REMOT E SENSING OF VEGETATION DYNAMICS AND CHANGE IN SEMI ARID ECOSYSTEMS By JESSICA ELIZABETH STEELE A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2014
Â© 2014 Jessica Elizabeth Steele
To strong coffee, good friends, Jillian Michaels, simplynoise, and red wine
4 ACKNOWLEDGMENTS This diss ertation was made possible by the immense support of my advisor , Dr. Jane Southworth, who offered years of research support, advice, comment s, and provided valuable feedback on my work . She persuaded me to pursue my research interests in an enthusiastic ma nner, which I found to be motivating and encouraging. I wish to express my sincere gratitude to Dr. Michael Binford, who has been a tremendous academic and moral support to me throughout my graduate studies at the University of Florida . He helped me collec t much of the data used in this dissertation, and always made time for me to discuss methods, models, and ethics in scientific research. Dr. Greg Kiker devoted much of his time to helping me develop my research proposal and forged many contacts in South Af rica that led to collaborations and work in Addo Elephant Park, which would not have been possible without him. I would also like to express my deepest appreciation to Dr. Stephanie Bohlman for her tremendous input toward improving this work ; t his disserta tion would not have this shape and form without her vast expertise, knowledge, and willingness to help. This research was suppor ted by NASA proposal NNX09AI25G; a University of Florida National Science Foundation Interdisciplinary Graduate Education and R esearch Traineeship grant ; and the University of Florida Center for African Studies, Department of Geography, Graduate Student Council, and College of Liberal Arts and Sciences. I also wish to acknowledge South African National Parks and Nelson Mandela M etropolitan University for their support of this work. I am most grateful for my experiences as a University of Florida National Science Foundation Interdisciplinary Graduate Education and Research Traineeship Fellow. The IGERT was extremely rewarding in t hat I worked with a group of bright and talented
5 people who ultimately taught me creative quantitative solutions to interdisciplinary problems and much more . Thank you, Dr. Ben Bolker and Dr. Craig Osenberg for your unfailing support and generosity. Thank you to Cohort 3, Dr. Tim othy Fullman, Dr. Michael Hyman, Dr. Cameron Browne, and Kristen Sauby, for your collaborative efforts and many lessons learned during our Workshop year . Thank you to my colleagues at the University of Florida for your support while I was working o n my PhD. I am deeply appreciative of Forrest Stevens , who was generous with his time, taught me innumerable things , and offered great advice . I also wish to thank Dr. Christian Russell, Dr. Pete r Waylen, Dr. Narcisa Pricope, Dr. Pinki Mondal, Dr. Shylock Muyengwa, Dr. Jason Blackburn, Dr . Br ian Child, Sarah Graves, Jake Hightower, Jonathan Trinastic , and many others . M any thanks to Desiree Price, Rhonda Black, and Julia Williams for their continued help during my time at UF. My sincerest appreciation to th e people at CLAS IT for persistent technical support throughout my time at UF; this work was made possible by their hard work and continuous assistance. Most importantly , thank you to my family and friends who were unfailing in their support of me . I could not hav e finished this PhD without you ; thank you for your amazing support , love, and f riendship over so many years.
6 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ .. 4 LIST OF TABLES ................................ ................................ ................................ ............ 8 LIST OF FIGURES ................................ ................................ ................................ .......... 9 LIST OF ABBREVIATIONS ................................ ................................ ........................... 11 ABSTRACT ................................ ................................ ................................ ................... 13 CHAPTER 1 INTRODUCTION ................................ ................................ ................................ .... 15 Vegetation Dynamics and Land Cover ................................ ................................ ... 15 Semi Arid Ecosystems and Remote Sensing ................................ ......................... 17 Imaging Spectrometry (Hyperspectral Imaging) ................................ ...................... 20 Dissertation Summary ................................ ................................ ............................. 22 2 ESTIMATING VEGETATION TRAITS IN SEMI ARI D SPECIES USING IN SITU SPECTRA ................................ ................................ ................................ ............... 26 Background ................................ ................................ ................................ ............. 26 Data and Methods ................................ ................................ ................................ .. 30 Study Area ................................ ................................ ................................ ........ 30 Trait Data ................................ ................................ ................................ .......... 32 Spectral Data ................................ ................................ ................................ .... 32 In situ data c ollection ................................ ................................ ................. 32 Spectral data processing ................................ ................................ ............ 33 Statistical Data Analyses ................................ ................................ .................. 35 Cluster analysis ................................ ................................ .......................... 35 Linear discriminant analysis ................................ ................................ ....... 36 Partial least squares regression ................................ ................................ . 37 Results ................................ ................................ ................................ .................... 37 Spectral Data ................................ ................................ ................................ .... 37 Spectral Clusters ................................ ................................ .............................. 38 Linear Discriminant Analysis ................................ ................................ ............ 38 Partial Least Squares Regression ................................ ................................ .... 39 Discussion ................................ ................................ ................................ .............. 40 Conclusions ................................ ................................ ................................ ............ 42 3 MULTIPLE ENDMEMBER SPECTRAL MIXTURE MODELING OF HYPERION IMAGERY IN THE CAPRIVI, NAMIBIA ................................ ................................ ... 54
7 Background ................................ ................................ ................................ ............. 54 Data and methods ................................ ................................ ................................ .. 57 Study Area ................................ ................................ ................................ ........ 57 Hype rion and Data Processing ................................ ................................ ......... 58 Field Data Collection ................................ ................................ ........................ 60 SMA and MESMA ................................ ................................ ............................ 60 MESMA Runs ................................ ................................ ................................ ... 62 Results ................................ ................................ ................................ .................... 64 Discussion ................................ ................................ ................................ .............. 65 Conclusions ................................ ................................ ................................ ............ 68 4 CHANGE DETECTION USING VEGETATION INDICES IN THE ADDO ELEPHANT NATIONAL PARK: 2009 2011 ................................ ......................... 79 Background ................................ ................................ ................................ ............. 79 Data and Methods ................................ ................................ ................................ .. 83 Study Area ................................ ................................ ................................ ........ 83 Remote Sensing Analyses ................................ ................................ ............... 84 Vegetation Indices ................................ ................................ ............................ 85 Climatological Data ................................ ................................ .......................... 8 6 Elephant Movement Data ................................ ................................ ................. 87 Results ................................ ................................ ................................ .................... 87 Vegetation Index Spatial Patterns and Change ................................ ................ 87 NDVI ................................ ................................ ................................ .......... 87 NDVIc ................................ ................................ ................................ ........ 88 NDWI ................................ ................................ ................................ ......... 88 Elephant Movement and Vegetation Indices ................................ .................... 89 NDVI ................................ ................................ ................................ .......... 89 NDVIc ................................ ................................ ................................ ........ 89 NDWI ................................ ................................ ................................ ......... 90 Discussion ................................ ................................ ................................ .............. 90 Conclusions ................................ ................................ ................................ ............ 92 5 CONCLU DING REMARKS ................................ ................................ ................... 104 Research Overvi ew ................................ ................................ .............................. 104 Contribution of the Study ................................ ................................ ...................... 109 LIST OF REFERENCES ................................ ................................ ............................. 111 BIOG RAPHICAL SKETCH ................................ ................................ .......................... 125
8 LIST OF TABLES Table page 2 1 In situ spectral data collection species list. ................................ ......................... 50 2 2 Cluster Analysis results for 4 cluster solution using continuum removed leaf and branch spectral means showing species membership i n clusters. .............. 51 2 3 LDA results showing pseudo R 2 value , model significance , and independent variable importance. ................................ ................................ ........................... 51 2 4 LDA results showing tests of equality of group means . ................................ ...... 52 2 5 PLS Regression results showing latent factors and variance explained . ............ 52 2 6 PLS Regression results showing the i mportance of reflectance band contributions to the latent factors in the model results. ................................ ....... 53 4 1 NDVI statistical values for different elephant use areas, botanical reserves and the entire park for 2009 and 2011. ................................ ............................. 101 4 2 NDVIc statistical values for different elephant use areas, botanic al reserves and the entire park for 2009 and 2011. ................................ ............................. 102 4 3 NDWI statistical values for different elephant use areas, botanic al reserves and the entire park for 2009 and 2011. ................................ ............................. 103
9 LIST OF FIGURES Figure page 1 1 Interrelationships among factors influenc ing savanna vegetation dynamics. ..... 25 2 1 Conceptual diagram of the vegetation features that influence spectral refl ectance to create optical types. ................................ ................................ ..... 43 2 2 Study area map for fiel d data collection training sites. ................................ ........ 44 2 3 Spectral d ata collection example for Lonchocarpus capassa . ............................ 45 2 4 Hierarchical cluster analysis species classification tree for vegetation species showing levels of vegetation clusters based on in situ spectral reflectance. ...... 46 2 5 In situ spectral reflectance measurements for vegetati on species collected in Namibia . ................................ ................................ ................................ ............. 47 2 6 Conti nuum removed spectral reflectance measurements for in situ spectra collected in Namibia . ................................ ................................ .......................... 48 2 7 First derivative spectral reflectance measurements for in si tu spectra collected in Namibia . ................................ ................................ .......................... 49 3 1 Study area map locating the training site data and Hyperion image subse t in the Caprivi Strip, Namibia . ................................ ................................ .................. 70 3 2 Hyperspectral versus multispectral data for Colophospermum mopane vegetation in the study area derived by Hyperion data (hyperspectral) and Hyperion data convolved to Landsat bands (multispectral). ............................... 71 3 3 False color image of the study area showing the fraction proportions of NPV, GV, and soil as derived by the Hyperion MESMA model and Landsat MESMA model. ................................ ................................ ................................ ... 72 3 4 Scatterplot of root mean square error associated with the Hyperion MESMA model versus the Landsat MESMA model. ................................ ......................... 73 3 5 Root Mean Squared Error images comparing the error assoc iated with MESMA models for the Hyp erion model and Landsat model. ............................ 74 3 6 Difference images showing the spatial distribution of fractional cover differences calculated by subtracting the Landsat convolved model from t he Hyperion model for the NPV , GV , and soil fraction s . ................................ .......... 75 3 7 Relationships between NPV cover in the field a nd NPV endmember images for Hyperion bands and Landsat convolved bands. ................................ ............ 76
10 3 8 Relationships between soil cover in the field an d soil endmember image s for Hyperion bands and Landsat convolved bands. ................................ ................. 77 3 9 Relationships between GV cover in the field and GV end member images for Hyperion bands and Landsat convolved bands. ................................ ................. 78 4 1 The Addo Elephant National Park in the Eastern Cape, South Africa ................ 95 4 2 SPOT5 image of the study area. ................................ ................................ ........ 96 4 3 Monthly climate data across the selected study periods in 2009 and 2011 and their antecedent conditions f or t emperature a nd p recipitation. .................... 97 4 4 Changes in Vegetation Indices ac ross the park from 2009 to 20 1 1 for NDVI, NDVI c and NDWI. ................................ ................................ ............................... 98 4 5 NDVI distribution for 2009 and 2011 for the study area. ................................ ..... 99 4 6 Vegetation Degradation in Main Camp, Addo Elephant National Park . ............ 100
11 LIST OF ABBREVIATIONS AENP Addo Elephant National Park ASD Analytical Spectral Devices AVHRR Advanced Very High Re solution Radiometer AVIRIS Earth Observing One Mission CAO Carnegie Airborne Observatory CASI Compact Airborne Spectrographic Imager CSR Competitive, Stress tolerant, Ruderal DESDynI Deformation, Ecosystem Structure and Dynamics of Ice ETM Enhanced T hematic Mapper FLAASH Fast Line of Sight Atmospheric Analysis of Spectral Hypercubes GIS Geographic Information System GPS Global Positioning System GV Green Vegetation HyMap Hyperspectral Mapper HyspIRI Hyperpectral Infrared Imager IPCC Intergovern mental Panel on Climate Change KAZA Kavango Zambezi Trans Boundary Conservation Area K S Kolmogorov Smirnov LDA Linear Discriminant Analysis LiDAR Light Detection and Ranging LMA Leaf Mass Per Unit Area MESMA Multiple Endmember Spectral Mixture Analy sis MIR Mid Infrared MODIS Moderate resolution Imaging Spectroradiometer
12 MODTRAN Moderate Resolution Atmospheric Transmission NASA National Aeronautics and Space Administration NDVI Normalized Difference Vegetation Index NDVIc Corrected Normalized Di fference Vegetation Index NDWI Normalized Difference Water Index NPV Non photosynthetic Vegetation PFT Plant Functional Type PLS Partial Least Squares PV Photosynthetic Vegetation RMSE Root Mean Squared Error SANParks South Africa National Parks SL A Specific Leaf Area SMA Spectral Mixture Analysis SPOT SWIR Shortwave Infrared USGS United States Geological Survey VI Vegetation Index VNIR Visible and Near Infrared
13 Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy REMOTE SENSING OF VEGETATION DYNAMICS AND CHANGE IN SEMI ARID ECOSYSTEMS By Jessica Steele August 2014 Cha ir: Jane Southworth Cochair: Greg Kiker Major: Geography providing the only means of collecting rapid, comprehensive surveys of the globe. As such, data collected by remote sensin g platforms have the potential to increase our understanding of functional diversity and biodiversity in ecosystems from space. The work presented here uses data from handheld and spaceborne multispectral and hyperspectral instruments ( imaging spectrometer s ) to study vegetation dynamics in semi arid ecosystems in southern Africa. These ecosystems are significant in local, regional, and global contexts, population and most of its rangeland, livestock and wil d herbivore biomass . Semi arid ecosystems cover more than half the area of the African continent and landscape c hanges have direct consequences to livelih oods, biogeochemical cycles, climate , and biodiversity . T he high spatial complexity exhibited in the semi arid ecosystems of southern Africa poses challenges to remote sensing techniques used for characterizi ng vegetation parameters. As such, m ultiple data sources are needed to estimate
14 vegetation attributes in semi arid ecosystems, and progress in remote sensing in these areas will come from e mploying multiple data products, including hyperspectral datasets. This work uses several approaches to better understand vegetation dynamics and change, addressing factors such as 1.) the link between biochemical tr aits and in situ reflectance spectroscopy, 2.) fractional land cover derived from linear unmixing models to compare hyperspectral and multispectral data , and 3.) the use of vegetation indices to monitor degradation in a national park . The results presented herein represent important steps toward better understanding and management of semi arid landscapes using satellite based estimates of vegetation parameters.
15 CHAPTER 1 INTRODUCTION Vegetation Dynamics and Land Cover Vegetation is important at every sca le. It provides global ecosystem services such as climate mitigation and carbon sequestration; regional and local resources such as agriculture, building materials and fuelwood; and it forms the basis of primary production in the terrestrial food chain. As such, understanding the causes and consequences of vegetation change has important significance at multiple scales. Vegetation dynamics reflect the interactions of numerous factors including climate, abiotic and biotic components, disturbance history, and anthropogenic influence. Vegetation is central to large number of critical biophysical processes and is a primary element of global terrestrial land cover 06), and land cover change at any spatiotemporal scale can be linked to changes observed at other scales (Lambin et al. 2006). Understanding the spatial and temporal patterns of land cover is a critical precursor to testing drivers of change, predicting ch ange into the future, and extrapolating to other areas. Land cover is intricately tied to land use, and most current land cover change is primarily caused by human activity. The collaboration between environmental, human, and remote sensing/GIS scientists has increased in recent decades, prompting a field in its own right: land change science (Turner et al. 2007; Lambin et al. 2006; Gutman et al. 2004). In order to comprehend land cover change on a global scale, scientists generally work within represe ntat ive classifications of land use situations, and seek to identify critical regions
16 or hot spots of change in terms of ecological value or vulnerability to land cover change (Lambin et al. 1999). This is a particularly interesting challenge in the landscapes of southern Africa where ecosystems are fundamentally important as they contribute significantly to biodiversity and human well being. Savannas act as major contributors to global net primary productivity (NPP) and are important factors in carbon cycles ( Williams et al. 2007 ). Toward that end, the IPCC forecasts for climate change impacts, adaptation, and vulnerability predict decreases in precipitation and increases in temperature in southern Africa, with approximately half of the semi arid parts of south ern Africa at moderate to high risk of desertification. Critical mainstays of local livelihoods, including the agricultural and tourism sectors, are under threat from climate variability and change and other stresses (Boko et al. 2007). Both variability and change in climate will affect the complex relationships between fire, grazers and browsers, and human activities in semi arid savannas, driving changes in vegetation patterns across southern Africa. Understanding these changes will be critical to envir onmental, social, and economic sustainability at multiple scales. Furthermore, the role of nitrogen in terrestrial ecosystem carbon dynamics is well known and its importance as the most important nutrient for plant growth long recognized (Heimann and Reich stein 2008; LeBauer and Treseder 2008; Knyazikhin et al. 2013 ). As such, quantifying nitrogen as an indicator of different plant materials, especially as it relates to photosynthesis, has been a primary goal in studies of ecosystem dynamics. Researchers ha ve been studying nitrogen at multiple scales from labs and greenhouses to field collection of leaf chemistry to data returned from airborne
17 and spaceborne instruments (Card et al. 1988; Mutanga et al. 2004; Martin et al. 2008; Ramoelo et al. 2013; Knyazikh in et al. 2013 ) understanding that insight into how this element is distributed on landscapes is useful to land managers interested in wildlife, livestock, and farming (Scholes and Walker 1993; Mutanga et al. 2004; Ramoelo et al. 2011 ), as well as to scien tists studying large scale carbon and nitrogen cycling (Ustin 2013; Ollinger 2011; Yuan 2007). Foliar nitrogen has been related to canopy and stand level traits including soil quality, NPP, and plant growth (Martin et al. 2008 and others). Given its import ance, researchers have been working on methods of detecting and predicting nitrogen using remote sensing, a complex task across a range of scales. This is particularly challenging in semi arid ecosystems where data sources are scarce and landscape characte ristics make some remote sensing methods untenable. Semi Arid Ecosystems and Remote Sensing Savannas occupy the continuum between forest and grassland, and a defining feature of this ecosystem is the coexistence of trees and grasses in the landscape. They cover more than half the area of the African continent and are significant in socio most of its rangeland, livestock and wild herbivore biomass (Scholes and Walker 1993; Sank aran et al. 2005; Sankaran and Anderson 2009 ; Hill et al. 2011; Hanan and Lehmann 2010 ). African savannas are highly heterogeneous mixed woody herbaceous systems, and the mechanisms controlling tree grass coexistence and the relative proportions of these c over types remain uncertain. Researchers have created multiple conceptual models to describe the interrelationships of similar sets of core variables governing savanna dynamics, generally agreeing that both positive and negative feedback loops are importa nt in determining structure and func tion (Sankaran and
18 Anderson 2009 ). The most important variables driving savanna patterns and processes are summarized in Figure 1 1 (Adapted from Hanan and Lehmann, 201 0 ). Interrelationships among climate, vegetation, d isturbance, and humans are shown in Figure 1 1. Humans graze animals on the landscape as economic capital and are responsible for controlled and uncontrolled fires. Humans have positive and negative interactions with browsers in that many browsing species eat crops planted for subsistence agriculture, but humans consume meat from browsing species and benefit from tourism based jobs dependent on conserving large herbivores. Grass biomass may compete for soil moisture with juvenile trees, but adult trees red uce soil moisture for grass growth (Walter 1971). Fire has negative effects on juvenile trees and may also reduce the long term survival of adult trees. Browsers and humans impact the survival and growth of juvenile and adult trees through consumption (Han an and Lehmann 201 0 ). Alongside these processes, anthropogenic land use pressures including fuelwood harvesting, pastoralism, and agriculture have led to changes in the relative abundance and condition of herbaceous and woody plant species (Scholes and Wal ker 1993; Asner et al. 2000; Sankaran and Anderson 2009 ). Studies of land use impacts on vegetation in semi arid ecosystems are frequently restricted to local scales with limited spatial and temporal context (Asner et al. 2000). Field data associated with measuring vegetation changes are typically limited to a locality, and consequently not readily applicable to regional or global extents. Furthermore, field based sampling methods are prohibitively expensive and time consuming at large spatial scales, and such methods are unsuitable for long term monitoring over large areas. Remote sensing datasets with larger temporal and spatial
19 extents have become an indispensable tool for ecological and conservation related applications (Kerr and Ostrovsky 2003; Ustin et al. 2004; Gillanders et al. 2008). Remote sensing holds a central role in relating information obtained at one scale to patterns and processes that manifest at another scale, providing the only means of viewing large areas of Earth at regular intervals and recording optical data which can be interpreted through the selective absorption and reflectance of light by plants (Ollinger 2011). Given the limitations of local field studies, remote sensing with satellites is an obvious tool for broader scale sampl ing. The availability of historical data in concert with data covering large spatial extents makes remote sensing data very useful (Asner et al. 1998; Mutanga et al. 2009; Nagendra and Gadgil 1999). For some applications, remote sensing may be the only da ta source available for measuring and monitoring environmental change (Kerr and Ostrovsky 2003; Turner et al. 2003). Turner et al. (2001) noted remote sensing is a handy tool at the landscape scale, and at regional and global scales it almost becomes a nec essity. The ability to translate remote sensing data into meaningful ecological information is increasing (Turner et al. 2001). The high spatial complexity exhibited in the semi arid ecosystems of southern Africa poses many challenges to using remote sensi ng techniques for characterizing vegetation like species classification, identification, biochemical concentration, and biomass (Mutanga et al. 2009). Studies using these techniques often grapple with achieving a high enough spatial resolution to resolve m ultiple spatial patterns observed at scales relevant to tree grass interactions (Asne r et al. 2000; Meyer et al. 2010 ). Tree grass systems are inherently difficult to classify using discrete techniques due to their heterogeneity (Hill and Hanan 201 0 ), and savannas exhibit a wide range of spatial
20 arrangements of trees, grasses, and shrubs, which are heterogeneous at scale s smaller than mos t pixel sizes (Meyer et al. 2010 ). Furthermore, open canopy systems are significantly influenced by highly variable refl ectance effects such as multiple scattering from bright soil and standing litter, two major components of the land surface. Thus it is necessary to explore data analyses and data sources when studying land cover change in savannas that maintain surface att ributes that affect the character of the land cover. Multispectral data at various resolutions have been used to retrieve woody cover and fractional cover of NPV, PV and bare ground in savannas. Phenology and unmixing methods in other regions have been use d to add extra dimensionality to estimate subpixel cover proportions of ecosystem structure (Hill and Hanan 201 0 ; Asner et al. 2005). But progress in the remote sensing of savannas will come from fusing data across scales and using multiple data products, including hyperspectral datasets. Imaging Spectrometry (Hyperspectral Imaging) Imaging spectrometry is a unique type of remote sensing because surface radiance is sampled in contiguous, narrow spectral bands also known as hyperspectral bands. Reflectance spectra from 400 2500 nanometers are often measured in the laboratory or field using spectro radio meters employing fiber optic cables to collect radiance data. Reflectance standards (white reference, dark reference) are used to normalize these measurement s, allowing data to be converted to reflectance. Hyperspectral sensors are most frequently used on aircraft but NASA included an imaging spectrometer (Hyperion) on its EO 1 satellite launched in November of 2000. A irborne and spaceborne sensors provide an image data cube with a reflectance spectrum for each picture element within the image (Im and Jensen 2008; Asner 1998). Hyperspectral information has been used for discriminating tree species in
21 environmentally complex landscapes (Clark et al. 2005; Nagend ra 2001), and researchers have linked vegetation across scales using imaging spectroscopy by testing and developing relationships of plant properties for use across broader bands (Gao 1996). With reference to southern Africa, studies in the Kruger National Park have used field spectroscopy and airborne imaging spectroscopy data to better understand foliar chemistry and map species (Skidmore et al. 2010; Mutanga and Skidmore 2004; Mutanga et al. 2004), but significant challenges exist in developing upscaling techniques to spaceborne spectrometers like Hyperion (Mutanga et al. 2009). Few studies have tested the potential of Hyperion data in estimating vegetation parameters (Mutanga et al. 2009). Some biophysical variables can be measured directly by a remote s ensing system, including vegetation pigments (e.g. chlorophyll a and b ) and leaf nutrient concentrations (Jensen 2005). Leaf optical properties are influenced by the concentration of chlorophyll and other biochemicals, water content, and leaf structure. Th ese characteristics are variable and the reflectance of vegetation is a result of complex, changing processes within the leaves, canopy, and stand (Kumar et al. 2001). The relative importance of tissue, canopy, and landscape factors on pixel level reflecta nce has been found to shift with plant composition and phenology (Asner et al. 2000) . As well, light scattering should be affected at multiple scales by t he expression of plant traits, which vary based on selective pressure for resources, and integrate pla nt functional strategies (Ollinger 2011). Trait driven differences in spectral signatures can be subtle, hence the need for hyperspectral signatures, and evaluating these at a scale
22 that is both ecologically meaningful and functionally visible to a sensor is quite challenging (Roelofsen et al. 2013). Airborne hyperspectral sensors such as AVIRIS, CASI, HyMap, and CAO have successfully mapped vegetation species. These platforms have many advantages, most notably their spatial resolution and minimal atmospher ic influence, but they are unavailable to many research applications due to high cost and limited geographical 1 platform offers a broad scale, low cost, repeatable collection of data. Hyperion origina ted as an experimental mission that has been extended to provide hyperspectral data to the public via the USGS up to present day. The main disadvantages of using Hyperion data include its moderate spatial resolution and atmospheric errors introduced as the signal travels from target to sensor, but studies have shown Hyperion data produce improved vegetation classifications and species discrimination (e.g. Thenkabail et al. 2004; Townsend and Foster 2003). These promising results alongside its advantages mak e Hyperion an attractive albeit challenging data source to assess in semi arid ecosystems . Dissertation Summary This dissertation investigate s the potential of using multispectral and hyperspectral datasets to improve estimates of vegetation parameters in the context of semi arid ecosystems through a series of three independent research manuscripts (Chapters 2 4). Chapter 2 examines the potential of mapping traits and aggregated functional types by examining relationships between in situ reflectance spect ra of key savanna vegetation species and corresponding plant trait data, including foliar nitrogen. Chapter 2 research applies two different approaches to linking plant trait data to reflectance spectra. First, k means and hierarchical cluster algorithms w ere applied to
23 reflectance spectra to determine spectral clusters in the data . Then a linear discriminant analysis was employed to explore the extent to which measured plant traits could determine optical clusters. Second, plant traits were regressed again st hyperspectral bands using partial least squares regression to explore how much variance in the trait data could be explained by wavelengths across the full spectrum. Chapter 3 uses reference spectra in linear unmixing models with a Hyperion image to der ive fractional abundances of land surface properties in the Caprivi Region in Namibia. Field data collected within one week of the Hyperion image acquisition were georeferenced and used to extract image endmembers with known species compositions. Multiple spectral libraries and various iterations of endmember selection were tested at differing levels of complexity to identify the optimal set of endmembers available to unmix the image. Furthermore, the endmembers were convolved to Landsat bands and tested in unmixing models to determine the added benefit of hyperspectral wavelengths in characterizing this landscape. Chapter 4 research quantifies changes in vegetation parameters in the Addo Elephant National Park, South Africa, (AENP) using SPOT data, providin g a set of continuous variables to be tested in models linking ecosystem drivers to observed change. This chapter is less experimental, producing applied science for the purpose of informing management, as it examine s for the first time a park wide charact erizati on of vegetation changes in AENP. Geography is concerned with studying the variability of phenomena over space and time. As well, phenomena are located in a region and place, be they ecological, human, cultural, demographic, or political in nature. In seeking to understand the
24 arrangement and interrelations of phenomena across the globe, geography is a unifying discipline that addresses the differentiation of human and physical earth structures and processes as they are formed by individual elements. A main focus within this is studying global terrestrial flora to better understand ecosystem patterns and processes at a range of scales. The concept of scale can be confusing to the extent that it has multiple referents, but identifying the correct scale of phenomena is a central problem spatially, temporally, and thematically. Issues of scale have always been central to geographic theory and research, and geographers investigate scale related problems as well as attempt to link processes and phenomena ac ross scales (Montello 2001). This dissertation contributes to the discipline by studying the variability of vegetation dynamics over space and time in southern Africa at multiple scales. This work analyzes multispectral and hyperspectral remote sensing dat asets, in concert with ancillary data, to examine landscape components and properties of vegetation in semi arid ecosystems. Remote sensing and GIS are fundamental geographical tools that, along with statistical analyses, comprise the methods of this disse rtation. The analyses in this work contribute to our understanding of both the spatial distribution of vegetation and how it is changing over time in both communal lands and highly managed conservation areas. The results of individual research papers arisi ng from this work both elucidate the need for more data and support further analyses linking these datasets to data with broader spatial and temporal ranges. These studies play a part in better understanding ecosystem properties and processes in semi arid regions in southern Africa, as well as inform studies utilizing similar remote sensing and GIS data in other places based on the applicability and limitations found herein.
25 Figure 1 1 Interrelationships among factors influencing savanna vegetation dynam ics. Arrows indicate interactions among system components with symbols showing neutral, negative, and + positive interac tions. For example, grazers reduce green biomass (G), but increased grass biomass can be neutral or deter grazers if it decreases fo rage quality ( ) or increases the presence of predators ( ). Adapted from Hanan and Lehmann, 201 0
26 CHAPTER 2 ESTIMATING VEGETATION TRAITS IN SEMI ARID SPECIES USING IN SITU SPECTRA Background Global and regional ecosystem models do a poor job of si mulating vegetation atmosphere interactions and need improved vegetation parameters, including leaf level and whole plant scales, to better account for photosynthesis and the carbon cycle (Wanxiao et al. 2008). Biochemical, morphological, physiological, an d phenological traits of plants have been measured and used by ecologists to study ecosystem functioning, community ecology, and ecosystem services (Homolova et al. 2013; Kattge et al. 2011; Lavorel et al. 2011; Diaz et al. 2004). Useful traits have been i dentified for influencing ecosystem functioning at various scales, including traits related to whole plants (growth form, plant height, photosynthesis rate), leaves (specific leaf area, leaf nitrogen and phosphorus, leaf phenology), and seeds (seed mass, g ermination efficiency) (Homolova et al . 2013; Lavorel and Garnier 2002 ; Diaz et al. 2004). At the ecosystem level, nitrogen availability is a limiting factor in primary production and its importance in plant growth is well known (LeBauer and Treseder 2008) . Therefore an improved understanding of canopy properties in general, and nitrogen content specifically, would provide direct information about photosynthesis and the terrestrial carbon cycle to improve our understanding of the relationship between the cl imate system and ecosystems (Us tin 2013; Knyazikhin et al. 2013 ). Researchers have become increasingly interested in retrieving information on plant traits and canopy properties using remotely sensed data. These data are appealing given that field measurem ents are limited to small areas and produce a snapshot in time of data for a limited number of species (Homolova et al. 2013). As
27 well, remote sensing offers the potential of standardized measurement globally and the ability to predict characteristics of communities for which direct measurement is not possible (Homolova et al. 2013; Martin et al. 2008). The most significant advantage of estimating traits via remote sensing is its ability to produce spatially explicit, continuous maps of traits repeatedly, and thus far nitrogen has been the most successfully estimated trait by optical remote sensing (Homolova et al. 2013). Optical diversity, the variation in spectral reflectance for different materials, serves as a proxy for other vegetation features that va ry with species diversity ( Figure 2 1 ). Therefore the combined effects of structural, phenological, and physiological properties of vegetation should be expressed in different reflectance signals, and many researchers have linked plant traits to canopy ref lectance by focusing on optical diversity (Ustin and Gamon 2010). B iodiversity estimation through optical diversity has been carried out successfully based on multiple traits evident in vegetation optical properties, including nitrogen, phosphorus, chlorop hyll, and carotenoids (Asner et al. 2009a, Asner and Martin 2009). Ollinger et al. (2008) and Martin et al. (2008) derive d relationships between canopy reflectance and canopy nitrogen, which was shown to distinguish broad groups of tree species. Skidmore e t al. (2010) mapped Colophospermum mopane trees and gr asses in the Kruger National Park, South Africa using airborne imaging spectrometry and foliar biochemicals (nitrogen and polyphenols). Researchers have also examined the relationship between canopy ref lectance and leaf structural properties such as specific leaf area (SLA ) (Asner et al. 2009a, Lymburner et al. 2000). SLA and its inverse, leaf mass per unit area (LMA) , were found to be associated with variations in leaf optical
28 properties, which translat e d to changes in canopy reflectance (Ollinger et al. 2008, Lymburner et al. 2000). The absorbing and scattering properties of natural surfaces are defined by their chemical bonds, water content, and three dimensional structure, which determine their reflec tanc e spectra (Ustin et al. 2004). These characteristics are variable and the reflectance of vegetation is the result of a set of complex, changing processes within the leaves, canopy, and stand. Portions of the electromagnetic spectrum are absorbed and re flected differently based on wavelen gth and plant characteristics. Researchers have explored various portions of the electromagnetic spectrum to differentiate plant constituent properties. For exa mple, intervals centered around 680, 850, 1650, and 2200 nm are useful for discriminating among different kinds of leaves (Kumar et al. 2001). The relative importance of tissue, canopy, and landscape factors on pixel level reflectance has been found to shift with plant composition an d phenology (Asner et al. 2000) . The expression of plant traits, which vary based on selective pressure for resources, should integrate their functional strategies and affect light scattering at multiple scales (Ollinger 2011). The concept of plant functional types (PFTs) was popula rized by J.P. Grime in the 1970 s. Grime et al. (1977) found plant responses to and effects on the environment to covary among species based on plant traits, and sets of traits were found to recur in group s of species . Further research into plant traits and func tional groups has shown trait values to vary along environmental gradients (Liu et al. 2012 ). Scientists are still exploring links between plant traits and functional type, including traits that are associated with plant response to environmental factors , and traits that determine the
29 effects of plants on ecosystem function (Lavorel and Garnier 2002). A generalization of plants into functional types is a necessary input to regional and global models to simplify modeling ecosystem processes and enable the pr ediction of environmental change (Ustin and Gamon 2010). Schmidtlein et al. (2012) mapped the three primary strategy types (CSR : competitive, stress tolerant, ruderal ) using airborne HyMap data and field plot data in Germany. They found many plant traits can be linked to optically relevant attributes and successfully mapped. Ollinger (2011) reviewed plant spectral properties with respect to leaf and canopy scales, noting that various plant traits show persistent and stable relationships to canopy reflecta nce including percent foliar nitrogen and leaf mass per unit area. In addition to linking traits to reflectance by seeking there exists a vast amount of information amassed on plant traits globally to be explored within the context of exis ting systems of plant functional types and remote sensing (Schmidtlein et al. 2012) . The extent to which plant traits and their response to environmental covariates ( disturbance, resource limitations, and other environmental factors ) can be generalized for predictive models will depend on detecting functional types with remote sensing (Ustin and Gamon 2010). These ideas are especially relevant in semi arid ecosystems, where woody herbaceous systems have experienced a change in plant fun ctional types in rece nt decades observed as an increase in woody plants (Archer 2010). This shift has brought with it a general consensus that woody plant encroachment is related to reductions in ecosystem functions and processes including decreased biodiversity, changes in th e spatial distribution of soil resources, decreases in annual NPP, and reductions in pastoral production (Oba et al. 2000; Van Auken 2000; Knapp et al. 2008; Van Auken
30 2009; Archer 2010). However, the phenomenon of woody plant encroachment is difficult to understand clearly in a mixed life form biome where plant species exist in multiple forms. The link between plant traits and functional types in savannas was considered by Skarpe (1996) in the Kalahari Desert, where precipitation variables related to soil moisture were found to be the main determinants for the distribution of PFTs on the landscape, but has not yet been approached within an imaging spectroscopy framework as this research considers. The objective of this study was to examine relationships be tween existing plant trait data and in situ reflectance spectra. Cluster algorithms were used to estimate optical clusters in field spectra and linear discriminant analysis employed to test whether traits would discriminate between optical clusters. Succes sful discrimination of reflectance groups by traits could serve as functional types on the landscape. Furthermore, spectral data and trait data were used in partial least squares regression models to examine what proportion of the variance in the trait dat a could be explained by the reflectance bands at wavelengths known to be associated with biochemical traits in plants . Field spectra collected in Namibia were processed and trait data from a range of sites in southern Africa were procured to answer the fol lowing research questions: 1.) What is the relationship between plant traits and in situ vegetation spectra, and 2.) Do plant trait data discriminate species clustered by spectral reflectance? Data and Methods Study Area The study area ( Figure 2 2 ) is located with in the Zambezi Region, until 2013 known as the Caprivi , in the extreme northeast of Namibia. This area is part of the
31 larger Kalahari Acacia Baikiaea Woodland ecoregion in the center of southern Africa. Vegetation is dominated by Zambezian Baikiaea and Mop ane woodlands as described by White (1983). These woodlands are comprised of Baikiaea plurijuga dominated forest, Mopane woodlands, scrub woodlands, and secondary grassland (White 1983). Other woody constituents of these dense communities include Acacia ni grescens , Acacia tortilis , Acacia erioloba , Combretum apiculatum , Combretum collinum , Dichrostachys cinerea , Peltophorum africanum , Piliostigma thonningii , Sclerocarya birrea , and Terminalia sericea . The herbaceous component of the landscape is variable an d dependent on rainfall totals, approximately 700 mm annually (White 1983; Low and Rebelo 1996; Gaughan and Waylen 2012). The study area is also within KAZA , one of the largest transfrontier conservation areas in Africa, which combines protected lands from five countries in an effort to conserve natural resources at the regional scale (van Aarde and Jackson 2007) . This is an extremely important region in southern Africa ecologically, socially, and economically, supporting many communities by a variety of li velihood activities including subsistence agriculture, pastoral grazing, controlled hunting, and wildlife based tourism. Various land use changes have occurred in recent decades, opening deb ate as to whether persistent l and degradation occurred and is occu rring (Thomas et al. 2000). Land cover change has direct effects locally, as well as implications in broader scale ecological processes. For these reasons, a better mechanistic understanding of vegetation is needed in this area, which would ultimately allo w for better management in agricultural and conservation lands, and improve estimations of nutrient cycling and
32 regional to global scale land models. Figure 2 2 includes sites of field data collection in Namibia located within a Hyperion image used in Chap ter 3 of this dissertation. Trait Data The largest and most comprehensive organization compiling plant traits globally is the TRY database (Kattge et al. 2011), hosted at the Max Planck Institute for Biogeochemistry in Jena, Germany. The TRY database has d ata for 661 plant trait measurements associated with the species for which field data were collected for this study (Table 2 1). A research proposal submitted to the TRY Steering Committee procured these traits, and data were released upon its approval. Th ese data were collected over a range of dates and sites across southern Africa by disparate research groups, and trait data were not available equally across all species. Of the full dataset, I omitted t raits if they were not measured frequently across all species of interest ; and s pecies of interest for which spectra were collected were omitted when little trait information was available . This produced a dataset of 11 common traits measured across 26 vegetation species (Craine et al. 2009; Reich et al. 200 9; Kattge et al. 2009; Kerkhoff et al. 2006; Han et al. 2005; Wright et al. 2004). These traits were germination efficiency, leaf compoundedness, leaf width, leaf 15 N abundance, leaf N dry mass, mycorrhizal association, nitrogen fixing capability, plant he ight, seed mass, specific stem density, and woodiness. Spectral Data In s itu data c ollection Field observations were made between 3 May and 24 June 2011. Sites for recording reflectance spectra were chosen opportunistically to select for the best represent ations of vegetation species and soil types within all major land covers.
33 Reflectance spectra from 350 nm to 2500 nm were measured for all materials in the field using an ASD (Analytical Spectral Devices, Inc., Boulder, CO, USA) Field Spec Pro spectroradio meter. Measurements were calibrated against a white Spectralon TM panel and internal dark current for each new material, every 4 5 minutes, or when lighting conditions changed. For each sample, 25 50 spectra were recorded and each spectrum used in analyse s was the mean of the individual measurements. We used a bare optical cable with a field of view of approximately 20.5 degrees. Spectral measurements collected at the leaf level were made for leaves removed from the plant, where the field of view of the se nsor was fully occupied by a single layer of leaves and measurements made immediately. Spectral measurements collected at the branc h level were made using a field of view instrument to approximate the 20.5 degree foreoptic view and reflectance was recorded by pointing the foreoptic cable at the densest sunlit area of the tree or shrub. The angle of inclination varied for these measurements based on the density of leaves on the tree or shrub and position of the sun. Bare ground or sky was present in these me asurements as Figure 2 3c shows. Table 2 1 lists the species for which spectral measurements were recorded. Figure 2 3 shows an example of spectral data collection at the leaf and branch levels for Lonchocarpus capass a . Spectral data p rocessing The ASD Fie ld Spec Pro spectroradiometer measures spectral radiance between 350 and 2500 nm, with a sampling interval of 2 nm and a spectral resolution of 10 nm. The instrument includes three spectrometers to sample visible and near infrared (VNIR) and two shortwave infrared (SWIR1 and SWIR2) regions of the electromagnetic spectrum (Hatchell, 1999). Standard preprocessing of these data yields reflectance values for each consecutive nanometer between 350 and 2500 nm. Spectral bands
34 exhibiting spectral inconsistencies o r strong atmospheric absorption (350 390 nm, 1350 1440 nm, 1790 1990 nm, 2360 2500 nm) were removed from the data (Thenkabail et al. 2004). Data transformations are commonly used to enhance spectral data and resolve spectral features of interest. C ommon spectral enhancements include derivative analysis and continuum removal, and these techniques are used to separate noise from between parameters of interest and spectral data (Curran et al. 2001; Kumar and Skidmore 1998). Derivative analysis enhances subtle spectral details while suppressing that illumination differences and atmospheric effects can be reduced using spectral derivatives and derivatives can better discriminate between materials than raw reflectance (Kumar and Skidmore 1998; Ph ilpot 1991). Also recently researchers have been using continuum removal to isolate regions of the spectrum known to be associated with biochemicals found in plants (Curran et al. 2001). In continuum removal, a mathematical function described as an appare nt continuum is used to isolate absorption features for analysis of a spectrum, where the continuum represents the absorption other than that of interest. Continuum removal enhances the wavelength band information by removing this slope (Mutanga and Skidmo re 2003), standardizing the isolated absorption features, and improving comparison (Pu et al. 2003). First derivative and continuum removed spectra were computed using the mean spectral reflectan ce for each vegetation species, from both
35 leaf and branch lev el spectral measurements, and the resulting data set was used as independent variables in statistical models. Statistical Data Analyses Cluster a nalysis Continuum removed and first derivative leaf and branch spectra l means were used in cluster analyses. Dif ferent clustering algorithms partition data differently, and determining an appropriate cluster analys Bootstrapping, resampling methods, and partitioning data are techniques used to assess the stability and validity of cluste rs , which can be problematic when working with high d imension datasets (Jain 2008). For this study, two cluster approaches were performed in tandem using SPSS version 22 (IBM Corp. 2013) to create spectral clusters of vegetation species. First, hierarchica l cluster analysis was used to identify groups of species within similar clusters joined in a classification tree (Figure 2 4). The dendrogram allows the analyst to see how clusters are joined hierarchically, providing multiple solutions of groups. Each ob ject starts as its own cluster and as separation criteria are relaxed; the most similar clusters are joined until all clusters are linked. C luster solutions for 2 10 clusters were selected to define membership. The analysis also produces a table showing ho w cluster membership per species changes as the number of clusters increases, where a good solution is indicated by a jump or gap in cluster membership (IBM Corp. 2013). Solutions of 4, 5, and 6 clusters were the best solutions using continuum removed spec tra, and 3 clusters was the best cluster solution using first derivative spectra.
36 Second, these cluster solutions were used as the k value in k means cluster analysis to compare species membership for the two algorithms. K means cluster analysis divides n observations into k clusters, where each observation belongs to the cluster with the nearest mean. K means is an unsupervised classification used to determine how the data are organized by finding the centers of natural clusters in the dat a (Hartigan and W ong 1979). The results from K means u sing 3, 4, 5, and 6 clusters were compared to the groupings from the hierarchical cluster analyses. The same species were clustered consistently using both algorithms, and therefore groups of species present in 3, 4, 5, and 6 cluster solutions were tested as groups in discriminan t analysis models. Linear discriminant a nalysis Linear d iscriminant analysis (LDA) is a statistical tool used to predict a categorical dependent variable by one or more continuous variables. The analysis is used for determining whether a set of variables is effective a t predicting category membership where t he groups are known a priori. Given a set of independent variables, LDA attempts to find linear combinations of those variables that best dis criminate between groups of cases. The linear combinations are called discriminant functions and have the form, D ik = b 0k + b 1k x i1 pk x ip ( 2 1) where d ik is the value if the k th discriminant function for the i th case p is the number of pred ictors b jk is the value if the j th coefficient of the k th function x ij is the value of the i th case of the j th predictor .
37 function coefficients; tests of equality of group me lambda; and classification results. All LDA models were run in SPSS version 22 using the full set of trait data to discriminate between 3, 4, 5, and 6 cluster solutions. Partial least squares r egression PLS Regression is an eigenvector analysis that reduces the full reflectance spectrum to a smaller set of independent factors, using the corresponding trait data in the spectral decomposition process (Martin et al. 2008). The procedure estimates PLS regression models using a combination of principal components analysis and multiple regression. PLS extracts a set of latent factors to explain as much of the covariance as possible between the independent and dependent variables. Then a regression step predicts values of the depe ndent variables using the decomposed independent variables. Model diagnostics for determining fit include proportion of variance explained by each latent factor; factor weights and loadings; independent variable importance; and regression parameter estimat es (IBM Corp. 2011). PLS regression models were run on the full set of data using all traits and spectral information, and then on a subset of data using bands known to be associated with protein absorption related to nitrogen compounds (1500, 1510, 1680, 1690, 1730, 1740, 2130, 2170, 2180, 2240, 2290, 2300, 2350) and nitrogen traits (leaf 15 N abundance, leaf N dry mass). Results Spectral Data Figure 2 5 shows the spectral measurements of 7 species listed as dominant in this region by White (1983).Looking a t these reflectance spectra, we would expect to see differences in traits, and possibly positive results linking these species to
38 hyperspectral imagery, because the signatures show clear differences in absorption and reflectance across the electromagnetic spectrum. Differentiation can be seen in the NIR region (700 1300 nm) where reflectance has been related to plant phenology and nitrogen (Im and Jensen 2008; Kumar et al. 2001); the MIR region (1400 1770 nm) which is characterized by lignin, cellulose, protein, and nitrogen absorption (Kumar et al. 2001); as well as the SWIR region (2000 2340 nm) which has also been related to cellulose, protein, and nitrogen absorption (Kumar et al. 2001). Continuum removed vegetation spectra for the same dominant sp ecies show clear differences in transformed spectral reflectance in the visible portion of the spectrum (400 700 nm), the NIR wavelengths between 1100 1300 nm, the MIR wavelengths between 1470 1740 nm, and SWIR bands from 2168 2348 nm (Figure 2 6). Furthermore first derivative spectral transformations show differences across the electromagnetic spectrum in the wavelengths between 530 800 nm; 1090 1200 nm; 1460 1560 nm, and 2015 2320 nm (Figure 2 7). Spectral Clusters In both the hierarchical and k means cluster analyses, outliers were seen for every cluster solution. Species that were consistent outliers included Pecheul loeschea leubnitziae , Terminalia sericea , Ximenia americana , and Combretum collinum . The species distribution for the 4 clu ster solution using continuum removed spectral means for leaf and branch measurements is shown in Table 2 2. This cluster solution provided the best results in the LDA models. Linear Discriminant Analysis A total of 12 LDA models were run to predict group membership for vegetation species based on trait data. These models used four cluster solutions (3, 4, 5, 6
39 clusters) for continuum removed spectra and one cluster solution (3 clusters) for first derivative spectra. The LDA models were unable to predict gr oup membership at a statistically significant level using trait data . The best model was for the 4 cluster solution using continuum removed spectra and results are reported in Tables 2 3 and 2 4 . The model chose two of the eleven traits to maximize discrim ination between groups: germination efficiency and leaf 15 N abundance. The approximate R 2 value for the model was 0.37, indicating the traits had little effect on the model outcome (Table 2 3 Lambda was 0.625, indicating the traits are unable to p redict group membership at a statistically significant level. Furthermore, traits were not able to discriminate between groups ( Table 2 4 ) as the clusters were not significantly different from each other for any trait. Lastly, the standardized coefficients for the traits are reported in Table 2 3 , indicating Leaf 15 N abundance had the highest loading in the model. This model was only able to predict classification correctly in one of the four clusters , cluster 4 in Table 2 2 . Partial Least Squares Regressio n PLS regression models using nitrogen specific first derivative transformed spectral bands (1500, 1510, 1680, 1690, 1730, 1740, 2130, 2170, 2180, 2240, 2290, 2300, and 2350) and nitrogen traits (leaf 15 N abundance, leaf N dry mass) performed slightly bett er than models run using the full suites of traits and both transformations of spectral data. This model accounted for 40.7% of variance in the trait data using all 5 latent factors with an associated R 2 value of 0.331 (Table 2 5 ). This model also accounte d for 81.2% of variance in the spectral data, utilizing almost all of the information available in the bands. The most important bands in the model were 1680,
40 1740, 2170, 2180, and 2240 (Table 2 6 ), and rerunning the analyses with this subset of bands did not yield better results. Discussion It was expected that species would cluster spectrally based on nitrogen and leaf structure as those traits have been successfully linked to reflectance data (Asner 1998; Dennison 2002; Martin et al. 2008). These models were constructed and tested using the most comprehensive trait and field spectra datasets assembled to date for these species globally, but the fact that the models did not link traits to reflectance well may be due to limitations of these datasets. First, the trait data were collected from different plants at disparate sites across different times and seasons than were the spectral data. Analyses utilizing trait data collected on the same species for which reflectance spectra are recorded may yield better results. Second, key traits that have been linked to optical reflectance in the literature (chlorophylls a and b, lignin, tannins) were not accounted for in this dataset. Simply stated, more trait data for these species are needed, and data collection that focus es on leaf/canopy chemistry and structure will likely yield the best results . Lastly, the idea of clustering species based on reflectance values makes sense when considering the convergence of reflectance with biochemical traits (as described in Olli nger 2011), but cluster analyses force data into groups where natural clusters may not exist in the data . The mean spectrum for each vegetation species was used to compute the model inputs , and models that incorporate a measure of the intraspecies variabil ity in the reflectance signal may yield better results. Furthermore many of the species considered in this work do not exhibit dense canopy cover or are mostly senesced in the dry season when the spectral data, and some of the trait data, were collected. S uccessful PLS predictions (e.g. Martin et al. 2008) have used field data
41 collected during the peak of the growing season. PLS models are also expected to benefit from a more robust trait dataset, and from regressing traits collected at the same time as ref lectance measurements. Linking trait expressions to hyperspectral reflectance is important because trait estimation in communities of species is difficult, and estimating trait data using spectral reflectance offers an approach that has immense potential, namely to apply to any group of plant species for which reflectance data have been collected. It is challenging to aggregate trait values to a community value that is ecologically meaningful and also corresponds to f unctional attributes that are visible to a sensor and influence the spectral signature (Roelofsen 2013). Successful prediction of biochemistry using reflectance data opens the potential for estimating canopy chemistry at larger scales, and providing quantitative estimates of foliar chemistry at sites for which there is no possibility of conducting field data collections (Martin et al. 2008). Scaling signals from the field to canopy level is still not resolved. One attempt by Roelofsen et al. (2013) aggregated traits measured in situ and biomass h arvested in the field to test the relationship between traits and canopy spectral signal across broad vegeta tion types in the Netherlands. Lignin, phenol, and tannin were accurately modeled and validated at the leaf level, but did not scale up to the canop y surface. They found nutrient related traits to be moderately correlated with spectral data with R 2 values up to 0.74. Issues of scaling to the canopy level are complicated by multiple scattering effects as the signature of the canopy is driven by many fa ctors, not just leaf biochemistry, including background litter and bare soil. These factors are not included in trait values, and numerous studies have shown them to be consistently confounding
42 factors in spectrum trait associations (Numata et al. 2008; Ro elofsen et al. 2013). As well, stronger relationships between reflectance and mass measurements or reflectance and leaf surface traits found by Roelofsen et al. suggest that it is important to further explore the ecological theory guiding how trait values relate to structural measures such as Leaf Area Index and Specific Leaf Area (Roelofsen et al. 2013). Incorporating 3 D structure may help aggregate trait values to an ecologically meaningful community value . Finally, developing better relationships betwee n traits and hyperspectral reflectance may allow for the generation of new indices that ecologists, biologists, and biogeographers can use to extract information about the Earth using sensors that encompass larger spatial and temporal scales. Conclusions T his work emphasizes the need for more trait data to be collected for key savanna species. Spectral data collected during the wet season may yield better results, and spectral and trait data collected from the same plants may produce stronger relationships. Partial Least Squares Regression outperformed other statistical methods used in this chapter, explaining 41% of nitrogen variability using nitrogen related reflectance bands. These results are encouraging given the datasets were collected at various times and places in southern Africa. This work represents the first comprehensive set of in situ reflectance data for these key savanna species, and will be made available to researchers via the TRY database.
43 Figure 2 1. Conceptual diagram of the vegetatio n features that influence spectral reflectance to create optical types. Adapted from Ustin and Gamon 2010.
44 Figure 2 2. Study area map for field data collection training sites as teal triangles on the image . Background imagery is Hyperion shown in R,G,B = NIR (773 nm), Red (671 nm), Green (529 nm).
45 (a) (b) (c) Figure 2 3. Spectral data collection example for Lonchocarpus capassa , or Rain Tree, The tree measured at this site is shown in full (a); with leaves removed for leaf measurement (b); and thr ough a field of view instrument for branch measurement (c).
46 Figure 2 4. Hierarchical cluster analysis species classification tree for vegetation species showing levels of vegetation clusters based on in situ spectral reflectance.
47 Figure 2 5. In situ sp ectral reflectance measurements for vegetation species collected in Namibia. Vegetation species shown represent dominant species in this region as described by White (1983).
48 Figure 2 6. Continuum removed spectral reflectance measurements for in situ spec tra collected in Namibia. Vegetation species shown represent dominant species in this region as described by White (1983).
49 Figure 2 7. First derivative spectral reflectance measurements for in situ spectra collected in Namibia. Vegetation species shown r epresent dominant species in this region as described by White (1983).
50 Table 2 1. In situ spectral data collection species list. Scientific Name Common Name Acacia erioloba Camel Thorn Acacia tortilis Umbrella Thorn Albizia versicolor Large leaved Fals e thorn Baikiaea plurijuga Rhodesian Teak, Zambian Teak Baphia massaiensis Sand Camwood Berchemia discolor Brown Ivory, Wild Almond Boscia albitrunca Shepherd's Tree Burkea africana Wild Syringa Colophospermum mopane Mopane Combretum collinum Bushwi llow Combretum elaeagnoides Oleaster Bushwillow Combretum hereroense Russett Bushwillow Combretum imberbe Leadwood Diospyrus lycioides Bluebush, Star apple Diospyrus mespiliformis Jackal Berry Ficus sycomorus Sycamore Fig Guibourtia coleosperma Larg e False Mopane Lonchocarpus capassa Rain Tree, Apple leaf Pecheul loeschea leubnitziae Wild Sage Peltophorum africanum Weeping Wattle Phyllanthus reticulatus Potato Plant Piliostigma thonningii Monkey Bread, Wild Bauhinia Sclerocarya birrea Marula T erminalia prunioides Lowveld Cluster leaf Terminalia sericea Silver Cluster leaf Ximenia americana Small Sourplum Ximenia caffra Large Sourplum Ziziphus mucronata Buffalo Thorn Grass
51 Table 2 2. Cluster Analysis results for 4 cluster solution usi ng continuum removed leaf and branch spectral means showing species membership in clusters. L signifies leaf spectra; B signifies branch spectra. Cluster 1 Cluster 2 Cluster 3 Cluster 4 Burkea africana L Acacia erioloba L Albizia versicolor L Acacia eriol oba B Berchemia discolor B Acacia tortilis B Boscia albitrunca B Acacia tortilis B Berchemia discolor L Acacia tortilis L Diospyrus lycioides B Albizia versicolor B Combretum elaeagnoides B Acacia tortilis L Diospyrus lycioides L Baphia massaiensis B C ombretum elaeagnoides L Burkea africana B Ficus sycomorus B Baphia massaiensis L Combretum hereroense L Boscia albitrunca L Ximenia americana B Combretum hereroense B Combretum hereroense L Combretum collinum B Ximenia americana L Combretum hereroense L Colophospermum mopane B Combretum hereroense B Combretum imberbe B Colophospermum mopane L Colophospermum mopane B Combretum imberbe L Diospyrus mespiliformis B Colophospermum mopane B Colophospermum mopane L Diospyrus mespiliformis L Lonchocarpus c apassa B Colophospermum mopane L Ficus sycomorus L Peltophorum africanum B Ficus sycomorus B Guibourtia coleosperma B Sclerocarya birrea B Ficus sycomorus L Guibourtia coleosperma L Ximenia caffra L Lonchocarpus capassa B Guibourtia coleosperma L Z iziphus mucronata B Lonchocarpus capassa L Lonchocarpus capassa L Peltophorum africanum L Sclerocarya birrea L Terminalia prunioides B Terminalia prunioides L Terminalia sericea L Ziziphus mucronata L Table 2 3 . LDA results showing pseud o R 2 value (Canonical Correlation); model Linear Discriminant Analysis Model Diagnostics Canonical Correlation 0.612 (R 2 = 0.37 ) Wilks' Lambda 0.625 Standardized Canonical Discriminant F unction Coefficients (Variable Importance in Model) : Germination efficiency 0.370 Leaf 15N abundance 0.965
52 Table 2 4 . LDA results showing tests of equality of group means indicating differences in independent variables are not statistically s ignificant between groups. Tests of Equality of Group Means Wilks' Lambda F df1 df2 Sig. Germination efficiency 0.646 1.645 1 3 0.29 Leaf compoundedness .a Leaf 15 N abundance 0.625 1.800 1 3 0.272 Leaf width 0.702 1.274 1 3 0.341 Leaf N dr y mass 0.680 1.412 1 3 0.320 Mycorrhizal association 0.938 0.200 1 3 0.685 N fixing 0.833 0.600 1 3 0.495 Plant height 0.957 0.133 1 3 0.739 Seed mass 0.664 1.520 1 3 0.305 SSD 0.876 0.425 1 3 0.561 Woodiness .a .a Cannot be computed because this variable is a constant Table 2 5 . PLS Regression results showing latent factors and variance explained. X variance accounts for spectral information in the bands; Y variance accounts for trait data. Proportion of Variance Explained Latent Fa ctors Statistics X Variance Cumulative X Variance Y Variance Cumulative Y Variance (R square) Adjusted R square 1 0.419 0.419 0.121 0.121 0.101 2 0.183 0.602 0.168 0.289 0.255 3 0.107 0.709 0.068 0.357 0.31 4 0.041 0.75 0.045 0.402 0.342 5 0.062 0. 812 0.005 0.407 0.331
53 Table 2 6 . PLS Regression results showing the importance of reflectance band contributions to the latent factors in the model results. Variable Importance in the Projection Variables Latent Factors 1 2 3 4 5 Band 1500 1.215 0. 789 0.711 0.67 0.684 Band 1510 1.276 0.827 0.745 0.703 0.705 Band 1680 0.926 1.824 1.645 1.601 1.591 Band 1690 0.265 0.95 1.023 0.967 0.961 Band 1730 1.306 0.921 0.914 0.923 0.921 Band 1740 1.27 1.326 1.346 1.279 1.279 Band 2130 0.543 0.894 0.855 0.8 09 0.813 Band 2170 0.923 0.947 1.408 1.347 1.342 Band 2180 1.629 1.378 1.352 1.597 1.594 Band 2240 0.099 1.302 1.211 1.145 1.147 Band 2290 0.141 0.094 0.551 0.52 0.526 Band 2300 1.059 0.698 0.635 0.698 0.72 Band 2350 0.836 0.708 0.637 0.704 0.701 Cu mulative Variable Importance
54 CHAPTER 3 MULTIPLE ENDMEMBER SPECTRAL MIXTURE MODELING OF HYPERION IMAGERY IN THE CAPRIVI, NAMIBIA Background The spectral properties of vegetation are determined by their biophysical and chemical characteristics, as well a s their structural attributes (Asner 1998; Ollinger 2011; Hill 2004 ) . Deriving accurate estimates of biophysical properties, such as vegetation type, cover, leaf area index, and biomass, remains a challenge in many ecosystems (Numata et al. 2008), particul arly in savanna ecosystems (Hill et al. 2013; Hill and Hanan 2010; Guerschman et al. 2009). Remote sensing in savannas is made difficult by the complexity of materials, including diverse soils, grass, shrubs and trees, on the landscape, which includes subs tantial spectral diversity at fine spatial scales. This is further complicated by seasonal patterns in which some life forms are deciduous or dry up and others remain green. Remote sensing techniques like classification and estimates of biochemistry and bi omass are considerably challenging in semi arid regions (Mutanga et al. 2009). As well, the spatial extent of green vegetation (GV), non photosynthetic vegetation (NPV), and bare soil is notoriously difficult to resolve in these areas using satellite imag ery due to fine scale heterogeneity on the ground and spectral ambiguity between NPV and soil (Numata et al. 2008). Measuring the fractional cover of GV, NPV, and soil is important for ecosystem modeling in savannas, notably as the differences in these con stituents control the ability of the landscape to capture and retain water (Hill et al. 2013; Guerschman et al. 2009). This is directly related to patch structure in savannas, and remote sensing has great potential for characterizing patch structure in the se systems (Hill et al. 2013; Asner et al. 2011; Ustin et al. 2004). The amounts of GV
55 and NPV in ecosystems also determine other important landscape processes such as nutrient exchange, carbon uptake, and fire frequency and intensity (Guerschman et al. 20 09). Savannas are inherently difficult to classify using discrete techniques (Hill and Hanan 2010) and majority rule classifiers introduce errors to land surface estimates, leading to difficulties in identifying materials in a pixel (Nagendra and Rocchini 2008). Alternative techniques are needed to estimate land surface parameters in savannas (Hill and Hanan 2010). Estimating vegetation fractions using broad band sensors, such as Landsat, has fallen short of accurately distinguishing between NPV and soil, and is inadequate to inform changes in vegetation properties on the spatial scale of land use and land cover change in savannas (Guerschman et al. 2009; Asner and Lobell 2000). Broad band sensors do not offer the advantages of multiple wavelengths across s pectral regions as do hyperspectral sensors, specifically bands associated with resolving the spectral ambiguity between NPV and soil. In areas of sparse vegetation cover, the SWIR2 region (2000 2400 nm) has shown enhanced capabilities for discriminating between these fractions, and has been found to be beneficial when analyzing land degradation in semi arid landscapes (Asner and Heidebrecht 2003; Huete et al. 2003). Therefore, hyperspectral sensors can be expected to outperform broad band sensors in sava nna ecosystems, but there are relatively few studies in savannas using hyperspectral data (Mutanga et al. 2009). Increased heterogeneity on the landscape is accompanied by increased dimensionality in spectral space. The most common technique used to separa te spectra within mixed pixels is spectral mixture analysis or SMA (Somers et al. 2011).
56 Successful SMA is based on appropriate endmember selection, where all spectra in the image fall within the boundaries of the endmembers. Selecting good endmembers inc ludes identifying the number and type of endmembers found in a scene, and their corresponding signatures to encompass the variability within cover types (Somers et al. 2011; Elmore et al. 2000; Sabol et al. 1992). Endmember signatures can have a large degr ee of spectral variability resulting from the range of cover types occurring within broad classes. This can reduce the accuracy of SMA pixel proportion estimates (Theseira et al. 2010), leading to the premise of multiple endmember mixture modeling, or MESM A: combinations of two or three classes selected from a larger number of well discriminated classes will result in better proportional estimates than using two or three aggregated classes. In MESMA, all endmember combinations are applied to each pixel in a scene and the combination of endmembers that best models each composite pixel signature is identified (Theseira et al. 2010; Roberts et al. 1998). MESMA has been used with multispectral datasets to resolve fractional covers with modest results but MESMA has seen greater successes using hyperspectral datasets, including modeling complex environments such as urban areas and shrub communities (Franke et al. 2009; Roberts et al. 2003). Airborne hyperspectral data are not globally accessible or cost effective for many applications, but spaceborne hyperspectral data from the Hyperion instrument have been useful in a variety of applications including agriculture, forestry, and managed pastures (Datt et al. 2003; Goodenough et al. 2003; Numata et al. 2008). Due t o challenges associated with remote sensing in savannas, the added spectral information of Hyperion may provide
57 better estimates of the land surface, especially in separating NPV and soil, which could then be linked to other datasets for analyses at broade r temporal and spatial scales. The objective of this study was to test the ability of hyperspectral data to provide better estimates of land surface materials, i.e. fractional abundances of vegetation cover, as compared to broad band multispectral data in a savanna ecosystem. Information derived from Hyperion hyperspectral data should improve the quantification of savannas and outperform land cover characterization by multispectral sensors such as Landsat. Research questions for this work include: 1. ) To what e xtent do MESMA models account for variable endmember conditions in a semi arid savanna ecosystem using Hyperion data, and 2.) Do spaceborne hyperspectral data provide better estimates of land surface parameters than b road band multispectral sensors ? Data and m ethods Study Area The study area (Figure 3 1) is located within the Zambezi Region, until 2013 known as the Caprivi, in the extreme northeast of Namibia. This area is part of the larger Kalahari Acacia Baikiaea Woodland ecoregion in the center of southern Africa. Vegetation is dominated by Zambezian Baikiaea and Mopane woodlands as described by White (1983). These woodlands are comprised of Baikiaea plurijuga dominated forest, Mopane woodlands, scrub woodlands, and secondary grassland (White 1983). Other wo ody constituents of these dense communities include Acacia nigrescens , Acacia tortilis , Acacia erioloba , Combretum apiculatum , Combretum collinum , Dichrostachys cinerea , Peltophorum africanum , Piliostigma thonningii , Sclerocarya birrea , and Terminalia seri cea . The herbaceous component of the landscape is variable
58 and dependent on rainfall totals, approximately 700 mm annually (White 1983; Low and Rebelo 1996; Gaughan and Waylen 2012). The study area is also within KAZA, one of the largest transfrontier cons ervation areas in Africa, which combines protected lands from five countries in an effort to conserve natural resources at the regional scale (van Aarde and Jackson 2007). This is an extremely important region in southern Africa ecologically, socially, and economically, supporting many communities by a variety of livelihood activities including subsistence agriculture, pastoral grazing, controlled hunting, and wildlife based tourism. Various land use changes have occurred in this area in recent decades, ope ning debate as to whether persistent land degradation occurred and is occurring (Thomas et al. 2000). Woodland conversion is prominent among these threats, and has important implications for biogeochemical cycles, climate, and biodiversity in regional and global contexts. In this area fire, extensive herding, and anthropogenic activities all modify vegetation (Ringrose 1997). The study of such conversions necessitates the use of remotely sensed data for the production of maps and the development of a system by which classification units are well defined to promote comparison across dates and places (Duadze 2004). This has not yet been quantified at this spatial scale. Hyperion and Data Processing Multispectral instruments integrate large, noncontiguous regio ns of the spectrum into broad bands and a single number represents the radiometric dynamics of a large region of the spectrum. Alternatively, imaging spectrometers (also called hyperspectral imagers) collect reflected solar energy from 350 to 2510 nm with 150 to 500 contiguous bands of 5 to 10 nm bandwidths (Figure 3 2). Imaging spectrometers provide an image data cube with a detailed spectrum of reflected solar energy for each pixel within the
59 image (Im and Jensen 2008; Asner 1998; Ustin et al. 2004). When considering the savanna landscape, with mixed communities and variable plant densities, it is clear why more spectral information is desirable. Studies employing hyperspectral information have been used to great effect for discriminating landscapes despit e environmental complexity (Clark et al. 2005; Nagendra 2001) and to date only a few studies have tested the potential of imaging spectroscopy in estimating vegetation parameters in southern Africa (Mutanga et al. 2009). EO 1 satellite is the first spaceborne hyperspectral instrument to acquire both visible/NIR (400 1000 nm) and shortwave infrared (900 2500 nm) spectral data (Pearlman et al. 2003). Hyperion has an image swath approximately 7.5 km wide and acquires data in 242 bands with spatial resolution of 30 m and spectral resolution of 10 nm. However, only 198 of these bands are radiometrically calibrated (Datt et al. 2003). T his study used a Hyperion image acquired over the Caprivi, Namibia on 25 May 2011 (Figure 3 1). Ground reference data were obtained in the field within one week of image acquisition and in a previous field effort in 2009. Hyperion data were inspected band by band for vertical stripes (streaking), line drop, and spectral smile (Datt and Jupp 2003 ). The image wa s then processed to correct out of range data, bad lines, vertical stripe removal (destreaking), spectral smile, and bad bands in the data (Datt and Jupp 2003). All preprocessing steps were completed in ENVI version 4.7 (ENVI: Research Syste ms, Boulder, CO) using the Hyperion Data Processing instructions and ENVI add on developed by the CSIRO Office of Space Science & Applications Earth Observation Center (Datt and Jupp 2003). Post processing, 150 bands were retained and used in the analyses. The image was
60 geometrically registered using a Landsat ETM+ image from the University of Maryland Global Land Cover Facility (http://glcf.umd.edu/data/landsat/) and ERDAS Imagine AutoSync (Intergraph Corporation 2009). Hyperion data were processed to appa rent reflectance using the MODTRAN based ENVI atmospheric correction module FLAASH v 4.7 (Fast Line of sight Atmospheric Analysis of Spectral Hypercubes) (ITT Visual Information Solutions, Inc.). Field Data Collection Field samples were collected for 60 si tes randomly generated within the footprint of available Hyperion imagery. A GPS point was recorded at the center of each training site and we walked through the surrounding 90 by 90 meter area, visually estimating the following landscape features: land co ver type; human disturbance; percent cover of tree, shrub, herbaceous, and bare ground cover; and percent cover of species composition, noting dominant and secondary species. Canopy cover was estimated using a spherical densiometer along a cross sectional transect 20 m long in a representative part of the training site. Within the plot the dominant tree species were found and height, GPS location, and diameter at breast height were measured. As well, 18 vegetation transects were recorded randomly within 6 o f the training sites. Transects were approximately 30.5 meters in length and employed the line intercept method. Species identification was recorded at 0.30 m intervals for both ground cover and canopy cover crossing the measuring tape at each interval. SM A and MESMA SMA has been used to derive endmember fractions in a wide variety of vegetation analyses, measuring vegetation response and producing superior results when compared to traditional vegetation indices (Dennison and Roberts 2003; Elmore
61 et al. 200 0). A mixed pixel signal (r) can be described as the linear combination of pure spectral signatures of its components, weighted by their subpixel fractional cover (Adams et al. 1986): (3 1) where M is a matrix with columns corresponding to the spectral signals of specific endmembers, f is a column vector [f 1 ,...,f m ] T designating cover fractions occupied by m at cannot be modeled using these endmembers (Somers et al. 2011). The accuracy of SMA is often quantified based on the fit between modeled and observed mixed spectral signals. The error related to the residual values of each pixel can be calculated, and th performance in depicting the different components of each pixel can be evaluated by the root mean squared error (RMSE): RMSE= (3 2) i is the error related to the residual values of each pixel (Roberts et al. 1998). SMA endmembers are selected to minimize RMSE and residual errors in the model, and crea te a meaningful measure of fractional abundance throughout most of an image. The method is straightforward and transferable across sensors and through time (Adams et al. 1995, Roberts et al. 1998). However, the model is most accurate when the exact number of endmembers required to account for the spectral variability in the image is utilized (Sabol 1992). Too few endmembers puts the unmodeled endmember into the fractions, increasing model error, and too many endmembers makes the model sensitive to noise fro m the atmosphere, instrument, and natural variability in spectra (Roberts et al. 1998). This method can
62 provide accuracy in relatively homogenous ecosystems, but can result in significant errors for complex landscapes (Asner and Lobell 2000). As an extensi on of simple SMA, MESMA accounts for endmember variability by allowing them to vary in number and type for each image pixel (Roberts et al. 1998). The general MESMA technique begins with a series of candidate two endmember models and then evaluates each mo del based on selection criteria to construct models that incorporate more endmembers where necessary. The best fit model (lowest RMSE) is then assigned on a pixel by pixel basis (Roberts et al. 1998). MESMA Runs Analyses for this study were performed in EN VI version 4.7 (ENVI: Research Systems, Boulder, CO). Image endmembers were selected by extracting reflectance spectra from pixels in the Hyperion image at training sites where ground reference data were collected. These data were imported into a spectral library and put into one of three broad surface types (GV, NPV, and soil) to run in MESMA models. Although we knew each training site contained a heterogeneous mixture of endmembers, each training site was designated as one endmember (based on its dominant endmember) for unmixing purposes. We assume MESMA chose the training sites with the most dominant signature of each cover type. Models were run using the VIPER Tools software package developed for ENVI (Halligan 2011) in unconstrained, constrained, and pa rtially constrained modes. Constraints in the model can include thresholds for minimum and maximum fractions for each pixel and for RMSE. Fractional constraints are imposed on a model to increase accuracy in the fractions, for example by forcing them to su m to one and be non negative (Heinz and Chang 2001). The RMSE threshold sets the highest allowable error assigned to each pixel during model selection. Pixels
63 that exceed the constraints remain unmodeled in the image, but MESMA run in unconstrained mode wi ll create a model for each pixel in the image regardless of whether it creates extremely positive (>100%) or negative (<0%) fractions or high RMSE values. Default values of 0.05, 1.05, and 0.025 for minimum fraction, maximum fraction, and RMSE threshold r epresent values used in the literature (Halligan, 2002; Roberts et al. 2003) and were used for constrained models. Partially constrained models employed a value of 1.50 for maximum fraction of each EM, but left the other thresholds unconstrained. The resul ts from running the full spectral library were examined to determine which endmembers were modeling the majority of GV, NPV, and soil classes in the image, thus being the most representative and exhibiting the least confusion with other endmembers in that class. MESMA was subsequently rerun using the best subset of endmembers. The best model for the scene was chosen based on endmembers that modeled the lowest RMSE across the image with reasonable constituent cover fractions, i.e. summed to one or near one w ithout excessive negative fractions. These endmembers were convolved to ETM+/Landsat 7 bands and used to unmix the image in order to address the second research question and evaluate improved land cover characterization by hyperspectral data over a multisp ectral sensor. Gauging potential gains from using hyperspectral data is important when considering the previously described practical limitations on vegetation discrimination in this region. Comparing these two datasets allows us to evaluate whether the hi gh spectral resolution of Hyperion is useful and necessary for estimating GV, NPV, and soil.
64 Results A total of 2,870 four endmember models were evaluated using the full spectral library to initially determine the optimal subset of endmembers. The chosen e ndmembers generated 27 four endmember models containing the best fractional cover representation of the landscape and modeled 100% of the pixels for GV, NPV, and soil. These models were used to create fraction images showing the spatial structure of GV, NP V, and soil on the landscape for the Hyperion bands and the convolved Landsat bands (Figure 3 3). The full set of Hyperion bands performed better at estimating NPV and soil than did the Landsat convolved bands, and RMSE error in the Hyperion model was up t o 8 times smaller than that of the Landsat model (Figure 3 4). The RMSE image associated with the Landsat model shows persistent high error throughout the image, with the highest error in areas of NPV and soil dominated land cover (Figure 3 5 ). Error in th e Hyperion model is primarily associated with noise in the data not remo ved by preprocessing (Figure 3 5 ), whereas the endmember fractions were modeled with relatively low error. Unlike the Landsat model, the Hyperion model had relatively high error in ve getation dominated areas, rather than NPV and soil, although the RMSE in the vegetation was still lower in the Hyperion than Landsat model. Endmembe r difference images were computed by subtracting the Landat model from the Hyperion model for each fractiona l component. Figure 3 6 show s the areas of difference between models by fractional component . In wetter areas, there were greater discrepancies in soil fractions where the Hyperion data modeled more NPV and Landsat modeled more soil. There was a higher dif ference in NPV prediction between models for areas of low GV cover, and lower difference in soil prediction between models for areas of lower GV
65 cover. Vegetation endmember differences were nearly random except for both models agreeing on wet GV areas and disagreeing on an area of low GV in the right middle portion of the image, where Hyperion modeled NPV and Landsat modeled a mix of all three fractional covers. In order to more quantitatively analyze the relationships between endmember abundance images and field cover components, model fractions of GV, NPV, and soil were plotted against field fractions of the same constituents as measured in the 60 plots. Hyperion was better able to estimate field measurements of NPV and soil than Landsat (Figures 3 7, 3 8 ) . Neither model performed well at estimating field measurements of vegetation (Figure 3 9 ). The Hyperion bands were far superior to the Landsat bands at teasing out fractional proportions with low associated error and able to resolve the spectral ambiguity between NPV and soil on the landscape. Discussion Our finding that compared to broad band multispectral bands, hyperspectral images were better able to capture the spatial heterogeneity of vegetation properties in different landscapes is consistent with p revious studies (Numata et al. 2009; Thenkabail et al. 2004; Miura et al. 2003). Hyperion data produced higher NPV fractions and lower soil fractions compared to Landsat data in areas of higher GV, and the opposite trend was found in areas with lower GV co ver. Numata et al. (2009) found similar results in pastures in Brazil and noted a high soil fraction in green grass to be a problem with Landsat data. Our results support their finding that hyperspectral data can reduce the incorrect prediction of NPV as s oil. Surprisingly, both models failed to model vegetation well. Green vegetation has a strong spectral signal, dominated by high chlorophyll absorption and high infrared
66 reflectance, providing the basis for indices like NDVI. However, even NDVI calculated from both hyperspectral and Landsat bands had low correlation with percent vegetation estim ated in the field. Okin et al. (2001) found the vegetation signal in semi arid and arid areas too faint amid a dominant, bright soil background to reliably estimate vegetation type and cover. They found intraspecies spectral variation, coupled with nonlinear mixing and a large soil background to be the main reasons behind high error associated with vegetation estimates. In our study, estimates of GV cover were likely overestimated in the field data, which were collected during the dry season when few green leaves were present on trees and shrubs. In this case, observers are likely to overestimate vegetation coverage based on the presence of trees and shrubs, rather tha n on the cover of green foliage itself. Percent cover was predominantly NPV given the time of year field data were collected. This will continue to present a problem in this region for studies seeking to discriminate finer levels of vegetation as cloud fre e imagery and field season efforts tend to coincide with the dry season in southern Africa. Better GV convergence in this ecosystem may be achieved with more quantitative field data that measures green foliage instead of estimates of percent cover. In addi tion, aerial photos or imagery with higher spatial resolution, such as QuickBird or IKONOS, taken during the same period of the year could help map vegetation more accurately in this region. The green vegetation chosen by the most pixels in the MESMA model was from a field plot covered in dense mopane shrubland, but we cannot say that pixels classified as GV contain mopane plants, only that mopane shrubland offered the strongest signal for green vegetation on the landscape during the field season as a dense cover with green leaves.
67 Including a time series of images, particularly an image taken during the wet season, would most likely improve the land surface characterization. The green up of NPV dominated areas from the dry to wet season would aid in separat ing NPV from soil and also give information on peak coverage of green foliage. Additionally, a wet season image could be used to provide complementary ground information to the dry season a spectral time series to define endmembers have produced increased subpixel accuracy when plant phenological cycles were reflected in temporal reflectance signals (Dennison et al. 2007). Nonetheless, sources of error in this study may be strongly related to the scale of Hyperion data as compared to the scale of variability in vegetation properties on the ground, which our field data collection confirmed is smaller than the 30 m pixel resolution of Hyperion. This study points to the need for more hyperspec tral data from high resolution sources such as airborne hyperspectral, because of its superior ability to distinguish NPV and soil. Despite some limitations, our classification provides a better characterization of the fractional cover of GV, NPV, and soil than has been previously achieved in this region that can be used to look at functional attributes of the landscape like patch structure, estimates of carbon storage, and fire potential. Distinguishing NPV from soil is valuable in fire danger assessment a s NPV is an important component of fuel loads (Roberts et al. 2003). As well, Hyperion data clearly outperformed Landsat data better to create better models. These results suggest land cover maps in this region will be improved by utilizing narrow band hyp erspectral datasets. Broad band remote sensing is limited in its ability
68 to accurately estimate vegetation because its coarse spectral resolution leads to ambiguity between NPV and soil backgrounds (Roberts et al. 1993; van Leeuwen and Huete 1996). The hig h structural variability in savannas necessitates very detailed spectral measurements in order to quantitatively characterize the landscape as vegetation and soil components (Asner et al. 2011), suggesting pixels were misclassified because training site da ta did not capture all of the variability within the image. Indeed land cover in this area is very complex and stratified and it is possible that more field measurements incorporating quantitative foliage estimates, species diversity, plant phenology, and temporal variability would improve subpixel classifications in this region. Nonetheless, for complex land surfaces such as savannas, better estimates of chemical and biophysical properties are expected to be achieved using hyperspectral data. As such, stud ies like this would certainly benefit from a more stable global imaging spectrometer with more complete coverage and less problems associated with noise in the data. This would allow for the addition of temporal signatures and time series data to be used i n routine mapping of GV, NPV, and soil, resulting in improved measurements for semi arid regions globally Conclusions This study investigated the ability of moderate resolution spaceborne hyperspectral data to characterize the land surface in a semi arid e cosystem using MESMA models and endmembers derived from field data collected at the time of satellite overpass. These data were able to resolve GV, NPV, and soil, across the landscape reasonably well with low error. C onvolved Landsat data were unable to re solve NPV and soil differences on the landscape, and the Landsat model exhibited much higher error. The spectral ambiguity between NPV and soil was better resolved
69 using hyperspectral data; and the strongest relationships between modeled and field data wer e found in these fractions. None of the models captured all of the endmembers across the scene. Pixels were misclassified because training site data did not capture all of the variability within the image .
70 Figure 3 1. Study area map locating the train ing site data and Hyperion image subset in the Caprivi Strip, Namibia. The false color Hyperion composite image is shown for R, G, B equal to Band 42=NIR (773 nm); Band 32=Red (671 nm); and Band 18=Green (529 nm), respectively.
71 Figure 3 2. Hyperspectral versus multispectral data for Colophospermum mopane vegetation in the study area derived by Hyperion data (hyperspectral) and Hyperion data convolved to Landsat bands (multispectral). 0 10 20 30 40 50 60 70 356 500 1,000 1,500 2,000 2,500 % Reflectance Wavelength (nm) Colophospermum mopane Vegetation Hyperion hyperspectral
72 Figure 3 3. False color image of the study area showing the f raction proportions of NPV, GV, and soil as derived by the (a) Hyperion MESMA model, and (b) Landsat MESMA model. Wetland Areas Low GV cover
73 Figure 3 4 . Scatterplot of root mean square error associated with the Hyperion MESMA model versus the Landsat MESMA model.
7 4 Figure 3 5 . Roo t Mean Squared Error images comparing the error associated with MESMA models for (a) Hyperion model and (b) Landsat model.
75 Figure 3 6 . Difference images showing the spatial distribution of fractional cover differences calculated by subtracting the Lands at convolved model from the Hyperion model for (a) NPV fraction, (b) GV fraction, and (c) soil fraction.
76 (a) (b) Figure 3 7 . Relationships between NPV cover in the field and NPV endmember images for (a) Hyperion bands and (b) Landsat convolved bands. y = 1.0624x 0.0463 RÂ² = 0.2206 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Hyperion % NPV Field % NPV Hyperion Model vs. Field NPV% y = 0.2191x + 0.2969 RÂ² = 0.0044 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 0 0.2 0.4 0.6 0.8 1 1.2 Landsat % NPV Field % NPV Landsat Model vs. Field NPV%
77 (a ) (b) Figure 3 8 . Relationships between soil cover in the field and soil endmember images for (a) Hyperion bands and (b) Landsat convolved bands. y = 0.6129x + 0.1105 RÂ² = 0.1874 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Hyperion % soil Field % soil Hyperion Model vs. Field Soil% y = 0.4516x + 0.1025 RÂ² = 0.033 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 0.0 0.2 0.4 0.6 0.8 1.0 1.2 Landsat % soil Field % soil Landsat Model vs. Field Soil%
78 (a) (b) Figure 3 9 . Relationships between GV cover in the field and GV endmember images for (a) Hyperion ba nds and (b) Landsat convolved bands. y = 0.4548x + 0.2766 RÂ² = 0.0742 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 0.0 0.2 0.4 0.6 0.8 1.0 Hyperion % Veg Field % Veg Hyperion Model vs. Field GV% y = 0.3257x + 0.4472 RÂ² = 0.0201 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 0.0 0.2 0.4 0.6 0.8 1.0 Landsat % veg Field % Veg Landsat Model vs. Field GV%
79 CHAPTER 4 CHANGE DETECTION USING VEGETATION INDICES IN THE ADDO ELEPHANT NATIONAL PARK: 2009 2011 Background Conventional field based sampling methods are becoming prohibitively expensive and time consuming at large spatial scales, and such methods are often unsuitable for long term monitoring over large areas. Researchers are becoming increasingly aware of the linkages between land use and land cover change, and the associated impacts on human environment processes. As such, remote sensing datasets with larger temporal and spatial extents have become a useful tool for ecological and conservation related applications (Kerr and Ostrovsky 2003; Ustin et al. 2004; Gillanders et al. 2008). Remote sensing holds a central role in relating information obtained at one scale to patterns and processes that manifest at another scale, providing the only means of viewing large areas of Earth at regular intervals and recording optical data which can be interpreted through the selec tive absorption and reflectance of light by plants (Ollinger 2011). Given the limitations of local field studies, remote sensing with satellites is an obvious tool for broader scale sampling. The increasing availability of historical data in concert with d ata covering large spatial extents makes remote sensing a very useful data source (Asner et al. 1998; Mutanga et al. 2009; Nagendra and Gadgil 1999). And for some applications, remote sensing may be the only data source available for measuring and monitor ing environmental change (Kerr and Ostrovsky 2003; Turner et al. 2003). The high spatial complexity exhibited in the arid and semi arid ecosystems of southern Africa poses many challenges to remote sensing techniques that have been applied to characterize vegetation variables such as species classification or
80 identification, biochemical concentration, and biomass in this region (Mutanga et al. 2009). Consequently, studies of land use impacts on vegetation in arid and semi arid ecosystems have frequently bee n restricted to local scales with limited spatial and tempora l context (Asner et al. 2000). These studies often grapple with achieving a high enough spatial resolution to resolve multiple spatial patterns observed at scales relevant to vegetation dynamics (Asner et al. 2000; Meyer et al. 2010 ). In addition, spectral response patterns alone can be inadequate in heterogeneous environments (Tanser and Palmer 1999). As such, there is a necessity to test alternative data analyses and data sources when studyin g l and cover change in arid and semi arid regions such that subtle changes in surface attributes affecting the character of the land cover are not lost in discrete classification techniques. Other efforts toward studying vegetation in arid ecosystems have foc used on assessing vegetation condition, typically employing vegetation indices that represent spatial and temporal variability in canopy greenness or leaf area index (Asner et al. 20 00). There can be significant problems interpreting vegetation indices in arid environments given the highly variable reflectance effects of standing litter and bare soil, which are major componen ts of the aboveground biomass. Quantifying and removing these effects has not been fully accomplished, and both of these factors are g enuine obstacles in using vegetation indices in arid ecosystems to infer vegetation condition from remote sensing measurements (Asner et al. 2000). The African elephant ( Loxodonta africana ) is a major driver of vegetation change and ecological processes i n sub Saharan Africa. The influence of elephant on
81 ecological systems has been studied for nearly 60 years (Kerley et al. 2008) , and they influence a range of ecological patterns and processes at various spatial and temporal scales (Kerley and Landman 2006 ). As numbers increase in many southern African protected areas, there are concerns about the negative impacts that high densities of ele phant may have on vegetation. The consequences of high elephant abundance on vegetation structure and composition, and therefore biodiversity, are of serious concern in conservation (Cumming et al. 1997; Kerley and Landman 2006; Owen Smith et al . 2006). Improved understanding of habitat use by elephants enables better management decisions through adaptive management as con tinued monitoring efforts flag unacceptable vegetation impacts that could lead to loss of biodiversity or critical ecosystem functions (Grant et al. 2011). Vegetation degradation in the Eastern Cape region of South Africa has been observed for decades and attributed to a number of factors including overgrazing, anthropogenic activity, and large herbivores (Kerley et al. 1995; Kakembo 2001). Particular attention has been paid to the impacts of increasing numbers of elephant ( Loxodonta Africana ) on vegetation in southern Africa, as they are recognized agents of change across a wide range of ecosystems. Studies have shown elephant decrease biodiversity, damage vegetation, and decrease canopy cover (Cumming et al. 1997; Conybeare 2004). Their influence on the la ndscape has resulted in severe impacts to thicket ecosystem endemic to the Eastern Cape of South Africa (Johnson et al. 1999; Lombard et al. 2001 ; Landman and Kerley 2014 ) . Elephant impacts affect vegetation dynamics at a range of spatial and temporal scales, and elephant distribution has been
82 linked to proximity of available surface water (ChamailleÂ´ Jammes et al. 2007; Smit et al. 2007), in some cases allowing elephant to overutilize vegetation near artificial waterholes (Loarie et al. 2009; Landman et al. 2012) and affect habitat available to and utilized by other large herbivores (Landman and Kerley 2014). Elephant access in the park has recently changed. Elephant have b een utilizing Main Camp in AENP since its inception, but recently an internal fence was removed making new resources in Colchester accessible to elephant as well. From an ecological standpoint, biodiversity loss in the newly available area is a concern and comparing these two areas in the park allows for immediate feedback to managers. This work explores vegetation dynamics at a more regional, park wide scale to consider the use of satellite derived vegetation indices (VIs) in estimating vegetation paramete rs and monitoring change over time. Scientists have been extracting and modeling vegetation biophysical variables using remotely sensed data since the 1970s. Most of these have been vegetation indices , which are dimensionless measures that indicate the rel ative abundance and activity of green vegetation. The objectives of this study were to estimate vegetation parameters using remotely sensed imagery and vegetation indices to assess how vegetation has changed in the Addo Elephant National Park ( AENP ) . Remot e sensing images covering the Main Camp and Colchester areas in AENP from 2009 and 2011 were processed and analyzed to answer the following research questions: 1.) Are there vegetation changes in the Main Cam p/Colchester areas between 2009 and 2011 as ind icated by vegetation indices , with these years selected based on image quality and fence removal allowing elephant new access to Colchester, and
83 2.) Is vegetation change in this area associated only with parts of th e park that elephant can access? Data and Methods Study Area Changes in vegetation structure and composition due to high elephant abundance are a concern in AENP, located in the Eastern Cape province (Figure 4 1), and have been monitored at the f ield scale since the early 1970 s (Kerley and Landma n 2006). The AENP was established in 1931 to preserve the diminished Eastern Cape elephant population, one of four in South Africa that survived into the 20th century (Kerley and Boshoff, 1997). AENP is the only national park in South Africa containing suc culent thicket vegetation , as well as a number of rare and endemic plant species. In addition to conserving its elephant population, the park maintains biodiversity objectives that include preserving an intact population of succulent thicket (Johnson et al . 1999; both incidentally and intentionally, and these sites used as a benchmark to compare with vegetation from elephant use areas. Field researchers documented veget ation degradation in the elephant use areas as early as the 1970 s, including decreases in plant biomass and changes in species biomass in the elephant use areas was about half that of adjacent areas (Kerley et al. 1995). To date, vegetation change is visible on the ground and has been monitored at the field scale over a temporal period of nearly 40 years (reviewed in Kerley and Landman 2006). Field efforts began in 1971 w ith Penzhorn et al. (1974) finding significant decreases in plant biomas s (55%) in elephant use areas.
84 Elephant use areas have grown in step with park expansion over the years to reduce elephant density and associated influences on vegetation (Kerley and L andman 2006). Despite this, research efforts in the park continued to show decreases in species richness, vegetation density, and cover when compared to areas in the park that remained inaccessible to elephant (Stuart Hill 1992; Moolman and Cowling 1994; L ombard et al. 2001). Most recently, Landman et al. (2012) found intensive utilization by elephant caused a replacement of the shrub thicket community with grasses, noting that specific species responded differently to elephant effects, but that grass cover decreased exponentially with distance from artificial waterholes. Remote Sensing Analyses Two images were processed for this analysis, a SPOT5 image from March 5, 2009 and a SPOT4 image acquired March 22, 2011. These images were chosen as the best quality , least cloud covered images falling as close to anniversary dates as possible, and representing time periods before and after an internal fence in the park was removed (August 2010), opening the Colchester region i n the park to elephant. Figure 4 2 shows the false color composite image (Bands 3, 2, and 1) of the study a rea from March 5, 2009. Figure 4 2 includes artificial water points in the park, as well as the elephant use areas and botanical reserves referenced throughout this paper. The internal fence that was removed in the park ran horizontally between the Main Camp and Colchester areas delineated in figure. The 2009 image was provided geometrically corrected by SANParks, and this was verified using GPS ground control points collected in the field ac ros s the study area in June 2012. This image was then used as a reference to geometrically correct the 2011 image using ERDAS Imagine AutoSync (Intergraph Corporation 2009). Both images were atmospherically corrected to provide
85 at surface apparent reflecta nce, clipped to the study area, and the 2009 SPOT5 image was resampled to 20 meters using nearest neighbor interpolation to match the coarser resol ution of the 2011 SPOT4 image. Clouds were manually digitized and removed . Atmospheric correction was perform ed using the QUAC algorithm in ENVI 4.7 and band by band histogram matching as the images came from different sensors with different levels of correction. The QUAC approach estimates the gain and offset directly from the in scene spectral data and referenc e material spectral libraries in ENVI. The method is described in Bernsetin et al. (2012) and has been found to retrieve accurate reflectance spectra in many conditions as compared to other atmospheric correction methods (Bernstein et al. 2012). After runn ing both images through QUAC, the images were intercalibrated by band by band histogram matching to make sure the changes between the images were due to land surface changes and not changes in the sensor or atmosphere. The histograms for each band from the 2011 image were matched to the 2009 image bands . This was because the 2009 image was radiometrically and geometrically corrected prior to these analyses and the 2011 image was raw. Vegetation Indices The normalized difference vegetation index (NDVI) and t wo indices that utilize the mid infrared band, a corrected normalized difference vegetation index (NDVIc) and a normalized difference water index (NDWI) were ca lculated across the study area for the aforementioned time periods using SPOT data. Although tra ditionally vegetation parameters have been estimated using the visible and near infrared bands, studies have shown that reflectance data from the mid infrared region (1500 1750nm) highlight the sensitivity of vegetation reflecta nce to water content, modi fy background reflectance in open areas, and can be used in indices that employ the
86 greenness/chlorophyll based bands along with the MIR bands (Everitt et al. 1989; Nemani et al. 1993; Vescovo and Gianelle 2008). NDVI is an indication of green lea f biomass and green leaf area. It is an important index because interannual changes in vegetation growth and activity can be monitored (Jensen 2005). This index indicates photosynthetic activity in vegetation and is calculated as the difference between the near inf rared and visible reflectances divided by the sum of the two (Sellers 1985; Jensen 2005). This is expressed as: NDVI = (IR R)/(IR + R), ( 4 1 ) where IR is infrared reflectance and R is red reflectance. NDVIc , or corrected NDVI, was developed to ut ilize the mid infrared band and linearize relationships among parameters using the MIR band . It has been found to modify NDVI background reflectance in moderately open to sparse canopies (Nemani et al. 1993). This is expressed as: NDVIc = NDVI (1 mir mir min mir max mir min )), ( 4 2 ) where NDVI is calculated based on Eq. 4 1 and MIR is mid infrared reflectance. NDWI, the normalized difference water index, uses the normalized difference between the NIR and MIR bands t o determine water content in vegetation (Hardisky et al. 1983; G ao 1996). This is expressed as: nir mir nir mir ). ( 4 3 ) Climatological Data Climatological data were obtained from South African National Parks for station [0055447A7] Addo Elephant Park, measured daily at 08:00. These data were used to calculate average monthly totals for the month of image acquisition plus preceding twelve months as shown in Figure 4 3 . In general, temperature trends did not vary
87 greatly in the study area for the time periods preceding image acquisition but precipitation was greater for nearly every month preceding the 2011 image capture. Elephant Movement Data The elephant population in AENP has increased from 22 individuals in 1954 to more than 400 a t present (Kerley and Landman 2006). Elephant cows representing the main family groups were collared in 2010 by SANParks Scientific Staff in cooperation with Nelson Mandela Metropolitan University. Elephant location was recorded 3 times per day until late 2011 when it was increased to hourly. Data received from SANParks for six collared elephants from January March 2011 (n=1,563) were used in this study to determine elephant u se areas . These elephant use areas were created using the minimum bounding geome try function in ArcMap which forms a convex polygon enclosing an input feature, in this case all location points of individual elephants recorded by their collars ( Figure 4 2 ). Botanical reserves, other fenced off areas in the park, and pixels with clouds were excluded from elephant polygons. Each collared elephant was considered individually to account for potential variation in VI values , given that different elephants utilized differ ent portions of the park (Fig. 4 2 ) . Image statistics were calculated fo r each VI for the botanical reserves and elephant use areas in both images , as well as the whole park . Image differencing was used to determine change over time for each VI. Results Vegetation Index Spatial Patterns and Change NDVI Figure 4 4a highlights N DVI change based on +/ 2 standard deviations from the mean between 2009 and 2011. Areas of NDVI increase over this time period occur
88 predominantly in Main Camp nearest artificial water points. Decreases in NDVI occurred in the Colchester area and to some degree in all of the botanical reserves. Park wide, there was a mean shift in NDVI from 0.55 to 0.51 between dates (Figure 4 5). The significance of this difference was tested using a two sample Kolmogorov Smirnov test (KS test). The KS test determines whe ther two datasets differ significantly, making no assumptions about the distribution of the data (Kirkman 1996). In this case, it tests whether the two years of NDVI values have a statistically significant difference in distribution and therefore meaningfu l change in NDVI. The KS test was done using individual pixels in the images for each year and showed a statistically significant difference in NDVI values for 2009 and 2011 (P<0.05). NDVIc NDVIc values were lower than NDVI values across the study area for both images, but similar spatial patterns were seen for NDVIc (Figure 4 4 b). Areas of greatest NDVIc increase over this time period occurred predominantly in Main Camp near water points and NDVIc showed greater increases in vegetation productivity in this area as compared to NDVI. Decreases in NDVIc occurred in the Colchester area and all of the botanical reserves. However there was greater variability in NDVIc change among botanical reserves than for NDVI, i.e. more pronounced increases or decreases. Par k wide, image statistics show a decrease in mean NDVIc from 0.338 to 0.290 between dates. This shift in distribution was tested using the two sample KS test on individual pixels for each year and found to be statistically significant (P<0.05). NDWI Figure 4 4 c shows the patterns of change in NDWI across the study region. Between 2009 and 2011, vegetation water content increased across the study area.
89 Mean NDWI values increased from 0.042 to 0.121. This shift in distribution was tested using the two sample K S test on individual pixels for each year and found to be statistically significant (P<0.05). NDWI changes nearly match NDVIc changes, with increases in water vegetation lining up with increases in photosynthetic activity and decreases in water vegetation matching decreases in photosynthetic activity. Elephant Movement and Vegetation Indices NDVI Summary statistics for NDVI are presented in Table 4 1 . Within the elephant use areas, mean NDVI change varied, with the highest increase s occurring in polygons as sociated with elephants centered on Main Camp (noted in green). However, NDVI values were higher in individual years in elephant polygons extending into the Colchester region of the park. Thus, although an increase in NDVI occurred predominantly in Main Ca mp where elephants have had the longest influence historically, the highest NDVI values in each image occur in the Colchester area of AENP which elephants have only recently accessed. In the botanical reserve areas, individual pixels increased in NDVI valu e between dates but a ll botanical reserves decreased in mean NDVI over time. NDVIc Summary statistics for NDVIc are presented in Table 4 2. As seen with NDVI, elephant use polygons that include Colchester decreased in NDVIc over time but also had the highe st NDVIc values within individual years. Elephant use polygons centered on Main Camp increased in mean NDVIc between dates. Mean NDVIc values decreased over time for all botanical reserves.
90 NDWI NDWI summary statistics are reported in Table 4 3. With the e xception of the Southern Botanical Reserve, mean NDWI values increased for all botanical reserves and elephant use polygons between 2009 and 2011. As well, there was a large increase in vegetation water content park wide from 0.042 in 2009 to 0.121 in 2011 . The greatest values of NDWI occurred in elephant use polygons centered on Main Camp and are associated with the greenest pixels of NDVI and NDVIc. Discussion Interpreting these results is temporally scale dependent. Two vegetation signals emerge dependin g on whether we consider individual years or change over time; these two scales offer several insights into vegetation ecology in AENP. Within individual years, the elephant use polygons centered on Main Camp exhibited lower NDVI and NDVIc than the rest of the park. These polygons line up with areas of piosphere degradation around the water points in Main Camp where succulent thicket vegetation has disappeared. The vegetation productivity in these areas is lower than that of Colchester and all of the botani cal reserves. Colchester has been accessible to elephants for the shortest period of time; a fence prohibiting their movement into this region was recently dropped, allowing them access to previously unavailable resources (August 2010 versus 1955). Only th ree of six collared elephants extend into this region and they spend less time there, creati ng a smaller impact than in Main Camp. The Colchester section has higher values of NDVI and NDVIc in individual years, showing higher vegetation productivity becaus e it has experienced less vegetation loss and less impact by herbivores. Similarly, herbivores have not influenced the botanical reserves in
91 the park at all and their NDVI and NDVIc values are also higher than degraded areas in Main Camp in individual year s. Looking at change over time, flushes of vegetation in degraded areas are coupled with an overall decrease in vegetation productivity park wide and increased vegetation water content from higher precipitation (Figures 4 4, 4 5, and 4 3). The vegetation i ncreases near artificial waterholes, and in much of Main C amp, are likely due to flush es of grass and forb species where thicket has already disappeared. Elephant respond to green vegetation flushes after rainfall events, and they spend time in open patche s when there is adequate green grass and forbs on disturbed areas after rainfall (Landman et al. 2012; personal communication SANParks scientific staff, 2012) . These areas line up with the greatest use by elephants in the park historically and values of in creased vegetation productivity between images. The decreases in vegetation productivity seen in botanical reserves, Colchester, and park wide are more pronounced when considering rainfall preceding the imagery (Figure 4 3). The relationship between rainfa ll variability and NDVI variability has been studied across Africa and at a global scale, leading to the understanding that the two are correlated when precipitation falls within a certain range (summarized in Wang et al. 2003). Nicholson et al. (1990) exp lored this relationship and found in most cases that NDVI is best correlated to the concurrent plus two antecedent months of preci pitation. For this study, 252 mm of rainfall fell during this time period associated with the 2011 image, compared to just 11 1 mm for the associated period in 2009. Because 2011 was a wetter year, higher NDVI and NDVIc would be expected, yet both indices were lower. Higher precipitation values are reflected in higher NDWI values in 2011.
92 Space for time substitution has been wide ly used to infe r future trajectories of change. F rom this perspective we can surmise that as elephants access the southern part of the park more and into the future, intact thicket vegetation will decrease and will be detectable at this scale in decreased NDVI and NDVIc values, both as compared to less impacted parts of the park and park wide. As well, we are likely to see more variable vegetation signals over time on degraded and impacted areas as grass and forb flushes increase and decrease with precipita tion events. The decrease in vegetation signal park wide cannot be explained solely by elephant impacts because it is occurring in the botanical reserves and in Colchester. Conclusions This study employs a multi year analysis to examine change over time an d two individual years to examine variability over space. These years were chosen to match available elephant data , and the patterns emerging at both temporal scales match field studies at the local scale and observations made b y SANParks scientific servic es (Figure 4 6). These findings justify continued monitoring at this scale using matching resolution imagery into the future. Furthermore , analyses examining the relationship between precipitation and vegetation indices over a longer temporal scale may hel p explain the lower vegetation productivity in areas of the park that are less impacted by elephant and completely fenced off from animal impacts. For example, given that NDVI is correlated with precipitation and location dependent, analyses between these two variables may deduce the scale at which temporal variation in precipitation influences NDVI in AENP. A sensor such as MODIS may not provide the spatial scale appropriate for linking elephant to vegetation parameters, but the hypertemporal nature of the se
93 data may allow for testing relationships between precipitation and NDVI at different temporal scales. conflict between the need to conserve a viable population of elephants and the need to maintain biodiversity and ecosystem structure (Johnson et al. 1999). AENP is the only national park in South Africa containing succulent thicket, as well as a number of rare and endemic plant species (Johnson et al. 1999; Lombard et al. 200 1 ). Patterns of elephant on the landscape have been linked to vegetation productivity, surface water, and season ( Loarie et al . 2009a,b ; Marshal et al. 2011; Smit et al. 2007 ). Many researchers have studied the relationship between elephant and surface wat er distribution, and the potential in manipulating surface water to induce spatial change in elephant impacts (Smit 2013, and others) . This is a feasible management option in AENP given all water points in the park are artificial. However, it is important to consider the likelihood that vegetation will decrease around water points in the southern part of the park as they have in Main Camp, and that the degraded areas will not recover succulent thicket or other endemics without sustained replanting efforts. Continued work in AENP includes using these results in mixed effects resource selection models to link GPS collar data to remote sensing i ndices and ancillary environmental variabl es (after Gillies et al. 2008). This will allow for more explicit relationsh ips between elephants and the environment and enable predictive modeling. P relimi nary models show NDVIc and water points to be significantly correlated with elephant location (unpublished data). Previous research has found elephant response to environmenta l factors to be scale dependent, with vegetation characteristics driving
94 habitat selection at coarse spatial scales and surface water being dominant at fine spatial scales (de Knegt et al. 2010). The reserve managers are faced with numerous challenges in their pursuit of economic development and ecological sustainability, not least of which is the need for information on which to make ecological decisions (Langholz and Kerley 2006). It is critical to better understand whether surface water manipulation wou ld be an effective strategy in shifting the elephant distribution in AENP to create respite for rare and endemic plant species, and degraded areas in general.
95 Figure 4 1. The Addo Elephant National Park in the Eastern Cape, South Africa
96 (a) (b) Figure 4 2 . SPOT5 image of the study area. RGB = 3,2,1 (NIR, R, G). Polygons indicating elephant use areas (a), botanical reserves (b), and Park Sections (b) are highlighted.
97 (a) (b) Figure 4 3 . Monthly climate data across the selected study periods in 2009 and 2011 and their antecedent conditions for (a) Temperature and (b) Precipitation .
98 (a) (b) (c) Figure 4 4 . Changes in Vegetation Indices across the park from 2009 to 2011, for (a) NDVI, b) NDVI c and (c) NDWI .
99 Figure 4 5. NDVI distribution for 2009 and 2011 for the study area .
100 (a) (b) Figure 4 6 . Vegetation Degradation in Main Camp, Addo Elephant Natio nal Park (a) Fenced in Botanical Reserve a rea next to elephant use area , (b) Vegetation degradation near an artificial waterhole in Main Camp .
101 Table 4 1. NDVI statistical values for different elephant use areas, botanical reserves and the entire pa rk, f or 2009 and 2011 . NDVI 2009 MIN MAX MEAN STD RANGE Elephant AG408 0.685 0.838 0.479 0.109 1.522 AG409 0.142 0.929 0.600 0.130 1.071 AG410 0.685 0.882 0.499 0.127 1.567 AG411 0.685 0.929 0.514 0.130 1.613 AG412 0.685 0.849 0.477 0.111 1.534 AG414 0.685 0.838 0.455 0.103 1.522 Botanical Reserve Western Botanical Reserve 0.017 0.821 0.480 0.124 0.805 Southern Botanical Reserve 0.072 0.780 0.594 0.094 0.708 Southeast Matyholweni 0.060 0.874 0.618 0.098 0.814 Northwest Botanical Reserv e 0.118 0.895 0.491 0.147 1.013 Kleinvlakte 0.009 0.870 0.636 0.142 0.861 Whole Park 0.685 0.929 0.549 0.139 1.613 NDVI 2011 MIN MAX MEAN STD RANGE Elephant AG408 0.486 0.775 0.506 0.099 1.261 AG409 0.291 0.825 0.544 0.124 1.116 AG 410 0.486 0.825 0.510 0.103 1.310 AG411 0.486 0.825 0.514 0.104 1.310 AG412 0.486 0.775 0.494 0.103 1.261 AG414 0.486 0.741 0.487 0.099 1.227 Botanical Reserve Western Botanical Reserve 0.055 0.730 0.471 0.101 0.674 Southern Botanical Reserve 0.076 0.666 0.513 0.064 0.591 Southeast Matyholweni 0.203 0.660 0.532 0.098 0.457 Northwest Botanical Reserve 0.183 0.760 0.445 0.118 0.943 Kleinvlakte 0.392 0.775 0.549 0.127 1.167 Whole Park 0.636 0.825 0.510 0.128 1.460
102 Table 4 2. NDVIc s tatistical values for different elephant use areas, botanical reserves and the en tire park, for 2009 and 2011 . NDVIc 2009 MIN MAX MEAN STD RANGE Elephant AG408 0.635 0.596 0.257 0.099 1.231 AG409 0.137 0.825 0.411 0.139 0.963 AG410 0.635 0.723 0 .287 0.134 1.358 AG411 0.635 0.723 0.295 0.138 1.358 AG412 0.635 0.596 0.254 0.104 1.231 AG414 0.635 0.584 0.248 0.102 1.219 Botanical Reserve Western Botanical Reserve 0.137 0.675 0.350 0.097 0.813 Southern Botanical Reserve 0.000 0.630 0.465 0.107 0.630 Southeast Matyholweni 0.090 0.836 0.479 0.089 0.746 Northwest Botanical Reserve 0.137 0.643 0.360 0.126 0.781 Kleinvlakte 0.005 0.592 0.340 0.085 0.587 Whole Park 0.635 0.936 0.338 0.144 1.571 NDVIc 2011 MIN MAX MEAN STD RANGE Elephant AG408 0.458 0.580 0.276 0.088 1.038 AG409 0.206 0.605 0.332 0.120 0.811 AG410 0.458 0.605 0.287 0.097 1.063 AG411 0.458 0.605 0.287 0.099 1.063 AG412 0.458 0.580 0.265 0.092 1.038 AG414 0.458 0.540 0.265 0.090 0.998 Botanic al Reserve Western Botanical Reserve 0.206 0.527 0.313 0.074 0.733 Southern Botanical Reserve 0.002 0.509 0.352 0.064 0.507 Southeast Matyholweni 0.078 0.482 0.388 0.081 0.404 Northwest Botanical Reserve 0.206 0.519 0.297 0.097 0.725 Kleinvlakte 0.385 0.592 0.326 0.100 0.977 Whole Park 0.635 0.605 0.290 0.117 1.240
103 Table 4 3. NDWI statistical values for different elephant use areas, botanical reserves and the entire park, for 2009 and 2011 . NDWI 2009 MIN MAX MEAN STD RANGE Elephant AG4 08 0.711 0.467 0.046 0.084 1.178 AG409 0.497 0.475 0.098 0.148 0.972 AG410 0.711 0.844 0.013 0.124 1.555 AG411 0.711 0.475 0.005 0.127 1.186 AG412 0.711 0.467 0.043 0.085 1.178 AG414 0.711 0.844 0.058 0.082 1.555 Botanical Reserve Wes tern Botanical Reserve 0.245 0.433 0.054 0.085 0.678 Southern Botanical Reserve 0.253 0.389 0.143 0.107 0.642 Southeast Matyholweni 0.253 0.352 0.181 0.093 0.604 Northwest Botanical Reserve 0.383 0.432 0.091 0.089 0.815 Kleinvlakte 0.322 0.472 0.120 0.119 0.794 Whole Park 0.711 0.844 0.042 0.146 1.555 NDWI 2011 MIN MAX MEAN STD RANGE Elephant AG408 0.305 0.562 0.094 0.086 0.867 AG409 0.335 0.479 0.139 0.089 0.814 AG410 0.305 0.604 0.103 0.091 0.909 AG411 0.335 0.562 0. 103 0.092 0.897 AG412 0.305 0.562 0.089 0.090 0.867 AG414 0.305 0.604 0.081 0.082 0.909 Botanical Reserve Western Botanical Reserve 0.169 0.380 0.120 0.060 0.549 Southern Botanical Reserve 0.185 0.317 0.101 0.062 0.501 Southeast Matyholweni 0 .104 0.356 0.226 0.093 0.460 Northwest Botanical Reserve 0.182 0.428 0.135 0.075 0.610 Kleinvlakte 0.212 0.631 0.230 0.143 0.843 Whole Park 0.432 0.631 0.121 0.096 1.063
104 CHAPTER 5 CONCLU DING REMARKS Research Overview The aim of this dissertation was to test methods of quantifying vegetation parameters in semi arid ecosystems in southern Africa, with particular focus on areas of national and transnational parks. These areas are created as a type of land use for which the state provides land in respons e to the increase in human and livestock populations across southern Africa (Child 2004). Ecological monitoring and mapping in these regions is critical from both social and economic standpoints, as southern African landscapes are experiencing management i ntensification, both in protected and unprotected areas. Furthermore, the timely monitoring of changes in the distribution of land cover is crucial for effective management efforts, regional sustainable development practices, and global change studies (Fer reira et al. 2007; Child 2004). In southern Africa, the availability and distribution of landscape resources like nutrients and water are directly linked to economic returns and local livelihoods through grazing animals, exploiting wildlife, and subsistenc e agriculture. Resource selection, distribution, and population dynamics of wildlife and domesticated species are linked to vegetation productivity at the landscape level, and detailed information on vegetation at this scale is crucial for managing herbivo res. Therefore a better understanding of the spatial complexity of the landscape is important to meet multiple goals, and becomes especially relevant in light of future climate projections for this region (Ramoelo et al. 2013; Smit 2013; Langholz and Kerle y 2006 ; Boko et al. 2007).
105 Remote sensing methods for ecological monitoring and mapping have myriad advantages, most notably their synoptic view and temporal repetition. Toward that end, a great deal of work has been accomplished to understand land surface dynamics in semi arid ecosystems using global scale, coarse resolution time series data from the AVHRR and MODIS sensors (Hill and Hanan 201 0 ). In recent years, hyperspectral remote sensing has furthered our understanding of ecosystems, including savannas , by providing estimates of canopy chemistry; better fractional covers of NPV, soil, and green vegetation; and linking to multispectral sensors for broader temporal and spatial coverage. Imaging spectroscopy has been used at multiple scales to better under stand the links between plant properties and reflectance. The results from Chapter 2 show a need for more trait data to be collected for key species in southern Africa, specifically structural and biophysical traits known to be related to distinct portion s of the electromagnetic spectrum. Linking traits to reflectance spectra would vastly improve our ability to study communities of plants, make comparisons across ecosystems, and parameterize broader scale systems models. Biochemical estimation in southern Africa will be extremely useful for informing vegetation productivity estimates and forage quality, carbon cycle inputs, and potentially species mapping. As well, successfully linking vegetation to reflectance bands through biochemistry opens the door towa rd generalizing relationships to groups of plants to estimate productivity and carbon cycling at new sites and sites for which field collections are not possible. Looking at vegetation composition through the lens of functional types is necessary for regio nal and global modeling.
106 Chapter 3 highlights the necessity of hyperspectral imagery, specifically bands found in the shortwave infrared spectrum, to estimate retrievals of plant litter. These measurements improve predictions of ecosystem processes includ ing disturbance, plant physiology, and biogeochemistry (Treuhaft et al. 2004; Asner 1999; Asner 2003). Furthermore, quantifying dry plant residues in semi arid ecosystems informs research on wildfires, climate change, and desertification (Roberts et al. 19 98; Ustin and Costick 1999; Asner and Green 2001). At their current scale, hyperspectral images offer the greatest benefit when linked to information from multispectral sensors and products to improve upon coarser resolution time series global analyses. In this way, assimilated datasets can offer improved land surface characterization. Toward that end, more collaboration is needed to link remote sensing studies in semi arid landscapes across southern Africa. Researchers work at disparate scales, deriving i mportant information from successes and limitations in paramount for the advancement of remote sensing to take successes in modeling quantitative retrievals of the land sur face from one locality to test their generalizability in areas similar in species, soil, climate, and animals. For example, data collection and analysis has been fruitful in the Kruger National Park, offering many insights into the workings of savannas. Ma ny of the data and relationships found between field and satellite scales in KNP should be tested in similar regions that are larger and less data rich like Namibia, Botswana, Angola, and Zambia. In fact, parks have an advantage of providing relatively und isturbed areas in southern Africa to research and tease out ecological relationships in the
107 absence of human disturbance and development. Toward that end, Chapter 4 demonstrates the need to view the results from such data analyses temporally and spatially within the context of field scale monitoring and climate variables. As vegetation decreases in places like the Addo Elephant National Park, reserve managers will need to make difficult decisions affecting plant biodiversity, herbivore densities, and econom ic returns. Continued monitoring at a park wide scale in the case of AENP will offer insight into how green vegetation densities are changing over time both in the park and in surrounding areas with similar landscapes but different tenure. Collectively, th is study highlights the limitations of working in data scarce regions and the need for reliable synoptic global imaging. A more complete set of plant traits for linking important biochemicals to specific wavebands would be crucial for testing in hyperspect ral image models. The spatial resolution of Hyperion is limiting in many ecosystems, including the semi arid savannas of Namibia, where pixels are far too mixed to link in situ field reflectance data to 013) where they collected imaging spectroscopy data on the same plants for which trait data were measured, scaling up the canopy level was not possible. Scaling techniques will continue to be pursued until the limitations of linking field and satellite sca le data are overcome. Looking at vegetation composition through the lens of plant functional types is necessary to inform regional and global modeling efforts by defining how and where changes are occurring. This is particularly challenging in environment s where plant species exist in multiple forms. Biochemical estimation using hyperspectral data offers promise in this area but more
108 comprehensive datasets and better data area. It is not uncommon for different vegetation type s on the ground to have simila r spectra in remotely sensed imagery , and this makes it difficult to get accurate classification s. This results in researchers testing a suite of methodologies and pursuing improved classification methods is always in vogue (Xie et al. 2008 ). The future of remote sensing will undoubtedly be ever more accurate measurements of the environment and its processes over space and time. Data fusion across scales using multispectral, hyperspectral, and structural (LiDAR) individual and assimilated datasets will beco me frequent, yielding results that would be impossible using standard techniques (Goetz 2009; Ustin et al. 2004; Hill and Hanan 201 0 ). This potential for quantitative estimates of key surface properties relies on planned future satellite sensor combination s to capture 8 day LiDAR and radar (DESDynI) measurements, and return monthly hyperspectral and thermal imagery (HyspIRI) of the globe (Hill and Hanan 201 0 ). We have seen results from the AVIRIS sensor prove hyperspectral data capable of returning informat ion not obtainable by multispectral instruments, and the need for a stable, pointable hyperspectral imager in orbit that can produce AVIRIS exists but a strong advocate is need ed to ensure the necessary resources and information in the future will push the remote sensing platform forward, both scientifically and politically.
109 Contribution of the S tudy This dissertation contributes to our knowledge of land change science, spatial ecology, and ecosystem function in semi arid ecosystems by using remote sensing and GIS techniques to examine vegetation properties and distribution over space and time. Th e results herein help stakeholders understand the landscape dynamics of these ecosystems, as they make clearer key processes to measure and monitor over time, i.e. vegetation productivity, degradation, and recovery (Chapter 4); and fractional components of the landscape (Chapter 3). This is achieved by quantifying land cover via multispectral and hyperspectral analyses using both discrete and continuous methods. This work elucidates the relative achievements afforded by different types of remote sensing dat a and pushes forward the field by using imaging spectroscopy data from handheld and satellite platforms to better understand the strengths and limitations of remote sensing in this type of environment. Field spectroscopy data cannot be readily scaled to th e 30 m resolution of Hyperion pixels, but results using Hyperion data itself, in concert with field training data, would likely benefit any ecosystem with a large soil/NPV background component as we saw in Chapter 3. Chapter 2 represents the first attempt at linking plant trait data to in situ reflectance spectra for species in this region, elucidating the need for more comprehensive trait data for these species and contributing a set of reflectance spectra for a set of key savanna species to the TRY databa se. These analyses can be used in collaboration with and to inform work currently underway in multiple fields including environmental science, landscape ecology, environmentally modeling, and conservation assessment. The
110 applicability of remote sensing and GIS methodologies goes far beyond the scope of this dissertation and, in combination with other ecological datasets, will greatly improve our ability to monitor and manage the environment at multiple scales. Toward that end, imaging spectroscopy is the fu ture of remote sensing both locally as researchers continue to understand relationships between plant properties and reflectance spectra, and more broadly as researchers successfully scale up to canopy properties and dynamics by specifically incorporating structural measures to account for light scattering processes and ultimately achieve more accurate vegetation measurements.
111 LIST OF REFERENCES Adams, J.B., Smith, M.O., Johnson, P.E. (1986) Spectral mixture modeling: A new analysis of rock and soil types at the Viking Lander 1 site. Journal of Geophysical Research 91 , 8090 8112. Adams, J.B., Sabol, D.E., Kapos, V., Filho, R.A., Roberts, D.A., Smith, M.O., Gillespie, A.R. (1995) Classification of Multispectral Images Based on Fractions of Endmembers: Applic ation Land Cover Change in the Brazilian Amazon. Remote Sensing of Environment 52 , 137 154. Archer, S. (2010) Rangeland conservation and shrub encroachment: new perspectives on an old problem. In: Toit R, Kock, J., Eds. Rangelands or Wildlands? Livestock a nd wildlife in semi arid ecosystems. Oxford, England: Blackwell Publishing, pp. 53 97. Asner, G.P. (1998) Biophysical and biochemical sources of variability in canopy reflectance. Remote Sensing of Environment 64 , 234 253. Asner, G.P., Townsend, A.R., Bust amante, M.M.C. (1999) Spectrometry of pasture condition and biogeochemistry in the Central Amazon. Geophysical Research Letters 26 , 2769 2772. Asner, G.P., Lobell, D.B. (2000) A biogeophysical approach for automated SWIR unmixing of soils and vegetation. R emote Sensing of Environment 74 , 99 112. Asner, G.P., Wessman, C.A., Bateson, C., Privette, J.L. (2000) Impact of tissue, canopy, and landscape factors on the hyperspectral reflectance variability of arid ecosystems. Remote Sensing of Environment 74 (1) , 69 84. Asner, G.P., Green, R.O. (2001) Imaging spectroscopy measures desertification in the Southwest US and Argentina. Eos, Transactions of the American Geophysical Union 82 , 601 606. Asner, G.P., Borghi, C., Ojeda, R., (2003) Desertification in Central Arg entina: Regional changes in ecosystem carbon nitrogen from imaging spectroscopy. Ecological Application 13 , 629 648. Asner, G.P., Heidebrecht, K.B. (2003) Imaging spectroscopy for desertification studies: comparing AVIRIS and EO 1 Hyperion in Argentina dry lands. Geoscience and Remote Sensing, IEEE Transactions on 41( 6) , 1283 1296. Asner, G.P., D.E. Knapp, A.N. Cooper, M.C.C. Bustamante, and L.O. Olander. (2005) Ecosystem structure throughout the Brazilian Amazon from Landsat Observations and Automated Spect ral Unmixing. Earth Interactions 9, 1 31.
112 Asner GP, Levick SR, Kennedy Bowdoin T, Knapp DE, Emerson R, Jacobson J, Colgan MS, Martin RE. (2009a) Large scale impacts of herbivores on the structural diversity of African savannas. Proceedings of the Nation al Academy of Sciences, USA 106, 4947 4952. Asner GP, Martin RE, Ford AJ, Metcalf DJ, Liddell MJ. (2009b) Leaf chemical and spectral diversity in Australian tropical fore sts. Ecological Applications 19, 236 253. Asner GP, Martin RE. (2009) Airborne spectran omics: mapping canopy chemical and taxonomic diversity in tropical forests. Frontie rs in Ecology and Environment 7, 269 276. Asner, G.P. and D. B. Lobell (2000) A biogeophysical approach for automated SWIR unmixing of soils and vegetation. Remote Sens. En viron. 74 , 99 112. Asner, G.P., S. R. Levick, and I.P.J. Smit. ( 2010 ) Remote sensing of fractional cover and biochemistry in savannas. In Hill, M.J. and N.P. Hanan, eds. Ecosystem function in savannas: measurement and modeling at landscape to global scales . CRC Press . Bateson, C. A., Asner, G. P., & Wessman, C. A. (2000). Endmember bundles: A new approach to incorporating endmember variability into spectral mixture analysis. Boko, M., I. Nia ng, A. Nyong, C. Vogel, A. Githeko, M. Medany, B. Osman Elasha, R. Tabo and P. Yanda. (2007) Africa. Climate Change 2007: Impacts, Adaptation and Vulnerability. Contribution of Working Group II to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change, M.L. Parry, O.F. Canziani, J.P. Palutikof, P.J. van der Linden and C.E. Hanson, Eds., Cambridge University Press, Cambridge UK, 433 467. Card, D.H. D.L. Peterson, P.A. Matson, J.D. Aber (1998) Prediction of leaf chemistry by the use of visible and near infrared reflectance spectroscopy. Remote Sensing of Environment 26, 123 147. ChamailleÂ´ Jammes S, Valeix M, Fritz H. (2007) Managing heterogeneity in elephant distribution: interactions between elephant population density and surface wate r availability. Journal of Applied Ecology 44, 625 633. Child, B., Ed. (2004) Parks in Transition: Biodiversity, Rural Development, and the Bottom Line. Earthscan, 278 pp. Conybeare, A.M. (2004) Elephant impacts on vegetation and other biodiversity in the broadleaved woodlands of S Central Africa. Biodiversity in the four corners area: technical reviews, Timberlake, J.R.&Childes, S.L.(Eds). Harare, Zimbabwe: Biodiversity Foundation for Africa, Bulawayo/Zambezi Society, Vol 2, 477 508.
113 Cumming, D.H.M.,Fenton , M.B.,Rautenbach, I.L.,Taylor, R.D.,Cumming, G.S.,Cumming, M.S.,Dunlop, J.M.,Ford, G.A.,Hovorka, M.D.,Johnston, D.S.,Kalcounis, M.,Mahlangu, Z.&Portfors, C.V.R. (1997) Elephants, woodlands and biodiversity in southern Africa. South African Journal of Scie nce 93 , 231 236. Curran, P.J., Dungan, J.L., Peterson, D.L. (2001) Estimating the foliar biochemical concentration of leaves with reflectance spectrometry testing the Kokaly and Clark methodologies. Remote Sensing of Environment 76, 349 359. Datt, B. McVic ar, T.R., van Niel, T.G., Jupp, D.L.B, Perlman, J.S. (2003) Preprocessing EO 1 Hyperion hyperspectral data to support the application of agricultural indexes. Geoscience and Remote Sensing, IEEE Transactions on 41, 1246 1259. Datt, B., Jupp, D. (2003) Hype rion Data Processing Workshop: Hands on processing instructions. CSIRO Office of Space Science and Applications, Earth Observation Center. de Knegt, H.J., van Langevelde, F., Skidmore, A.K., et al. (2011) The spatial scaling of habitat selection by African Elephants. Journal of Animal Ecology 80, 270 281, 2011. Dennison, P.E., Roberts, D.A. (2003) Endmember selection for Multiple Endmember Spectral Mixture Analysis using Endmember Average RSME. Remote Sensing of Environment 87, 123 135. Dennison, P.E., D.A. Roberts, S.H. Peterson (2007) Spectral shape based temporal compositing algorithms for MODIS surface reflectance data. Remote Sensing of Environment 109, 510 522. Diaz, S., Hodgson, K., Thompson, M., Cabido, J.H.C., Cornelissen, A., et al. (2004) The plan t traits that drive ecosystems: evidence from three continents. Journal of Vegetation Science, 15, 295 304. Duadze S.E.K. 2004. Land use and land cover study of the Savannah ecosystem in the Upper West region (Ghana) using remote sensing. ZEF Bonn, Univers ity of Bonn, Germany. Elmore, A. J., Mustard, J. F., Manning, S. J., & Lobell, D. B. (2000). Quantifying vegetation change in semiarid environments: Precision and accuracy of spectral mixture analysis and the normalized difference vegetation index. Remote Everitt, J.H., Escobar, D.E., Richardson, A.J. (1989) Estimating grassland phytomass production with near infrared and mid infrared spectral variables. Remote Sensing of Environment 30, 257 261.
114 Ferreira, N.C., L.G. Ferre ira, F. Miziara (2007) Deforestation hotspots in the Brazilian Amazon: Evidences and causes as assessed from remote sensing and census data. Earth Interactions 11, 1 16. Franke, J., Roberts, D.A., Halligan, K., Menz, G. (2009) Hierarchical Multiple Endmemb er Spectral Mixture Analysis (MESMA) of hyperspectral imagery for urban environments. Remote Sensing of Environment 113, 1712 1723 Gao, B C. (1996) NDWI a normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sensing of Environment 58, 257 266. Gaughan, A.E., Waylen, P.R. (2012) Spatial and temporal precipitation variability in the Okavango Kwando Zambezi catchment, southern Africa. Journal of Arid Environments 82, 19 30. Gillanders, S.N. Coops, N.C., Wulder, M .A., Gergel, S.E., Nelson, T. (2008) Multitemporal remote sensing of landscape dynamics and pattern change: describing natural and anthropogenic trends. Progress in Physical Geography 32(5) , 503 528. Gillies, C.S., Hebblewhit M., Nielsen S.E., Krawchuk M.A ., Aldridge C.L., Frair J.L., Saher D.J., Stevens C.E., Jerde C.L. (2006) Application of random effects to the study of resource selection by animals. Journal of Animal Ecology 75, 887 898. Goetz, A.F.H. (2009) Three decades of hyperspectral remote sensing of the Earth: A personal view. Remote Sensing of Environment 112, S5 S16. Goodenough, D.G., Dyk, A., Niemann, K.O., Pearlman, J.S., Chen, H., Han, T., Murdock, M., West, C. (2003) Processing Hyperion and ALI for Forest Classification. Geoscience and Remot e Sensing, IEEE Transactions on 41(6), 1321 1331. Grant, R.C.C., Peel, M.J.S., Bezuidenhout, H. (2011) Evaluating herbivore management outcomes and associated vegetation impacts. Koedoe 53(2), 15 pp. Grime, J.P. (1977) Evidence for the existence of three p rimary strategies in plants and its relevance to ecological and evolutionary theory. American Naturalist 111, 1169 1194. Guerschman, J.P., M.J. Hill, L.J. Renzullo, D.J. Barret, A.S. Marks, E.J. Botha (2009) Estimating fractional cover of photosynthetic ve getation, non photosynthetic vegetation and bare soil in the Australian tropical savanna region upscaling the EO 1 Hyperion and MODIS sensors. Remote Sensing of Environment 113(5), 928 945.
115 Gutman, G., Janetos, A., Justice, C., Moran, E. (2004). Land chang e science. surface . New York: Kluwer Academic Publishers, 459 pp. Halligan, K. Accessed December 2012. Halligan, K.Q., 2002, Multiple Endmember Spectral Mixt ure Analysis of Vegetation in the Northeast Corner of Yellowstone National Park, Masters Thesis, University of California, Santa Barbara, 64 pp. Hanan N.P. and C.E.R. Lehmann (2010) Tree Grass Interactions in Savannas: Paradigms, Contradictions, and Concep tual Models . In Hill, M.J. and N.P. Hanan, eds. Ecosystem function in savannas: measurement and modeling at landscape to global scales. CRC Press. Hardinsky, M.A., Klemas V., Smart R.M. (1983) The influence of soil salinity, growth form, and leaf moisture on spectral radiance of Spartina alterniflora canopies. Photogrammetric Engineering and Remote Sensing 49, 77 83. Hartigan and Wong (1979) A K Means Clustering Algorithm. Journal of the Royal Statistical Society. Series C (Applied Stati stics), Vol. 28, No. 1 (1979), pp. 100 108 . Hatchell, D.C., Ed. (1999) Analytical spectral devices. 3 rd ed. Boulder, Analytical Spectral Devices, 140 pp. Heimann, M., M. Reichstein. (2008) Terrestrial ecosystem carbon dynamics and climate feedbacks. Natur e 451 (7176), 289 292. Heute, A.R., Miura, T., Gao, X. (2003) Lands cover conversion and degradation analyses through coupled soil plant biophysical parameters derived from hyperspectral EO 1 Hyperion. Geoscience and Remote Sensing, IEEE Transactions on 41, 1268 1276. Hill, M.J. (2004) Grazing agriculture: Managed pasture, grassland, and rangeland S.L. Ustin (Ed.), Manual of Remote Sensing volume 4 Remote sensing for natural resource management and environmental monitoring, John Wiley & Sons, Hoboken, NJ, 44 9 530 . Hill and Hanan ( 2010 ) Current Approaches to Measurement, Remote Sensing, and Modeling in Savannas: A Synthesis In: Ecosystem Function in Savannas, (M. J. Hill & N. P. Hanan, Eds.), CRC Press, pp. 515 545 . Hill, M. J., RomÃ¡n, M., O. and Schaaf, C., B . (2010 ) Biogeography and dynamics of global tropical and subtropical savannas. In: Ecosystem Function in Savannas, (M. J. Hill & N. P. Hanan, Eds.), CRC Press, pp. 3 37 .
116 Homolova, L., Malenovsky, Z., Clevers, J., Garcia Santos, G., Schaepman, M.E. (2013) Review of optical based remote sensing for plant trait mapping. Ecological Complexity 15, 1 16. IBM Corp. (2013) IBM SPSS Statistics for Windows, Version 22.0. Armonk, NY: IBM Corp. Im, J. Jensen, J.R. (2008) Hyperspectral Remote Sensing of Vegetation. Geo graphy Compass 2, 1943 1961. IPCC(1990) The IPCC Scientific assessment of climate change. In Report of Working Group I, Houghton, J.T., Jenkins, G.J., Epharaums, J.J. Eds. Cambridge University Press, Cambridge UK. Jain, A.K. (2008) Data Clustering: 50 Year s Beyond K Means. King Sun Fu Prize lecture delivered at the 19 th International Conference on Pattern Recognition, Tampa, FL. Jensen, J.R. (2005) Introductory Digital Image Processing: A Remote Sensing Perspective, Third Edition. Prentice Hall, Upper Saddl e River, N.J., 316p. Johnson, C.F., Cowling R.M. and Phillipson P.B. (1999) The flora of the Addo Elephant National Park, South Africa: are threatened species vulnerable to elephant damage? Biodiversity Conservation 8, 1447 1456. Kakembo, V. (2001) Trends in Vegetation Degradation in Relation to Land Tenure, Rainfall, and Population Changes in Peddie District, Eastern Cape, South Africa. Environmental Management 28, 39 46. Kattge, S., Diaz, S., Lavorel, I.C. et al. (2011) TRY a global database of plant tr aits. Global Change Biology 17, 2905 2935. Kerley, G.I.H., Landman, M., Kruger, L., Owen Smith, N. (2008) Effects of Elephants on Ecosystems and Biodiversity. In Elephant Management: A Scientific Assessment for South Africa, R.J. Scholes and K.G. Mennell e ds, Johannesburg: Wits University Press, pp. 146 205. Kerley, G.I.H. and Landman, M. (2006 ) The impacts of elephants on biodiversity in the Eastern Cape Subtropical Thickets. South African Journal of Science 102 , 395 402. Kerley, G.I.H. and Boshoff, A.F. ( 1997) A proposal for a Greater Addo National Park: a regional and national conservation and development opportunity. Terrestrial Ecology Research Unit, University of Port Elizabeth. Kerley, G.I.H., Knight M.H., and DeKock M. (1995) Desertificatoin of Subtr opical Thicket in the Eastern Cape, South Africa: are there alternatives? Environmental Monitoring and Assessment 37, 211 230.
117 Kerr, J.T. and Ostrovsky, M. (2003) From space to species: ecological applications of remote sensing. Trends in Ecology and Evolu tion 18, 299 305. Kirkman, T.W. (1996) Statistics to Use: Kolmogorov Smirnov test. (Accessed 20 Dec 2013) Knapp, A.K., J.M. Briggs, S.L. Collins, S.R. Archer. (2008) Shrub encroachment in North American grassland s: shifts in growth form Kumar, L., and Skidmore, A.K. (1998) Use of Derivative Spectroscopy to Identify Regions of Differences between some Australian Eucalypt Species. Paper to 9 th ARSPC, UNSW, Sydney, Australia. Kumar, L., Schmidt, K., Dury, S.J., Skid more, A. (2001) Imaging spectrometry and vegetation sciences. In Imaging Spectrometry Basic Principles and Prospective Applications. van der Meer and de Jong, Eds., Kluwer Academic Publishers, Dordrecht, The Netherlands, pp. 111 155. Lambin, E.F., H. Geist , and R.R. Rindfuss. (2006) Introduction: local processes with global impacts. Pages 1 8 in E.F. Lambin and H. Geist, eds. Land use and land cover change: local processes and global impacts. Springer, Berlin, Germany. Lambin, E.F., Baulies, X., Bockstael, N. et al. (1999) Land Use and Land Cover Change (LUCC) Implementation Strategy. A Core Project of the International Geosphere Biosphere Programme and the International Human Dimensions Programme on Global Environmental Change (IGBP Report 48/IHDP Report 10). IGBP Secretariat: Stockholm & IHDP Secretariat: Bonn. Landman, M., Schoeman, D.S., Hall Martin, A.J., Kerley, G.I.H. (2012) Understanding Long Term Variations in an Elephant Piosphere Effect to Manage Impacts. PlosOne 7, 1 11. Landman, M., Kerley, G.I .H. (2014) Elephant both increase and decrease availability of browse resources for black rhinoceros. Biotropica 46(1), 42 49. Langholz, J.A. and G.I.H. Kerley. (2006). Combining conservation and development on private lands: an assessment of ecotourism ba sed private game reserves in the Eastern Cape. ACE Report No. 56. Lavorel, S., Grigulis, K., Lamarque, P., Colace, M., et al. (2011) Using plant functional traits to understand the landscape distribution of multiple ecosystem services. Journal of Ecology 9 9, 135 147. Lavorel, S., Garnier, E. (2002) Predicting changes in community composition and ecosystem functioning from plant traits: revisiting the Holy Grail. Functional Ecology, 16, 545 556.
118 LeBauer, D.S., Treseder, K.K. (2008) Nitrogen limitation of net primary productivity in terrestrial ecosystems is globally distributed. Ecology 89 (2), 371 379. Liu X, Swenson NG, Wright SJ, Zhang L, Song K, et al. (2012) Covariation in Plant Functional Traits and Soil Fertility within Two Spec ies Rich Forests. PLoS ON E 7(4), e34767 Loarie, S.R., van Aarde, R.J. and Pimm, S.L. (2009) Fences and artificial water affect African savannah elephant movement patterns. Biological Conservation 142, 3086 3098. Lombard, A.T., Johnson, C.F., Cowling, R.M., Pressey, R.L. (2001) Pro tecting plants from elephants: botanical reserve scenarios within the Addo Elephant National Park, South Africa. Biological Conservation 102, 191 203. Low, A.B., Rebelo, A.G. (1998) Vegetation of South Africa, Lesotho and Swaziland. Department of Environme ntal Affairs and Tourism, Pretoria. 2 nd ed. Lymburner, L., P. Beggs, and C. Jacobson. (2000) Estimation of Canopy Average Surface Specific Leaf Area Using Landsat TM Data. Photogrammetric Engin eering and Remote Sensing 66(2), 183 191. Marshal, J.P., Raj ah, A., Parrini, F., Henley, M., Henley, S.R., and Erasmus, B.F.N. (2011) Scale dependent selection of greenness by African elephants in theh Kruger private reserve transboundary region, South Africa. European Journal of Wildlife Research 57(3), 537 548. M artin, M.E., L.C. Plourde, S.V. Ollinger, M.L. Smith, B.E. McNeil (2008) A generalizable method for remote sensing of canopy nitrogen across a wide range of forest ecosystems. Remot e Sensing of Environment 112(9), 3511 3519. Meyer, K.M., Wiegand, K., Ward, D. (2010) Spatially Explicit Modeling of Savanna Processes. In Hill, M.J. and N.P. Hanan, eds. Ecosystem function in savannas: measurement and modeling at landscape to global scales. CRC Press Moolman, H.J., and Cowling, R.M. (1994) The impact of elephant and goat grazing on the endemic flora of South African Succulent Thicket. Biological Conservation 68, 53 61. Mutanga O., van Aardt J., Kumar, L. (2009) Imaging spectroscopy (hyperspectral remote sensing) in southern Africa: an overview. South African Jour nal of Science 105, 193 198. Mutanga O. and Skidmore A.K. (2004) Integrating imaging spectroscopy and neural networks to map tropical grass quality in the Kruger National Park, South Africa. Remote Sens. Environ. 90(1), 104 115.
119 Mutanga O., Skidmore A.K. and Prins H.H.T. (2004) Predicting in situ pasture quality in the Kruger National Park, South Africa using continuum removed absorption features. Remote Sensing of Environment 89(3), 393 408. Nagendra, H., Rocchini, D. (2008) High resolution satellite ima gery for tropical biodiversity studies: the devil is in the detail. Biodiversity and Conservation 17(14), 3431 3442. Nagendra, H., Gadgil, M. (1999) Biodiversity assessment at multiple scales: Linking remotely sensed data with field information. Proc Natl Acad Sci USA 96 (16), 9154 9158. Nagendra (2001) Using remote sensing to assess biodiversity. International Jour nal of Remote Sensing 22(12), 2377 2400. Nemani, R., Pierce, L., Running, S. (1993) Forest ecosystem process at the watershed scale: sensitivity to remotely sensed leaf area index estimates. International Journal of Remote Sensing 14, 2519 2539. Numata, I., Roberts, D.A., Chadwick, O.A., Schimel, J.P., Galvao, L.S., Soares, J.V. (2008) Evaluation of hyperspectral data for pasture estimate in the Br azilian Amazon using field and imaging spectrometers. Remote Sensing of Environment 112(4),1569 1583 Nicholson, S.E., Davenport, M.L., Malo, A.R. (1990) A Comparison to the Vegetation Response to Rainfall in the Sahel and East Africa Using Normalized Veget ation Index from NOAA AVHRR. Climatic Change 17, 209 241. Okin, G.S., Roberts, D.A., Murray, B., and Okin, W.J., 2000, Practical limits on hyperspectral vegetation discrimination in arid and semiarid environments, Remote Sens. Env. 77 (2), 212 225 Ollinge r, S.V. (2011) Sources of variability in canopy reflectance and the convergent properties of plants. New Phytologist 189(2), 375 394. Ollinger SV, Richardson AD, Martin ME, Hollinger DY, Frolking SE, Reich PB, Plourde LC, Katul GG, Munger JW, Oren R et al. (2008) Canopy nitrogen, carbon assimilation, and albedo in temperate and boreal forests: functional relations and potential climate feedbacks. Proceedings of the National Academy of Sciences, USA 105 , 19335 19340. Owen Smith, N., Kerley, G.I.H., Page, B., Slotow, R., and van Aarde, R.J. (2006) A scientific perspective on the management of elephants in the Kruger National Park and elsewhere. S. Afr. J. Sci. 102, 389 394. Pearlman, J.S., Barry, P.S., Segal, C.C., Shepanski, J., Beiso, D., Carman, S.L. (2003) Hyperion, a space based imaging spectrometer. IEEE Transactions on Geoscience and Remote Sensing 41, 1160 1173
120 Knyazikhin, Y., M.A. Schull, P. Stenberg, M. Mottus et al. (2013) Hyperspectral remote sensing of foliar nitrogen content. Proc Natl Acad of Sci USA 110 (3), E185 E192. Oba, G., Post, E., Syvertsen, P.O., Stenseth, N.C. (2000) Bush cover and range condition assessments in relation to landscape and grazing in southern Ethiopia. Landscape Ecology 15, 535 546. Penzhorn, B.L., Robbertse P.J., and Olivi er M.C. (1974) The influence of the African elephant on the vegetation of the Addo Elephant National Park. Koedoe 17, 137 158. Philpot, W. (1991) The Derivative Ratio Algorithm: Avoiding Atmospheric Effects in Remote Sensing. Geoscience and Remote Sensing 29, 350 357. Ramoelo, A., Skidmore, A., Schlerf, M., Heitgonig, I., Renaud, M., Cho, M. (2013) Savanna grass nitrogen to phosphorous ratio estimation using field spectroscopy and the potential for estimation with imaging spectroscopy. International Journal of Applied Earth Observation and Geoinformation 23, 334 343. Ramoelo, A., A.K. Skidmore, M. Schlerf, R. Mathieu, I.M.A. Heitkonig (2011) Water removed spectra increase the retrieval accuracy when estimatinv savanna grass nitrogen and phosphorus concentrat ions. ISPRS Journal of Photogrammetry and Remote Sensing 66, 408 417. Ringrose, S., Vanderpost, C., Matheson, W. 1997: Use of image processing and GIS techniques to determine the extent and possible causes of land management/fenceline induced degradation p roblems in the Okavango area, northern Botswana. International Journal of Remote Sensing 18(11), 2337 2364 Roberts, D.A., Smith, M.O., Adams, J.B. (1993) Green vegetation, nonphotosynthetic vegetation, and soils in Aviris data. Remote Sensing of Environmen t 44, 255 269. Roberts, D. A., Gardner, M. Church, R., Ustin, S., Scheer, G., Green, R.O. (1998) Mapping chaparral in the Santa Monica Mountains using multiple endmember spectral mixture models. Remote Sensing of Environment 65, 267 279. Roberts, D.A., Den nison, P.E., Gardner, M., Hetzel, Y., Ustin, S.L., and Lee, C. (2003) Evaluation of the potential of Hyperion for fire danger assessment by comparison to the AVIRIS. Geoscience and Remote Sensing, IEEE Transactions on 41(6), 1297 1310. Roelofsen, H.D., van Bodegon, P.M., Kooistra, L. Witte, J P.M. (2013) Trait Estimation in Herbaceous Plant Assemblages from in situ Canopy Spectra. Remote Sensing 5, 6323 6345.
121 Sabol, D. E., Adams, J. B., & Smith, M. O. (1992). Quantitative subpixel spectral detection of targ ets in multispectral images. Journal of Geophysical Research, Sankaran, M., T.M. Anderson (2009) Management and restoration in African savannas: interactions and feedbacks. In R. Hobbs and K. Suding, Eds. New Models for Ecosystem Dun amics, Island Press, Washington, pp. 136 155. Sankaran, M., et al. (2005) Determinants of woody cover in African savannas. Nature 438(7069)846 849. Schmidtlein, S., H. Feilhauer, and H. Bruelheide. (2012) Mapping of plant strategy types using remote sensin g. Journal of Vegetation Science 23 , 395 405. Scholes, R.J., Archer, S.R. (1997) Tree grass interactions in savannas. Annual Review of Ecology and Systematics 28, 517 544. Scholes, R.J. B.H. Walker. (1993) An African Savanna. Synthesis of the Nylsvley St udy. Cambridge University Press, United Kingdom. Sellers, P.J. (1985) Canopy reflectance, photosynthesis and transpiration. International Journal of Remote Sensing 6, 1335 1372. Skarpe, C. (1996) Plant Functional Types and Climate in a Southern African Sav anna. Journal of Vegetation Science 7, 397 404. Skidmore, A.K., Ferwerda, J.G., Mutanga, O., van Wieren, S.E., et al. (2010) Forage quality of savannas: simultaneously mapping foliar protein and polyphenols for trees and grass using hyperspectral imagery . Remote Sensing of Environment 114(1), 64 72. Smit, I.P.J. (2013) Systems approach towards surface water distribution in Kruger National Park, South Africa. Pachyderm 53, 91 98. Smit, I.P.J., Grant C.C., Devereaux B.J. (2007) Do artificial waterholes infl uence the way herbivores use the landscape? Herbivore distribution patterns around rivers and artificial surface water sources in a large African savanna park. Biological Conservation 136, 85 99. Somers, B., Asner, G.P., Tits, L., Copplin, P. Endmember var iability in spectral mixture analysis: A review. Remote Sensing of Environment 115(7), 1603 1616. Stuart Hill, G.C., & Aucamp, A.J. (1993) Carrying capacity of the succulent valley bushveld of the Eastern Cape. African Journal of Range and Forage Science 1 0, 1 10.
122 Thenkabail, P. S., Enclona, E.A., Ashton, M.S., Legg, C., De Dieu, M. J. (2004) Hyperion, IKONOS, ALI, and ETM+ sensors in the study of African rainforests. Remote Sensing of Environment 90, 23 43. Theseira, M. A., Thomas, G., Taylor, J. C., Gemme ll, F., & Varjo, J. (2003). Sensitivity of mixture modeling to endmember selection. International Journal of Remote Theseira, M. A., Thomas, G., & Sannier, C. A. D. (2002). An evaluation of spectral mixture modeling applied to a semi a rid environment. International Journal of Thomas, D.S., D. Sporton, and J. Perkins. 2000. The environmental impact of livestock ranches in the Kalahari, Botswana: natural resource use, ecological change and human response in a d ynamic dryland system. Land Degradation and Development 11 , 327 341. Townsend, P. Foster, J., (2003) Mapping forest composition in the Appalachians using data from EO 1 Hyperion, Landsat, and AVIRIS. Available online at: http://eo1.gsfc.nasa.gov/new/valida tionReport/index.html Treuhaft, R.N., B.E. Law, and G.P. Asner (2004). Forest Attributes from Radar Interferometric Structure and Its Fusion with Optical Re mote Sensing, BioScience, 54(6), 561 571. Turner, B., Lam bin, E. and Reenberg, A. (2007) The emergen ce of land change science for global environmental change and sustainability. Proc Natl Acad Sci USA 104 (52), 20666 20671. Turner, B.L. II. (2006) Land change as a forcing function in global environmental nd: An Encyclopedia of Land Use and Land Cover Change. Vol. 1 (A K). Westport: Greenwood Press, pp. xxv xxxii. Turner DP, Ollinger SV, Kimball JS. (2004) Integrating remote sensing and ecosystem process models for landscape to regional scale analysis of the carbon cycle. BioScience 54, 573 584. Turner, W., Spector, S., Gardiner, N., Fladeland, M. Sterling, E., Steininger, M. (2003) Remote sensing for biodiversity science and conservation. Trends in Ecology and Evolution 18(6), 306 314. Turner, W., E.J. Sterling, and A.C. Janetos (eds.). (2001) Special section on contributions of remote sensing to biodiversity conservation: A NASA approach. Conservation Biology, 15 (4), 832 953. Ustin, S.L. (2013) Remote sensing of canopy chemistry. Proc Natl Acad Sci USA 110 (3), 804 805.
123 Ustin, S.L., L.A. Costick (1999) Multispectral remote sensing over semi arid landscapes for resource management. In Hill, M.J., R.J.Aspinall, Eds. Spatial Information for Land Manage ment. Gordon and Breach, UK, 97 112. Ustin, S.L. & Gamon, J.A. (2010) Remote sensing of plant functional types. New Phytologist 186 , 795 816. Ustin, S.L., Roberts, D.A., Gamon, J.A., Asner, G.P., Green, R.O. (2004) Using imaging spectroscopy to study ecosystem processes and properties Bioscience, 54, 523 534 va n Aarde, R.J., Jackson, T.P. (2007) Megaparks for metapopulations: addressing the causes of locally high elephant numbers in southern Africa. Biological Conservation 134, 289 297. Van Auken, O.W. (2009) Causes and consequences of woody plant encroachment i nto western North American grasslands. Journal of Environmental Management 90(10), 2931 2942. Van Auken, O.W. (2000) Shrub invasions of North American semiarid grasslands. Ann. Rev. Ecol. Syst. 31, 197 215. van Leeuwen, W.J.D., Huete, A.R. (1996) Effects of standing litter on the biophysical interpretation of plant canopies with spectral indices. Remote Sensing of Environment 55, 123 138. Veganzones, M. A., & Grana, M. (2008). Endmember extraction methods: a short review. Knowledge Based Intelligent Information and Engineering Systems, 12 th Vescovo, L., Gianelle, D. (2008) Using the MIR bands in vegetation indexes for the estimation of grasslands biophysical parameters from satellite remote sensing in the Alps region of Trentino (Italy). Advances i n Space Research 41, 1764 1772. Walter, H. (1971) Ecology of Tropical and Subtropical Vegetation. Oliver and Boyd, Edinburgh, UK. Wang, J., Rich P.M., Price K.P. (2003) Temporal responses of NDVI to precipitation and temperature in the central Great Plains , USA. International Journal of Remote Sensing 24, 2345 2364. Wanxiao,S., Liang, S., Gang, X., Hongliang, F., Dickinson, R. (2008) Mapping plant functional types from MODIS data using multisource evidential reasoning. Remote Sensing of Environment 112,1010 1024. White, F. (1983) Vegetation of Africa a descriptive memoir to accompany the UNESCO/AETFAT/UNSO vegetation map of Africa, UNESCO. Paris, France, 356 pp.
124 Williams, C.A., N.P. Hanan, J.C. Neff, R.J. Scholes, J.A. Berry, A.S. Denning, D.F. Baker. (200 7) Africa and the global carbon cycle. Carbon Balance and Management 2:3, 1 13. Xie, Y., Sha, Z., Yu, M. (2008) Remote sensing imagery in vegetation mapping: a review. Journal of Plant Ecology 1(1) , 9 23. Yuan, J., Zheng, N., Limin, L., Shihua, L. (2007). Nitrogen content estimation using Hyperion hyperspectral image based on normalized band depth method. SPIE Proceedings 6787, MIPPR 207.
125 BIOGRAPHICAL SKETCH Jessica Steele started her graduate studies at the University of Florida in 2009, where she worked in collaboration with South Africa National Parks, the UF Quantitative Spatial Ecology, Evolution, and Environment IGERT, NASA Grant N9875664 , Namibia Ministry of Tourism and Environment, and the Botswana Ministry of Tourism and Environment. Her primary a cademic interest lies in combining field studies and satellite remote sensing to estimate land use and land cover change ; statistical modeling ; spatial ecology ; and the use of imaging spectroscopy to link vegetation properties at multiple scales . Before at tending the University of Florida , Jessica earned her Bachelor of Science degree in Applied Ecology and Environmental Sciences from Michigan Technological University in Houghton, Michigan. She served as an AmeriCorps ula shortly thereafter. In 2005, she earned her Master of Arts degree in Environmental Science and Policy from Clark University in Worcester, Massachusetts.