, and tropical forests cover is about 2 billion ha. Tropical forests play a significant role in supporting the earth's life and natural ecosystems. But many conservation and protection efforts have not been effective, as they are being cleared in many countries for timber and expansion of agricultural land. Few undisturbed tropical forests remain today, and unsustainable selective logging (SL) is probably the single biggest factor contributing to the global degradation of tropical forests. The amount of forest degradation that is not detected using currently available remote sensing (RS) techniques is unknown. Many methods used to map SL in tropical forests using low/medium spatial resolution datasets have a high rate of false detections. As a result, reliable and operational methods for monitoring SL in tropical forests ought to be utilized. Recently, very high resolution (VHR) RS datasets have caught the interest of researchers studying SL in tropical forests. Therefore, this study was aimed to apply spectral-texture analysis approach to detect canopy gaps caused by illegal logging of Ocotea usambarensis (East African camphor) in Mt. Kenya Forest Reserve (MKFR) in Chuka, Tharaka Nithi County, Kenya using the VHR WorldView-3 (WV-3) satellite data. Several features were derived from the WV-3 data— however, a large number of features results to longer computing time, and might result in reduced classification accuracy. Therefore, feature ranking measures—the mean decrease accuracy, the mean decrease Gini, and the pairwise feature correlation were used. The Jeffries- Matusita distance, the transformed divergence index, the G-statistic, and the Euclidean distance were used to calculate the separability of the forest landscape classes. First, the study reviewed and discussed RS techniques used to map SL in the tropical forests. Second, the threatened trees species (TS) in the selectively logged MKFR were mapped. Third, gaps in the forest canopy were detected and quantified using two approaches—initially, only spectral features were used to detect gaps in the forest canopy. A total of 55 spectral features were extracted from the WV-3 dataset—23 means (of 15 vegetation indices–VIs and 8 visible-near-infrared– VNIR bands), and 23 standard deviations–SDs (of 15 VIs and 8 VNIR bands). Also extracted were 8 ratios (of 8 VNIR bands), and 1 brightness feature (average of the means of bands 1 to 8). The study also explored the potential of rich textural features combined with color to model canopy gaps using GLCM-, LBP-, and MLBP-based rotation-invariant feature descriptors derived from WV-3 imagery. Due to their excellent performance and clear logic, two advanced machine learning (ML) classification models—the random forest (RF) and support vector machine (SVM) models were used to identify and classify canopy gaps in the spectral and texture domains of the WV-3 data. During the training process the learning parameters of RF (mtry and ntree) and SVM (γ and C) algorithms were optimised to obtain the best possible settings. Finally, the study reviewed Kenya’s forest policy and law on participatory forest management (PFM). The best tree species classification results reported F1-scores of 68.56 ± 2.6% and 64.64 ± 3.4% for RF and SVM, respectively. The RF and SVM models used to map canopy gaps using the spectral features reported average overall accuracies (OAs) of 92.3 ± 2.6% and 94.0 ± 2.1%, respectively. Average kappa coefficients (ĸ) were 0.88 ± 0.03 for RF and 0.90 ± 0.02 for SVM. The user’s accuracy (UA) and producer’s accuracy (PA) were in the range of 84– 100%. The OA for the classification of canopy gaps using textural/spectral features reported values between 80 (RF, block F’s MLBP/ASM) and 95.1% (SVM, block E’s MLBP/CON). The average OA scores were 84.68 ±3.1, 84.54 ±2.5, 84.86 ±3.0, 86.46 ±3.9, 87 ±4.0, and 85.44 ±3.7 for image blocks A, B, C, D, E, and F, respectively, for the RF classifier, and 85.44 ±3.6, 87.2 ±1.8, 86.3 ±4.3, 89.84 ±2.8, 87.28 ±4.5 and 86.12 ±3.6 for the SVM classifier. The iii UA and PA were in the range of 67-75% and 77-100% for the univariate LBP and MLBP models, respectively. Texture fused with colour resulted to higher classification accuracies. Overall, the approach used in this study demonstrated improved ability of VHR satellite data and ML classification models to accurately map fine canopy gaps resulting from SL. Knowledge about where canopy gaps are located and how they are distributed is critical in accurate estimation of carbon densities of forests, and also for managing the proliferation of invasive species, among other applications. LiDAR datasets acquire the three-dimensional (3- D) structure of forest vegetation—repeat surveys can thus detect the removal of individual trees. The integration of optical images and LiDAR data may boost canopy gap classification. iv Preface This doctoral thesis describes work undertaken as part of a program of study at the School of Geography, Archaeology, and Environmental Studies, University of the Witwatersrand, Johannesburg from January 2018 to January 2022, under the supervision of Professor Elhadi Mohammed I. Adam (School of Geography, Archaeology, and Environmental Studies, University of the Witwatersrand, Johannesburg). This document is the result of my investigations, therefore I have exercised reasonable care to ensure that this work is original, and where otherwise stated, all used sources of literature are listed at the end of each chapter. This doctoral thesis has not previously been submitted/accepted in substance for any degree and is not being concurrently submitted in candidature for any degree. Colbert Mutiso Jackson Date: 11/11/2022 As the candidate’s supervisor, I certify the above statement and have approved this thesis for submission. Prof. Elhadi Mohammed I. Adam ___________________ Date: _____________________ A0037623 Typewriter 11.11.2022 v Declaration 1-Plagiarism I, Colbert Mutiso Jackson, declare that: 1. The research reported in this thesis, except where otherwise indicated, is my original research. 2. This thesis has not been submitted for any degree or examination at any other university. 3. This thesis does not contain other persons’ data, pictures, graphs, or other information unless specifically acknowledged as being sourced from other persons. 4. This thesis does not contain other persons’ writing unless specifically acknowledged as being sourced from other researchers. Where other written sources have been quoted, then: a. Their words have been re-written, but the general information attributed to them has been referenced. b. Where their exact words have been used, then their writing has been placed in italics and inside quotation marks and referenced. 5. This thesis does not contain text, graphics, or tables copied and pasted from the Internet unless specifically acknowledged and the source is detailed in the thesis and the References section. Signed vi Declaration 2-Publication and manuscripts 1. Jackson, C.M., and Adam, E. (2020). Remote sensing of selective logging in tropical forests: current state and future directions. iForest - Biogeosciences and Forestry, 13, 286–300. 2. Jackson, C.M., and Adam, E. (2021). Machine learning classification of endangered tree species in a tropical submontane forest using WorldView-2 multispectral satellite imagery and imbalanced dataset. Remote Sensing, 13, 4970. 3. Jackson, C.M., and Adam, E. (2022). A machine learning approach to mapping canopy gaps in an indigenous tropical submontane forest using WorldView-3 multispectral satellite imagery. Environmental Conservation, 49, 255–262. 4. Jackson, C.M., and Adam, E. (in review). Feature extraction and classification of WorldView-3 imagery using GLCM - and MLBP-based rotation-invariant feature descriptors for detection of canopy gaps. Tropical Conservation Science. 5. Jackson, C.M., and Adam, E. (in review). An assessment of forest policy and law on participatory forest management for sustainable forest management in Kenya. Restoration Ecology. vii Dedication To my dear wife, Charity, and our adorable children, Lulu and Lamont for their constant support and prayers for my success viii Acknowledgments This thesis represents the culmination of a process made possible by the contributions of numerous institutions and individuals, for which I am grateful. I would like to express my heartfelt appreciation to the people and organizations listed below for their contributions to this project. The research was carried out at the University of the Witwatersrand's School of Geography, Archaeology, and Environmental Studies, Faculty of Science, in Johannesburg, South Africa. I am extremely grateful to this institution for covering my tuition and providing me with research facilities as well as the time to work on this study. I am extremely grateful to my supervisor, Prof. Elhadi Mohammed I. Adam, for accepting me as a Ph.D. student and guiding me through the entire Ph.D. process. Throughout my time at the University of the Witwatersrand, he advised this thesis and mentored me. His broad knowledge and deep insights have been a valuable asset to my success. I enjoyed working on such an interesting and complex topic as the use of very high resolution (VHR) WorldView- 3 imagery to detect selective logging in Kenya's Mount Kenya Forest Reserve. Thank you for your moral support, encouragement, and inspiration to complete the research. It was always a relief to share taxing issues with you because you always had a solution and a very encouraging attitude. Your suggestions and criticisms influenced the final shape of the thesis. Thank you, Prof. Elhadi Mohammed I. Adam, for allowing me to teach GEOG 3017: Introduction to GIS and Remote Sensing as a lab assistant, tutor, and lecturer. I also want to thank the leadership, lecturers, and support staff in the School of Geography, Archaeology, and Environmental Studies for their help and enabling learning environment. I would like to thank the Kenya Forest Service leadership at the Head Office in Nairobi, particularly the office of the Chief Conservator of Forests, for allowing me to collect data in Mount Kenya Forest Reserve. Mr. James Mburu, Tharaka Nithi County Ecosystem Conservator, and Mr. John Maina, Chief Forester, Chuka Forest Station, deserve special thanks for making it possible for me to visit Mt. Kenya Forest Reserve. I would also like to thank Mwebia and Keter for showing me around the research area and providing armed security while I was collecting field data. They were extremely dedicated to their work. They were extremely helpful and made a significant contribution to this research. I am eternally grateful to both. Thank you for agreeing to be my field assistant, Mr. David Kithinji. You worked tirelessly. Many thanks to my family for their unwavering support while I was studying. Thank you from the bottom of my heart for your love, care, understanding, and encouragement. Your help has been unwavering. You have always shown your affection. I am especially grateful to ix Charity, my dear wife, for her moral support. Thank you, Charity, for being a devoted wife, a true friend, and supportive during all of the difficulties we faced while I was away at school. My Ph.D. would not have been completed without her unwavering support. My success is her reward. Despite coming from different countries, backgrounds, cultures, and age groups, I am grateful to my fellow Ph.D. candidates for the challenges and "home feeling" they provided. This provided me with a reasonably comfortable and conducive study environment. While I may not be able to thank everyone, I would like to express my gratitude to everyone who contributed to the success of my studies in some way. Finally, I thank God for the gifts of life, good health, and wisdom, which allowed me to study uninterrupted. x Table of Contents Abstract ..................................................................................................................................... ii Preface ...................................................................................................................................... iv Declaration 1-Plagiarism ......................................................................................................... v Declaration 2-Publication and manuscripts ......................................................................... vi Dedication ............................................................................................................................... vii Acknowledgments ................................................................................................................. viii Table of Contents ..................................................................................................................... x List of figures ......................................................................................................................... xvi List of tables........................................................................................................................... xix Abbreviations and acronyms ............................................................................................... xxi 1. CHAPTER ONE ............................................................................................................... 1 General introduction ............................................................................................................... 1 1.1 Status of tropical forests ................................................................................................... 2 1.2 Defining Selective logging and canopy gap ..................................................................... 2 1.2.1 Selective logging ....................................................................................................... 2 1.2.2 Canopy gap ................................................................................................................ 2 1.3 Detection of canopy gaps in tropical forests .................................................................... 3 1.4 Research objectives .......................................................................................................... 5 1.5 Scope of the study ............................................................................................................ 6 1.6 The study area .................................................................................................................. 7 1.7 Thesis overview.............................................................................................................. 10 2. CHAPTER TWO ............................................................................................................ 12 Literature review ................................................................................................................... 12 Abstract ................................................................................................................................ 13 2.1 Introduction .................................................................................................................... 14 2.2 Methods .......................................................................................................................... 17 xi 2.2.1 Database search ....................................................................................................... 17 2.2.2 Content analysis ....................................................................................................... 17 2.3 Results of literature review ............................................................................................ 18 2.3.1 Publication details.................................................................................................... 18 2.3.2 Geographical information ........................................................................................ 21 2.3.4 Spatial scale ............................................................................................................. 26 2.3.5 Temporal scale ......................................................................................................... 26 2.3.6 Methods employed to map and characterize selective logging in tropical forests .. 29 2.3.7 Accuracy assessment ............................................................................................... 31 2.4. Discussion ..................................................................................................................... 33 2.4.1 Application of remote sensing ................................................................................. 33 2.4.2 Geographic distribution of the scientific activity .................................................... 35 2.4.3 Multi-temporal monitoring of selective logging ..................................................... 36 2.4.4 Remote sensing techniques applied to selective logging in tropical forests............ 37 2.4.5 Accuracy assessment ............................................................................................... 41 2.5. Conclusion ..................................................................................................................... 41 3. CHAPTER THREE ........................................................................................................ 43 Machine learning classification of endangered tree species using WorldView-2 imagery and imbalanced dataset ......................................................................................................... 43 Abstract ................................................................................................................................ 44 3.1 Introduction .................................................................................................................... 45 3.2 Materials and methods ................................................................................................... 48 3.2.1 Acquisition and pre-processing of WorldView-2 satellite data ............................... 48 3.2.2 Field data collection ................................................................................................. 49 3.2.3 Spectral separability ................................................................................................ 51 3.2.4 Training of random forest and support vector machine classifiers ......................... 52 3.2.5 Class imbalance ....................................................................................................... 53 xii 3.2.6 Measures of model performance ............................................................................. 54 3.3 Results ............................................................................................................................ 55 3.3.1 Spectral separability between tree species ............................................................... 55 3.3.2 Optimization of random forest and support vector machine ................................... 58 3.3.3 Relative importance of variables ............................................................................. 60 3.3.4 Model performance .................................................................................................. 62 3.3.5 The spatial distribution of the endangered tree species ........................................... 68 3.4 Discussion ...................................................................................................................... 70 3.4.1 Spectral separability between the tree species ......................................................... 70 3.4.2 Relative importance of variables ............................................................................. 71 3.4.3 Class imbalance ....................................................................................................... 71 3.4.4 Model Performance ................................................................................................. 73 3.4.5. The Spatial Distribution of Endangered Tree Species ............................................ 74 3.5 Conclusions .................................................................................................................... 74 4. CHAPTER FOUR .......................................................................................................... 76 A machine learning approach to map canopy gaps in tropical forests using WorldView-3 multispectral imagery ............................................................................................................ 76 Abstract ................................................................................................................................ 77 4.1 Introduction .................................................................................................................... 78 4.2 Materials and methods ................................................................................................... 80 4.2.1 Acquisition and pre-processing of satellite data ...................................................... 80 4.2.2 Acquisition of field data .......................................................................................... 81 4.2.3 Feature extraction and selection .............................................................................. 82 4.2.4 Spectral separability ................................................................................................ 85 4.2.5 Pairwise feature comparison .................................................................................... 86 4.2.6 Morphological filtering ............................................................................................ 86 4.2.7 Image classification ................................................................................................. 87 xiii 4.2.8 Measures of model performance ............................................................................. 88 4.3. Results ........................................................................................................................... 88 4.3.1. Explanatory Power of the features extracted from WorldView-3 VNIR Bands .... 88 4.3.2. Optimization of Random Forest and Support Vector Machine .............................. 91 4.3.3 Spectral separability ................................................................................................ 92 4.3.4 Pairwise feature comparison .................................................................................... 94 4.3.5 Logging feature detectability ................................................................................... 96 4.3.6 Model Performance ................................................................................................. 96 4.3.7 Classification maps .................................................................................................. 97 4.4. Discussion ..................................................................................................................... 99 4.4.1 Relative importance of variables ............................................................................. 99 4.4.2 Pairwise feature comparison .................................................................................. 100 4.4.3 Spectral Separability .............................................................................................. 100 4.4.4 Model performance ................................................................................................ 101 4.4.6 Challenges, limitations, and the way forward ....................................................... 102 4.5 Conclusions .................................................................................................................. 103 5. CHAPTER FIVE .......................................................................................................... 104 Feature extraction and classification of canopy gaps using GLCM - and MLBP-based rotation-invariant feature descriptors derived from WorldView-3 imagery ................. 104 Abstract .............................................................................................................................. 105 5.1 Introduction .................................................................................................................. 106 5.2 Materials and methods ................................................................................................. 109 5.2.1 Acquisition and pre-processing of satellite data .................................................... 109 5.2.2 Acquisition of field data ........................................................................................ 109 5.2.3 Feature extraction and selection ............................................................................ 111 5.2.4 Similarity and separability between training signatures ........................................ 117 5.2.5 Image classification ............................................................................................... 117 xiv 5.2.6 Morphological filtering .......................................................................................... 118 5.2.7 Measures of model performance ........................................................................... 119 5.3 Results .......................................................................................................................... 119 5.3.1 Similarity and separability between training signatures ........................................ 119 5.3.2 Optimization of Random Forest and Support Vector Machine ............................. 120 5.3.3 Model performance ................................................................................................ 122 5.3.4 Image classification ............................................................................................... 125 5.4 Discussion .................................................................................................................... 126 5.5 Conclusion .................................................................................................................... 127 6. CHAPTER SIX ............................................................................................................. 129 An assessment of Kenya’s forest policy and law on participatory forest management for sustainable forest management ........................................................................................... 129 Abstract .............................................................................................................................. 130 6.1 Introduction .................................................................................................................. 131 6.2 Methods ........................................................................................................................ 135 6.2.1 Sample design ........................................................................................................ 135 6.2.2 Data analysis .......................................................................................................... 136 6.3 Results .......................................................................................................................... 137 6.3.1 Characterization of respondents ............................................................................ 137 6.3.2 Stakeholders analysis ............................................................................................. 139 6.3.3 Participation of CFAs in conservation and management of MKFR ...................... 141 6.3.4 Participation of CFAs in the planning, implementation, evaluation, and monitoring of forest conservation in MKFR ..................................................................................... 143 6.3.5 Policy and institutional framework and conservation of MKFR ........................... 150 6.4 Discussion .................................................................................................................... 153 6.4.1 Consequences of illegal activities in Mt. Kenya Forest Reserve .......................... 154 6.4.2 Stumbling blocks in forest governance reform affecting conservation of MKFR 155 6.4.3 Policy failure and recommendations ..................................................................... 159 xv 6.4.4 Institutional challenges and recommendations ...................................................... 160 6.4.5 Other challenges .................................................................................................... 162 6. 5 Conclusion ................................................................................................................... 164 7. CHAPTER SEVEN ...................................................................................................... 165 The remote sensing of forest canopy gaps in a selectively logged sub montane tropical forest reserve: A synthesis ................................................................................................... 165 7.1 Introduction .................................................................................................................. 166 7.2 Remote sensing of selective logging in tropical forests ............................................... 167 7.3 Classification of endangered tree species using an imbalanced dataset ....................... 168 7.4 Mapping canopy gaps using narrow-band vegetation indices ..................................... 169 7.5 Enhancing the detection and mapping of canopy gaps ................................................ 171 7.6 Forest policy and institutional framework on participatory forest management ......... 172 7.8 The future ..................................................................................................................... 175 REFERENCES ..................................................................................................................... 176 Appendix I: Questionnaire for Community Forest Associations ....................................... 194 Appendix II: An interview schedule for KFS, KWS, and NEMA Officials ...................... 201 Appendix III: An interview schedule for KFS Forest Guards ........................................... 205 Appendix IV: Observation Checklist ................................................................................. 207 xvi List of figures Figure 1.1. Map of the study area in Mt. Kenya Forest Reserve (MKFR) showing the location of (a) Chuka Forest in MKFR, and (b) Chuka Forest and the adjacent locations. .. 8 Figure 1.2. The climate of the study area; mean monthly rainfall (bars) and mean monthly temperature (red dotted line). .................................................................................. 9 Figure 2.1. Temporal distribution of published articles where remote sensing analysed selective logging in tropical forests. ..................................................................... 19 Figure 2.2. Affiliation of lead authors in published articles where remote sensing analysed selective logging in tropical forests. ..................................................................... 20 Figure 2.3. Spatial distribution of studies where remote sensing analyzed selective logging in tropical forests. ...................................................................................................... 22 Figure 2.4. Sensors used to assess selective logging (SL) in tropical forests. (a) Frequency of optical sensors; (b) frequency of RADAR sensors. .............................................. 24 Figure 2.5. Validation methods used in selective logging (SL) studies in tropical forests. .... 33 Figure 3.1. Samples of endangered tree species in Mt. Kenya Forest Reserve in Chuka, Tharaka Nithi County, Kenya. .............................................................................. 51 Figure 3.2. Tree species mean spectral signatures derived from the WorldView-2 data. ...... 56 Figure 3.3. Box-whisker plots showing tree species (TS) mean reflectance values derived from WorldView-2 (WV-2) data. AlG–Albizzia gummifera, AnG–Anthocleista grandiflora, MK–Macaranga kilimandscharica. NB–Newtonia buchananii, OWV– other woody vegetation, PA–Prunus Africana, SD–Shadow, SG–Syzygium guineense, ZG–Zanthoxylum gilletii. ................................................................... 57 Figure 3.4. Optimization by 10-fold cross-validation: (a) random forest parameters (mtry and ntree), and (b) support vector machine parameters (gamma and cost). ................ 60 Figure 3.5. The relative importance of WorldView-2 bands in discriminating between different tree species as measured by RF classifier. ............................................................ 61 Figure 3.6. The mean decrease accuracy (MDA) values show the relationship between each tree species and WorldView-2 spectral bands as measured by the RF classifier. AlG–Albizzia gummifera, AnG–Anthocleista grandiflora, MK–Macaranga kilimandscharica. NB–Newtonia buchananii, OWV–other woody vegetation, PA– Prunus Africana, SD–Shadow, SG–Syzygium guineense, ZG–Zanthoxylum gilletii. ............................................................................................................................... 62 xvii Figure 3.7. Visualizing tree species (dis)similarity using multidimensional scaling: (a) random forest (b) support vector machine. ........................................................................ 63 Figure 3.8. Species-level F1-score for the combined sampling technique: (a) random forest; (b) support vector machine. AlG–Albizzia gummifera, AnG–Anthocleista grandiflora, MK–Macaranga kilimandscharica. NB–Newtonia buchananii, OWV– other woody vegetation, PA–Prunus Africana, SD–Shadow, SG–Syzygium guineense, ZG–Zanthoxylum gilletii. ................................................................... 66 Figure 3.9. Model-level F1-score for the datasets: (a) random forest; (b) support vector machine. ................................................................................................................ 67 Figure 3.10. Classification maps of tree species in Mt. Kenya Forest Reserve obtained using (a) random forest and (b) support vector machine. ............................................... 69 Figure 4.1. A true-color image showing (a) two uncut Ocotea trees in 2014 WorldView-2 image (1.89 m); (b) fresh and older canopy gaps created after cutting the camphor trees in 2019 WorldView-3 multispectral image (1.20 m); (c) the two gaps in 2019 pansharpened WorldView-3 image (0.3 m). ......................................................... 82 Figure 4.2. False-color image (752-RGB) for the study area. Yellow, blue, and green polygons represent the reference vegetated gaps, shaded gaps, and forest canopy, respectively. .......................................................................................................... 85 Figure 4.3. The relative importance of metrics derived from pansharpened Worldview-3 bands in discriminating newly created, shadowed, and vegetated gaps, and forest canopy as measured by random forest classifier using (a) mean decrease analysis; and (b) mean decrease in Gini. .......................................................................................... 90 Figure 4.4. Optimization by 10-fold cross-validation: (a) random forest parameters (mtry and ntree), and (b) support vector machine parameters (gamma and cost). ................ 91 Figure 4.5. Mean (spectral reflectance curves) and standard deviations (short lines) of shaded gaps, vegetated gaps, and forest canopy extracted from the best performing variables of WorldView-3 imagery pixels. ........................................................... 92 Figure 4.6. The mean decrease accuracy values showed the relationship between each forest landscape feature and the metrics extracted from pansharpened WorldView-3 bands as measured by the random forest classifier. .............................................. 94 Figure 4.7. The means of the chlorophyll absorption ratio index, the red edge position index, and the modified chlorophyll absorption ratio index extracted from 2014 pansharpened WorldView-2 data showing sites before logging events (a); and the means of the chlorophyll absorption ratio index, the red edge position index, and xviii the modified chlorophyll absorption ratio index extracted from 2019 pansharpened WorldView-3 data after known logging events (b). ............................................. 96 Figure 4.8. Final thematic maps showing results of supervised pixel-based classification for (a) random forest, and (b) support vector machine. .............................................. 98 Figure 5.1. A subset of the WorldView-3 image of the study area: (a) band 8; (b) colour composite of bands 8, 5, and 3 (RGB). ............................................................... 110 Figure 5.2. The spatial relationships of the pixels, where D represents the distance from the central pixel (nc,mc). .......................................................................................... 115 Figure 5.3. The relative importance of the LBPc,j fused with grey level co-occurrence matrix (GLCM) features in discriminating between vegetated gaps, shaded gaps, and tree crowns by random forest model. ......................................................................... 122 Figure 5.4. Subset classification results of image block D based on the Multivariate texture- based (853-RGB) classification of canopy gaps from WorldView-3 image based on MLBPc and MHOM distribution—(a) random forest, and (b) support vector machine. .............................................................................................................. 126 Figure 6.1. Gender and age distribution among respondents from the questionnaire for the Community Forest Associations around Chuka Forest, Tharaka Nithi County, Kenya. ................................................................................................................. 137 Figure 6.2. Participation of the community forest association members around Chuka forest, Chuka, Tharaka Nithi County. ............................................................................ 138 Figure 6.3. Overall monthly income of the community forest association members around Chuka forest, Chuka, Tharaka Nithi. .................................................................. 139 xix List of tables Table 2.1: Information extracted from the literature on the remote sensing techniques applied in detecting and monitoring selective logging in tropical forests. .......................... 18 Table 2.2: Distribution of researchers in remote sensing of selective logging in tropical forests. .................................................................................................................................................. 21 Table 3.1: List of tree species, codes, family, leafy phenology, diameter at breast height (DBH), number of individual tree crowns, and composition of training and test data for each species used to map trees in Mt. Kenya Forest Reserve. .......................... 50 Table 3.2: Tree species’ inter-specific spectral separability as calculated by the Jeffries- Matusita (J-M) distance (Equation (3.1)). AlG–Albizzia gummifera, AnG– Anthocleista grandiflora, MK–Macaranga kilimandscharica. NB–Newtonia buchananii, OWV–other woody vegetation, PA–Prunus Africana, SD–Shadow, SG– Syzygium guineense, ZG–Zanthoxylum gilletii. ..................................................... 58 Table 3.3: The random forest (RF) and support vector machine (SVM) model optimization parameters. .............................................................................................................. 59 Table 3.4: Confusion matrices for (a) random forest and (b) support vector machine algorithms, for the combined sampling technique. AlG–Albizzia gummifera, AnG– Anthocleista grandiflora, MK–Macaranga kilimandscharica. NB–Newtonia buchananii, OWV–other woody vegetation, PA–Prunus Africana, SD–Shadow, SG– Syzygium guineense, ZG–Zanthoxylum gilletii. ..................................................... 65 Table 3.5: McNemar’s test results to compare random forest (RF) and support vector machine (SVM) classification models using the combined sampling technique. ................. 68 Table 3.6: Area coverage in hectares and percentage of the classes using random forest and support vector machine classifiers. ......................................................................... 70 Table 4.1: Composition of train and test data for vegetated and shaded gaps, and forest canopy for the study area. .................................................................................................... 82 Table 4.2: The features used to detect canopy gaps in Mt. Kenya Forest Reserve: 23 means (of the 15 vegetation indices and 8 visible–near-infrared bands), 23 standard deviations (of the 15 vegetation indices and 8 visible–near-infrared bands), 8 ratios (of the 8 visible–near-infrared bands) and 1 brightness feature. ........................................... 84 Table 4.3: Spectral separability as calculated by: (a) transformed divergence index (Equation (4.3)), and (b) Jeffries-Matusita distance (Equation (4.4)). ..................................... 93 xx Table 4.4: Pairwise correlations between the best-performing variables extracted from the WorldView-3 visible–near-infrared bands (B1-8), computed from the reference samples. CARI - chlorophyll absorption ratio index, MCARI - modified chlorophyll absorption ratio index, CRI 1 - carotenoid reflectance index 1, CRI 2 - carotenoid reflectance index 2, NPCI - normalized pigment chlorophyll index, PSRI - plant senescence reflectance index, REPI - red edge position index, and the Shadow detection index (SDI), and ARI - anthocyanin reflectance index. .......................... 95 Table 4.5: Confusion matrices for (a) random forest algorithm, and (b) support vector machine algorithm for the best-performing model. ............................................................... 97 Table 4.6: The McNemar’s test results comparing random forest and support vector machine classification models. .............................................................................................. 99 Table 5.1: Reference data and their description. ................................................................... 111 Table 5.2: Texture measures used in this study, their abbreviations, and equations. ........... 116 Table 5.3: The random forest and support vector machine model optimization parameters of the Multivariate Local Binary Pattern model. ....................................................... 121 Table 5.4: Confusion matrices for random forest classifier and support vector machine classifier for the composites of homogeneity (MLBP/HOM), contrast (MLBP/CON), entropy (MLBP/ENT), angular second moment (MLBP/ASM), and correlation (MLBP/COR) models for image block D. .......................................... 123 Table 5.5: The overall accuracy results of the six image blocks. ......................................... 125 Table 6.1: Analysis of main stakeholders and their roles and responsibilities in Mt. Kenya Forest Reserve (MKFR). ....................................................................................... 140 Table 6.2: Participation of community forest associations (CFAs) in conservation and management of MKFR; SDis = strongly disagree, Dis = disagree, N = neutral, A = agree, SA = strongly agree, SD = standard deviation, f = frequency, (%) = percentage. ............................................................................................................. 142 Table 6.3: Participation of community forest association (CFA) members based on the four stages of forest conservation and management; SDis = strongly disagree, Dis = disagree, N = neutral, A = agree, SA = strongly agree, SD = standard deviation, f = frequency, (%) = percentage. ................................................................................ 145 Table 6.4: Policy and institutional framework and conservation of Mt. Kenya Forest Reserve (MKFR); SDis = strongly disagree, Dis = disagree, N = neutral, A = agree, SA = strongly agree, SD = standard deviation, f = frequency, (%) = percentage. ......... 151 xxi Abbreviations and acronyms AGB: Above-Ground Biomass AIGF: Area-Integrated Gap Fraction ALS: Airborne Laser Scanning ARI: Anthocyanin Reflectance Index ASM: Angular Second Moment AutoMCU: Automated Monte Carlo Unmixing CARI: Chlorophyll Absorption Ratio Index CCA: Contextual Classification Algorithm CFA: Community Forest Association CIPFA: The Chartered Institute of Public Finance and Accountancy CLAS: Carnegie Landsat Analysis System CON: Contrast COR: Correlation CRI-1: Carotenoid Reflectance Index 1 CRI-2: Carotenoid Reflectance Index 2 CV: Cross-Validation DAP: Digital Aerial Photograph ENT: Entropy EO: Earth Observation ETM+: Enhanced Thematic Mapper Plus GFW: Global Forest Watch GLCM: Grey Level Co-occurrence Matrix GoK: Government of Kenya GPS: Global Positioning System GSFD: Gap Size-Frequency Distribution HOM: Homogeneity HR: High Resolution HYDICE: Hyperspectral Digital Collection Experiment IEA: Institute of Economic Affairs ITC: Individual Tree Crown J-M: Jeffries-Matusita ĸ: Kappa KFS: Kenya Forest Service KWS: Kenya Wildlife Service LAI: Leaf Area Index LBP: Local Binary Pattern LDA: Linear Discriminant Analysis LiDAR: Light Detection and Ranging MCARI: Modified Chlorophyll Absorption Ratio Index MDA: Mean Decrease in Accuracy MDG: Mean Decrease in Gini MDS: Multi-Dimensional Scaling MENR: Ministry of Environment and Natural Resources MKFR: Mt. Kenya Forest Reserve ML: Machine Learning MLBP: Multivariate Local Binary Pattern MSAVI: Modified Soil-Adjusted Vegetation Index xxii MSAVIaf: Modified Soil Adjusted Vegetation Index aerosol free NBR: Normalized Burn Ratio NDFI: Normalized Difference Fraction Index NDVI: Normalized Difference Vegetation Index NEMA: National Environment Management Authority NIR: Near Infra-Red NPCI: Normalized Pigment Chlorophyll Index NPV: Non-Photosynthetic Vegetation OA: Overall Accuracy OBIA: Object-Based Image Analysis OLI: Operational Land Imager OOB: Out-Of-Bag PA: Producer’s Accuracy PFM: Participatory Forest Management PRI: Photochemical Reflectance Index PSND: Pigment-Sensitive Normalized Difference PSRI: Plant Senescence Reflectance Index RADAR: Radio Detection and Ranging RBF: Radial Basis Function RENDVI: Red Edge Normalized Difference Vegetation Index REPI: Red-Edge Position Index RF: Random Forest RIL: Reduced Impact Logging RGB: Red, Green, Blue RMSE: Root Mean Square Error RS: Remote Sensing SAM: Spectral Angle Mapper SAR: Synthetic Aperture Radar SDI: Shadow Detection Index SFM: Sustainable Forest Management SIFI: Secretariat for International Forestry Issues SIPI: Structurally Insensitive Pigment Index SL: Selective Logging SRred/green: Simple ratio red/green SMA: Spectral Mixture Analysis SVM: Support Vector Machine TD: Transformed Divergence TM: Thematic Mapper TS: Tree Species UA: User’s Accuracy UNESCO: United Nations Educational, Scientific and Cultural Organization. VARI: Visible Atmospherically Resistant Index VHR: Very High Resolution VI: Vegetation Index VNIR: Visible and Near Infra-Red WV-3: WorldView-3 WV-2: WorldView-2 1 1. CHAPTER ONE General introduction 2 1.1 Status of tropical forests Tropical forests are home to more than half of the Earth's species diversity and provide several essential biological services (Solberg et al., 2008). They control global weather patterns and, more importantly, play an important role in the global carbon cycle (Da Ponte et al., 2015), storing large amounts of carbon while producing large amounts of the Earth's oxygen (Solberg et al., 2008). Many indigenous cultures and peoples rely on tropical forests and the land they occupy (Solberg et al., 2008). Despite substantial attempts to create novel methods to identify deterioration, little is known about the pace and severity of tropical forest degradation. However, some researchers estimate that unsustainable selective logging (SL) is destroying about 20% of humid tropical forests (Asner et al., 2009). This uncertainty, in turn, has made it difficult to understand and estimate the effects of forest degradation on tropical biodiversity (Borrego and Skutsch, 2014). Thus, reliable and operational systems, such as remote sensing (RS), are required to monitor selective logging (SL) in tropical forests. 1.2 Defining Selective logging and canopy gap 1.2.1 Selective logging According to Laurance et al. (2014), the term "selective logging" involves the felling of trees in large-scale for industrial use, as well in small-scale for local use. In this thesis report, selective logging is defined as a type of timber extraction in which a tree(s) from a selected high-value species, are harvested without necessarily displaying any logging infrastructure (Asner et al., 2005; Andersen et al., 2014). 1.2.2 Canopy gap The concept of "gaps" and their proper definition and measurement have been the subject of some discussion. Different authors have proposed various methods for defining and measuring canopy gaps. The extended gap concept is used by Runkle (1981) in defining a canopy gap as covering the stems of surrounding trees. But for RS purposes a tree trunk's location may not be easily determined from the top, because the location of the stem's base may not necessarily correspond with the centre of the crown (Gaulton and Malthus, 2010). According to Gaulton and Malthus (2010), ‘a gap boundary is defined by a line at ground level (the drip-line) located vertically beneath the inner most point reached by the foliage of a tree crown at any level, at that point on the gap perimeter.’ A ground level line that runs vertically beneath the innermost 3 point of a tree crown is how Gaulton and Malthus (2010) define a gap boundary. Brokaw (1982) defines a canopy gap as a 'hole' in a forested area reaching an average height of about 2 m, and the smallest size of a gap—20 to 40 m2. Gaps are defined by Fox et al. (2000) in relation to the difference in height from the bordering canopy. A gap, according to Koukoulas and Blackburn (2004), is a 'hole' in the canopy of forest due to the removal of a tree(s). Gaulton and Malthus (2010) state that a gap has a minimum area ≥5 m2 and height ≤ 10 m. Getzin et al. (2014), mapped 1 m2 canopy gaps using unmanned aerial vehicle (UAV) data—pixel size of 7 cm. Gaps in a forest measure at least width of 2 m, a maximum vegetation height of 3 m, and a minimum area of 50 m2 (Bonnet et al., 2015). Therefore, the minimum size of a gap and gap closure height, normally varies from forest to forest, and affected by factors such as the size of tree crowns, understory height, and advance regeneration, among others (Gaulton and Malthus, 2010). Canopy gaps are usually small (<1000 m2) (Betts et al., 2005). In the past, the gap's size was determined by first measuring its length and width from the ground, and then computing its area using a circle or ellipse's formula (e.g. Runkle, 1992). Ground-based techniques can be challenging, especially in rough terrain. Furthermore, the results of ground surveys can be subjective—depending on the surveyor (e.g. Nakashizuka et al., 1995). Additionally, in field surveys it can be challenging to distinguish fine canopy gaps and the ones where the understory vegetation is dense. However, suggestions have been proposed to increase the ground surveys’ objectivity—such as measuring the height of canopy at specific locations on a predetermined grid (Brokaw and Grear, 1991) or conducting regular line- intersect sampling (Battles et al., 1996). Nonetheless, these techniques require a sizable amount of time and manpower (Nakashizuka et al., 1995). The most appropriate solution was to observe forests from above rather than below through RS. Satellite and airborne data with very high-resolution (VHR) are appropriate for precisely delineating forest canopy gaps, as well as individual tree crowns. Only gaps with a minimum area of 100 m2 were considered in this study since it is such gaps that are significant for carbon dynamics (Negrón-Juárez et al., 2011), and to rule out small gaps that were not likely caused by felled Ocotea usambarensis (East African camphor) trees. 1.3 Detection of canopy gaps in tropical forests A challenging task is the detection and delineation of a gap (Vepakomma et al., 2008). According to Nyamgeroh (2015), there have been few studies on canopy gap delineation, and 4 even fewer are the methods to delineate canopy gaps. Forest canopy gaps have been studied using Landsat data with either the mono-temporal (e.g. Asner et al., 2005 and Negrón-Juárez et al., 2011)) or the multi-temporal approaches (e.g. Wang et al., 2019). However, Landsat’s pixel size (30 m) is incapable of detecting small canopy gaps (Clark et al., 2004). Additionally, the detectability of the effects of SL on Landsat images is 1 to 3 year after the disturbance event (Costa et al., 2019). High resolution (HR) multispectral sensors, e.g. SPOT and IKONOS were used to solve the spatial resolution problem (Clark et al., 2004). Different mapping methods depend on their ability to identify gaps in the forest canopy. Small canopy gaps can be detected using very high resolution (VHR) optical and light detection and ranging (LiDAR) data (Asner et al., 2013). Calibrated VHR synthetic aperture radar (SAR) data have been promising in detecting canopy gaps created through the removal of individual trees (Mitchell et al., 2017). Traditional aerial photography was successfully used for mapping canopy gaps before the introduction of HR optical data (Malahlela et al., 2014). But these methods fall short in capturing crucial biochemical traits like chlorophyll content, which are crucial for canopy gap mapping (Malahlela et al., 2014). Furthermore, its data acquisition is cost-intensive, but technological advances have revitalized its use through UAVs. Spaias et al. (2016) and Ota et al. (2019) used UAV data acquired before and after logging to map small- scale canopy gaps. An increase in shadow was used as an indicator of the location of logging events by Spaias et al. (2016). According to Ota et al. (2019), change of above-ground biomass (AGB) was related to disappearing tree crowns due to SL. Spaias et al. (2016) used hyperspectral data—its application is limited because it is computationally demanding particularly if cloud computing resources were not available. A canopy structure can be described in three-dimensional (3-D) by use of a canopy height model (CHM). Several studies, e.g. Asner et al. (2013), Boyd et al. (2013), Kent et al. (2015), Wedeux and Coomes (2015), and Dalagnol et al. (2019) have applied CHMs. The CHMs were derived from LiDAR data from a single acquisition to delineate canopy gaps with exceptional accuracy and consistency in tropical areas. Andersen et al. (2014) used repeat LiDAR data and applied a simple differencing of LiDAR CHMs to detect disappearing tree crowns. Airborne LiDAR data have been used to acquire precise measurement of AGB stock and change—several studies used a single acquisition, e.g. D'Oliveira et al. (2012), Ellis et al. (2016), Melendy et al. (2018), and Pearson et al. (2018). Others used before and after logging datasets, e.g., Englhart et al. (2014), Silva et al. (2017), Pinagé et al. (2019), and Rex et al. (2020). Nonetheless, airborne LiDAR’s coverage of relatively small spatial extents and high data acquisition costs limits its application, especially in tropical areas. 5 Descals et al. (2017) proposed an automated processing workflow in bi-temporal VHR SmallSat imagery to enable the detection of SL. Dalagnol et al. (2019) compared the potential of VHR satellite data to that of airborne LiDAR to detect tree loss. According to Yang et al. (2015), broad-band multispectral images can yield promising results for identifying canopy gaps. The visual interpretation of VHR multi-date satellite imagery is a promising way to detect and quantify canopy gaps created by logging activities with fairly low uncertainty (Clark et al., 2004). Nonetheless, spatially precise validation data is not readily available, and automated approaches based on VHR satellite data to accurately map canopy gaps over expansive areas are lacking (Dalagnol et al., 2019). In order to quantify illegal logging in a tropical forest, this study was aimed to evaluate whether gaps in forest canopies can be detected and mapped using the very high resolution (VHR) WorldView-3 multispectral imagery. High resolution (HR) WorldView-2 image and Google Earth were used to provide historical data. Accurate quantification of canopy gaps from disappearing tree crowns has a crucial contribution in calculating carbon densities of forests, as well as modelling the effects of forest degradation on tropical biodiversity. Selective logging (SL), a major cause of forest degradation in tropical forests, results to long-term carbon emissions into the atmosphere. The accuracy of carbon estimates can be improved, provided canopy gaps are accurately identified and correctly estimate trees in per unit area of forest. This can be used by forest managers to implement on-the-ground conservation and restoration projects in Mt. Kenya Forest Reserve (MKFR). 1.4 Research objectives The primary focus of this study was to use a remote sensing (RS) technique—the spectral- texture analysis approach—to explore the potential of the very high resolution WorldView-3 (WV-3) multispectral dataset to identify and quantify canopy tree loss associated with selective logging (SL) in a closed-canopy submontane tropical forest. The gaps in the study area were vegetated—the canopy gaps had low vegetation inside them—an initial stage of vegetation recovery from SL disturbance. Two advanced machine learning (ML) classification models— the random forest (RF) and support vector machine (SVM) models were used to identify and classify canopy gaps in the spectral and texture domains of the WV-3 multispectral imagery. The specific objectives were: 6 1. To review the status, trends, potentials, and challenges and recommended future directions in remote sensing (RS) techniques used to map selective logging (SL) in the tropical forests from 1992 to 2019; 2. To use species-level tree classification to map threatened trees species in a selectively logged sub-montane heterogeneous tropical forest—MKFR; 3. To explore the potential of spectral features from the visible-near-infrared bands and vegetation indices to model canopy gaps caused by logging of Ocotea trees; 4. To apply a texture-spectral analysis approach—using GLCM-, LBP-, and MLBP-based rotation-invariant feature descriptors—to exploit the potential of the very high resolution WorldView-3 multispectral imagery to model canopy gaps; and 5. To assess how inadequacies in Kenya’s forest policy and law on participatory forest management led to illegal activities in forests in Kenya. 1.5 Scope of the study This study used multispectral RS data to discriminate among tree species under the threat of SL in Mt. Kenya Forest Reserve (MKFR) in Chuka, Tharaka Nithi County, Kenya. The applications of spectral features extracted from the WV-3 VNIR bands and vegetation indices (VIs) were used to detect canopy gaps resulting from SL of Ocotea usambarensis (East African camphor). Fused texture and spectral features extracted using GLCM-, LBP-, and MLBP-based models were also explored to detect gaps in forest canopy. Two machine learning (ML) classification algorithms—random forest (RF) and support vector machine (SVM) were used. The study collected a field dataset of georeferenced canopy gaps in the study site, and historical data—Google Earth and WorldView-2 datasets. The Participatory Forest Management (PFM) policy and law were then evaluated in order to determine the efficacy of changes made to Kenya's policies, legislation, and institutional arrangements to support transitions toward Sustainable Forest Management (SFM). Mt. Kenya Forest Reserve (MKFR) is an Afromontane tropical forest ecosystem that supports a diverse range of biodiversity and provides valuable ecosystem services, as well as being a major forested water catchment area in Kenya (MENR, 2016). For decades, MKFR has been subjected to illegal logging for its commercially valuable reserves of indigenous timber, as well as other illegal activities that have resulted in reduced ecosystem services (KWS, 2010). 7 1.6 The study area The forest reserve was established in 1932 and placed under the management of the Forest Department (now known as the Kenya Forest Service) with the primary goal of preserving and developing the forest. This included creating plantations to replace harvested indigenous stands, regulating resource access, and preserving a forest industry (KWS, 2010). Because of its remarkable ecosystems and natural beauty, the forest reserve was designated a UNESCO World Heritage Site in 1997 (NEMA, 2010). Mt. Kenya Forest Reserve (MKFR) is located in Central Kenya and includes parts of the counties of Meru, Tharaka-Nithi, Embu, Kirinyaga, and Nyeri. Chuka Forest, located in Tharaka-Nithi County, is part of the Mt. Kenya ecosystem and encompasses approximately 21,740 ha (Figure 1.1). Six locations—Kiang'ondu, Magumoni, Mitheru, Mugwe, Mwonge, and Thuita—make up the forest-adjacent area. Chuka Forest is located between longitudes 37019'0"E and 37036'0"E, and latitudes 0011'0"S and 0019'30"S. The MKFR spans a range of elevation, slope, and aspect positions. The bedrock of the study area is formed by phonolites from volcanic events about 2 million years ago, but the inorganic part of the soils was formed from volcanic ashes and pyroclastic rocks (Lange et al., 1997). In the 2700-3300 m zone, dark andosols with low bulk density and high organic matter content predominate. The topsoil has a high silt content, whereas the B-horizon has a high clay content (Lange et al., 1997). 8 Figure 1.1. Map of the study area in Mt. Kenya Forest Reserve (MKFR) showing the location of (a) Chuka Forest in MKFR, and (b) Chuka Forest and the adjacent locations. Annual rainfall follows a bimodal distribution—March-May are the long rains whereas October-December are the short rains (Figure 1.2). Rainfall is light on the lower slopes but heavy at higher altitudes (KFS, 2010). The highest rainfall (about 2500 mm) occurs between 2750 and 3750 m above sea level, while above 4500 m, most precipitation falls as snow or hail, and frosts are common above 2500 m above sea level (KFS, 2010). January and February are the driest months, and the trade wind system has the greatest impact on the windward side (Wooller et al., 2002). Large diurnal oscillations characterize the climate, but the annual 9 amplitude in the monthly means is small (Lange et al., 1997). The eastern and southern sides are wetter (2500 mm), while the northern side is drier (900 mm) (Nyongesa and Vacik, 2019). Figure 1.2. The climate of the study area; mean monthly rainfall (bars) and mean monthly temperature (red dotted line). Mt. Kenya shows a marked vegetational gradient dictated by altitude and rainfall amount. The lower tree line of the forest belt is due to agricultural and pastoral activities (Lange et al., 1997). In the southern parts of the mountain, cultivated land extends up to nearly 1800 m, in the eastern and western approximately 2400 m, and in the northern it is about 2900 m. About 3400 m in the south and west of the mountain lies the upper tree line whereas in the north the tree line is at 3000 m (Lange et al., 1997). The vegetation on the lower slopes of Mt. Kenya is montane forest. Characteristic tree species in the indigenous forest include Podocarpus latifolia, Nuxia congesta, Newtonia buchananii, Calodendrum capense, Croton megalocarpus, Juniperus procera, Ocotea usambarensis, and Olea europaea spp africana. Then next to the montane forest is the bamboo forest (which extends from 2,550 m), scrub, and moorland (from 3,000 to 3,500 m). Characteristic species in this zone include Hagenia abyssinica and Hypericum spp. The moorland gives way to Afro-Alpine vegetation. The bare rock, ice, and snow are at the highest altitude (Nature Kenya, 2019). Ocotea usambarensis, normally targeted for its hardwood which has excellent decay and insect resistance, also forms the evergreen sub-montane forests on the extremely humid southern, south eastern, and eastern slopes at 1500 to 2500 m. They never form pure stands and prefer humid Niti- and Acrisols 0 5 10 15 20 25 0 50 100 150 200 250 300 350 400 450 Jan Feb Mar Apr May Jun Jul Aug Sept Oct Nov Dec T em p er at u re ( °C ) P re ci p it at io n ( m m ) 10 (Bussmann and Beck, 1995). Nyayo Tea Zone, established by Legal Notice No. 265 of 1986 provides a buffer zone to check against human encroachment into MKFR. 1.7 Thesis overview The thesis aims to achieve the main objectives of this study through five research papers that have been submitted to international peer-reviewed journals, three of which have already been published, one is being reviewed, and the other one is being prepared for journal submission. Each was written as a stand-alone article that can be read independently of the other parts of the thesis but draws separate conclusions that are linked to the overall research objectives and questions. As a result, there are overlaps and replications in the sections "Introduction" and "Method" in the various chapters. When one considers the rigorous peer-review procedure and the fact that the various chapters are papers that can be read independently without losing the overall context, this problem is of relatively minor importance. The thesis is divided into seven chapters. The second chapter conducts a pantropical meta-analysis to determine the current state of knowledge on remote sensing techniques used to detect and monitor SL in tropical forests. The chapter examines and discusses the current state of RS techniques used to detect and monitor SL disturbances in tropical forests. The analyses focused on the years 1992-2019. The number of studies on RS for SL has steadily increased over time, necessitating their review to inform current research and forest management techniques. A variety of peer-reviewed articles are discussed to assess the applicability and accuracy of various methods in various situations. Major issues with existing approaches are identified, and future needs were discussed. Using RF and SVM classifiers and WorldView-2 (WV-2) multispectral imagery, Chapter 3 examines the effects of imbalanced data on identifying and mapping tree species under threat in a selectively logged sub-montane heterogeneous tropical forest. Three datasets derived using three techniques—oversampling, undersampling, and combined sampling techniques were compared with the original dataset. The chapter lists the key spectral bands that were determined to be crucial in mapping the endangered tree species in the research area. Chapter 4 detects canopy gaps caused by illegal logging of Ocotea trees using very high-resolution multispectral satellite data and machine learning algorithms, as well as historical data from WV-2 and Google Earth. Thus, the role of vegetation indices derived from WorldView-3 data in mapping canopy gaps was investigated. 11 The Grey-Level Co-Occurrence Matrix (GLCM), which was fused with the Multivariate Local Binary Pattern (MLBP) to detect canopy gaps, is the focus of Chapter 5. It is nearly 50 years old and was used for the analysis of grey level texture based on statistical approaches. Three WorldView-3 (WV-3) bands' worth of spectral and textural data were combined to enhance classification outcomes. Chapter 6 assesses Kenya's forest policy and law on Participatory Forest Management (PFM) to know the extent to which adjustments in laws, regulations, and institutional structures intended to speed up the transition to sustainable forest management (SFM) have failed. Several steps have been proposed to simplify and rationalize Kenya's forest policies and laws. The study is summarized in Chapter 7. The findings and conclusions from the preceding chapters are summarized. Some suggestions for future research on the use of remote sensing to map canopy gaps in tropical forests are made. At the end of the thesis, there is a single reference list. 12 2. CHAPTER TWO Literature review This chapter is based on: Jackson, C.M., and Adam, E. (2020). Remote sensing of selective logging in tropical forests: current state and future directions. iForest - Biogeosciences and Forestry, 13, 286–300. 13 Abstract This chapter reviewed and discussed the status of remote sensing (RS) techniques applied in detecting and monitoring selective logging (SL) disturbance in tropical forests. The number of studies on remote sensing (RS) for selective logging (SL) has grown steadily over the years, therefore, there was need for an updated review to guide forest management practices and current research. The analyses concentrated on the period 1992-2019. Accurate and precise detection of selectively logged sites in a forest is crucial for analysing the spatial distribution of forest disturbances and degradation. Remote sensing (RS) can be used to monitor SL activities over tropical forests, which otherwise require labour-intensive and time-consuming field surveys. Peer-reviewed articles were searched in ScienceDirect®, Web of Science®, and Scopus® web-based databases containing the largest abstract and citation databases of peer- reviewed literature. The key terms used were (“Selective Logging” OR “Canopy Gap” OR “Canopy Opening” OR “Deforestation” OR “Forest Degradation” OR “Forest Disturbance”) AND “Tropical Forests” AND “Remote Sensing” in the title, abstracts and the keywords in the search. About 5546 articles were found in ScienceDirect, 2395 in Web of Science, and 888 articles in Scopus (as of June 20, 2019). After considering the relevant articles, abstracts of 328 articles were read through—information was extracted from 110 relevant articles that discussed the RS of SL in tropical forests. The literature review showed that there was increased publication activity over the years. This could be related to the availability of information, awareness about RS technology, and growing attention around the monitoring of SL and related activities in tropical forests. A large number of studies (54%) used low/medium spatial resolution datasets to map SL in tropical forests. Only 28% and 18% used very high resolution (VHR) and high-resolution (HR) sensors, respectively. Low and medium spatial resolution datasets are too coarse to detect less intensive SL. Many of the methods for mapping SL in tropical forests using low/medium spatial resolution datasets have a high rate of false detections. Recently, VHR satellite and unmanned aerial vehicle- (UAV-) based datasets have caught the interest of researchers mapping disappearing tree crowns due to SL tropical forests. Keywords: Tropical Forest Disturbance. Selective Logging. Forest Degradation. Forest Canopy Gaps. Disturbance Mapping. Remote Sensing. Forest Monitoring. 14 2.1 Introduction Tropical forests are housing more than half of the diversity of life on earth and offering several vital biological applications such as pest control, pollination, and seed dispersal (Solberg et al., 2008). Tropical forests regulate global weather patterns, and more importantly, play a key role in the global carbon cycle (Da Ponte et al., 2015), keeping large amounts of carbon and producing great supplies of the Earth’s oxygen (Solberg et al., 2008). Tropical forests and the land they occupy support numerous indigenous cultures and peoples (Solberg et al., 2008). Despite the remarkable values and functions of tropical forests, many conservation and protection efforts have not been effective, as a result, the forests are being cleared in many countries for timber and agricultural land expansion (Gibson et al., 2011). Forest disturbance from selective logging (SL) and associated degradation by forest fires may have enduring effects on forest dynamics and composition (Asner et al., 2006), thus interfering with forest health and the availability of essential ecosystem functions and services (Gibson et al., 2011). A study in the Brazilian Amazon showed an increase in total forested areas affected by SL and forest fires from approximately 11,800 to 35,600 km2 in 1992 and 1999, respectively (Matricardi et al., 2013). Asner et al. (2005) indicated that each year, SL can expand over as much forested area as does deforestation, with logged areas ranging in size from 12,075 to 19,823 km2 between 1999 and 2002. The study showed that between 1999 and 2004, about 76% of all timber harvest practices led to high levels of canopy damage sufficient to leave forests prone to drought and fire (Asner et al., 2005). Currently, few truly undisturbed tropical forests exist, and arguably the single most important cause of tropical forest degradation worldwide is unsustainable SL (Miettinen et al., 2014). More than 400 million hectares of natural tropical forests have been degraded since 1980 (Edwards et al., 2014). Selective logging (SL) is a form of extraction of timber where a group of trees from selected species, the high- value tree species, are removed from the forest (Andersen et al., 2014). Global demand for precious and rare tropical timber, such as ebony and rosewood, is expected to continue to grow, and the international market has c. 50⎯90% of tropical wood harvested illegally (Nellemann, 2012). Thus, there is a need for sustainable forest management (SFM). Sustainable forest management (SFM) is the process of managing a forest to reduce forest degradation and deforestation by ensuring the sustainability of forest resources, protection, and conservation of genetic diversity and to ensure the sustainable exploitation of the biological resources, and enhancing the full valuation of forest goods and services (Poudyal et al., 2018). However, SFM is suffering from a variety of obstacles related to (Chia et al., 2020): 15 • governance issues (e.g. poorly defined tenure and resource rights, inadequate transparency and accountability, corruption, and limited involvement of relevant actors in the formulation of management plans); • economic issues (e.g. high opportunity cost of maintaining forests, high transaction cost for better forest management, low financial returns from improved forest management, and unattractive incentives); • regulatory and legislative issues (e.g. poor regulatory framework, lack of political will and incentive to implement regulations, and unrealistic legislation); and • knowledge and capacity issues (e.g. poor understanding on the benefits to improve forest management, inadequate financial and material resources, and limited human resources to enforce and monitor regulations). However, researchers and practitioners should not give up the SFM idea, but rather the effort to enhance it must be redoubled and refined. Selective logging (SL) practices determine the outcome of SFM (Poudyal et al., 2018), thus an important component of SFM is monitoring of the forest status. Reliable and operational systems for monitoring SL in tropical forests have to be utilized (Hirschmugl et al., 2017). Such a system should be able to provide an estimate of baseline forest conditions in a spatially explicit fashion, departures from which can be used to assess current and previous trends of forest degradation and deforestation (Verbesselt et al., 2010). Remote sensing (RS) can be used to assess and monitor SL over tropical forests (Anwar and Stein, 2012), in a spatially and temporally continuous manner (Banskota et al., 2014). Accurate and precise detection of selectively logged sites in a forest is crucial for analysis of the spatial distribution of forest disturbances and degradation (Anwar and Stein, 2012). The RS methods that have been developed to detect SL in tropical forests detect intensive timber harvest (> 20 m3 ha-1). Therefore, medium spatial resolution datasets like Landsat are normally considered too coarse to detect less intensive SL (Hethcoat et al., 2018). Pinage et al. (2016) showed that the intensity of canopy impacts may vary according to the SL activity. High-intensity logging causes high forest damage that is long-lasting, and detectable on satellite imagery and vice versa. Soil fraction images obtained from spectral mixture modelling of multispectral or hyper-spectral data serve as a suitable approach for the detection of SL (Souza et al., 2005; Matricardi et al., 2010). Forests with obvious selective logging have well-defined logging infrastructure and extensive canopy degradation. Forests where subtle SL is taking place normally show less canopy perforation or visible infrastructure. Therefore, RS techniques may not easily 16 differentiate them from undisturbed forests (Asner et al., 2004). Methods required for monitoring SL at high temporal resolution are not available. Some of the existing methods for mapping SL mostly come with numerous false detections, and existing techniques for minimizing them either impair the temporal accuracy or increase the omission error for the forest disturbance (Hamunyela et al., 2016). The extent of forest degradation as a result of SL using currently available techniques is unknown (Hethcoat et al., 2018). New change detection approaches can improve the detection of SL to achieve accurate mapping and quantification of forest loss (Hamunyela et al., 2016). Over the last decades, there has been a rapid growth in the number of studies that investigated the use of RS for SL (Hethcoat et al., 2018). Providing an overview of the RS data and techniques that have been used in SL to identify the challenges and opportunities is essential. Such an overview would be useful practically in forest management and scientifically by highlighting the priorities and remaining research gaps for further investigation. Several review studies have been analyzed to evaluate the application of RS in mapping deforestation and forest degradation in tropical forests. Solberg et al. (2008) analysed the state of the art of RS techniques, detailing the relevant sensors and algorithms, usable datasets, and information on the leading institutions for research and development on techniques that might lead to operational monitoring of tropical forests. Miettinen et al. (2014) critically discussed available approaches for large-area forest degradation monitoring with satellite RS data at high to medium spatial resolution, in Southeast Asia. Da Ponte et al. (2015) provided an overview of the RS-based studies of tropical forest dynamics in Latin America, categorizing the existing studies based on selected sensors and data analysis methodologies. The review has considered both large-scale as well as small-scale forest changes solely induced by anthropological activities. Mitchard (2016) provides a review of earth observation (EO) methods for detecting and measuring forest change in the tropics. The study describes current and emerging EO technologies, and how these can be used to map forests and forest changes. Hirschmugl et al. (2017) reviewed the current state of the art in RS-based monitoring of forest disturbances and forest degradation. This literature reviews focus on Europe’s temperate forests and Africa’s tropical evergreen forests, using optical EO data. Mitchell et al. (2017) reviewed the RS approaches for monitoring forest degradation in support of countries' measurement, reporting, and verification (MRV) systems for reducing emissions from deforestation and forest degradation, conservation of existing forest carbon stocks, sustainable forest management and enhancement of forest carbon stocks (REDD+). The paper reviewed forest degradation that leads to canopy gaps that are detectable using RS, taking into consideration both forests within 17 and outside the tropics. Therefore, there was need for an updated overview of the methods used in detecting and monitoring SL in tropical forests only. This chapter had the main focus laid on types of sensors and methodologies of data analysis, and the major challenges and further research needed to explore the use of RS in monitoring small-scale forest disturbance due to SL in tropical forests. The analyses concentrated on the period from 1992 to 2019. 2.2 Methods 2.2.1 Database search The peer-reviewed articles published between 1992 and 2019 were searched in the web-based databases ScienceDirect®, Web of Science®, and Scopus®. They contain the largest abstract and citation databases of peer-reviewed literature, and the University of Wisconsin⎯Madison rates ScienceDirect®, Web of Science®, and Scopus® among the top ten web-based databases (UW Madison Libraries, 2019). The two most well-established databases that cover the widest scope of English-language literature, while focusing on peer-reviewed publications are Scopus® and Web of Knowledge® (Kleinschroth et al., 2016). Neither database is inclusive but complements the other. The key terms used were (“Selective Logging” OR “Canopy Gap” OR “Canopy Opening” OR “Deforestation” OR “Forest Degradation” OR “Forest Disturbance”) AND “Tropical Forests” AND “Remote Sensing” in the title, abstracts and the keywords in the search. About 5546 articles were found in ScienceDirect, 2395 in Web of Science, and 888 articles in Scopus (as of June 20, 2019). Bibliographies of the articles were iteratively scanned until no new relevant articles were identified. All abstracts were scrutinized to filter out irrelevant articles. The remaining articles were read and retained only if they were relevant. Articles with irrelevant titles were excluded. After bringing the relevant articles together, the abstracts of 328 articles were read through. Finally, information was extracted from 110 relevant articles that explicitly discussed the RS of SL in tropical forests. 2.2.2 Content analysis For the selected 110 articles, the same set of attributes for analysis was extracted (Table 2.1). In particular, noted were publication details, reference data availability, and the geographical location of the study area (i.e. if the study was conducted within the tropics). To characterize the sensor applied in each study, the name of the sensor, properties, and also the platform were recorded. The image processing techniques used to detect and quantify the areas under SL have 18 been discussed. The auxiliary data used, the measure of accuracy, and the level of accuracy achieved were also investigated. Table 2.1: Information extracted from the literature on the remote sensing techniques applied in detecting and monitoring selective logging in tropical forests. Attribute Description Publication details The year of publication, journal type, and national affiliation of the authors and co- authors. Reference data How reference data used for model calibration and validation were collected. Geographical location Location of the study area. If no location was given in the paper, Google Maps was used to determine the approximate location, or Landsat WRS2 reference was used to identify the centre location of the Landsat footprint used in the analysis. Sensor The sensor type used in the study. Platform The platform of the sensor, i.e., space-borne, air-borne, or unmanned aerial vehicle. Spatial properties The sensor resolution used: coarse resolution (>100 m), medium resolution (10-100 m), high resolution (5-10 m), and very high resolution (<5 m); the spatial scale of the studies. Temporal properties The temporal resolution of the sensor, and temporal scale of the remote sensing data (single date, bi-temporal, multi-temporal, or time series analysis), and sensor archive. Spectral properties The spectral properties used in the analysis: Vegetation Index (if a single vegetation index or band was used, then the index name was noted), Multi (if multi-spectral bands or multiple spectral indices were used), Hyper (if hyperspectral bands or multiple narrow- band indices were used), Lidar, and SAR. Techniques The techniques employing satellite data used to detect and quantify selective logging. Accuracy measure The measure of accuracy reported in the study. Accuracy Level of accuracy, according to accuracy measure. 2.3 Results of literature review 2.3.1 Publication details Watrin and Rocha (1992) provided some of the first RS estimates of the area affected by SL (Souza et al., 2005). After the publication of the second article by Stone and Lefebvre (1998), the application of RS in SL analyses of tropical forests rose steadily (Figure 2.1). However, no relevant articles were published in 1993-1997, 1999, 2001, and 2011. The first major publication activity took place between 2002 and 2010, with an average of four papers per year, 19 while the second considerable publication activity occurred between 2012 and 2019, with an average of nine articles per year. This shows that publication activity has more than doubled between 2012 and 2019. This could be related to the availability of information, awareness about remote sensing technology, and growing attention around the monitoring of SL and related activities in tropical forests. Figure 2.1. Temporal distribution of published articles where remote sensing analysed selective logging in tropical forests. The results from the content analysis showed that using RS to examine SL in tropical forests is being accepted by a rising number of scientific disciplines; the articles are published in about fifty different scientific journals. About half of publications are appearing in journals specifically focusing on RS applications (Souza and Barreto, 2000; Souza et al., 2003, 2005; Asner et al., 2005; Koltunov et al., 2009; Matricardi et al., 2010, 2013; Heiskanen et al., 2015; Dalagnol et al., 2019), while the other half is distributed among other journal categories (Asner et al., 2004; Furusawa et al., 2004; Burivalova et al., 2015; Ellis et al., 2016; Sofan et al., 2016). The distribution between the journal categories has articles published in information and communication technology, ecological, cartography, interdisciplinary, and natural hazards-oriented journals. The majority of the articles’ first authors have been affiliated with institutions located in Europe (De Wasseige and Defourny, 2004; Rahm et al., 2013; Wang et al., 2019), North y = 3E-05x4 - 0.0023x3 + 0.0628x2 - 0.3147x + 0.5656 R² = 0.6371 0 2 4 6 8 10 12 14 19 92 19 93 19 94 19 95 19 96 19 97 19 98 19 99 20 00 20 01 20 02 20 03 20 04 20 05 20 06 20 07 20 08 20 09 20 10 20 11 20 12 20 13 20 14 20 15 20 16 20 17 20 18 20 19 N u m b er o f st u d ie s Years 20 America (Stone and Lefebvre, 1998; Wang et al., 2005; Ellis et al., 2016), South America (Souza et al., 2003; Graça et al., 2015; Conde et al., 2019), and Asia (Furuisawa et al., 2004; Sofan et al., 2016; Qu et al., 2018). Institutions in Africa do not have any representation (Figure 2.2). Only 0.7% of the articles have African co-authorship (Burivalova et al., 2015; Descals et al., 2017; Kankeu et al., 2016). Figure 2.2. Affiliation of lead authors in published articles where remote sensing analysed selective logging in tropical forests. Overall, 510 researchers from universities and research institutions around the world appear in the 110 articles (Table 2.2). The United States of America has 146 researchers who are from American universities and research institutions. The most featured institutions are Michigan State University (Matricardi et al., 2005, 2007, 2010), Carnegie Institution of Washington, Stanford University (Asner et al., 2004; Broadbent et al., 2008), Woods Hole Research Center (Laporte and Lin, 2003; De Grandi et al., 2015) and United States Department of Agriculture (USDA) Forest Service (Andersen et al., 2014). Brazil has contributed 111 researchers. Universities or research institutions hosting researchers at the time a study was being conducted and/or published, determined which countries the researchers were listed under. Eraldo A.T. Matricardi, for example, is listed under the USA because he published an article (Matricardi et al., 2007) on selective logging while at the Department of Geography, Michigan State University. In Matricardi et al. (2013), the same author is listed under Brazil because at the time of the publication of the article he was at the Department of Forestry, University of Brasilia. 21 Table 2.2: Distribution of researchers in remote sensing of selective logging in tropical forests. Country Lead author Co-author Country Lead author Co-author Country Lead author Co-author USA 26 120 France 2 16 PNG 0 2 Brazil 26 85 Austria 2 7 Norway 0 1 Italy 7 35 Netherlands 2 7 Laos 0 1 UK 6 30 F. Guiana 2 5 Cameroon 0 1 Germany 6 18 Mexico 1 5 Laos 0 1 Finland 6 9 Switzerland 1 2 Brunei 0 1 Japan 5 14 China 1 0 Bolivia 0 1 Australia 4 9 Malaysia 1 1 Argentina 0 1 Indonesia 4 9 Puerto Rico 0 3 Madagascar 0 1 Canada 3 6 Singapore 0 3 S. Africa 0 1 Belgium 3 1 Peru 0 2 Nepal 0 1 2.3.2 Geographical information This study established the location of the study areas within the tropical region, where latitude and longitude coordinates were indicated. In the cases where there were no coordinates provided in the paper, Google Maps™ was used to find the approximate location, or Landsat World Reference System 2 (WRS2) was used to spot the centre location of the Landsat scene used in the analysis. As indicated in Figure 2.3, the scientific activities’ geographic distribution is uneven in the tropical forests. Significantly more research has been done in South America, and about eighty-nine percent occurred in the Brazilian Amazon, specifically in Para, Rondonia and Mato Grosso states, where most of the disturbance has taken place. In Asia, the majority of the research was conducted in Kalimantan (Borneo). In Africa such studies were conducted in Cameroon. Cameroon has the highest percentage of previously logged forests than its Congo Basin neighbours since it has a higher population density (De Grandi et al., 2015). 22 Figure 2.3. Spatial distribution of studies where remote sensing analyzed selective logging in tropical forests. 2.3.3 Sensors used to assess selective logging in tropical forests Overall, twenty-six different sensors were used in the studies reviewed. Optical, RADAR (radio detection and ranging), and LiDAR (light detection and ranging) are identified as the three types of EO data, each with different characteristics. A significant number of articles utilized optical sensors (Figure 2.4a), such as the Landsat sensor (Costa et al., 2019), while others used the RapidEye (Franke et al., 2012; Deutscher et al., 2013), IKONOS (Read et al., 2003; Furusawa et al., 2004), Satellite Pour l’Observation de la Terre (SPOT)-4 (Guitet et al., 2012; Sofan et al., 2016), and SPOT-5 (Pithon et al., 2013). Very high resolution (VHR) and high resolution (HR) data were usually utilized as a single dataset or in conjunction with Landsat imagery as reference data. GeoEye-1 (Dalagnol et al., 2019), QuickBird (Hirschmugl et al., 2014), Pleiades (Langner et al., 2018), and the Chinese-Brazil Earth Resources Satellite 2B (CBERS-2B) High-Resolution Camera (HRC)—panchromatic, 2.5 m resolution)—(Anwar and Stein, 2012) were used as reference data. Wang et al. (2005) used IKONOS 1-m pansharpened imagery to validate canopy fractional cover maps resulting from Landsat Enhanced Thematic Mapper Plus (ETM+) data. In the published articles, the most recent Sentinel-2 (Lima et al. 2019), Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER)—(Broadbent et al., 2006), the MODerate Resolution Imaging Spectroradiometer (MODIS)—(Neba et al., 2014), and SmallSat (Descals et al., 2017) were also used to detect SL in tropical forests. Lima et al. (2019) used Sentinel-2 and Landsat-8 23 Operational Land Imager (OLI) images for monitoring SL in the Brazilian Amazon. Landsat- 8 detected 36.9% more area of SL than Sentinel-2 data. Logging infrastructure was better mapped from Sentinel-2 (43.2%) than Landsat-8 (35.5%) data. Neba et al. (2014) reported that SL reduces aboveground biomass (AGB) stock, and through linear regression modelling discovered that logging roads and normalized difference vegetation index (NDVI) values derived from MODIS 250 m can indirectly determine above-ground biomass (AGB) logged. Unmanned digital aerial photographs (DAPs – Ota et al., 2019), and LiDAR data (Englhart et al., 2013; Kent et al., 2015; Melendy et al., 2018). Qu et al. (2018) estimated leaf area index (LAI) from LiDAR height percentile metrics and compared it with MODIS product in a selectively logged tropical forest area in Eastern Amazonia. Wedeux and Coomes (2015) employed airborne laser scanning (ALS) data to measure the canopy of old-growth and selectively logged peat swamp forest across a peat dome in Central Kalimantan, Indonesia. The tropical areas are usually characterized by high cloud cover, thus RADAR sensors that can infiltrate clouds have been favoured in many cases (Figure 2.4b). Rauste et al. (2013) developed a technique to map SL in the northern Republic of Congo using ALOS PALSAR imagery acquired before and after the logging activities and attained an overall accuracy of 70.4%. The same technique was applied for TerrSAR-X data and achieved an overall accuracy of 53.6%. Lei et al. (2018) developed a new approach using TanDEM-X data to detect and quantify SL events in Tapajos National Forest, south of Santarem, Pará in the Brazilian Amazon region. A comparison of TanDEM-X results with ALOS-2 data qualitatively matches, confirming both the location and the epoch of the disturbance event. In the assessment of SL in tropical forests, a majority of the studies used medium spatial resolution sensors (52.2%), followed by VHR sensors (28%), HR sensors (18%), and coarse resolution (1.8%). 24 Figure 2.4. Sensors used to assess selective logging (SL) in tropical forests. (a) Frequency of optical sensors; (b) frequency of RADAR sensors. Regarding the sensors and area covered by each study, Monteiro et al. (2003) used Landsat Thematic Mapper (TM) and Enhanced Thematic Mapper Plus (ETM+) to detect area affected by logging in three study areas covering 900 km2, located in Sinop, Cláudia and Marcelândia in the state of Mato Grosso, Brazil. Field data were used to test the accuracy of the spectral mixture models to estimate the total area affected by logging (recent, old, and repeated). Marcelândia reported the greatest accuracy (80%), followed by Cláudia (73%) and Sinop (69%). The lowest area affected by logging was reported in Sinop (10,731 ha), followed by Marcelândia (19,391 ha) and Cláudia (25,276 ha). Souza et al. (2003) mapped forest canopy damage associated with logging and burning in a study site approximately 1600 km2 in Paragominas, northeast of Pará, and achieved an overall accuracy of 86% of the forest degradation map (non-forest, degraded forest, and logged forest). As well, a high correlation (R2 = 0.97) was observed between the total live AGB of degraded forest classes and the non- 25 photosynthetic vegetation (NPV) fraction image. The NPV fraction improved the ability to map old