Citation
Spectrally-based tree species identification and mapping using high-resolution satellite imagery

Material Information

Title:
Spectrally-based tree species identification and mapping using high-resolution satellite imagery
Creator:
Cross, Matthew David ( author )
Place of Publication:
Denver, CO
Publisher:
University of Colorado Denver
Publication Date:
Language:
English
Physical Description:
1 electronic file (162 pages) : ;

Subjects

Subjects / Keywords:
Biomass energy ( lcsh )
Biomass energy ( fast )
Genre:
Academic theses. ( fast )
Academic theses. ( lcgft )
bibliography ( marcgt )
theses ( marcgt )
non-fiction ( marcgt )
Academic theses ( fast )
Academic theses ( lcgft )

Notes

Review:
A better measure of above-ground biomass for tropical forests can improve assessments of the effect of forests on the global climate through the carbon cycle. Estimates of above-ground biomass vary significantly, and achieving a more accurate measure would allow scientists to accurately assess the role of trees on the global climate and how this can be proactively addressed through efforts such as carbon offset programs. A first critical step in achieving this goal is an accurate assessment of tree species, as each species represents a different biomass measure. This dissertation utilized high-resolution remote sensing imagery to achieve tree species differentiation in a complex tropical forest assemblage. La Selva Biological Station in Costa Rica was the chosen study cite due to its advantageous proximity to a representative biome of diverse lowland Atlantic tropical forests in Central America. In addition, the facility has access to established canopy towers for spectral data acquisition and field validation of individual tree canopies. ( ,, )
Review:
The first step in the research effort was the development of a validated data and image-processing schema demonstrating the capability of current meter-scale satellite technology to identify specific tree species within a tropical forest. This included the process of properly calibrating and correcting field-acquired tree spectral data and WorldView-3 image data for viewing and illumination geometry. Assessments of three current atmospheric compensation methods for correcting recent WorldView-3 satellite imagery established the most accurate compensation process for a tropical forest setting. Next, we statistically evaluated image bands and image-derived spectral vegetation indices (two specifically developed for tropical forest environments) to determine which are the most effective for tree species discrimination. In addition, an object-oriented classification used full tree crowns extracted through a segmentation process, and a series of pre-processed image bands and spectral vegetation indices, to delineate species in a controlled study area. A verification of classification accuracy demonstrated the defined process could accurately differentiate tropical tree species in a controlled study area. Accurate classification of tropical tree species is critical for studies of forest habitat, forest composition, biomass, and ultimately a better understanding of the role trees play in climate variability through carbon uptake.>
Bibliography:
Includes bibliographical references.
System Details:
System requirements: Adobe Reader.
Statement of Responsibility:
Matthew David Cross.

Record Information

Source Institution:
University of Colorado Denver
Holding Location:
Auraria Library
Rights Management:
Copyright Matthew David Cross. Permission granted to University of Colorado Denver to digitize and display this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.
Resource Identifier:
on10762 ( NOTIS )
1076274585 ( OCLC )
on1076274585
Classification:
LD1193.E53 2018d C76 ( lcc )

Downloads

This item is only available as the following downloads:


Full Text

PAGE 1

SPECTRALLY BASED TRE E SPECIES IDENTIFICA TION AND MAPPING USI NG HIGH RESOLUTION SATE LLITE IMAGERY by MATTHEW DAVID CROSS B.S. University of Nebraska Lincoln, 1985 M.A. University of Nebraska Lincoln, 1988 A thesis submitted to the Faculty of the G raduate School of the University of Colorado in partial fulfillment of the requirements for the degree of Doctor of Philosophy Civil Engineering 2018

PAGE 2

ii 2018 MATTHEW DAVID CROSS ALL RIGHTS RESERVED

PAGE 3

iii This thesis for the Doctor of Ph ilosophy degree by Matthew David Cross has been approved for the Civil Engineering Program by Rafael Moreno Chair Wesley Marshall, Advisor Jacek Grodecki Ted Scambos Austin Troy Date: May 12, 2018

PAGE 4

iv Cross, Matthew David (PhD, Civil Engineering ) Tropical Tree Species Delineation Utilizing High Resolution Satellite Imagery for Improved Above Ground Biomass Estimates Thesis directed by Associate Professor Wesley Marshall ABSTRACT A better measure of above ground biomass for tropical forests can impr ove assessments of the effect of forests on the global climate through the carbon cycle. Estimates of above ground biomass vary significantly, and achieving a more accurate measure would allow scientists to accurately assess the role of trees on the globa l climate and how this can be proactively addressed through efforts such as carbon offset programs. A first critical step in achieving this goal is an accurate assessment of tree species, as each species represents a different biomass measure. This disse rtation utilized high resolution remote sensing imagery to achieve tree species differentiation in a complex tropical forest assemblage. La Selva Biological Station in Costa Rica was the chosen study cite due to its advantageous proximity to a representa tive biome of diverse lowland Atlantic tropical forests in Central America. In addition, the facility has access to established canopy towers for spectral data acquisition and field validat ion of individual tree canopies The first step in the research e ffort was the development of a validated data and image processing schema demonstrating the capability of current meter scale satellite technology to identify specific tree species within a tropical forest. This included the process of properly calibratin g and correcting field acquired tree spectral data and WorldView 3 image data for viewing and illumination geometry. As sessments of three current atmospheric compensation methods for correcting recent WorldView 3 satellite imagery

PAGE 5

v established the most acc urate compensation process for a tropical forest setting. Next, we statistically evaluated image bands and image derived spectral vegetation indices (two specifically developed for tropical forest environments) to determine which are the most effective fo r tree species discrimination. In addition, a n object oriented classification used full tree c rowns extracted through a segmentation process and a series of pre processed image bands and spectral vegetation indices to delineate species in a controlled st udy area A verification of classification accuracy demonstrated the defined process could accurately differentiate tropical tree species in a controlled study area. Accurate classification of tropical tree species is critical for studies of forest habit at, forest composition, biomass, and ultimately a better understanding of the role trees play in climate variability through carbon uptake. The form and content of this abstract are approved. I recommend its publication. Approved: Wesley Marshall

PAGE 6

vi DEDI CATION To Mary and Mackenzie for their unwavering patience and support. I could not have done this without both of you.

PAGE 7

vii ACKNOLEDGEMENTS I would first like to recognize my committee. They have been extremely helpful in assisting me to get the stage I a m as a researcher. I am forever grateful for their guidance and support. Others that hav e assisted me along the way include Dr. Jennifer Balch, who enlightened me on forest complexity, Dr. Deb Thomas, for understanding and support, and Dr. Brian Brady fo r getting me started on the right path for success In addition Dr. John Wyckoff for being a mentor and friend, Dr. Fabio Pac ifici for scientific guidance Dr. Kenneth Enge l brecht for convincing me to pursue a PhD and my parents for their support I w ould also like to thank the staff at L a Selva Biological Station, who include Orlando Vargas Ramirez, Ricardo Sandi Sagot Dr. Diego Dierick, Bernal Matarrita Carranza, and Danilo Brenes Madrigal for their support in Costa Rica and their friendship. Sp e cial tha nks goes out to Dr. Ted Scambos friend and colleague.

PAGE 8

viii TABLE OF CONTENTS CHAPTER I. INTRODUCTION ................................ ................................ ................................ ........... 1 Overview ................................ ................................ ................................ ........................... 1 Background and Literature Review ................................ ................................ ................... 3 Study Site ................................ ................................ ................................ .......................... 7 Research Approach ................................ ................................ ................................ ............ 9 Expected Contributions ................................ ................................ ................................ ... 11 References ................................ ................................ ................................ ....................... 13 II. VALIDATING THE USE O F METRE SCALE MULTI S PECTRAL SATELLITE IMAGE DATA FOR IDENT IFYING TROPICAL FORE ST TREE SPECIES .... 17 Abstract ................................ ................................ ................................ ........................... 17 Introduction ................................ ................................ ................................ ..................... 18 Background ................................ ................................ ................................ ..................... 18 Project Approach and Methodology ................................ ................................ ................ 22 Overall Approach ................................ ................................ ................................ ........ 22 Data and Imagery Collection ................................ ................................ ...................... 24 Data and Imagery Preparation ................................ ................................ .................... 28 Analysis of Atmospheric Co mpensation Procedures ................................ .................. 33 Results and Discussion ................................ ................................ ................................ .... 37 Image Data Extraction for P. macroloba and Z. longifolia ................................ ......... 37 Statistical Assessment of 3 Atmospheric Compensation Procedures ......................... 38 Extension of the Statistical Assessment to all Species Collected ............................... 42 Conclusions and Recommendations ................................ ................................ ................ 49 References ................................ ................................ ................................ ....................... 52

PAGE 9

ix III. DETERMINING EFFECTIV E METER SCAL E IMAGE DATA AND SPECTRAL VEGETATION INDICIES FOR TROPICA L FOREST TREE SPECIES DIFFERENTIAT ION ................................ ................................ .................. 57 Abstract ................................ ................................ ................................ ........................... 57 Introduction ................................ ................................ ................................ ..................... 57 Overall Approach ................................ ................................ ................................ ............ 62 ASD Data and WorldView 3 Image Processing ................................ ............................. 64 ASD Da ta Collection and Processing ................................ ................................ ......... 64 WorldView 3 Imagery Acquisition and Preparation ................................ .................. 69 Tree Species Identification Analysis and Resu lts ................................ ......................... 72 Identification Analysis ................................ ................................ ................................ 72 Identification Results ................................ ................................ ................................ .. 72 Tree Sp ecies Differentiation Analysis and Results ................................ ....................... 74 Differentiation Analysis ................................ ................................ .............................. 74 Differentiation Results ................................ ................................ ................................ 82 Conclusions and Recommendations ................................ ................................ ................ 89 References ................................ ................................ ................................ ....................... 91 IV. CLASSIFICATION OF TR OPICAL FOREST TREE S PECIE S USING METER SCALE IMAGE DATA ................................ ................................ ................................ 97 Abstract ................................ ................................ ................................ ........................... 97 Introduction ................................ ................................ ................................ ..................... 98 Overa ll Approach ................................ ................................ ................................ .......... 102 WorldView 3 Imagery and Ground Truth Data Preparation ................................ ......... 104 Imagery Acquisition and Preparation ................................ ................................ ....... 104 Ground Truth Data Collection and Processing ................................ ......................... 107 Arboretum Classification Analysis ................................ ................................ ............... 110

PAGE 10

x Data Selection ................................ ................................ ................................ ........... 111 Objec t based Classification ................................ ................................ ...................... 117 Arboretum Classification Results ................................ ................................ .................. 122 Conclusions and Recommendations ................................ ................................ .............. 126 References ................................ ................................ ................................ ..................... 128 V. CONCLUSION AND FUTUR E RESEARCH ................................ .......................... 134 Conclusion ................................ ................................ ................................ ..................... 134 Future Research Efforts ................................ ................................ ................................ 136 Determination of Forest Biomass ................................ ................................ ............. 136 Forest Diversity Assessments ................................ ................................ ................... 139 Spatial Variability and Quantity ................................ ................................ ............... 140 References ................................ ................................ ................................ ..................... 142 VI. APPENDIX ................................ ................................ ................................ .................. 145

PAGE 11

xi LIST OF TABLES CHAPTER II TABLE 1. Ten tree species collected for this study in May 2016 from La Selva Biological Station, Costa Rica ... 23 2. ASD data collection parameters during acquisition in May 2016 25 3. WorldView 3 imagery specifications and specific image information 28 4. BRF Correction factors for the species collected from the ASD data .. 30 5. BRF Correction factors for the WorldView 3 Imagery ... 33 6. WorldV iew 3 imagery values subsets for P. macroloba and Z. longifolia vs. ASD Data ... 37 7. The results of a Single Sample t test comparing the atmospheric correction output per band, per tree species, to the ASD Data ground truth 40 8. WorldView 3 average imagery reflectance values for ten tropical tree c rowns vs. ASD Data .. .. 43 9. The results of a Single Sample t test for nine tree species, comparing the ASD data to the averaged WorldView 3 image c rown r eflective response .. 48 CHAPTER III TABLE 1. Ten tree species collected for this study in May 2016 from La Selva Biological Station, Costa Rica 6 5 2. ASD data collection parameters during acquisition in May 2016 ... 6 6 3. BRF Correction factors for ASD data collections. Values represent both b ackscatter 8

PAGE 12

xii 4. WorldView 3 imagery s pecifications and specific image 70 5. BRF Correction factors for the WorldView 3 Imagery 70 6. ASD data and WorldView 3 imagery values for ten tropical trees ....... 7 1 7. The results of a two tailed Single Sample t test for 9 tree species, comparing the ASD data to the averaged WorldView 3 image c rown refl ective response 7 3 8. The multiple applications of WorldView 3 imagery 7 6 9. 3 bands for the seven tree species in the study 8 3 10. The output of the Wilks seven tree species in the study 8 4 11. F test significance values for discriminating the seven tree spec ies in the study 8 5 12. Correlation matrix of the most significant independent variables 8 5 13. The results of the Error Matrix analysis using the best independent variables for discriminating tree species ... 8 7 CHAPTER IV TABLE 1. Wor ldView 3 imagery specifications and specific image information 5 2. BRF Correction factors for the WorldView 3 Imagery 10 6 3. Six tree species extracted for the classification ground truth in May 2017 from various locatio ... 10 7 4. test significance values for discriminating the six tree species in the study 11 4

PAGE 13

xiii 5. Correlation matrix of the independen t variables from Table 4 11 5 6. Error Matrix Results Classification 1. The classification results for all rule set inputs equally weighted across all bands/SVIs for all tree species 12 3 7. Error Matrix Classification 2. The classifica tion results for all rule sets specified by tree species ... 12 3

PAGE 14

xiv LIST OF FIGURES CHAPTER I FIGURE 1. La Selva Biological Station, Costa Rica 8 CHAPTER II FIGURE 1. The view of P. macroloba and Z. longifolia ta ken from the vantage point of the canopy towers 27 2. ASD spectral reflectance data for the ten tree species in this study before and after the defined corrections 31 3. Crown r eflectance for ASD Data values for the ten tree species in this study convolved to WorldView 3 imagery band values 32 4. Pictures of P. macroloba and Z. longifolia 34 5. ASD Data vs. Worl dView 3 imagery with AComp, FLAASH and QUAC atmospheric compensation for P. macroloba 6. ASD Data vs. WorldView 3 imagery wi th AComp, FLAASH and QUAC atmospheric compensation for Z. longifolia 7. ASD Data vs. WorldView 3 imagery for C. elastica, I. marginata, L. procera, O. floribun da and P. macroloba 8. ASD Data vs. WorldView 3 imagery for R. Kunthiana, S. microstachyum, S. mombin, W. regia and Z. longifolia 46

PAGE 15

xv CHAPTER III FIGURE 1. One of two canopy towers used for data collection 6 4 2. ASD Data values for the ten tree species and the Canopy Average convolved to WorldView 3 imagery band values ... 6 8 CHAPTER IV FIGURE 1. A true color composite WorldView 3 image of the Holdridge Arboretum 10 4 2. C rown data extractions for two trees in La Selva 0 8 3. Average reflectance the six tree species used as ground control for this study ... 1 0 4. Average WorldView 3 band reflectivity in the visible EM spectrum for the seven tree species in this study 11 6 5. Average WorldView 3 band reflectivity in the near infrared EM spectrum and SVI values for the seven tree species in this study 11 6 6. False color composite of the Arboretum study area .. 11 9 7. The re sult of the segmentation process for the Arboretum 1 20 8. The output of Classification 2 using custom rule sets for each species 12 4

PAGE 16

1 CHAPTER I INTRODUCTION Overview late 1800s. Worldwide carbon in the atmosphere has increased from a pre industrial level of 283 ppm (Stanhill 1982) to present day concentration s of 399 ppm as of July 2014 (Tans and Keeling 2014) An increase in carbo n, manifested in the atmosphere, has led to an increase in worldwide temperatures that are influencing local, regional and worldwide climatic conditions (Silva and Anand 2013; J. Hansen et al. 2008; IPCC 2007) A known control for carbon within the ecosystem is vegetation, where through the normal process of photosynthesis plants uptake and store carbon (Yu et al. 2014) Forest removal through natural fires or anthropogenic bu rning from slash and burn logging releases stored carb on to the atmosphere. The removal of forest also has the additive effect of reducing carbon uptake from the atmosphere, which assists in sustaining and increasing carbon concentrations in the atmosphere (Ghommem, Hajj, and Puri 2012) The control of carbon through veget ation is vital to maintaining appropriate levels of greenhouse gas in the atmosphere. Tropical forests of the world are currently under the most threat of deforestation, and this process has a significant effect on climate change (Gibbs et al. 2007) as tropical deforestation has contributed as much as 25% of the annual global greenhouse gas emissions (Houghton 2005) Measuring biomass accurately in tropical regions in a systematic and cost effective way is important for understanding the effect it will have on a changing climate, specifically estimating carbon content of the standing biomass, leading to a better understanding the role forests play in the global carbon cycle (Barbosa, Broadbent, and

PAGE 17

2 Bitencourt 2014; Gibbs et al. 2007; Jenkins et al. 2004) In addition, the results will likely have important implications reg arding carbon offset formulas (Cifuentes Jara and Henry 2014) Identifying and mapping tropical trees at the species level from space would support a better understanding of the response of forests to climate change through improved assessment of biomass, and ultimately, forest carbon uptake. Significant uncertainties exist in current biomass estimates, particularly in Central and South American tropical forests, due to the complexity and variations among forest stands (Barbosa, Broadbent, and Bitencourt 2014) Biomass can vary drastically among different tree species, and a n accurate tree species differentiation and b etter forest characterization are critical for accurate biomass estimates (Wulder et al. 2004) Depending on the forest structure and trees analyzed, current biomass estimates range from 65 92% accuracy (Dubayah et al. 2010; van Leeuwen and Nieuwenhuis 2010; Lefsky et al. 2005; Lim et al. 2003) These estimates used contemporary measurement techniques, such as remote sensing (passive imagery, LiDAR, etc.) with varying resolutions an d precision (M. C. Hansen and Loveland 2012; Evans et al. 2009) This dissertation will investigate the utilization of the most recent remote sensing technologies to determine the first step in the realization of improved biomass estimates identification of tropical trees at the species level in a complex tropical forest assemblage. A series of remote sensing processing steps and statistical analysis procedures will verify that the image data used for this analysis can differentiate tree species. Once verified the data products will be the input into a classification procedure and validated for accuracy. After the establishment of a classification process within the controlled study area, the schema might be applicable to a larger complex forest setting in calculating above

PAGE 18

3 ground biomass. Realizing a more accurate above ground biomass assessment will lead to a better understanding of how carbon concentrations in the atmosphere could be reduced by the reduction in defo restation (J. Hansen et al. 2008) as carbon emissions have a direct impact on the global climate (Gibbs et al. 2007) Background and Literature Review The scientific community have utilized various approaches to identify forest species through remote sensing applications. Varying imaging systems, such as the coarse resolution the Landsat legacy (Landsat satellites 1 through 7) imagery (Lu 2005) as well as high resolution (meter scale) imagery from commercial sensors (Katoh 2004; Wulder et al. 2004) have been used for tree identifica tion and biomass studies. Many studies have used satellite imagery for tree type identification spanning both temperate locations (Carleer and Wolff 2004) and tropical areas such as Cen tral and South America (Li et al. 2011) Most of the local to regional studies of forest mapping have included some type of remotely sensed imagery as the scale of the study area grows beyond what is cost effective for in situ surveys (Li et al. 2011; M. C. Hansen et al. 2008) Many studies over the last two decades ha ve utilized remotely sensed imagery for mapping and monitoring land cover types, specifically forest type, at the regional scale (Li et al. 2011; M. C. Hansen et al. 2008; Foody, Boyd, and Cutler 2003) Landsat imagery has proven its versatility in mapping a wide array of land cover types and environmental variables in a variety of environments (Vogelmann et al. 1998) It has been especially useful in this regard due to its lengthy record, worldwide coverage, appropriate spatial resolution (30 meters) for regional studies and excellent radiometric characteristics (12 bit resolution) for vegetation analysis and land cover characterization (M. C. Hansen and Loveland 2012) Several studies have advanced the use

PAGE 19

4 of Landsat legacy imagery (pre Landsat 8) combined with other imagery sources, such as MODIS (M. C. Hansen et al. 2008) and active Radar data, wh ich provides land cover textural information Li et al. (2011) used a combination of Landsat legacy imagery and environme ntal/textural information of landscapes in Brazil to provide a solution for identification of forest types. This process was successful in both homogeneous forests and forests having complex stand structures. However, due to the complexity of the forest canopy in bo th tree type and overall extent and past limitations with the sensors ability to discern between subtle variations in surface reflectivity, most studies have reduced any detailed forest identification to only basic tree types. This would incl ude simple differentiations between conifer, deciduous, grassland/savanna, and bare ground or, if specific tree species identification is attempted, it is only over small controlled areas where the tree types/species can be verified (Carleer and Wolff 2004; Katoh 2004) This has severely limited the accuracy of any study regarding accurate forest measurements or biomass There is a recognition within the scientific community that better delineation of tree types, possibly at the species level, could lead to a significant improvement in biomass estimates within the natural forest (UN REDD 2015; Cifuentes Jara, Morales, and Henry 2013; Cole and Ewel 2006) because of varying biomass between species The first step in improving the process of using remotely sensed imagery in a tropical forest setting is determining if spectral reflectance measurements can achieve some separability between tree species. Cochrane (2000) collected data from a ground radiometer to determine if the instrument could differentiate eleven specif ic tree species in Brazil. The visible and near IR responses were the most effective in differentiating species, with an accuracy of separability

PAGE 20

5 achieved of 94%. These results show that the spectrometer achieves a successful delineation of tree species, and possibly the same result with remotely sensed imagery if the data was radiometrically precise and had the appropriate band wavelength combinations. In a similar study, van Aardt et al. (2001) used a spectrometer to asses s its ability in a temperate forest. Ground based spectrometer data were acquired (0.350 um 2.500 um -the approximate spectral range for many contemporary multispectral imaging systems) for loblolly pine (Pinus taeda), virginia pine (Pinus virginiana) shortleaf pine (Pinus echinata), scarlet oak (Quercus coccinea), white oak (Quercus alba), and yellow poplar (Liriodendron tulipifera). The ability of the pure spectral signal to differentiate between species ranged from 62 to 100 percent accuracy. A s imulation dataset for Landsat TM imagery (generated from the spectrometer data) generated various accuracies from 45 to 96 percent. It also identified that most of the species differentiation occurred in the visible to near infrared portions of the EM spe ctrum, and less in the longer IR wavelengths. These and other studies identified that traditional spectral locations currently used for a spaceborne sensor platforms would be ideal for tree species differentiation. This outcome is encouraging; if one ca n achieve accurate surface reflectivity values from traditional spectral band locations, then individual species differentiation is achievable with current remote sensing technologies. There have also been attempts at moving past sensor resolution limita tions by using more advanced imaging systems (Zhang et al. 2006; Clark, Roberts, and Clark 2005) with the thought that improved spatial resolution and radiometric resolution can improve the differentiation of forest species. Pu and Landry (2012) realized a 16 18% increase in accuracy of species identification using newer Worldview 2 data as compared to older,

PAGE 21

6 archived IKONOS data using the same species identification schema. Imagery from the Hyperspectral Digital Imagery Collection Experiment (HYDICE) measured tree crown level spectral variatio n in a controlled forest in Costa Rica (Zhang et al. 2006) Because of the high spatial resolution (1m) of the data, several pixels of image data fall within individual tree crowns within the st udy area. The intra crown variation within species was large and using the pure spectral reflectance pixel values yielded modest results. Further analysis showed that assimilating information within the tree crown was the most promising approach, as the analysis process was more capable of modeling the intra crown variation and differentiating between tree species. Following on from the studies above, other studies have focused on evaluating the most effective classification technique for tree species id entification. Myint et al. (2014) evaluated five different classification schemes on their effectiveness in identifying changes in Mangroves u sing Landsat TM imagery. The study identified that bands 2, 5 and 7 from Landsat TM were the most effective at differentiating mangrove forests from surrounding land cover types. A nearest neighbor classification and an object oriented scheme were the m ost effective, with overall accuracies of 84% and 88% percent respectively. Clark, Roberts, and Clark (2005) conducted a comparison of three specific classifi cation schemes (Maximum Likelihood, the SAM method, Object based process) for determining the best process to differentiate tropical tree spices. Processed HYDICE hyperspectral simulated the spectral bands of Landsat TM, ASTER, and IKONOS imagery. The c omparison showed that the Object based process and Maximum Likelihood processes both achieved good results (overall accuracies for both processes exceeded 88%) as compared with the SAM process, with the Object based process performing slightly better

PAGE 22

7 for a ll simulated sensor spectral ranges. Regarding sensor simulations, the ASTER and Landsat TM simulated bands performed better because of more bands existing in the SWIR, where much of the discrimination of tree species occurred in the study. The successfu l use of an object based approach has been documented (Blaschke 2010) and successfully implemented in several studies of different forest compositions (Tehrany, Pradhan, and Jebu 2014; Chubey, Franklin, and Wulder 2006) showing the validity of this approach and its comparability to more traditional classification sc hemes. The studies outlined above reflect the past ability of remotely sensed imagery to identify forest features or species within a complex forest environment with some accuracy. Some have achieved good results by : 1) focusing on one or few species, 2) limiting a study to a very specific area, 3) implementing hyperspectral imagery in a study, which is not readily available, and/or 4) focusing on general categories of groundcover or vegetation. The difficulty is in differentiating many tree species wi thin a complex forest environment regardless of the sensor system used Developing a process to successfully perform individual tree taxonomy identification (taking advantage of the improved spatial and radiometric information from recent sensor systems) would be of great benefit to a variety of applications (Campbell, James and Wayne 2011) including the calculation of an accurate biomass estimate. This is the basis of study within this dissertation. Study Site La Selva Biological Station, located within Bra ulio Carrillo National Park in north central Costa Rica (Figure 1), comprises over 1 600 hectares of tropical forests and other protected lands. Approximately 73% of the area is primary tropical rain forest. Since its establishment in 1954, it has becom e a premier site for ongoing research in lowland rain

PAGE 23

8 forests. In particular, work on climate change and its impact on biodiversity in tropical wet forests has become a significant area of study at the station. Nearly 240 published scientific papers refl ect the important research conducted at La Selva. Management of La Selva is under the direction of the Organization for Tropical Studies (OTS), a non profit consortium of nearly sixty research organizations and u niversities worldwide. Figure 1. La Selva Biological Station, Costa Rica. T he yellow outline defines the extent of La Selva. Areas designated in white are restricted study areas within La Selva.

PAGE 24

9 A well planned f ield study with accessible locations is essential to realizing success in the field (Fuller et al. 1998) and La Selva facilitate s this by providing excell ent research facilities with direct tropical forest accessibility. OTS manages three Costa Rican research stations representing ecologically diverse ecosystems, with La Selva as one of those sites (OTS 2015) The close relationship between the Univers ity of Colorado Denver, Department of Geography and Environmental Sciences, and La Selva Biological Station helped facilitate the collection of data used for the research in this study. In addition, the highly qualified staff at La Selva assisted our effo rts by providing access to critical infrastructure and facilities, as well as providing expertize about tree species and locations within the biological study area. Research Approach Three specific phases defined the research in this dissertation, each bu ilding upon the other as research segments are completed. The analyses in Chapter 2 initiated the first steps in the research to verify the use of WorldView 3 imagery in identifying tree species. Data collected of specific tree species crowns using ASD fi eld spectrometer established solid ground control for comparison with the WorldView 3 imagery. Removing variations in illumination based on the time of collection and atmospheric effects (Schow engerdt 2007) standardized both the ASD and WorldView 3 surface reflectivity measurements to an on nadir view The ASD data, integrated to specific WorldView 3 band ranges using spectral response curves pr ovided by DigitalGlobe, allowed a direct comparis on to the satellite image data. A statistical analysis confirmed the ability of the WorldView 3 data, using the ASD ground truth data, to differentiate tropical tree species.

PAGE 25

10 The focus of Chapter 3 was to follow on from the results of Chapter 2 by providi ng a statistically based objective determination of the most effective mul t i spectral image data in realizing the maximum separability between studies tree species In addition, an assessment of Spectral Vegetation Indices (SVIs), both traditional and n ewly developed, provided a complete assessment of the ability of the imagery to differentiate tree species in a complex tropical environment. Chapter 4 focused on classifying individual trees at the Holdridge Arboretum, a controlled study area within La Selva that has extensive ground truth information regarding species type, size and position. This will provide the verification information necessary to evaluate the classification schema used in the study. The statistical results from Chapter 3 provide d the necessary objective inputs into the classification of the WorldView 3 imagery. An Object based rule set classification process differentiated select tree species in the Arboretum. The object based classification approach requires a segmentation ana lysis of the spatial extent of each tree type within the AOI to provide a measure of tree crown extent, providing an object based approach for spatial delineation of tree types (Clark, Roberts, and Clark 2005) Spec tral measurements of the trees classified taken outside of the Arboretum at known sites in La Selva provided an independent measure of tree crown reflectivity for all bands and SVIs used in the classification schema. This spectral information was the basi s for the classification process performed. The verification of the classification analysis required the utilization of an Error (Confusion) Matrix, a standard assessment of classification error (Lillesand, Kiefer, and Chipman 2008; Senseman, Bagley, and Tweddale 1995; Congalton 1991) providing ov erall classification accuracy and errors of commission and o mission.

PAGE 26

11 Expected Contributions This dissertation research effor t improves on existing research through the utilization of state of the art remote sensing technologies, precise ground truth and unique processing schemas. The results will ultimately contribute to a better understanding of above ground biomass measureme nts at varying scales and complexity. The outcome of the research will include d the following: 1. A n established processing schema for spectrometer data and high resolution imagery in tropical environments; 2. A verification of the high resolution imagery to di fferentiate tropical tree species; 3. The establishment of a spectral reference library for select tropical tree species; 4. A objectively established list of the most impactful multi spectral image data and data products for tropical tree differentiation; 5. An e stablished object based rule set classification schema for tropical forest species discrimination ; and 6. A validation of the ability of high resolution imagery to identify ree species in a complex tropical forest assemblage. T he above results provide the g roundwork for improving an above ground biomass measurement within a study area. Ultimately, this research improvement could lead to improved biomass estimates at the local and regional level, which could then be further used to calculate total carbon and carbon uptake (Andrade, Brook, and Ibrahim 2008) and influence future estimates for carbon offsets (Cifuentes Jara and Henry 2014) In addition, there is the potential of the schema developed in this dissertation, with minor modifications due to local variables, could be implemented in other areas of the world in a var iety of other tropical forest settings. Other likely applications of this research at the regional level would be to create accurate assessments of species distribution, rates of regrowth and degradation

PAGE 27

12 (Gaston 2000) The results of the remote sensing research activity described in this dissertation could influence traditional forest classification processes, ultimately assist in better characterization of the complex tropical rainforest, and potentially influence policy changes for tropical forest management.

PAGE 28

13 References Aardt, Jan a N van, Randolph H Wynne, Richa rd G Oderwald, and James B Campbell. 2001. Photogrammetric Engineering and Remote Sensing 67 (12): 1 184. Carbon Sequestration of Silvopastoral Systems with Native Timber Species in the Dry Plant and Soil 308 (1 2): 11 22. Aboveground Biomass in Tropical Secondar International Journal of Forestry Research 2014: 1 14. doi:10.1155/2014/715796. ISPRS Journal of Photogrammetry and Remote Sensing 65 (1). Elsevier B.V.: 2 16. doi:1 0.1016/j.isprsjprs.2009.06.004. Campbell, James, B., and Randolph H. Wayne. 2011. Introduction to Remote Sensing 5th ed. New York, NY: The Guilford Press. Species Photogrammetric Engineering Remote Sensing 70 (1): 135 40. doi:10.14358/PERS.70.1.135. Based Analysis of Ikonos 2 Photogramme tric Engineering and Remote Sensing 72 (4): 383 94. doi:10.14358/PERS.72.4.383. Cifuentes Workshop on Tree Volume and Biomass Allome tric Equations in South and Central UN REDD Programme 92. Cifuentes UN REDD MRV Report 11, CATIE, Turialba, C osta Rica, Food & Agriculture Organization of the United Nations Remote Sensing of Environment 96 (3 4): 375 98. doi:10.1016/j.rse.2005.03.009. International Journal of Remote Sensing 21 (10): 2075 87. doi:10.1080/01431160050021303. Col Forest Ecology and Management 229 (1 3): 351 60. doi:10.1016/j.foreco.2006.04.017.

PAGE 29

14 assifications of Remote Sensing of Environment 37 (1): 35 46. doi:10.1016/0034 4257(91)90048 B. Dubayah, R. O., S. L. Sheldon, D. B. Clark, M. A. Hofton, J. B. Blair, G. C. Hurtt, and R. L. st Height and Biomass Dynamics Using Journal of Geophysical Research: Biogeosciences 115 (2): 1 17. doi:10.1029/2009JG000933. ete Return Lidar in Natural Resources: Recommendations for Project Planning, Data Remote Sensing 1 (4): 776 94. doi:10.3390/rs1040776. Tropi cal Forest Biomass from Landsat TM Data and Their Transferability between Remote Sensing of Environment 85 (4): 463 74. doi:10.1016/S0034 4257(03)00039 7. Fuller, R. M., G. B. Groom, S. Mugisha, P. Ipulet, D. Pomeroy, A. Katende, R. Bailey, and R Ogutu Biodiversity Assessment: A Case Study in the Tropical Forests and Wetlands of Sango Biological Conservation 86 (3): 379 91. doi:10.1016/S0006 3207(98)00005 6. Gasto Nature 405 (6783): 220 27. doi:10.1038/35012228. Ecolo gical Modelling 235 236. Elsevier B.V.: 1 7. doi:10.1016/j.ecolmodel.2012.04.005. Environmental Resear ch Letters 2 (2007): 45023. doi:10.1088/1748 9326/2/4/045023. Hansen, James, Makiko Sato, Pushker Kharecha, David Beerling, Robert Berner, Valerie Masson Delmotte, Mark Pagani, Maureen Raymo, Dana L. Royer, and James C. The Open Atmospheric Science Journal 2 (1): 217 31. doi:10.2174/1874282300802010217. Remote Sensin g of Environment 122. Elsevier Inc.: 66 74. doi:10.1016/j.rse.2011.08.024. Hansen, Matthew C., David P. Roy, Erik Lindquist, Bernard Adusei, Christopher O. Justice, Systematic M Remote S ensing of Environment 112 (5): 2495 2513. doi:10.1016/j.rse.2007.11.012. Houghton, R. 2005. Tropical Deforestation as a Source of Greenhouse Gas Emissions Tropical Deforestation and Climati c Change doi:10.1017/S0376892900029775.

PAGE 30

15 Intergovernmental Panel on Climate Change: Contribution of Working Group I to the Fourth Assessment Report of the Intergovernmental Panel on Climate Ch ange 1 18. Jenkins, Jennifer C, David C Chojnacky, Linda S Heath, and Richard a Birdsey. 2004. Based Biomass Regressions for North American USDA Forest Service Vol. GTR NE 319. ssifying Tree Species in a Northern Mixed Forest Using High Journal of Forest Research 9 (1): 7 14. doi:10.1007/s10310 003 0045 z. Parameters Using European Journal of Forest Research 129 (4): 749 70. doi:10.1007/s10342 010 0381 4. Lefsky, Michael A., David J. Harding, Michael Keller, Warren B. Cohen, Claudia C. Carabajal, Fernando Del Bom Espirito Santo, Maria O. Hunter, and R aimundo De Geophysical Research Letters 32 (22): 1 4. doi:10.1029/2005GL023971. Cover Classifi cation in a Moist Tropical Region of Brazil with Landsat Thematic Mapper International Journal of Remote Sensing 32 (23): 8207 30. doi:10.1080/01431161.2010.532831. Lillesand, Thomas M., Ralph W. Kiefer, and Jonathan W. Chipman. 2008. Remote Sens ing and Image Interpretation 6th Editio. John Wiley and Sons Ltd. Lim, K., P. Treitz, M. A. Wulder, B. St Progress in Physical Geography 27 (1): 88 106. doi:10.1191/0309133303pp360ra. L International Journal of Remote Sensing 26 (12): 2509 25. doi:10.1080/01431160500142145. g Change Photogrammetric Engineering and Remote Sensing 80 (10): 983 93. doi:10.14358/PERS.80.10.983. OTS Website www.ots.ac.cr/. Pu, Ruiliang, and Sha IKONOS and WorldView Remote Sensing of Environment 124. Elsevier Inc.: 516 33. doi:10.1016/j.rse.2012.06.011. Schowengerdt, Robert A. 2007. Remot e Sensing: Models and Methods for Image Processing 3rd Editio n Elsevier Inc.

PAGE 31

16 the Discrete Classification of Remotely Const ruction Engineering Research Lab (ARMY) no. CERL TR EN 95/04: 1 31. Global Ecology and Biogeography 22 (1): 83 92. doi:10.1111/j.1466 8238.2012.00783.x. 1877 Climatic Change 4 (3): 221 37. www.esrl.noaa.gov/gmd/ccgg/trends/. Comparative Assessment between Object and Pixel Based Classification Approaches Geoc arto International 29 (4): 351 69. doi:10.1080/10106049.2013.768300. UN UN REDD Website www.un redd.org/. Vog Environmental Monitoring and Assessment 51 (1 2): 415 28. Wulder, Michael A., Ronald J. Hall, Bioscience 54 (6): 511 21. doi:10.1641/0006 3568(2004)054[0511:HSRRSD]2.0.CO;2. Yu, G, Z Chen, S Piao, C Peng, P Ciais, Q Wang, Proc Natl Acad Sci U S A Natl Acad Sci U S A 111 (13): 4910 15. doi:10.1073/pnas.1317065111. Zhang, Jinkai, Benoit Rivard, Arturo Snche z Azofeifa, and Karen Castro Esau. 2006. and Inter Class Spectral Variability of Tropical Tree Species at La Selva, Costa Remote Sensing of Environment 105 (2): 129 41. doi:10.101 6/j.rse.2006.06.010.

PAGE 32

17 CHAPTER II VALIDATING THE USE O F METRE SCALE MULTI SPECTRAL SATELLITE IMAGE DATA FOR IDENT IFYING TROPICAL FORE ST TREE SPECIES Abstract Identifying and mapping tropical trees at the species level from space can support an improve d assessment of forest composition, forest carbon uptake, tree species distribution and preferred habitat as well as a better understanding of the response of forests to climate change. In this study, the development of a validated data and image processin g schema demonstrated the capability of current meter scale satellite technology (WorldView 3) to identify specific tree species within an unmanaged tropical forest. The experimental site, La Selva Biological Station in Costa Rica, provided access for fie ld validation and spectral data acquisition of individual tree canopies from established canopy towers. It is also a representative biome of diverse lowland Atlantic tropical forests in Central America. The process defined in this paper calibrated and co rrected field acquired ASD data for ten tree species and corrected WorldView 3 image data for viewing and illumination geometry. In addition, assessments of three current atmospheric compensation methods for correcting recent WorldView 3 satellite imagery established the most accurate compensation process for a tropical forest setting. Corrected reflectance in the satellite data matched the spectrometer data to 0.25% for visible bands and 0.5% for near infrared bands. This study shows that spectral dat a from the satellite and field spectrometer data are nearly equivalent when applying the appropriate atmospheric compensation, band response emulation, and viewing correction processes established in this study.

PAGE 33

18 Introduction The goal of the study was to es tablish a data processing protocol for achieving a good representation of the spectral characteristics of emergent trees in the tropical forest canopy through high resolution imagery. An important part of the image processing required to achieve a high le vel of spectral precision is accurately compensating for atmospheric effects, particularly in tropical environments where atmospheric attenuation is high due to water vapour and aerosols. This study included an analysis to determine the most accurate atmo spheric compensation processing for images of such regions. Ultimately, the research goal was to determine if high spatial and spectral resolution imagery could be effective at differentiating tree species within a complex tropical forest setting. La Sel va Biological because it offers quick access to a mostly undisturbed tropical forest setting, with infrastructure and research facilities that can support field wor k In addition, the facility has a technical support staff that is knowledgeable in field collection techniques and the forest species present in the region. In particular, access to spectral measurements is available from several above canopy towers, all owing observation from near nadir and even on nadir perspectives (using a canopy tower bridge) for comparison with satellite based imagery. It is an exceptional site for lowland tropical forest studies Background Remotely sensed imagery is an established data source for determining basic tree types, their spatial distribution, and overall assessments of forest health and above ground carbon content. Forest biomass studies have used various imaging systems, such as the coarse resolution Landsat legacy seri es (Landsat satellites 1 through 7) imagery (G. Li et al.

PAGE 34

19 2011) and high resolution (meter scale) imagery from commercial sensors (Katoh 2004; Wulder et al. 2004) Several studies have used satellite imagery for tree identification in a variety of areas, spanni ng both temperate locations (Carleer and Wolff 2004) and tropical areas such as Central and South America (G. Li et al. 2011) Due to the complexity of most tropical forest canopies in both number of tree species and varying composition, as well as sensor limitations in measuring canopy reflectivity, most studies have succeeded in identi fying only basic forest tree types (e.g., conifer or deciduous). This inherently limits the accuracy of any study of forest composition or biomass estimation. A number of studies have recognized that better delineation of trees, ideally at the species le vel, could lead to a significant improvement in biomass estimates for the tropical forest environment (UN REDD 2015; Cifuentes Jara, Morales, and Henry 2013; Cole and Ewel 2006) A key early step toward such an analysis is determining if tree species identification is achievable within a c omplex forest setting using common satellite image sensors. A study by van Aardt et al. (2001) assessed ground based spectral data of individual trees to determine the feasibility of distinguishing between six distinct tree sp ecies in the southeastern US. The results of the cross validation test for separability generated accuracies varying between 45 to 96 percent, with the high variation in accuracy due to a pixel based classification schema. Much of the spectral separabili ty between species occurred in the visible and near infrared bands, and less in the longer IR wavelengths, demonstrating that typical current satellite band locations in the spectrum have at least the potential for species discrimination within a complex f orest. In a similar study, Cochrane (2000) collected data from an ASD field spectrometer (350nm 1050nm) to determin e if eleven distinct tree species in Brazil could be differentiated. Root mean square differences between reflective

PAGE 35

20 responses for the eleven species determined their dissimilarity. The visible and near IR responses were the most effective in classifying species, achieving an accuracy of 94%. Pu and Landry (2012) observed a 16 18% increase in accuracy of species identifi cation using WorldView 2 multispectral imagery (1.84m pixels), as compared to IKONOS data (4m pixels) using the same species identification schema, and stated that even finer spatial resolution imagery would be more successful. There have been attempts at moving past the sensor limitations by using more advanced imaging systems (Zhang et al. 2006; M. L. Clark, Roberts, and Clark 2005) and innovative analysis of the data though contemporary spatial analysis techniques (Myint et al. 2014; J. Li 2004) Imagery from the Hyperspectral Digital Imagery Collection Experiment (HYDIC E) was used by Zhang et al. (2006) to examine tree crown level spectral variation in a controlled forest in northern Costa Rica. A comparison between the Spectral Angle Mapper (SAM) classification process and a wavelet transformed spectral domain process, regarding the differentiation of five tree species, showed the wavelet process to be more effective than the SAM process It was more capable of inco rporating the intra crown variation. Our processing approach, averaging multiple pixels and multiple spectral observations across the upper tree crowns, makes use of this conclusion. Li (2004) used hyperspectral imagery with a Discrete Wavelet Transform (DWT) based feature extraction to improve the differentiation of end member values in a linear pixel unmixing process. This assisted in determining the relative abundance of ground cover with each pixel, increasing the precision of the spatial domain of each feature extracted. The use of a wavelet based feature signal classification process improved end member extraction precision by more than 30% as compared to a Principal Components An alysis (PCA). Other

PAGE 36

21 studies attempted to determine the most efficient way to extract end members, or pure signal from the imagery. Most of the established processes are effective at this task, with Automatic Morphological Endmember Extraction (AMEE) typi cally performing slightly better (Martnez et al. 2006) The cited studies reflect the current ability of remotely sensed imagery to identif y forest features or species within a complex forest environment. Some achieved good results by: 1) focusing on one or a few species for detection; 2) limiting the study area; 3) applying hyperspectral imagery in their analysis (which at present is not pr esently available on satellite borne platforms); and/or 4) focusing on very general categories of groundcover or vegetation. The limitations of appropriate spatial and radiometric resolution, overall data availability, and appropriate data corrections cons trained the above studies. Here we present an improved approach, using an advanced meter scale satellite imaging system (WorldView 3), that can provide images of the canopy with sub meter resolution and high radiometric resolution. In this study, we des cribe a data processing protocol for achieving a good representation of the atmospherically compensated tropical forest canopy spectra, resolving ten individual tree crowns as pixel clusters through textural processing of WorldView 3 data. We determine th e most accurate atmospheric compensation process for this objective, and describe the appropriate processing steps Our study shows that WorldView 3 image data captures the defining spectral characteristics of a set of tropical tree species using our appr oach, as determined by in situ field spectrometry, and are potentially distinctive enough to support canopy mapping at the species level.

PAGE 37

22 Project Approach and Methodology Overall Approach This study assessed both the relative quality of the atmospheric cor rections, as well as the likelihood that tree species might be discriminated in a random area of similar forest using WorldView 3 imagery if one followed a similar measurement and data processing protocol The first methodological step was to establish a ccurate ground truth for comparison to the imagery collected. ASD field spectrometer acquisitions f rom ideally situated examples of ten major tree species of the Costa Rican tropical lowland forest were the first activity for this study (Table 1) followe d by a series of data acquisition corrections (white reference, attenuation, viewing geometry). The ten species exhibit a large variability regarding their composition, botanical family, and overall canopy structure. The accessibility of well exposed tre e c rowns from established canopy towers at La Selva constrained the selection of tree species for this study. This facilitated a far lower data acquisition cost compared to a drone or aircraft based approach. The next step for the project was the select ion of a cloud free WorldView 3 image of the study area during the dry season (November to March) from the DigitalGlobe image archive To facilitate a straightforward protocol for ASD data and WorldView 3 imagery processing, we initially selected two tree species from the original ten collected. This select tree group allowed a first assessment of the ability of the WorldView 3 imagery to identify tree species. ASD spectra was compared with image pixels acquired by WorldView 3 under similar viewing and en vironmental conditions (dry season, off nadir) after applying three current atmospheric compensation procedures; QUAC, (Bernstein et al. 2012) FLAASH, (Matthew et al. 2002) and AComp (Pacifici 2013) A single sample t test statistical analysis

PAGE 38

23 established a measure of the likelihood of a sample being a member of the population (ASD data value), based on the closeness of the values and the variance in the sampled whole crown reflectance values from the WorldView 3 imagery. After establishing the processing schema, a similar analysis of the remaining trees collected provided an assessment of the WorldView 3 imagery in differentiatin g multiple tree species in a complex tropical setting. Species Acquired Within La Selva Biological Station Tree Species Family Common Name Range Unique Characteristics Castilla elastica Moraceae Panama Rubber Tree Northern S. America, Central America S ource of latex Inga marginata Fabaceae Guabilla Northern S. America, Central America Edible fruit, leaves are used as an astringent Laetia procera Salicaceae Manga larga Central America Rare, restricted to old forests Ocotea floribunda Lauraceae Laure l espada South and Central America Wood used in fine carpentry, furniture Pentaclethra macroloba Fabaceae Pracaxi, Kuntze Northern S. America, Central America Seeds a source of cooking oil, bark for antiseptic Rhodostemonodaphne kunthiana Lauraceae Qui zarra Negro, Sweetwood Northern S. America to Costa Rica Used in construction Stryphnodendron microstachyum Fabaceae Cacha, Yellow Targuayugo South and Central America Bark used as an astringent, and for ink Spondias mombin Anacardiaceae Yellow Mombin Hogplum South and Central America Widely cultivated for its edible fruit Welfia regia Arecaceae Amargo Palm South and Central America Edible leaves, palm heart, used in construction Zygia longifolia Fabaceae Sotacaballo, Chiparo Central America Used for erosion control, esp. banks of rivers Table 1. Ten tree species collected for this study in May 2016 from La Selva Biological Station, Costa Rica. Species information abov e from (Castro 2014; STRI 2014; Condit, Perez, and Daguerre 2011)

PAGE 39

24 Tree c rowns have inherent variability within the canopy structure (i.e. varied leaf orientation, varied branch and leaf density, open areas, exposed branches, parasitic flora). For this reason, we performed an integration for both the satellite imagery (average of all pixels selected for a canopy) and field spectrometer data over a substantial part of the individual tree crowns, rather than extracting field spectra of individual leaves, or using single image pix els. This approach integrates within canopy variations within an individual tree crown. Both the field spectrometry data and the textural processing of the images constrained the analysis to the center of the tree crowns to reduce shadowing and species o verlap along the crown edges, giving a more complete and consistent measure of the tree canopy spectra (M. L. Clark, Roberts, and Clark 2005) The multi pixel representation of individual tree canop ies, and ground truth data that integrates over a single tree canopy, is the combination of meas urement techniques that provides a unique characteristic for each tree type and maximizes differentiation between species (Feret and Asner 2013; Cho et al. 2008) Data and Imagery Collection ASD Data Collection Protocol A collection of spectral reflectance data from ten tree species from established canopy towers at La Selva acquir ed using a portable Analytical Spectral Devices (ASD) FieldSpec Pro Spectroradiometer in May 2016 was the main field data set used for this study. The data acquisition geometry from the canopy towers mimicked near nadir satellite viewing and illumination geometry. The ASD instrument collects reflected energy spanning 300 2500 nm, with a sampling interval from 1.4 nm (350 1000 nm) to 2 nm (1000 2500nm). An 18 fore optic field of view port attached to the detector element masked c rown edges and mixe d tree areas while facilitating mean central c rown averaging during data acquisitions.

PAGE 40

25 The ASD instrument has a gain optimization procedure that adjusts the raw digital numbers collected (16 bit for the ASD Fieldspec Pro instrument) to maximize radiometric sensitivity and minimize saturation during the collection. A dark current adjustment before each collection session eliminated instrument background input into the fina l signal. The acquisition of a 500 sample spectrum average minimized variability in r eflectivity and ensured a consistent data collection (ASD 2002) This high resolution collection capability allows for very precise reflective measurements in the visible through the near infrared range of the spectrum (Hatchell 1999) All ASD data collected from two tower locations within La Selva Biological Station occurred in May 2016 at the end of the dry season (Table 2). The timing of the data collection minimized m id day cloud cover standing water on the leaves, and facilitated observations under viewing conditions similar to those of the satellite data acquisitions. Tree Species ASD Acq. Date ASD Acq. Time OZA AZ Target Dist. Sky Cond. C. elastica 5/10/16 11:55 am 30 180 12m Part. Cldy. I. marginata 5/10/16 12:08pm 25 90 14m Part. Cldy. L. procera 5/11/16 11:40am 40 90 15m Clear O. floribunda 5/11/16 11:25am 45 225 22m Clear P. macroloba 5/11/16 11:50am 40 200 16m Clear R. kunthiana 5/10/16 12:25pm 20 320 10m Clear S. microstachyum 5/10/16 2:27pm 40 220 11m Clear S. mombin 5/10/16 12:15pm 45 20 11m Clear W. regia 5/11/16 12:10pm 15 90 6 m Clear Z. longifolia 5/10/16 12:45pm 30 250 12m Clear T able 2 ASD data collection parameters during acquisition in May 2016. OZA refers to the Observation Zenith Angle, which is the angle of data collection off zenith. AZ refers to the Azimuth collection angle.

PAGE 41

26 The height of the towers (33 and 46 m) prov ided moderately off nadir observations of the tree canopies. Data collections occurred within 2 hours of local noon to minimize solar zenith angle variability between data acquisitions. All data collections included extensive ancillary data for the ASD d ata correction processes. This included tree species, time of acquisition, Observation Zenith Angle (OZA), Azimuth (AZ), distance to target, and sky conditions as well as GPS positions of the towers (not shown). Each spectra dataset consisted of four col lections averaging 500 spectra during the collection period. For each tree, a white reference (Spectralon) dataset initiated the collection, followed by two tree collections, and then a second white reference. Using the least affected white reference and tree collection data ensured the minimization of any variation in illumination due to cloud obscurations. To determine the ASD acquisition footprints for tree c rown target s we used the following equation: w here the Field Of View (FOV) is the diameter of the angle of the collection cone set by the fore optic (18 degrees for this study), h is the height above the target, and the calculated GFOV is the area of tree canopy collection (ASD 2002; Hatchell 1999) For an 18 fore optic, the GFOV calculated in equation (1) is approximately one third the distance (h) to the target. Ensuring that the s pectrometer collection would focus on an individual tree c rown a data collection distance between the measured tree c rown s and the instrument tower observing site typically ranged from 6 m to 22 m. Figure 1 illustrates the approximate collection for two tree species resulting from the slight movement of spectrometer over each tree canopy to ensure full canopy coverage.

PAGE 42

27 WorldView 3 Imagery Acquisition A nearly cloud free WorldView 3 off nadir image, acquired on November 11, 2014 was used for this stu dy. Table 3 outlines the specifics of the imagery collection. The criteria for image selection were that it had to be within the dry season, have minimal cloud cover and atmospheric water vapor over the Area of Interest (AOI), and an acquisition as close to the field data collection time as possible. Although it would have been preferable to have a clear sky satellite acquisition during the field expedition period, the November 11, 2014 image best fit the criteria necessary for a good spectral comparison. Despite the 18 month separation, we easily identified the tree canopies used for this study in the satellite image. Having a high spatial resolution dataset is critical to having multiple imagery pixels defining the central portions of each tree c rown Th is supports processing that integrates across an entire single tree c rown without mixture from adjacent trees. WorldView 3, in addition to high spatial resolution, has additional band information in the visible and near infrared po rtions of the spectrum ( Table 3 ). In addition, WorldView 3 imagery for the visible/near infrared (VNIR) bands has a radiometric resolution of 11 bits per pixel Figure 1. The view of P. macroloba (left) an d Z. longifolia (right) taken from the vantage point of the canopy towers. Areas depicted in yellow delineate the approximate area of the spectral data collection. Slight movement of the ASD fore optic during collection resulted in acquisition of the ave rage response of the entire tree c rown

PAGE 43

28 (DigitalGlobe 2014) increasing the sensitivity to spectral features. This is a significant improvement over most previous publically available satellite imagery. This higher radiometric resolution supports better differentiation of tree types by more precisely defining the reflectivity intensity level in each data band used (Campbell, Jam es and Wayne 2011) Da ta and Imagery Preparation A SD Data Conversion to On Nadir Absolute Surface Reflectance Converting the ASD data collected to a standard surface reflectance condition for direct comparison to the WorldView 3 imagery required multiple data processing steps. A description of these processes is below. Absolute Reflectance and Spectralon Correction Equation 2 converted the DN values of the ASD data collections to absolute reflectance : where R abs is the absolute reflectance of the target in question, I sample is the radiant intensity measured for the sample (tree crown ), and I ref is the radiant intensity from a Spectralon white Bands Spectral Range (nm) Resolution Panchromatic 450 800 Panchromatic 0.31m Costal 400 450 Multispectral 1.24m Blue 450 510 Dynamic Range 11 bits/pixel Green 510 580 Specific Image Information for Study Yellow 585 625 Date/time 11/11/2014, 15:52:28Z Red 630 690 Zenith View Angle 26.2 Red Edge 705 745 Azimuth View Angle 108.8 Near IR1 770 895 Cloud Cover 0.5% Near IR2 860 1040 Data Extent NW Corner 10.48 N, 84.14 W SE Corner 10.25 N, 83.99 W Table 3 WorldView 3 Imagery Specifications (DigitalGlobe 2014) and specific image information extracted from the metadata file from the imagery used in this study.

PAGE 44

29 reference panel under identical lighting conditions. As the white reference panel is not a perfect white Lambertian surface, we applied a correction for the white reference, derived as part of the periodic calibration for the ASD spectrometer (Anderson 2002) to each wavelength collected over the range 350 nm 2500 nm (ASD 2002) I ref in equation (2) included this correction to convert the tree spectra to absolute crown reflectance (R obinson and MacArthur 2011; Miura and Huete 2009; M. L. Clark, Roberts, and Clark 2005; R. N. Clark et al. 1999) Water Vapor Band Removal The elimination of water vapor absorption bands (1350 nm 1460 nm, 1790 nm 1960 nm and 2350 nm 2500 nm) allow ed a more accurate representation of the final absolute surface reflectance output Bachmann, and Heldens 2006; ASD 2002; Hatchell 1999) High water vapor absorption, typical of tropical locations, can result in very low reflective intensity values in some image bands. A smoothing pro cess is typically performed on an absolute reflectance output to remove other anomalies (Dorigo, Bachmann, and Heldens 2006) but was not applied in this study, as subtle variances in reflectivity could be important in the validation of the atmospheric compensation procedures analyzed. Bi Directional Reflectance Correction An important step often overlooked in remotely sensed imagery processing is the correction for variation in surface reflectivity defined by the Bidirectional Reflectance Distribution Function (BRDF) for surface reflectance (Roy et al. 2016; Gastellu Etchegorry et al. 1999) This is especially true for variable view angle, high spatial resolution imagery (Pacifici, Longbotham, and Emery 2014) The error generated by BRDF variability c an be significant, as most surfaces are

PAGE 45

30 neither perfectly specular nor perfectly diffuse, and this is the case for tree canopies (Breunig et al. 2015; Moura et a l. 2012) A recent study by Weyermann et al. (2014 ) found a measurement uncert ainty of up to 20% attributable to ground reflectance anisotropy. Reflectance variability can be as much as 30% in the near infrared portion of the spectrum, depending on the view angle and fore, side, or backscatter geometry. Several studies of tropical rainforest environments identified varied anisotropic effects in the near infrared through the visible range (Moura et al. 2012; Bousquet et al. 2005; Middleton 1992) Because of the field limitations in collecting near nadir top of c rown reflectance from each tree species (all data were collected from static towers and collection view angles were restricted), all ASD data were collected off nadir The application of Bidirectional Reflectance Factor (BRF) correction estimates (derived from Multi angle Imaging Spectroradiometer ( MISR) imagery over a mixed tree canopy in Brazil similar to the La Selva forest canopy) from Breunig et al. (2013) allowed the normalization of the ASD data to on nadir reflectance values (Table 4). ASD Spectrometer Data Collection Spectral Region Range (nm) C. elastica 30 OZA Backscatter I. marginata 25 OZA Backscatter L. procera 40 OZA Backscatter O. floribunda 45 OZA Backscatter P. macroloba 40 OZA Sidescatter Blue 350 499 0.8784 0.8701 0.9144 0.9324 0.9491 Green 500 628 0.9785 0.9 910 0.9473 0.9317 0.8135 Red 629 700 1.0543 1.0538 1.0495 1.0471 0.8981 Near IR 701 2500 0.9991 1.0134 0.9607 0.9415 0.8093 Spectrum Location Range (nm) R. kunthiana 20 OZA Forescatter S. microstachyum 40 OZA Sidescatter S. mombin 45 OZA Sides catter W. regia 15 OZA Backscatter Z. longifolia 30 OZA Sidescatter Blue 350 499 0.9161 0.9491 0.9702 0.9220 0.9070 Green 500 628 0.8793 0.8135 0.8031 0.9946 0.8344 Red 629 700 0.9161 0.8981 0.9008 1.0323 0.8926 Near IR 701 2500 0.8896 0.8093 0.7926 1.0081 0.8428 Table 4 BRF Correction factors for the species collected from the ASD data. We derived values from (Breunig et al. 2013) that represent both backscatter and sidescatter conditions. Dividing the observed reflectance by the BRF correction factors above derives the on nadir surface reflectance. OZA refers to Observation Zenith Angle.

PAGE 46

31 Supporting information from Gastellu Etchegorry et al. (1999) helped to characterize the variability of the BRF from a fore, side, and backscatter viewing condition. The zenith view angle for the ASD collection, azimuth illumination angle, and wavelength all factored into the BRF correction. Figure 2 shows the ASD reflectance results for the ten tree species before and after applying all of the previous corrections described. This study considers the corrected spectra as the ground truth for each tree species analyzed, as this represents the spectra for each species crown measured under near ideal growing and viewing conditions with all major corrections applied. Figure 2. ASD spectral reflectance data for the ten tree species in this study before and after the defined corrections.

PAGE 47

32 ASD Data Conversion to Equivalent WorldView 3 Imagery Bands The processing of spectra shown in Figure 2 facilitated direct comparison to the WorldView 3 imagery collect ed in the AOI by resampling the data into the band ranges of the Worldview sensor (Campbell, James and Wayne 2011) The spectral response curve values for the panchromatic, visible and near infrared bands, provided by DigitalGlobe (Kuester 2016) allowed the conversion of the ASD data to WorldView 3 bands. A MATLAB program convolved the ASD spectral data to the WorldView 3 band detected spectrum as a function of th e spectral response curves for each of the 8 WorldView 3 spectral bands (Milton and Choi 2004) The result was a derived crown reflectivity value generated from the ASD data for each of t he eight WorldView 3 spectral bands from visible to the near infrared (Figure 3) for the ten tree species in this study. Figure 3 Crown reflectance for ASD Data values for the ten tree species in this study convolved to WorldView 3 imagery band values. Lines between points are for clarity to show the trend in variability between bands.

PAGE 48

33 WorldView 3 Imagery Preparation A BRF correction was required to normalize all data to an on nadir view for direct comparison with the convolved ASD data due to the off nadir view angle and backscatter viewing condition of the WorldView 3 image. This required different correction values (Fabio M. Breunig et al. 2013) compared to the ASD data collection, as the observation zenith ang le and illumination angle were different during the WorldView 3 acquisition timeframe (Table 5). This image correction step brought the WorldView 3 imagery to a processing level that facilitated a comparison of three specific atmospheric compensation proce dures to determine which one is most appropriate for a tropical forest setting. Analysis of Atmospheric Compensation Procedures Two species from the tree collection ( Pentaclethra macroloba and Zygia longifo lia) were the focus for an initial study of atmospheric compensation procedures because of their prevalence within the La Selva, and their high variation in c rown and leaf structure (Figure 4) that leads to a large difference in overall crown reflectance a cross the EM spectrum WorldView 3 Imagery Collection Backscatter Condition Band Locations Wavelength Range (nm) BRF Factor 26.2 OZA Coastal 400 450 1.01 Blue 450 510 1.08 Green 510 580 1. 15 Yellow 585 625 1.17 Red 630 690 1.20 Red Edge 705 745 1.25 Near IR1 770 895 1.33 Near IR2 860 1040 1.34 Table 5 BRF Correction factors for the Worl dView 3 Imagery. Values were derived from (Breunig et al. 2013) and represent backscatter viewing conditions for the imagery. Dividing the observed reflectance by the BRF correction factors above derives the on nadir surface reflectance.

PAGE 49

34 (Figure 1). Both tree species belong to the family Fabaceae, or legumes, which are widely distributed in Central America. It is the third largest land plant family in terms of number of species, behind only the Orchidaceae and Aste raceae, with about 751 genera and some 19,000 known species (Christenhusz and Byng 2016) P. macroloba reaches heights of 30 35 m and a trunk diameter of 130 cm. Leaves are arranged in a spiral pattern, twice compound, and can be up to 30 cm long with 15 to 20 paired leaflets 2 10 cm long (Flores, 2002) The range for P. macroloba is from Nicaragua to the Amazon basin, defined by three distinct populations. One population ranges from the Amazon, through Venezuela and the Guianas, to Trinidad and Tobago. The second occurs in western Colombia and Darien Province in Panama. The third population is in western Panama, Costa Rica, and Nicaragua. It grows in lowland forest s and is prevalent in wetter climates (Flores 2002) P. macroloba is the predominant tree type within La Selva (Whitmore, Peralta, and Brown 1985) Z. longifolia is a medium sized tree that can grow up to 15 m high with a rounded crown and the trunk branched low to the ground. The leaves are bipinnate and alternate, 5 Figure 4. Pictures of P. macroloba (left) and Z. longifolia (right). Notice the significant variance in leaf structure for two species in the same family (Fabaceae).

PAGE 50

35 12 cm long and 1.5 4 cm wide (STRI 2014; Condit, Perez, and Daguerre 2011) Z. longifolia is native to only northern South America and Ce ntral America. It is common around wetland areas, and prefers streambanks (STRI 2014; Condit, Perez, and Daguerre 2011) It is quite prevalent within La Selva, as this location provides several wetland areas and two major rivers that define its wester n, northern and eastern borders. Comparison of Atmospheric Compensation Approaches We processed the WorldView 3 imagery using 3 atmospheric compensation procedures; QUAC, FLAASH and AComp (Bernstein et al. 2012; Matthew et al. 2002; Pacifici 2013) to compare which process works best in a tropical rainforest environm ent. The output of each procedure (surface reflectance) facilitated a direct comparison to the field gather ed ASD tree spectra as ground truth (Marcello et al. 2016) Harris Geospatial Solutions includes the QUAC atmospheric compensation process in their ENVI version 5.2 image processing software. Data input is sensor type (MODIS, IKONOS, Landsat TM, etc.), acquisition solar elevation angle, and sensor view angle. QUAC generates most of its atmospheric compensat ion parameters directly from the observed pixel spectra within the image (Marcello et al. 2016; Bernstein et al. 2012) and produces an approximate correction typically within +/ 15 % from a physics based compensation schema (ENVI 2009) The FLAASH compensation process, based on MODTRAN4 radiation transfer code (Felde et al. 2003) is also included in ENVI 5.2. Input variables include sensor type, sensor altitude, acquisition date and time, sensor zenith and azimuth, geographic atmospheric variability (polar, mid latitude, tropical, etc.), aerosol estimate s, and sensor look zenith and azimuth angles (Manakos et al. 2009) FLAASH can be used with hyperspectral imaging

PAGE 51

36 systems (Matt hew et al. 2002) and a variety of multispectral imagers (Landsat, ASTER, SPOT, etc.). Since no water vapor bands or humidity data were available, the conditions were set as cloud free tropical environment with a high visibility index (25 km visibility) and water vapor present in the atmospheric column (ENVI 2009 ) AComp is a proprietary atmospheric compensation process developed by DigitalGlobe for their WorldView sensor series imagery. AComp is similar to FLAASH in that it is a physics based compensation schema (Pacifici 2013) but is also similar to QUAC in that it uses observed in scene pixel spectra, not just scene statistics (Pacifici 2013) The generation of an Aerosol Optical Depth (AOD) map applies an atmospheric compensation for each pixel in scene for all imagery bands collected (Pacifici 2016) The AComp compensation product output is an atmospherically corrected image of surface reflectance values per pixel. AComp also compensates for water vapor in the atmospheric column. A single sample t test was used to compare the mean for each band of each atmospheric compensation procedure to the established value for each band from the ASD ground truth data (Manakos et al. 2009) determining if values collected for the WorldView 3 data (for each atmospheric compensation procedure performed) are significantly different from the ground truth collected. T his provided an objective measure of which procedure performed best in a tropical environment for use in the identification of P. macroloba and Z. longifolia After we established the most appropriate atmospheric compensation procedure, an additional appl ication of the single sample t test on all species collected allowed the best assessment of the ability of the corrected imagery to identify tropical tree species.

PAGE 52

37 Results and Discussion Image Data Extraction for P. macroloba and Z. longifolia We extracte d a subset of central c rown pixels from the WV 3 image to examine the integrated signature of the trees as compared to the ASD data acquired from the La Selva towers. A conservative manual extraction (only central canopy pixels, eliminating pixels near th e canopy edge) resulted in a central canopy subset comprising 13 and 18 total pixels for P. macroloba and Z. longifolia, respectively, providing a good representation of the characteristics of each tree c rown (Feret and Asner 2013; M. L. Clark, Roberts, and Clark 2005) Table 6 summarizes the subset of pixels, which includes the average reflectivity and standard deviation, and the ASD response for each tree c rown ASD Data Collection P. macroloba Canopy St atistics 13 WorldView 3 pixels represent the tree c rown WV 3 Bands ASD Data reflectance AComp Average AComp Std. Dev. FLAASH Average FLAASH Std. Dev. QUAC Average QUAC Std. Dev. Coastal 0.0145 0.0383 0.0066 0.1181 0.0124 0.0403 0.0010 Blue 0.0178 0.0226 0.0032 0.0502 0.0049 0.0433 0.0009 Green 0.0475 0.0478 0.0046 0.0690 0.0067 0.0520 0.0019 Yellow 0.0387 0.0342 0.0050 0.0457 0.0068 0.0424 0.0024 Red 0.0248 0.0204 0.0035 0.0269 0.0046 0.0410 0.0024 Red Edge 0.2004 0.2005 0.0223 0.2292 0.0264 0.1288 0 .0113 Near IR1 0.4012 0.4086 0.0477 0.4508 0.0543 0.3502 0.0359 Near IR2 0.4208 0.4156 0.0353 0.4475 0.0386 0.2903 0.0218 ASD Data Collection Z. longifolia Canopy Statistics 18 WorldView 3 pixels represent the tree crown WV 3 Bands ASD Data reflectanc e AComp Average AComp Std. Dev. FLAASH Average FLAASH Std. Dev. QUAC Average QUAC Std. Dev. Coastal 0.0158 0.0423 0.0068 0.1301 0.0118 0.0415 0.0009 Blue 0.0192 0.0295 0.0034 0.0635 0.0053 0.0463 0.0009 Green 0.0607 0.0606 0.0063 0.0887 0.0089 0.0582 0. 0025 Yellow 0.0446 0.0442 0.0047 0.0599 0.0061 0.0482 0.0022 Red 0.0254 0.0274 0.0036 0.0368 0.0045 0.0473 0.0024 Red Edge 0.2573 0.2530 0.0220 0.2908 0.0256 0.1543 0.0109 Near IR1 0.5220 0.5355 0.0502 0.5869 0.0540 0.4388 0.0357 Near IR2 0.5511 0.556 6 0.0380 0.5947 0.0399 0.3723 0.0226 Table 6 WorldView 3 imagery values subsets for P. ma croloba and Z. longifolia vs. ASD Data. Included are the number of pixels per tree c rown and the average and standard deviation values extracted for each WolrdView 3 band for each atmospheric correction procedure studied.

PAGE 53

38 Statistical Assessment of 3 Atmospheric Compensation Procedures The graphs below (Figures 5 and 6) show the comparison of the imagery, processed with QUAC, FLAASH and AComp to the ASD collected in situ for the same tree canopy and converted to equivalent WorldView 3 band locations ( Table 4 ). In each of the graphs for P. macroloba and Z. longifolia vertical bars showing the standard deviation provided a measure of reflectivity variability collected for the pixel gr ouping for each band. Figure 5 ASD Data vs. WorldVi ew 3 imagery with AComp (red), FLAASH (yellow) and QUAC (blue) atmospheric compensation for P. macroloba Vertical bars show the variation in reflectivity within the tree c rown The single horizontal bar shows the ASD derived band reflectivity for each W orldView 3 band location.

PAGE 54

39 The implementation of a single sample t test assisted in accessing the success of the atmospheric compensation process to the in situ ASD collected. Table 7 shows the results of this analysis. Setting a 95% significance level ensures a high degree of precision within the analysis. A p < 0.05 and a large t value suggest that significant differences exist between the ASD ground truth and the corrected reflectance measurement from WorldView 3, and therefore one can state the part icular measured value is not a good representation of the true Figure 6. ASD Data vs. WorldView 3 imagery with AComp (red), FLAASH (yellow) and QUAC (blue) atmospheric compensation for Z. longifolia Vertical bars show the variation in reflectivity within the tree c rown The single horizo ntal bar shows the ASD derived band reflectivity for each WorldView 3 band location.

PAGE 55

40 crown reflectance for the tree species studied. A high p value and low t value provided confidence that the atmospheric compensation procedure for a single band i n matching the ground truth da ta and ensured a determination of which atmospheric correction procedure was most likely to produce an accurate measure of the tree type in question. The high confidence level also reflects the precise nature of the imagery as the 11 bit radiometric resol ution lends itself to high degrees of precision when determining surface reflectance. Because of the high measurement repeatability (Hatchell 1999) when collecting ASD data, the high degree of averaging over a small acquisition period (500 acquisitions over the 2 minut e collection window), and consistent data collection protocols (minimizing collection P. macroloba Statistics = 0.05) AComp FLAASH QUAC WV 3 Bands t Sig (2 tailed) t Sig (2 tailed) t Sig (2 tailed) C oastal 13.109 0.000 30.115 0.000 97.743 0.000 Blue 5.333 0.000 23.857 0.000 107.015 0.000 Green 0.176 0.863 11.594 0.000 8.523 0.000 Yellow 3.229 0.007 3.715 0.003 5.449 0.000 Red 4.527 0.001 1.626 0.130 23.849 0.000 Red Edge 0.015 0.988 3.946 0.002 22.771 0.000 Near IR1 0.562 0.584 3.292 0.006 5.123 0.000 Near IR2 0.524 0.610 2.498 0.028 21.530 0.000 Z. longifolia Statistics = 0.05) AComp FLAASH QUAC WV 3 Bands t Sig (2 tailed) t Sig (2 tailed) t Sig (2 tailed) Coastal 16.439 0.000 41.214 0.000 123.143 0.000 Blue 12.637 0.000 35.744 0.000 123.306 0.000 Green 0.069 0.9 46 13.352 0.000 4.276 0.001 Yellow 0.391 0.701 10.647 0.000 7.110 0.000 Red 2.288 0.035 10.700 0.000 38.420 0.000 Red Edge 0.837 0.414 5.556 0.000 39.955 0.000 Near IR1 1.140 0.270 5.103 0.000 9.898 0.000 Near IR2 0.613 0.548 4.627 0.000 33.604 0.000 Table 7 The results of a Single Sample t test comparing the atmospheric correction output per band, per tree species, to the ASD Data ground truth. Values in bold are statistically significant.

PAGE 56

41 angle variance), it was anticipated there would be little variability in the ASD data (the ASD instrumentation does not report collection error). To confirm, we performe d a comparison between two consecutive ASD field collections for one P. macroloba c rown separated by ~5 minutes. The average difference in reflectivity for the total radiometric range collected (350 nm 2500 nm) between the two spectra was 0.960%. Becau se of the high reflectivity values in the near infrared (701 nm 1349 nm), the difference was larger as well, at 1.843%. The reflectivity was proportionally lower in the less reflective visible range (400 nm 700 nm) at 0.311%. Therefore, a single sampl e t test was appropriate for this analysis, using the ASD as the representative population value for comparison, and the WorldView 3 reflectivity measurements from the c rown as the tested sample. For the short wavelength Coastal and Blue bands, the WorldVi ew 3 imagery values do not match the ASD Data. Significance values of 0.000 and high t values for each atmospheric compensation procedure for both tree species studied support this observation. This is due to high scattering of shorter wavelengths in the EM spectrum due to water vapor (Lillesand, Kiefer, and Chipman 2008) The output of AComp compensation when applied to WorldView 3 imagery is shown to be a statistically accurate measure of the ASD Data collected when considering the Green, Yellow, Red, Red Edge, Near IR1, and Near IR2 bands (Table 6). The QUAC procedure performs slightly better in the visible spectru m as compared to FLAASH, but significantly under estimates reflectivity in the Near Infrared (Figures 5 and 6). FLAASH consistently over estimates reflectivity over the entire spectrum studied. The differences are more apparent in the red edge and inf rared bands, as the variations in reflective values are high (Table 6). Even though FLAASH produced

PAGE 57

42 statistically improved values for P. macroloba in the red edge and two Near IR bands, the values were nearly out of the acceptable bounds to match the ASD data, as FLAASH consistently overestimated the reflectivity for each tree species in those bands. Because FLAASH and QUAC differed significantly from the in situ ASD spectra in the infrared, there would be a significant impact in any analysis using tradit ional vegetation analysis tools (GVI, NDVI, etc.). For analysis requiring an absolute reflectivity measurement, such as in any contemporary classification schema (Del Frate et al. 2007) they would induce even larger errors. Even though the overall variance in the measured reflectance is high in longer wavelength WorldView 3 bands (Table 6), the unique characteristics of species within the forest canopy emerge (Cho et al. 2008) This uniqueness is evident when the imagery is corrected using the AComp procedure (Table 7). This strongly suggests that WorldView 3 imagery, with the utilization of the AComp atmospheric compensation process with the other appropriate corrections, is a viable option for identifying tree species in a complex tropical forest setting. Extension of the Statistical Assessment to all Species Collected With an established set of procedures for ASD Data processing and W orldView 3 image preparation, we extended the study to all species collected (Table 8) This processing included the ASD correction for look angle, illumination angle, white reference and bidirectional reflectance. In addition, a convolution to equivalen t WorldView 3 imagery bands completed the processing steps. Table 8 WorldView 3 average imagery reflectance values for ten tropical tree c rowns vs. ASD Data. Included are the number of pixels per tree c rown and the average and standard deviation values extracted for each WolrdView 3 band when using the ACOMP atmospheric compensation procedure.

PAGE 58

43 C. elastica ( Crown 4 pixels) I. marginata ( C rown 15 pixels) WV 3 Bands ASD Data reflectance AComp Average AComp Std. Dev. ASD Data reflectance AComp Average AComp Std. Dev. Coastal 0.0194 0.0430 0.0082 0.0117 0.0365 0.0052 Blue 0.0221 0.0222 0.0029 0.0147 0.0222 0.0046 Green 0.0472 0.0496 0.0041 0.0388 0.0431 0.0063 Yellow 0.0383 0.0300 0.0058 0.0293 0.0316 0.0039 Red 0.0268 0.0137 0.0061 0.0183 0.0204 0.0020 Red Edge 0.1925 0.1843 0.0 088 0.1887 0.1973 0.0262 Near IR1 0.3808 0.3863 0.0508 0.4416 0.4362 0.0781 Near IR2 0.3978 0.3986 0.0794 0.4593 0.4606 0.0717 L. procera ( C rown 3 pixels) O. floribunda ( C rown 8 pixels) WV 3 Bands ASD Data reflectance AComp Average AComp Std. Dev. ASD Data reflectance AComp Average AComp Std. Dev. Coastal 0.0199 0.0379 0.0045 0.0168 0.0377 0.0077 Blue 0.0224 0.0249 0.0027 0.0193 0.0230 0.0051 Green 0.0470 0.0480 0.0011 0.0510 0.0503 0.0111 Yellow 0.0372 0.0365 0.0056 0.0411 0.0360 0.0060 Red 0 .0245 0.0199 0.0015 0.0238 0.0202 0.0050 Red Edge 0.2029 0.2147 0.0092 0.2017 0.2114 0.0317 Near IR1 0.4323 0.4219 0.0356 0.4026 0.3978 0.0819 Near IR2 0.4471 0.4550 0.0176 0.4152 0.4224 0.0655 P. macroloba ( C rown 13 pixels) R. kunthiana ( C rown 5 pixels) WV 3 Bands ASD Data reflectance AComp Average AComp Std. Dev. ASD Data reflectance AComp Average AComp Std. Dev. Coastal 0.0145 0.0383 0.0066 0.0298 0.0343 0.0046 Blue 0.0178 0.0226 0.0032 0.0318 0.0238 0.0037 Green 0.0475 0.0478 0.0046 0.053 2 0.0485 0.0040 Yellow 0.0387 0.0342 0.0050 0.0464 0.0326 0.0043 Red 0.0248 0.0204 0.0035 0.0366 0.0182 0.0033 Red Edge 0.2004 0.2005 0.0223 0.1646 0.1544 0.0301 Near IR1 0.4012 0.4086 0.0477 0.3309 0.3393 0.0561 Near IR2 0.4208 0.4156 0.0353 0.3445 0 .3470 0.0345 S. microstachyum ( C rown 10 pixels) S. mombin ( C rown 7 pixels) WV 3 Bands ASD Data reflectance AComp Average AComp Std. Dev. ASD Data reflectance AComp Average AComp Std. Dev. Coastal 0.0427 0.0356 0.0054 0.0319 0.0321 0.0055 Blue 0.04 66 0.0205 0.0031 0.0423 0.0237 0.0051 Green 0.0698 0.0344 0.0035 0.0814 0.0440 0.0105 Yellow 0.0648 0.0239 0.0041 0.0726 0.0327 0.0059 Red 0.0526 0.0164 0.0049 0.0521 0.0198 0.0052 Red Edge 0.1804 0.1554 0.0299 0.2390 0.1966 0.0357 Near IR1 0.3349 0.3 341 0.0515 0.4487 0.4619 0.0861 Near IR2 0.3457 0.3486 0.0476 0.4479 0.4306 0.0829 W. regia ( C rown 1 pixel) Z. longifolia ( C rown 18 pixels) WV 3 Bands ASD Data reflectance AComp Average ASD Data reflectance AComp Average AComp Std. Dev. Coastal 0 .0415 0.0527 0.0158 0.0423 0.0068 Blue 0.0433 0.0285 0.0192 0.0295 0.0034 Green 0.0657 0.0603 0.0607 0.0606 0.0063 Yellow 0.0556 0.0426 0.0446 0.0442 0.0047 Red 0.0435 0.0346 0.0254 0.0274 0.0036 Red Edge 0.1887 0.1458 0.2573 0.2530 0.0220 Near IR1 0.3584 0.3519 0.5220 0.5355 0.0502 Near IR2 0.3700 0.2975 0.5511 0.5566 0.0380

PAGE 59

44 A descriptive statistical comparison was made of the processed ASD data to a manual extraction of WorldView 3 imagery pixels for each individual tree c rown studied. Th e manual extraction method matched the previous extraction process of P. macroloba and Z. longifolia (only central canopy pixels, eliminating pixels near the canopy edge). Depending on the size of the canopy collected, this resulted in varying pixel colle ctions for each tree c rown Table 8 summarizes the ASD data collection compared with the WorldView 3 imagery collection using AComp. Due to the results of the previous section, we only used the AComp corrected imagery for additional statistical tests. W. regia does not have a standard deviation computed for the WorldView 3 imagery because the overall c rown area is quite small for W. regia and the conservative pixel extraction resulted in only one pixel of imagery for the analysis. The results listed Ta ble 8 for all tree species supports the initial results using P. macroloba and Z. longifolia All WorldView 3 bands, excluding the Costal and Blue band, are statistically close to the ASD data collected. Humidity and particulates attenuate shorter wavele ngths, making the Costal and Blue band nearly unusable in the tropical environment (Lillesand, Kiefer, and Chipman 200 8) Figures 7 and 8 are the graphical representation of Table 8, showing the comparison of the imagery, processed with AComp, to the ASD collected in situ for the same tree c rowns and converted to equivalent WorldView 3 band locations. Error bars provi ded for each band a measure of reflectivity variability for each band within each tree species c rown No error bars were included for W. regia because of the identification of only one pixel of imagery for that particular species.

PAGE 60

45 Figure 7. ASD Data vs. WorldView 3 imagery for C. elastica, I. marginata, L. procera, O. floribunda and P. macroloba. Vertical error bars from the c rown mean show the variation in reflect ivity within the tree c rown The single black horizontal bars for each WorldView 3 band show the ASD data reflectivity measured for that specific band location.

PAGE 61

46 Figure 8. ASD Data vs. WorldView 3 imagery for R. Kunthiana, S. microstachyum, S. mombin W. regia and Z. longifolia Vertical error bars from the c rown mean show the variation in reflectivity within the tree c rown The single black horizontal bars for each WorldView 3 band show the ASD data reflectivity measured for that specific band loca tion. The data collection for W. regia does not have any error bars (only one pixel collected from the imagery).

PAGE 62

47 There is st rong agreement with the remaining six WorldView 3 bands between the ASD data and the imagery, with a closer match for tree c rowns with more pixels collected per canopy ( I. marginata, O. floribunda, P. macroloba, Z. longifolia ). Two exceptions to a success ful match across all species studied are S. microstachyum and S. mombin The strong difference in the visible portion of the spectrum (Green, Yellow, Red bands) between the ASD data and WorldView 3 imagery for these two species are seasonal in nature. Bo th S. microstachyum and S. mombin flower during the spring (Condit, Perez, and Daguerre 2011) which was during the time of the ASD data co llections in May of 2016. The ASD data clearly showed this characteristic, as the ASD reflectivity in the visible portion of the spectrum was significantly higher as compared to the WorldView 3 imagery values. The flowering condition of the tree during t he spring affected the Red Edge reflectivity, as the Red band value was elevating during the flowering stage, increasing the overall red edge reflectivity. The bands in the infrared portion of the spectrum (Near IR1, Near IR2) correlate quite well with t he ASD data, suggesting that overall tree structure and water content did not vary significantly between the flowering and non flowering plant stages. This result suggests that the near infrared bands of WorldView 3 could provide some consistent annual id entification of tree species, regardless of specific species reproductive cycles or other cycles throughout the year. To provide a more rigorous comparison between the ASD data and the WorldView 3 imagery, Table 9 below outlines the results of the single sample t test extended to all tree species in this study. W. regia was not included in the single sample t test because of the collection of only one data sample.

PAGE 63

48 The low significance values a nd high t values for the short wavelength Coastal and Blue bands for all trees matched the earlier results for P. macroloba and Z. longifolia The Green, Red Edge, Near IR1 and Near IR2 bands performed at a consistently with high significance values and l ow t values, demonstrating their importance in any tree species differentiation. The Yellow and Red band did not match as well as the other bands. This could be due to the very low reflectivity in the longer wavelengths of the visible portion of the spec trum, as compared to the Red Edge or near infrared bands, which comparatively can have a 10 fold higher reflectivity. The seasonality for S. microstachyum and S. mombin are C. elastica (df=3) I. marginata (df=14) L. procera (df=2) O. floribunda (df=7) P. macrolo ba (df=12) WV 3 Bands t Sig. t Sig. t Sig. t Sig. t Sig. Coastal 5.773 0.010 18.610 0.000 6.980 0.020 7.656 0.000 13.109 0.000 Blue 0.056 0.959 6.369 0.000 1.555 0.260 2.028 0.082 5.333 0.000 Green 1.167 0.328 2.681 0.018 1.543 0.263 0.176 0.866 0.176 0.863 Yellow 2.876 0.064 2.228 0.043 0.241 0.832 2.384 0.049 3.229 0.007 Red 4.282 0.023 4.256 0.001 5.353 0.033 2.045 0.080 4.527 0.001 Red Edge 1.889 0.155 1.281 0.221 2.222 0.156 0.867 0.414 0.015 0.988 Near IR1 0.219 0.841 0.267 0.7 94 0.508 0.662 0.165 0.874 0.562 0.584 Near IR2 0.020 0.985 0.070 0.945 0.776 0.519 0.308 0.767 0.524 0.610 R. kunthiana (df=4) S. microstachyu m (df=9) S. mombin (df=6) Z. longifolia (df=17) WV 3 Bands t Sig t Sig t Sig t Sig Coastal 2.231 0.089 4.190 0.002 0.058 0.956 16.439 0.000 Blue 4.899 0.008 26.598 0.000 9.580 0.000 12.637 0.000 Green 2.641 0.058 31.894 0.000 9.445 0.000 0.069 0.946 Yellow 7.194 0.002 31.851 0.000 17.995 0.000 0.391 0.701 Red 12.525 0.000 23. 180 0.000 16.470 0.000 2.288 0.035 Red Edge 0.759 0.490 2.641 0.027 3.136 0.020 0.837 0.414 Near IR1 0.335 0.754 0.045 0.965 0.406 0.699 1.140 0.270 Near IR2 0.164 0.878 0.193 0.852 0.553 0.600 0.613 0.548 Table 9. The results of a Single Sample t test for nine tree species, comparing the ASD data to the averaged WorldView 3 image c rown reflect ive response. Highlighted values are statistically significant. Significance Level for the test is 0.05. Significance values are 2 tailed, and df = degrees of freedom.

PAGE 64

49 evident through very high t values and low significance values in the visible port ion of the spectrum. The values in the Near IR1 and Near IR2 bands are significantly close showing that the plant structure did not change significantly when the tree was flowering was not significantly changed while flowering. The t test also showed tha t when more pixels were included in the c rown sample the ability of the imagery to match the ASD data went up significantly. The work of Cho et al. (2008) supports this observation, where unique characteristics in the canopy are significant in providing an reflectance measure for the tree species in question. This could indicate a limitation in the process as the tree c rown might have to be of a certain size to have enough WorldView 3 pixels for accurate species identification. Conclusions and Recommendations The research focus of this study was to determine if high spatial and spectral resolution imagery could be effective at dif ferentiating tree species within a complex tropical forest setting. WorldView 3 imagery, using the AComp atmospheric compensation procedure, and properly corrected for satellite look angle and bidirectional reflectance variations, has a demonstrated poten tial to measure individual tree c rown spectra at a level that should allow species mapping in a tropical forest setting. The correction with AComp and the other processing steps shows that the WorldView 3 data and ASD field spectral data are essentially eq uivalent within the limits of the broader and fewer spectral bands in the imagery. A key element of the data acquisition is integration of data (both types) across the central tree c rowns allowing the inclusion of specific and unique characteristics of t he tree structure into the differentiation process.

PAGE 65

50 The AComp atmospheric compensation process significantly outperformed QUAC and FLAASH processing in our study (Figures 5 and 6, Tables 7 and 9). In the coastal and blue satellite bands, significant at mospheric scattering compromised the quantification of surface reflectance for any of the three atmospheric compensation procedures. These species spectra appear to be separable in the imagery, which will facilitate any planned mapping studies in the regio n. In addition, the fully corrected ASD tree spectra would be a positive edition to any spectral library for tropical species. This study demonstrated that by having a strict data processing protocol for both ground truth data and imagery (corrections fo r sun and look orientation), and a proper atmospheric compensation, one can successfully differentiate trees species at the c rown level in a complex tropical landscape with WorldView 3 imagery. There is strong agreement between the WorldView 3 bands and t he convolved ASD data from the spectral range Green to Near IR2, suggesting the utilization of the a single tree canopy pixel average, as compared to individual pixels with a tree canopy, is more effective for discriminating individual tree species. Given the difference in average c rown reflectivity between the ten tree species studied (and the small variance in average c rown surface reflectance values per band), it is likely that several additional tree species could be discriminated using WorldView 3 ima gery. Delving into this promising imagery application requires more in situ data collected for multiple tree species. The identification of many tree species within a complex forest setting would be a significant step forward in the study of not only tro pical forests, but also varied forest ecosystems across the landscape. Developing a process to successfully perform individual tree taxonomy identification would be of great benefit to a variety of applications (Campbell, James and Wayne 2011)

PAGE 66

51 including the r ealization of an improved biomass estimate, forest resource assessments, forest response to climate change, and forest maturity. Studies by van Leeuwen & Nieuwenhuis (2010) also identified the high potential of LiDAR combined with optical remote sensing imagery, to contribute to such an as sessment to advance the accuracy of biomass models within forests stands. An improved schema could be developed from the research presented in this study for identification of tropical tree species, building on work of earlier studies (Feret and Asner 2013; Zhang et al. 2006; M. L. Clark, Roberts, and Clark 2005) as the basis for an improved biomass estimate for the tropical forest.

PAGE 67

52 References Field Spectroscopy Facility, Natural Environment Research Group TM 2002): 1 136. Bernstein, Lawrence S., Xuemin Jin, Brian Gregor, and Steven M. Adler Golden. 2012. Optical Engineering 51 (11): 1 11. doi:10.1117/1.OE.51.11.111719. Bousquet, La BRDF Measurements and Model for Specular and Diffuse Components Remote Sensing of Environment 98 (2 3): 201 211. doi:10.1016/j.rse.2005.07.005. Breunig, Fbio M., Lnio Soares Galvo, Joo Roberto dos Santos, Anatoly A. Gitelson, Anisotropy of Subtropical Deciduous Forest Using MISR and MODIS Data Acquired under Large Seasonal Variation International Journal of Applied Earth Observation and Geoinformation 35. Elsevier B.V.: 294 304. doi:10.1016/j.jag.2014.09.017. Breunig, Fabio M., Lenio Soares Galvao, Yhasmin Mendes Moura, and Rafaelo Balbinout. Results of the BRF Dependence of a Subtropical Semideciduous Anais XVI Simposio Brasileiro de Sensoriamento Remoto SBSR, Foz Do Iguau, PR, Brasil, 13 a 18 de Abril de 2013, INPE no. 1986: 6917 6922. Campbell James, B., and Randolph H. Wayne. 2011. Introduction to Remote Sensing 5th ed. New York, NY: The Guilford Press. Photogrammetric Engine ering Remote Sensing 70 (1): 135 140. doi:10.14358/PERS.70.1.135. Organization for Tropical Studies http://sura.ots.ac.cr/florula4/index.php. Cho, M.A., I. Sobhan, A.K. Skidmore, and J. de Leeuw. The International Archives of Photogrammetry, Remote Sensing and Spatial Information Science 37 (Part B7): 1 8. umber of Known Plants Phytotaxa 261 (3): 201 217. doi:10.11646/phytotaxa.261.3.1. Cifuentes Jara, M, Miguel Morales, and D Henry. 2013. Inventory of Volume and Biomass Tree Allometric Equations for Central and South America UN REDD MRV Report 11, CATIE, Turialba, Costa Rica, Food & Agriculture Organization of the United Nations

PAGE 68

53 Discrimination of Tropical Rain Forest Tree Species at Le Remote Sensing of Environment 96 (3 4): 375 398. doi:10.1016/j.rse.2005.03.009. Clark, Roger N, G A Swayze, K Eric Livo, Raymond F Kokaly, Trude V V King, J Brad Dalton, J Sam Vance, Barnaby W Rockwell, Todd Hoefen, and Robert R Mcdoug al. USGS Spectroscopy Lab 1 21. Classification of Hyperspectral D International Journal of Remote Sensing 21 (10): 2075 2087. doi:10.1080/01431160050021303. Forest Ecology and Management 229 (1 3): 351 360. doi: 10.1016/j.foreco.2006.04.017. Condit, Richard, Rolando Perez, and Nefertaris Daguerre. 2011. Trees of Panama and Costa Rica Princeton University Press. Neural Networks for Automatic Classification from High IEEE Transactions on Geoscience and Remote Sensing 45 (4): 800 809. doi:10.1109/TGRS.2007.892009. DigitalGlobe, DS WV3 09/14 2. Dorigo, Wouter, Marti 31. ITT Visual Information Solutions Version 4. (Copyright ITT Visual Information Solutions): 44. Felde, G.W., G.P. Anderson, T.W. Cooley, M.W. Matthew, S.M. adler Golden, a. Berk, and IGARSS 2003. 2003 IEEE International Geoscience and Remote Sensing Symposium. Proceedings (IEEE Cat. No.03CH37477) 1 (C): 90 92. doi:10.1109/IGARSS.2003.1293688. Feret, Jean IEEE Transactions on Geoscience and Remote Sensing 51 (1): 73 84. In J.A. Vozzo, Agricultural Handbook, USDA Forest Service 721: 601 604. Gastellu Etchegorry, J.P., P. Guillevic, F. Zagolksi, V. Demarez, V. Trichon, D. Deering, and and Radiation Regime of Tropical and Boreal Forests, Remote Sensing of Environment 68 (July 1998): 281. Analytial Spectral Devices, Inc. (ASD) no. 4th Ed.: 144.

PAGE 69

54 g Tree Species in a Northern Mixed Forest Using High Journal of Forest Research 9 (1): 7 14. doi:10.1007/s10310 003 0045 z. 3 Imagery Technical Note 1 WorldView 3 Instrument W orldView 12. Cover Classification in a Moist Tropical Region of Brazil with Landsat Thematic Mapper International Journal of Rem ote Sensing 32 (23): 8207 8230. doi:10.1080/01431161.2010.532831. Based Feature Extraction for Improved Endmember Abundance IEEE Transactions on Geoscience and Remote Sensin g 42 (3): 644 649. doi:10.1109/TGRS.2003.822750. Lillesand, Thomas M., Ralph W. Kiefer, and Jonathan W. Chipman. 2008. Remote Sensing and Image Interpretation 6th Editio. John Wiley and Sons Ltd. Manakos, Ioannis, Kiril Manevski, Chariton Kalaitzidis, and Dennis Edler. 2009. Comparison Between FLAASH and ATCOR Atmospheric Correction Modules on the Baisis of WorldView 2 Imagery and in Situ Spectroradiometric Measurements Mediterranean Agronomic Institute of Chania Marcello, Javier, Francisco Eugenio, Ulis es Perdomo, and Anabella Medina. 2016. Sensors (Switzerland) 16 (10): 1 18. doi:10.3390/s16101624. Martnez, Pablo J., Ros a M. Prez, Antonio Plaza, Pedro L. Aguilar, Mara C. Cantero, and Annals of Geophysics 49 (1): 93 101. doi:10.4401/ag 3156. Matthew, M. W., S. M. Adler Golden, A. Berk, G. Fe lde, G. P. Anderson, D. Gorodetzky, S. Proceedings Applied Imagery Pattern Recognition Workshop 2002 Janua: 157 163. doi: 10.1109/AIPR.2002.1182270. Journal of Geophysical Research Atmospheres 97 (D17): 18935 18946. doi:10.1029/92JD00879. Milton, E J, and K Choi. 2 Proceedings of the Annual Conference of the Remote Sensing and Photogrammetry Society 1 11. Calibration Methods f Sensors 9 (2): 794 813. doi:10.3390/s90200794.

PAGE 70

55 Moura, Yhasmin Mendes, Lnio Soares Galvo, Joo Roberto dos Santos, Dar A. Roberts, and I nter Remote Sensing of Environment 127. Elsevier Inc.: 260 270. doi:10.1016/j.rse.2012.09.013. Detection Approa Photogrammetric Engineering and Remote Sensing 80 (10): 983 993. doi:10.14358/PERS.80.10.983. Model for Mapping Relative Differences in Belowground Biomass and Root:Shoot Ratios Using Spectral Reflectance, Foliar N and Remote Sensing 7: 16480 16503. doi:10.3390/rs71215837. tomatic Atmospheric Compensation Algorithm for Very High Proc. JACIE 1 43. In Geoscience and Remote Sen sing Symposi um(IGARSS) 2016 IEEE International 1973 75. Physical Quantities for the Analysis of Multitemporal and Multiangular Optical Very I EEE Transactions on Geoscience and Remote Sensing 52 (10): 6241 6256. doi:10.1109/TGRS.2013.2295819. IKONOS and WorldView Re mote Sensing of Environment 124. Elsevier Inc.: 516 533. doi:10.1016/j.rse.2012.06.011. Post Processing Spectral Data in MATLAB Roy, D. P., H. K. Zhang, J. Ju, J. L. Gomez Dans, P. E. Lewis, C. B. Schaaf, Q. Sun, J. Li, H. Remote Sensing of Environment 176: 255 271. doi: 10.1016/j.rse.2016.01.023. STRI Website www.si.edu. UN from Deforestation and Forest Degradation UN REDD Website www.un redd.org/. van Aardt, Jan a N, Randolph H Wynne, Richard G Oderwald, and James B Campbell. 2001. Photogrammetric Engineering and Remote Sensing 67 (1 2): 1 184.

PAGE 71

56 European Journal of Forest Research 129 (4): 749 770. doi:10.1007/s10342 010 0381 4. Weyermann, Jrg, Alexander Damm, Ma thias Kneubhle r, and Michael E Schaepman. 2014 IEEE Transactions on Geoscience and Remote Sensing 52 ( 1 ): 616 627 Whitmore, T. C., R. Peral Journal of Tropical Ecology 1 (4): 375 378. Spatial Resolution Remotely Sensed D Bioscience 54 (6): 511 521. doi:10.1641/0006 3568(2004)054[0511:HSRRSD]2.0.CO;2. Zhang, Jinkai, Benoit Rivard, Arturo Snchez Azofeifa, and Karen Castro Esau. 2006. and Inter Class Spectral Variability of Tropic al Tree Species at La Selva, Costa Remote Sensing of Environment 105 (2): 129 141. doi:10.1016/j.rse.2006.06.010.

PAGE 72

57 CHAPTER III DETERMINING EFFECTIV E METER SCALE IMAGE DATA AN D SPECTRAL VEGETATION INDICIES FOR TROPICAL FOREST TREE SPECIES DIFFERENTIATION Abstract In this study, we evaluate several processing schemas using field collected spectral data that would be applicable to WorldView 3 satellite imagery for identifying and mapping tropical tree species. The study site, La Selva Biological Preserve in Costa Rica, provides the infrastructure to facilitate collection of proper ground truth data for image validation. An objective statistical analysis demonstrated that the World View 3 imagery, after applying a series of spectral and illumination image corrections, was able to accurately identify selected tree species This paper defines the image bands and image derived spectral vegetation indices that are the most effective for tree species discrimination. We show that corrected absolute reflectivity values from the Green, Red, Red Edge, and Near Infrared are good differentiators of tropical tree species c rowns We also evaluate 14 possible multi band vegetation indices and sh ow that two new indices developed here, using WorldView 3 image bands, have the highest discriminatory power for tropical tree species. A combination of both individual band information and vegetation indexes would significantly improve image based classi fication of tropical forests. Introduction Identifying and developing effective analytical tools for accurate tree species classification is an important step in studies of forest composition, biomass, and overall habitat. Specifically, an accurate iden tification of tree species would support a more accurate

PAGE 73

58 biomass measure (Wulder et al. 2004) leading to better estimates of carbon uptake from tropical forests and possible improvements in car bon offset protocols (Cifuentes Jara and Henry 2014) At present, large uncertainties exist in overall species composition of tropical forests (Barbosa, Broadbent, and Bitencourt 2014) significantly impacting any improvement in an overall biomas s measure. Here, we aim to establish an effective set of remotely sensed image derived data products for use in a classification schema for tropical forest species. This approach utilizes high resolution multi spectral imagery and employs a series of st atistical procedures to determine what bands and combinations of bands (within established and new vegetation indices) are most effective in differentiating tropical tree species. Viewing and illumination corrections for the field spectrometer data and Wo rldView 3 imagery, and an atmospheric compensation procedure for the imagery were essential for accurate assessments of the primary site for this study because it facilitated easy access to a dense tropical forest setting, as well as infrastructure to support all field data collection efforts. It is a premier site for tropical lowland Atlantic forest studies in Central America. There is a long, extensive history of research on the application of remote sensing imagery to vegetation analysis. Various imaging systems have been used, including lower resolution satellites such as the Advanced Very High Resolution Radiometer (AVHRR) and the Moderate Resolution Imagin g Spectroradiometer (MODIS), the Landsat legacy series (G. Li et al. 2011) Landsat 8 (Roy et al. 2016) and contemporary commercial high resolution imaging systems (Katoh 2004; Wulder et al. 2004) Regarding tree i maging and analysis, several studies have attempted to use imagery for tree type identification in a variety of

PAGE 74

59 locations (Carleer and Wolff 2004) including complex forest assemblages such as tropical fores ts (G. Li et al. 2011) Applications of remote sensing for use in a complex forest assemblage, such as a tropical forest environment, remains a major challenge. Several studies have attempted to use more advanced imaging systems (Zhang et al. 2006; M. L. Clark, Roberts, and Clark 2005) and innovative analysis techniques, such as incorporating the tree intra crown variations (Myint et al. 2014; J. Li 2004) Limitations in spatial and radiometric resolution, overall data availability, and appropriate data corrections have constrained the above studies. As an exampl e of data availability challenges and overall data collection, two studies listed here (Zhang et al. 2006; M. L. Clark, Roberts, and Clark 2005) used the exact same hyperspectral imaging system collection (HY DICE) for each study. More recently, al. 2010) used a spaceborne hyperspectral sensor (Hyperion) in an attempt to characterize tree species in southeastern Peru. Difficulties existed in characterizing the tr ee canopy as the Hyperion sensor spatial resolution (30m) is coarse In addition to using the sensor band information directly, the use of spectral vegetation indices (hereafter, SVIs) has been well documented for determining vegetation type, chlorophyll production, gross primary productivity, leaf are a index and biomass estimates (Gitelson et al. 2008; W. J. D. van Leeuwen and Orr 2006; Gao et al. 2000; Price 1993) Ove r the last 30 years, the use of spectral vegetation indices within remote sensing has become one of the most common approaches to understanding various biophysical factors related to vegetation (Xue and Su 2017) Researchers in (Foody, Boyd, and Cutler 2003) used vegetation indices in their study of the potential of using a common prediction schema in determining biomass in three distinct tropical regions (Brazil, Malaysia, and

PAGE 75

60 Thailand). Their pre dictive model incorporated ten vegetation indices that were highly correlated with tropical biomass. The results showed that the input of the indices assisted in the calculation of biomass within their three distinctive areas of study, but the use of the common schema between the sites was problematic, suggesting that site specific biomass calculations are currently more appropriate until better data and processes are available (Foody, Boyd, and Cutler 2003) Accurate measures of these biophysical parameters can be difficult in dense vegetation cover, as most traditional vegetation indices saturate, leaving little measurement variation (Ves covo et al. 2012) Some attempts have been made to minimize the saturation through scaling the original indices (Gitelson 2004) or the addition of other bands to help differentiate the signal and minimize atmospheric effects (Vescovo et al. 2012; Gitelso n, Kaufman, and Merzlyak 1996) Recent studies have shown that the red edge spectral region (680 740 nm) provides significant information regarding the chlorophyll production and tree structure, and appears to saturate less when imaging dense forest canop ies (Vescovo et al. 2012) This red edge reflectivity effectively gives a measure of the variance between the near infrared reflectance peak and a minimum of reflectivity in the red (Campbell, James and Wayne 2011) Changes in the red edge brightness ind icates changes in vegetation health/stress, seasonality and senescence, water content, leaf area, and biomass (Vescovo et al. 2012) The addition of the red edge band in any vegetation analysis schema would likely enhance its effectiveness for measuring vegetation and understanding its variability. There has been significant work on establishing an objective assessment of importance regarding the information acquired and derived from remote sensing imagery using various statistical tests. Researchers in (Vaiphasa et al. 2005) used a one way Analysis

PAGE 76

61 of Variance (ANOVA) to assess field spectrometer collections (FieldSpec Pro FR) of 16 ma ngrove species in Thailand. The results showed that there was a statistically significant association between vegetation type and reflectivity in the red edge and near infrared portions of the spectrum. In a later study, (Vaiphasa et al. 2007) employed a t test to select key wavelengths for discrimination of the same 16 species. Band locations in the red edge and near infrared had the highest discriminatory power. Several studies have used discriminant analysis to evaluate which variables have the highest discriminatory power regarding the differentiation of specific classes of information. A discriminant analysis procedure is associat ed with a classification of the input data to determine importance or relevance. Discriminant analysis has been used several times in remote sensing research for purposes of prediction of analyzed groups (Immitzer, Atzberger, and Kouk This can provide an initial assessment of group membership from the independent samples and insight into the importance of each independent variable. In addition, discriminant analysis, when performed for purposes of classification, can provide a first look into the predictive power of the independent variables. This information can assist in setting the initial conditions for standard image classification procedures. t indicates the discriminatory power of the model used, as well as the individual measure of the variables associated with the model al. 2010) o reduce the dimensionality of complex data sets by identifying a subset of variables that are most important for discriminating defined classes (Immitzer, Atzberger, and Koukal 2012) In a study by (Thenkabail et al. 2004) a discriminant analysis approach wa s used to assess the

PAGE 77

62 ability of imagery (Hyperion) for defining Land Use and Land cover. The results of the 9 bands were needed to discriminate between pristine and degraded fores t, and 23 bands from the Hyperion hyperspectral sensor were required to achieve good separability for all rainforest land cover types identified in the study area (Thenkabail et a l. 2004) More recently, a study by Immitzer, Atzberger, and Koukal ( 2012) used t WorldView 2 bands to verify the most important bands in classifying 10 tree species in a temperate forest in Austria. Four bands (Green, Red Edge, Near Infrared 1 and Near Infrared 2) had the greatest discriminatory powe r in differentiating tree species In this study, we present an improved approach, building on the efforts and analysis described above to objectively assess data inputs into a classification schema The goal is to define the best multi spectral data t o achieve a more accurate and consistent tree species identification within a complex tropical forest canopy If improvements can be made in in the identification of tree species, improvements in biomass estimates could be realized (UN REDD 2015; Cifuentes Jara, Morales, and Henry 2013) leading to improved estimates of stored carbon within the biosphere (Rosenqvist et al. 2003) Overall Approach The present study has two objectives: (1) to use imagery and ground truth data to validate the ability of the imagery to identify tree species in a tropical forest environment; and (2) to identify the multi spectral imagery bands and vegetation indices that are most effective for tree differentiation. We acquire spectra (both field spectrometer and satellite pixel clusters) that average full single tree c rown and its within c rown variations (M. L. Clark, Roberts, and Clark 2005) C rown field spectrometer data for ten major tree species of

PAGE 78

63 the Costa Rican tropical lowland forest were processed using a s eries of acquisition corrections (white reference, attenuation, viewing geometry), and convolved with band spectral response functions to match WorldView 3 image bands. The species selected for the study were constrained by their accessibility from observ ation towers at La Selva, as the goal was to acquire data from a position that was as near to nadir as possible. We then compared the field gathered spectra with image pixels acquired by WorldView 3 for the tree canopies under similar conditions (dry sea son, off nadir) after applying the AComp atmospheric compensation procedure (Pacifici 2013) A comparison of the ASD convolved spectra (ground tru th) to the WorldView 3 imagery, through a single sample t test, established a measure of confidence regarding how well the WorldView 3 imagery sample fit the field spectrometer range of data for a particular species. Once the WorldView 3 imagery, with the proper corrections, was shown to be capable of correctly identifying tree species, both WorldView 3 image bands and 14 select vegetation indices (8 traditional, 6 modified or created specifically for this study) were statistically analyzed using a discrim inant analysis The application of SVIs for determining various measures of vegetation cover is a well established remote sensing process. SVIs provide approximations of various vegetation parameters used in both scientific analysis and assessments of fo rest, grassland, or agricultural characteristics (Campbell, James and Wayne 2011) These include measures of overall greenness, biomass, water content, or health of a specific tree or plant types, and seasonal variations (Xue and Su 2017; Asner 1998)

PAGE 79

64 ASD Data and WorldView 3 Image Processing ASD Data Collection and Pro cessing We collected spectral reflectance values from the central c rowns of ten tree species in May 2016 (end of the dry season) accessible from established canopy observati on towers (Figure 1) Two canopy towers facilitated a near nadir perspective for collect ion An Analytical Spectral Devices (ASD) FieldSpec Pro Spectroradiometer collected reflected energy in 220 spectral bands spanning 300 2500 nm, with a sampling interval of 1.4 nm (350 1000 nm) and 2 nm (1000 2500nm). An 18 fore optic f ield of view port attached to the detector element masked crown edges and mixed tree areas while facilitating mean central c rowns averaging during data acquisitions. We applied an ASD dynamic range optimization that maximized the radiometric precision of the acquired data (based on the 16 Figure 1. One of two canopy towers used for data collection

PAGE 80

65 bit digitization for the ASD Fieldspec Pro instrument) and prevented saturation. The acquisition of a dark current value before each collection session eliminated instrument input into the final signal. A 500 sample spe ctrum average was taken during each acquisition to minimize variability in reflectivity and ensure consistent data collection (ASD 2002) This high resolution collection capability allows for very precise reflective measurements in the visi ble near infrared range of the spectrum (Hatchell 1999) Table 1 below outlines the tree species collected with the ASD instrument. Species Acquired Within La Selva Biological Station Tree Specie s Family Common Name Range Characteristics Castilla elastica Moraceae Panama Rubber Tree Northern S. America, Central America Source of latex Inga marginata Fabaceae Guabilla Northern S. America, Central America Edible fruit, leaves are used as an astringent Laetia procera Salicaceae Manga larga Central America Rare, restricted to old forests Ocotea floribunda Lauraceae Laurel espada South and Central America Wood used in fine carpentry, furniture Pentaclethra macroloba Fabaceae Pracaxi, Kuntze N orthern S. America, Central America Seeds a source of cooking oil, bark for antiseptic Rhodostemonodaphne kunthiana Lauraceae Quizarra Negro, Sweetwood Northern S. America to Costa Rica Used in construction Stryphnodendron microstachyum Fabaceae Cacha, Y ellow Targuayugo South and Central America Bark used as an astringent, and for ink Spondias mombin Anacardiaceae Yellow Mombin, Hogplum South and Central America Widely cultivated for its edible fruit Welfia regia Arecaceae Amargo Palm South and Central America Edible leaves, palm heart, used in construction Zygia longifolia Fabaceae Sotacaballo, Chiparo Central America Used for erosion control, esp. banks of rivers Table 1 Ten tree species collected for this study in May 2016 from La Selva Biological Station, Costa Rica, from (Castro 2014; STRI 2014; Condit, Perez, and Daguerre 2011)

PAGE 81

66 In addition, a Canopy Average reflectance ASD collection provided a baseline measureme nt, spanning many species (approximately 40 by 60 meters). This process consisted of panning the ASD fore optic over the canopy (not focusing specifically on any one tree) while maintaining a side scatter illumination condition for data consistency and ea se of correction. This provided a measure the distinctiveness of a particular tree species within the mean forest canopy from the overall the forest canopy. The canopy average data, collected at the time of the specific ASD data acquisitions for individu al tree canopies, is a reference dataset for this study. An average canopy measure acquired from other sources, such as a sample collection from the WorldView 3 imagery or averaging many known species reflectivity values from different band locations, wou ld be equally sufficient. The timing of the collections (within 2 hours of local noon) minimized late day cloud cover and solar zenith angle, facilitating collections similar to those of a typical satellite data acquisition (Table 2). Tree Species Acq Date Acq. Time OZA AZ Target Dist. Sky Cond. C. elastica 5/10/16 11:55am 30 180 12m Part. Cldy. I. marginata 5/10/16 12:08pm 25 90 14m Part. Cldy. L. procera 5/11/16 11:40am 40 90 15m Clear O. floribunda 5/11/16 11:25am 45 225 22m Clear P. macroloba 5/11/16 11:50am 40 200 16m Clear R. kunthiana 5/10/16 12:25pm 20 320 10m Clear S. microstachyum 5/10/16 2:27pm 40 220 11m Clear S. mombin 5/10/16 12:15pm 45 20 11m Clear W. regia 5/11/16 12:10pm 15 90 6 m Clear Z. longifolia 5/10/16 12:45pm 30 250 12m Clear Canopy Average 5/10/16 11:00am 40 ~45 >30m Clea r Table 2. ASD data collection parameters during acquisiti on in May 2016.

PAGE 82

67 A white reference (Spectralon) collection, followed by two tree collections, and then a second white reference provided data for correcting variations in illumination due to cloud attenuation. Data collection distance ranged from 6m to 22m. With the 18 fore optic, the acquisition spot on the target canopies ranged between 1.8 and 6.8 m. Equation 1 defines the absolute reflectance calculation: where R abs is the absolute reflectance of the target in question, I sample is the radiant intensity measured for the sample (tree), and I ref is the radiant intensity from a Spectralon white reference panel under identical lighting conditions. A white reference panel correction was included to correct the tree spectra to absolute reflectance (Robinson and MacArthur 2011; Miura and Huete 2009; M. L. Clark, Roberts, and Clark 2005; R. N. Clark et al. 1999) As tree canopies are non Lambertian a Bidire ctional Reflectance Factor (BRF) correction was required to convert data to an on nadir view (Moura et al. 2012; Bousquet et al. 2005; Middleton 1992) Variability in reflectance can be quite high (up to 30%) depending on view and illumination geometry (Weyermann et al. 2014 ) The process converts the data to a normalized on nadir reflectance (Roy et al. 2016; Gastellu Etchegorry et al. 1999) BRF correction values for different wavelength regions derived from (Breunig et al. 2013) for a mixed tree canopy in n orthern Brazil (similar to the La Selva forest canopy) and supporting information from (Gastellu Etchegorry et al. 1999) were used for this analysis (Table 3). The application of this correction to the ASD data provided a corrected data set that represented the on nadir reflectivity for each tree s pecies studied.

PAGE 83

68 Table 3. BRF Correction factors for ASD data collections. Values represent both backscatter and sidescatter conditions. The Canopy Average BRF correction was the same as for P. macroloba (40 OZA Sidescatter). ASD Spectrometer Data Collection Spectrum Location Range (nm) C. elastica 30 OZA Backscatter I. marginata 25 OZA Backscatter L. procera 40 OZA Backscatter O. floribunda 45 OZA Backscatter P. macroloba 40 OZA Sidescatter Blue 350 499 0.8784 0.8701 0.9144 0.9324 0.9491 Green 500 628 0.9785 0.9 910 0.9473 0.9317 0.8135 Red 629 700 1.0543 1.0538 1.0495 1.0471 0.8981 Near IR 701 2500 0.9991 1.0134 0.9607 0.9415 0.8093 Spectrum Location Range (nm) R. kunthiana 20 OZA Forescatter S. microstachyum 40 OZA Sidescatter S. mombin 45 OZA Sides catter W. regia 15 OZA Backscatter Z. longifolia 30 OZA Sidescatter Blue 350 499 0.9161 0.9491 0.9702 0.9220 0.9070 Green 500 628 0.8793 0.8135 0.8031 0.9946 0.8344 Red 629 700 0.9161 0.8981 0.9008 1.0323 0.8926 Near IR 701 2500 0.8896 0.8093 0.7926 1.0081 0.8428 Figure 2. ASD Data values for the ten tree species and the C rown Average convolved to WorldView 3 imagery band values. Lines between points are for clarity to show the trend in variability between bands. Figure adapted from Cross e t al., 2018 in review.

PAGE 84

69 The ASD spectra for each tree species and c rown average were convolved to the WorldView 3 band detected spectrum as a function of wavelength for each of the 8 WorldView 3 spectral bands (Cross et al. 2018; Milton and Choi 2004) The spectral response curve values for the panchromatic, visible and near infrared bands provided by DigitalGlobe (Kuester 2016) allowed the conversion of the data via a program developed in MATLAB. This conversion (Figure 2) facilitated a direct comparison of the ASD data to the WorldView 3 imagery. The dataset for the ten tree s pecies is unique in that it represents the full canopy reflective information for a particular tree species. Because of the processing schema followed and the incorporation of BRF correction, the derived data represents a unique dataset for tree species v erification. Unfortunately, because of the tower locations within La Selva, other tree species were not accessable on a near nadir basis, and the species collected were limited in their number and spatial scope. WorldView 3 Imagery Acquisition and Prepa ration The imagery used was a nearly cloud free, dry season DigitalGlobe An image acquisition during the ASD field collection would have been preferred, but that option was not available, and t he image selected was a best fit based on the criteria defined above. DigitalGlobe provided the imagery with the AComp atmospheric compensation procedure already applied. AComp is a proprietary atmospheric compensation process developed by DigitalGlobe f or use with their WorldView sensor series imagery (Pacifici 2013) The compensation uses an Aerosol Optical Depth (AOD) map is generated which app lies an atmospheric compensation for each pixel in scene for all imagery bands collected (Pacifici 2016) providing surfa ce reflectance.

PAGE 85

70 A B RF correction normalized the imagery to an on nadir view condition for direct comparison with the convolved ASD data (Table 5) This is especially true for variable view a ngle high spatial resolution imagery (Pacifici, Longbotham, and Emery 2014) This required specifi c correction values for the imagery due the specific illumination angle and backscatter view condition during image acquisition (Breunig et al. 2013) Bands Spectral Range ( nm ) Resolution Panchromatic 450 800 Panchromatic 0.31m Costal 400 450 Multispectral 1.24m Blue 450 510 Dynamic Range 11 bits/pixel Green 510 580 Specific Image Information for Study Yellow 585 625 Da te/time 11/11/2014, 15:52:28Z Red 630 690 Zenith View Angle 26.2 Red Edge 705 745 Azimuth View Angle 108.8 Near IR1 770 895 Cloud Cover 0.5% Near IR2 860 1040 Data Extent NW Corner 10.48 N, 84.14 W SE Corner 10.25 N, 83.99 W WorldView 3 Imagery Collection Backscatter Condition Band Locations Wavelength Range (nm) BRF Factor 26.2 OZA Coastal 400 450 1.01 Blue 450 5 10 1.08 Green 510 580 1.15 Yellow 585 625 1.17 Red 630 690 1.20 Red Edge 705 745 1.25 Near IR1 770 895 1.33 Near IR2 860 1040 1.34 Table 4. WorldView 3 imagery s pecifications (DigitalGlobe 2014) and specific image information extracted from the metadata file from the imagery used in this study. The image used in the study was nearly clo ud free (0.5%). Table 5. BRF Correction factors for the WorldView 3 Imagery. Values were derived from (Breunig et al. 2013) and represent backscatter viewing conditions for the imagery. Dividing the observed reflectance by the BRF correction factors above derives the on nadir surfac e reflectance.

PAGE 86

71 Table 6. ASD data and WorldView 3 imagery values for ten tropical trees. Included are the number of pixels per tree c rown (in parentheses) and the average and standard deviation values extracted for each WolrdView 3 band. The Canopy Ave rage response is also included at the end of the table. C. elastica (4) I. marginata (15) WV 3 Bands ASD Data reflectance Imagery Average Imagery Std. Dev. ASD Data reflectance Imagery Average Imagery Std. Dev. Coastal 0.0194 0.0430 0.0082 0.0117 0.0365 0.0052 Blue 0.0221 0.0222 0.0029 0.0147 0.0222 0.0046 Green 0.0472 0.0496 0.0041 0.0388 0.0431 0.0063 Yellow 0.0383 0.0300 0.0058 0.0293 0.0316 0.00 39 Red 0.0268 0.0137 0.0061 0.0183 0.0204 0.0020 Red Edge 0.1925 0.1843 0.0088 0.1887 0.1973 0.0262 Near IR1 0.3808 0.3863 0.0508 0.4416 0.4362 0.0781 Near IR2 0.3978 0.3986 0.0794 0.4593 0.4606 0.0717 L. procera (3) O. floribunda (8) WV 3 Bands ASD Data reflectance Imagery Average Imagery Std. Dev. ASD Data reflectance Imagery Average Imagery Std. Dev. Coastal 0.0199 0.0379 0.0045 0.0168 0.0377 0.0077 Blue 0.0224 0.0249 0.0027 0.0193 0.0230 0.0051 Green 0.0470 0.0480 0.0011 0.0510 0.0503 0.0111 Yellow 0.0372 0.0365 0.0056 0.0411 0.0360 0.0060 Red 0.0245 0.0199 0.0015 0.0238 0.0202 0.0050 Red Edge 0.2029 0.2147 0.0092 0.2017 0.2114 0.0317 Near IR1 0.4323 0.4219 0.0356 0.4026 0.3978 0.0819 Near IR2 0.4471 0.4550 0.0176 0.4152 0.4224 0.0655 P. macroloba (13) R. kunthiana (5) WV 3 Bands ASD Data reflectance Imagery Average Imagery Std. Dev. ASD Data reflectance Imagery Average Imagery Std. Dev. Coastal 0.0145 0.0383 0.0066 0.0298 0.0343 0.0046 Blue 0.0178 0.0226 0.0032 0.0318 0.0238 0.0037 Green 0.0475 0.0478 0.0046 0.0532 0.0485 0.0040 Yellow 0.0387 0.0342 0.0050 0.0464 0.0326 0.0043 Red 0.0248 0.0204 0.0035 0.0366 0.0182 0.0033 Red Edge 0.2004 0.2005 0.0223 0.1646 0.1544 0.0301 Near IR1 0.4012 0.4086 0.0477 0.3309 0.3393 0.0561 Near IR2 0.4208 0.4156 0.0353 0.3445 0.3470 0.0345 S. microstachyum (10) S. mombin (7) WV 3 Bands ASD Data reflectance Imagery Average Imagery Std. Dev. ASD Data reflectance Imagery Average Imagery Std. Dev. Coastal 0.0427 0.0356 0.0054 0.0319 0.0321 0.0055 Blue 0.0466 0.0205 0.0031 0.0423 0.0237 0.0051 Green 0.0698 0.0344 0.0035 0.0814 0.0440 0.0105 Yellow 0.0648 0.0239 0.0041 0.0726 0.0327 0.0059 Red 0.0526 0.0164 0.0049 0.0521 0.0198 0.0052 Red Edge 0.1804 0.1554 0.0299 0.2390 0.1966 0.0357 Near IR1 0.3349 0.3341 0.0515 0.4487 0.4619 0.0861 Near IR2 0.3457 0.3486 0.0476 0.4479 0.4306 0.0829

PAGE 87

72 Tree Species Identification Analysis and Results Identification Analysis A statistical comparison o f a subset of central c rown pixels from the WorldView 3 image to the ASD data acquired from the La Selva towers provided the basis of our analysis of the effectiveness of the WorldView 3 imagery to differentiate species. A single sample t test was used as an objective measure assessing the similarity of the WorldView 3 imagery to the ASD ground truth data (Manakos et al. 2009) Table 6 summarizes the subset of pixels, which includes the average reflectivity and standard deviation, and the ASD response for each tree canopy. Due to its very small c rown W. regia does not have a Standard Deviation value, as only one pixel of imagery could be identified for that tree species. The C rown Average is a lso included for comparison. Identification Results Results of the single sample t test (Table 7) provided a more rigorous comparison between the ASD data and the WorldView 3 imagery. W. regia was not included because of its extremely small sample size ( one pixel) and its close proximity to a La Selva canopy tower, which could potentially skew reflectivity values in the image sample. The results from Table 7 show that WorldView 3 bands show a significant statistical association to the W. regia (1) Z. longifolia (18) Canopy Ave. WV 3 Bands ASD Data reflectance Imagery Average ASD Data reflectance Imagery Average Imagery Std. Dev. Imagery Av erage Coastal 0.0415 0.0527 0.0158 0.0423 0.0068 0.0238 Blue 0.0433 0.0285 0.0192 0.0295 0.0034 0.0282 Green 0.0657 0.0603 0.0607 0.0606 0.0063 0.0561 Yellow 0.0556 0.0426 0.0446 0.0442 0.0047 0.0483 Red 0.0435 0.0346 0.0254 0.0274 0.0036 0.0349 Red Edge 0.1887 0.1458 0.2573 0.2530 0.0220 0.1916 Near IR1 0.3584 0.3519 0.5220 0.5355 0.0502 0.3824 Near IR2 0.3700 0.2975 0.5511 0.5566 0.0380 0.3954

PAGE 88

73 Table 7. The results of a two tailed Single Sample t t est for 9 tree species, comparing the ASD data to the averaged WorldView 3 image canopy reflective response (with AComp) Values in bold are st at istically significant, with = 0.05. df = degrees of freedom. ASD data, with the exception of the Costal and Blue bands. The Green, Red Edge, Near IR1, and Near IR2 bands have high significance values and low t values, verifying the strong agreement between the average canopy from WorldView 3 imagery and the convolved ASD ground truth data. The Yellow and Red band responses for our target species had lower agreement with the imagery, likely due to their low reflectivity or seasonal changes in chlorophyll content of the leaves. The Costal and Blue band variability and lack of correlation is mostly due to severe atmospheric attenuation in a humid tropical environment ( Lillesand, Kiefer, and Chipman 2008) The results for C. elastica and R. kunthiana show the impact of smaller canopies on C. elastica (df=3) I. marginata (df=14) L. procera (df=2) O. floribunda (df=7) P. macroloba (df=12) WV 3 Bands t Sig. t Sig. t Sig. t Sig. t Sig. Coastal 5.773 0.010 18.610 0.000 6.980 0.020 7.656 0.000 13.109 0.000 Blue 0.056 0.959 6.369 0.000 1.555 0.260 2.028 0.082 5.333 0.000 Green 1.167 0.328 2.681 0.018 1.543 0.263 0.176 0.866 0.176 0.863 Yellow 2.876 0.064 2.228 0.043 0.241 0.832 2.384 0.049 3.229 0.007 Red 4.282 0.023 4.256 0.001 5.353 0.033 2.045 0.080 4.527 0.001 Red Edge 1.889 0.155 1.281 0.221 2.222 0.156 0.867 0.414 0.015 0.988 Near IR1 0.219 0.841 0.267 0.794 0.508 0.662 0.165 0.874 0.562 0.584 Near IR2 0.020 0.985 0.070 0.945 0.776 0.519 0.308 0.767 0.524 0.610 R. kunthiana (df=4) S. microstachyu m (df=9) S. mombin (df=6) Z. longifolia (df=17) WV 3 Bands t Sig t Sig t Sig t Sig Coastal 2.231 0.089 4.190 0.002 0.058 0.956 16.439 0.000 Blue 4.899 0.008 26.598 0.000 9.580 0.000 12.637 0.000 Green 2.641 0.058 31.894 0.000 9.445 0.000 0.069 0.946 Yellow 7.194 0.002 31.851 0.000 17.995 0.000 0.391 0.701 Red 12.525 0.000 23.180 0.000 16.470 0.000 2.288 0.035 Red Edge 0.759 0.490 2.641 0.027 3.136 0.020 0.837 0.414 Near IR1 0.335 0.754 0.045 0.965 0.406 0.699 1.140 0.270 Near IR2 0.164 0.878 0.193 0.852 0.553 0.600 0.613 0. 548

PAGE 89

74 our statistical testing. Significance and t test scores improved for trees with more WorldView 3 pixel samples. Researchers in (Cho et al. 2008) suggested that unique characteristics in the crown are significant in delineating individual tree species. The results of the t test verify this observation, where sco res improved when including more WorldView 3 pixel samples per c rown The varied responses for both S. microstachyum and S. mombin in the visible and red edge portion of the spectrum can be explained by the seasonal flowering of both species in spring (Castro 2014; Condit, Perez, and Daguerre 2011) The May ASD acquisition recorded this seasonal variation, but was not present in t he November WorldView 3 imagery acquisition. The matching values in the near infrared portion of the spectrum between the ASD data and WorldView 3 imagery suggests the near infrared bands of WorldView 3 could provide some information for tree species iden tification regardless of variations in reflectivity due to a specific species reproductive cycle. Given the difference in average canopy reflectivity between the ten tree species studied and the small variance in average canopy surface reflectance values per band, it is likely that many additional tree species could be discriminated using WorldView 3 imagery. Tree Species Differentiation Analysis and Results Differentiation Analysis We performed a series of statistical analyses to determine which World View 3 bands (or combination of bands) and SVIs achieve the highest discriminatory power regarding the observed species in this study. In addition, an error matrix provided an assessment of which combination of image bands and SVIs correctly predicted tre e species. A description of the process and results of the statistical tests are below.

PAGE 90

75 Data U sed for the Analysis WorldView 3 Imagery. The placement of the WorldView 3 bands within the electromagnetic spectrum maximizes the differentiation of earth su rface features (Table 8) The passive reflectivity (corrected to true surface reflectance in this study) allows, through the combination of different bands, the detection of surface feature type s in the landscape (Campbell, James and Wayne 2011) Typically, for discriminating vegetation type the most important bands are in the mid to long wavelength visible range and near IR, showing the variation in chlorophyll activity (visible) and structural variability/water content (near infrared ). In addition to the legacy sensor bands (Blue, Gr een, Red, Near IR), WorldView 3 has available an additional coastal band, yellow band, red edge band, and Near IR band ( T able 4). Th ese additional high spatial and high radiometric resolution bands provides the potential for better differentiat ion of surf ace features, including vegetation (Wolf 2012) Chlorophyll production increases absorption in the blue and red (Lillesand, Kiefer, and Chipman 2008; Gitelson and Merzlyak 1998) The Yellow band can highlight chlorophyll variations among different species (Asner 1998) T he Red Edge band provides information on plant phenology and plant stress, and is sensitive to varying chlorophyll concentrations. Red Edge reflectance correlates strongly with other vegetation measures, such as Leaf Area Index (Delegido et al. 2013; Merton and Huntington 1999; Gitelson and Merzlyak 1998) The second Near IR band a llows a more detailed view of the vegetation structure and water content of vegetation. Table 8 below summarizes properties of WorldView 3 sensor bands for several applications.

PAGE 91

76 Table 8. The multiple applications of WorldView 3 imagery (Delegido et al. 2013; Anderson and Marchisio 2012; Campbell, James and Wayne 2011; M. L. Clark, Roberts, and Clark 2005; Haboudane et al. 2004; Elvidge and Chen 1995) Spectral Vegetation Indices We assess 14 different SVIs for discriminatory power in differentiating tree species within the Costa Rican rainforest Six of the indices are well established and have been widely used in previous studies to assess certain vegetation characteristics (LAI, biomass, etc.). Two more recent SVIs are specifically for use with the WorldView sensor bands. Six SVIs are new or modified indices developed for this study to take advantage of information gathered in the field and intended to maximize the ability of the WorldView 3 sensor. The details of the analyzed indices are below. The most widely implemented multispectral veg etation index is the Normalized Difference Vegetation Index (NDVI). The index exploits the strong variance between red and near IR in all vegetation, with greater variances between those bands showing higher levels (Lillesand, Kiefer, and Chipman 2008) or high levels of plant productivity. It Bands Applications Coastal 400 450 nm Chlorophyll absorpt ion W ater depth, coastline studies Atmospheric scattering Blue 450 510 nm Water body penetration S oil/vegetation discrimination Forest mapping Cultural feature identification Green 510 580 nm Vegetation reflectance/identification Cultural feat ure identification Yellow 585 625 nm Vegetation applications Turbidity measurements Vegetation/material feature identification Red 630 690 nm Sensitive to chlorophyll absorption Differentiation of vegetation types Cultural feature identification Red Edge 705 745 nm Vegetation analysis/condition Plant health C hlorophyll function Near IR1 770 895 nm Differentiation of vegetation types Vegetation vigor/biomass Vegetation structure Delineating water bodies Soil moisture discrimination Nea r IR2 860 1040 nm Vegetation analysis Biomass studies Urban materials differentiation

PAGE 92

77 has also been used for related applications, most recently as a proxy for Leaf Area Index (LAI) and global vegetation assessments (Xue and Su 2017) In some sensors, NDVI has a tendency to saturate in the near IR channel over dense forest, or for dense forest cover to show a uniformly high NDVI va lue. This can limit its usefulness for tropical forests. The WorldView Improved Vegetation Index (WV VI) is a modification of the NDVI using the Worldview 3 Near IR2 band. This has less tendency to saturate over dense green biomass (Bannari et al. 1995) An index that is similar to NDVI is the Green Normalized Difference Vegetation Index (GNDVI). It uses the Green band (typical range from 540 570 nm) instead of the Red band, and is typically more sensitive to chlorophyll variability than NDVI (Gitelson and Merzlyak 1998) It has also been used to estimate crop yield output based on crop health and vigor (Hatfield and Prueger 2010) The Wide Dynamic Range Vegetation Index (WDRVI) was developed to be useful in high biomass locations (such as tropical forests) where NDVI can saturate for some sensor s (Gitelson 2004) The dynamic range output of the index is widened to account for high index values in high biomass areas (Xue and Su 2017) A typical value used for the weighting coefficient ( a ) is 0.2.

PAGE 93

78 The Renormalized Difference Vegetation Index (RDVI) attempts to maximize the reflective difference between the red and infrared bands when imaging healthy vegetation It also has a reduced sensitivity to solar and viewing geometry variations (Roujean and Breon 1995) It is sensitive to various vegetation biophysical parameters and has a good correlation with vegetation biomass (Das and Singh 2012) The Modified Simple Ratio (MSR) was developed from RDVI to increase the sensitivity of the index to variable biological cha nges (Das and Singh 2012; Chen 1996) The Red Edge Greenness Vegetation Index (REGVI) is a modification of a previously used VI, the Red Edge Normalized Difference Vegetation Index (RENDVI). Instead of the Red band, the Green band is i ncluded to characterize the maximum reflectivity in the visible, while still characterizing the variance in the red to Near IR1 through the Red Edge measurement. It benefits from the red edge sensitivity to changes in canopy foliage and senescence (Merton and Huntington 1999) as well as a measure of vegetation stress (Sims and Gamon 2002) An index typically used for hyperspectral narrowband sensors, but applied in this study to the WorldView 3 multispectral imagery, is the Modified Triangu lar Vegetation Index (MTVI). The index combines bands in the areas of green peak reflectance, maximum

PAGE 94

79 chlorophyll absorption, and near infrared high reflectance based on plant cellular structure. It is sensitive to increasing canopy density through its m easurement of chlorophyll production. The index is typically used as a proxy for Leaf Area Index (LAI) measurements and biomass (Haboudane et al. 2004) The remaining SVIs used here are either previously published indexes that have been modified to utilize bands available from the WorldView 3 sensor (i.e., providing add itional information on chlorophyll or plant structure), or have been newly constructed for this study to maximize information extraction from the tropical forest. The first of the modified indices is the WorldView modified GNDVI (WV GNDVI). It is similar to GNDVI but instead uses the WorldView 3 Near IR2 band, which expands the index value range due a typically higher reflectivity value relative to Near IR1 band (Wolf 2012) The WorldView Yellow, Red Edge Vegetation Index (WV YREVI) is a new index that based on the two new WorldView 3 bands. The red edge reflectance intensity can be a measure of the slope between the red wavelengths (absorbed by chlorophyll) and the near infrared wavelengths (a measure of plant structure and water). The Yellow band is unique in that it sits on the slope between green reflectivity and chlorophyll abs orption within the red spectrum, and is sensitive to variations in canopy density (Asner 1998) The WorldView Average Canopy Reference Index (WV ACRI), developed specifically for this study, is a measure of reflective variation of a pixel or segment region

PAGE 95

80 relative to the average canopy intensity measured from the ASD spectrometer and convolved to match the WorldView 3 bands (Milton and Choi 2004) The peak infrared response is an average o f the Near IR1 and Near IR2 bands to minimize the weight of reflectivity variation in this region of the EM spectrum. In the equations below, ACnir1 refers to average canopy Near IR1 band, ACgrn refers to average canopy Green band, ACre refers to the aver age canopy Red Edge Band, etc. where; The Red Edge Vegetation Stress Index (RVSI) was originally derived to be used in a hyperspectral sensor setting (M. L. Clark, Roberts, and Clark 2005) but was modified in this study to be used with data from the WorldView 3 sensor The output of the index correlates with vegetation phenology and health by measuring changes in the red edge based on plant stress and variations in chlorophyll content during the growing season (Merton and Huntington 1999 ) It is also positively correlated with leaf biomass (Delegido et al. 2013; Perry and Roberts 2008) The WorldView Red Edge Vegetation Stress Index (WV RVSI) is a modification of the RVSI index described above, with the substitution of the Near IR2 band for the Near IR1 band. T his can better characterizes biomass variations within vegetation cover (Wolf 2012) It will also apply additional information from a spectral location that shows increased divergence between vegetation types (Marchisio, Pacifici, and Padwick 2010)

PAGE 96

81 As a follow on to the two previous Red Edge indexes above, the WorldView Red Edge Slope Weighted Index (WV RESWI), developed for this study, looks specifica lly at the Red Edge variability measured from the imagery. The formula provides a slope of the reflectivity from the red and near infrared wavelengths, calculated as the difference in reflectivity from the Near IR1 band and Red band divided by the distanc e in micrometers between the bands, multiplied by the Red Edge band reflectivity. This index provides a measure of Red Edge band intensity plus a component of the red edge slope. Variation in red absorption by chlorophyll and infrared reflectivity by plant structure are important reflective characteristics that define specific tree species (Xue and Su 2017; Lillesand, Kiefer, and Chipman 2008) Statistical Analysis The statistical analysis used only seven of the ten species originally studied. W. regia was not included due to small image sample size. Data for S. mombin and S. microstachyum were not included because of the minimal confidence in imagery detection due to the varying time of a cquisition of the ASD data (collected May June during flowering) relative to the image data from early November. (Castro 2014; Condit, Perez, and Daguerre 2011) A Discriminant Analysis (DA) procedure analyzed all 8 of the WorldView 3 band pixel values and all 14 Vegetation Indices for the defined seven tree species. A DA is typically performed to either maximize the discrimina ting power of a predictive function by determining the most effective independent variables for that task, or to determine the overall

PAGE 97

82 predictive classification of the categorical tree species through a confusion (error) matrix (Hair et al. 2010) In this study, the primary use of a DA was to p rovide objective clarity as to which WorldView 3 bands and Vegetation Indices are most important for characterizing/differentiating tropical tree species. A further verification of the discriminatory power of the data was assessed through an error matrix that verified the species classification accuracy (dependent variables) from the image information (independent variables) described above (Hair et al. 2010; Thenkabail et al. 2004) The DA process requires that the data exhibit multivar iate normality, i.e., that variances among group variables are the same across all level of predictors (homoscedasticity), and that multicollinearity exists between variables (predictive power decreases with increased correlation between predictor variable s). It is also assumed that all data is randomly sampled and the analysis can be sensitive to data outliers (Hair et al. 2010) Because a DA is resistant to violations of these assumptions, it is often used for imagery analysis, as some of these assumptions are typically not met, especially no rmality (Immitzer, Atzber ger, and Koukal 2012; Thenkabail et al. 2004) discriminatory power of the independent variables to define the dependent group variables Lambda for a value is small, the value is significant and a high discriminatory power exists that corresponds to a statistically greater separability (Thenkabail et al. 2004) Differentiation Results 3 showed the individual discriminatory power of each band in differentiat ing tree species. Table 9

PAGE 98

83 Table 9. 3 bands for the seven tre e species in the study. The table also includes the F test measure of significance Significance Level = 0 .05. discriminatory power in the bands expected to differentiate vegetation (Campbell, James and Wayne 2011) The Green, Red, Red Edge, and Near Infrared bands all had high discriminato test statistic values. All bands were significant exc ept the coastal band, which exceeded the defined above. The output of this test is below in Table 10. showed poor di scriminatory power in many of the traditional SVIs. This is likely due the lack of variation in net leaf cover or vegetation health in SVIs designed to detect sparse or stressed canopy in a healthy tropical forest. The traditional indices, GNDVI and MTV1 even though not specifically designed to differentiate species, perform well due to the use of the Green band in each index. Also, both tend to not saturate in dense vegetation cover like NDVI (Gitelson 2004) Band Locations F test Sig. Coastal 0.832 1.985 0.082 Blue 0.591 6.806 0.000 Green 0.468 11.171 0.000 Yellow 0.442 12.427 0.000 Red 0.438 12.603 0.000 Red Edge 0.389 15.47 0 0.000 Near IR1 0.470 11.087 0.000 Near IR2 0.369 16.830 0.000

PAGE 99

84 Table 10. The output o 14 spectral vegetation indices for the seven tree species in the study. The table also includes the F test measure of significance Signi ficance Level = 0 .05. The results for WDRVI were surprising, g iven the fact that the index was supposed to perform well in dense biomass areas (Xue and Su 2017) The indices that performed best were WV ACRI and WV RESWI, constructed specifically for this study. Both are specific to the WorldView 3 sensor and take advantage of extra bands in t he near infrared and red edge to better differentiate species. The WV ACRI, by its construction, represents a decorrelation stretch of the bands. In this application, the removal of a large part of correlation between tree species amplifies their particu lar reflective variations. Table 11 provides a summary of the results of the previous two tables, showing the or less, and the F test value threshold was set to > 10.0. It would be the recommended list of variables for input into a classification sche ma for accurate tree species identification. Vegetation Indices F test Sig. NDVI 0.766 3.007 0.012 WV VI 0.863 1.567 0.173 GNDVI 0.520 9.075 0.000 WDRVI 0.747 3.328 0.007 RDVI 0.532 8.648 0.000 MSR 0.666 4.923 0.000 REGVI 0.566 7.526 0.000 MTV1 0. 487 10.341 0.000 WV GNDVI 0.569 7.458 0.000 WV YREVI 0.643 5.461 0.000 WV ACRI 0.358 17.661 0.000 RVSI 0.720 3.818 0.003 WV RVSI 0.725 3.721 0.003 WV RESWI 0.392 15.242 0.000

PAGE 100

85 Table 11. F test significance values for discriminating the seven tree species in the study. Significance Level = 0 .05. Table 12. Correlation matrix of the most significant independent variables from Table 11. bands/indices from Table 11 determined the most de correlated data for the analysis (Table 12). H ighly correlated indexes indicate an information redundancy, which could reduce classification accuracy (Immitzer, Atzberger, and Koukal 2012; Thenkabail et al. 2004) In contrast, an identification schema comprised of the least correlated, high discriminatory bands and SVIs will be the simplest, most effective mea ns of species mapping. Both the Correlation Matrix of Table 12 (which identifies the most de correlated Bands F test Sig. Green 0.468 11.171 0.000 Yellow 0.442 12.427 0.000 Red 0.438 12.603 0.000 Red Edge 0.389 15.470 0.000 Near IR1 0.470 11.087 0.000 Near IR2 0.369 16.830 0.000 MTV1 0.487 10.341 0.000 WV ACRI 0.358 17.661 0.000 WV RESWI 0.392 15.242 0.000 WV 3 Bands and Indices Green Yellow Red Red Edge Near IR1 Near IR2 MTV1 WV ACRI WV RESWI Green 1.000 0.560 0.570 0.585 0.644 0.453 0.666 0.692 0.629 Yellow 1.000 0.363 0.697 0.496 0.441 0.505 0.656 0.604 Red 1.000 0.383 0.572 0.204 0.517 0.474 0.504 Red Edge 1.000 0.810 0.657 0.820 0.926 0.934 Near IR1 1.000 0.575 0.996 0.907 0.950 Near IR2 1.000 0.591 0.818 0.644 MTV1 1.000 0.916 0.952 WV ACRI 1.000 0.953 WV RESWI 1.000

PAGE 101

86 discriminatory power) determine a final best independent var iable grouping for discriminating tree species. An error matrix through the Discriminant Analysis determined the best combination of independent variables to yield the highest overall accuracy (Table 13). An analysis of several different combinations of variables assisted in determining the best fit grouping for accurate prediction of the seven tree species studied. The data combination included the following bands/indices: Green band, Red band, Red Edge band, Near IR2 band and WV RESWI. A Simultaneous Estimation process ensured that all of the identified variables were considered concurrently (Hair et al. 2010) Correct classification values are on the diagonal, incorrect classifications are on the off diagonal. It is important to state that the intent of the Error Matrix in Table 13 gener ated from the Discriminant Analysis is to inform image band and SVI choice only. Individual pixel values are representative of the variability of a larger diverse population, but the data are non independent and belong to each of the classes studied. The interpretation of output of the Error Matrix is not a fully independent analysis but is an insight into which input data would work best with an independent classification. The four bands used in the analysis (Green, Red, Red Edge, Near IR2, WV RESWI) o utlined in Table 13 yielded the highest overall accuracy (81.8%) and a very low group combination of independent variables. In addition, t he analysis produced a relatively high accuracy when using a small sample of pixels per canopy per species (Table 6). It follows

PAGE 102

87 Table 13. The results of the Error Matrix analysis using the best independent variables determined their inclusion in the analysis. that the application of these data inputs when using more pixel samples per tree crown will yield a higher overall accuracy than currently shown. The importance of the Red Edge is evident as both the Red Edge band and the WV RESWI index maximizes the Red Edge information from the WorldView 3 imagery in the error matrix analysis. Several other data groupings yielded poorer results but showed some promise. The substitution of WV ACRI for WV RESWI in the m atrix above yielded an overall accuracy of 78.8%. When it was included with WV RESWI, the accuracy increased to 80.3%. It did not exceed the accuracy of Table 13 due to the redundancy between these two indices (Table 12). Unexpectedly, the Yellow band d id not contribute to a high accuracy, although it had a relatively low correlation with the other bands and indices and more in other types of vegetation environments The Near IR1 band did not add any accuracy to the classification due to redundancy with the WV RESWI shown in the correlation. The calculation of WV RESWI uses Near IR1 (Equation 15). Bands/Index Used: Green, Red, Red Edge, Near IR2, WV RESWI Overall Accuracy 81.8 % Predicted Group Membership Input Data C. elastica I. marginata L. procera O. floribunda P. macroloba R. kunthiana Z. longifolia C. elastica 100.0 0 0 0 0 0 0 I. marginata 0 7 3.3 6.7 0 13.3 0 6.7 L. procera 0 0 66.7 0 33.3 0 0 O. floribunda 0 12.5 .0 75.0 12.5 0 0 P. macroloba 0 0 15.4 7.7 76.9 0 0 R. kunthiana 20.0 0 0 0 0 80.0 0 Z. longifolia 0 5.6 0 0 0 0 94.4

PAGE 103

88 Adding the blue band to the independent variable input for Table 1 3 increased the potentially indicate that the addition of the blue band was an important augmentation to species classification. However, the analyses from Table 7 indicate a very weak association between the reflectivity values of the Blue band and tree species in the study area. This is mostly due to high and variable blue band scattering in humid air, giving a wide variable response for each tree species even aft er atmospheric correction. This generates a false reflectivity separability between species in that spectral range, partly a result of our low ground truth sample population for each species (Figure 2). The statistical analysis demonstrated that World Vie w 3 image data can differentiate several distinct tree species in a complex tropical forest. Both WorldView 3 band data and value added VIs incorporating the additional WorldView 3 infrared, red edge, and yellow bands had a strong discriminatory power when differentiating species. The traditional indices (NDVI, GNDVI, WDRVI, etc.) did not perform as well as the modified or custom indices because the original intent of many of these traditional indices is for vegetation detection (Xue and Su 2017) not species identification. Ev en a traditional index, WDRVI, originally developed to minimize saturation in areas of dense vegetation, performed poorly in the tropical forest environment. The WV RESW I and the WV ACRI performed best regarding their discriminatory power for species dif ferentiation. These new indexes we re specifically developed for the WorldView 3 sensor, but could be adapted to other sensor systems provided they have bands near the World View 3 channels. The WV RESWI performed well in showing variability between the c hlorophyll production and plant structure. The construction of the WV ACRI

PAGE 104

89 is highly dependent on a canopy average measurement. C anopy average values, from either a spectrometer or a random average of pixel values, would be required i n order for this SVI to be appl icable to other study sites The error matrix accurately verified the selected imagery band and vegetation indices selected for a future in depth classification analysis of tree species in the complex tropical forest. The result is encouraging, as the relatively high overall accuracy for the classification suggest that the process defined within this study would be successful if applied to many more species, helping to define the density and distribution of tree species within the tropical fores t. In addition, other DA outputs, such as the canonical coefficients, might be useful as weights to the independent variables for maximizing the predictive value of those variables regarding tropical tree species. Conclusions and Recommendations This stu dy establishes an effective set of parameters derived from multi spectral image data for use in identification of tropical forest species. Tropical forests pose a significant challenge due to their complexity and the wide variations in species reflectance and abundance, even in in small study areas (Barb osa, Broadbent, and Bitencourt 2014) Our results show that WorldView 3 imagery, after atmospheric compensation and pre processing for viewing geometry and bidirectional reflectance variations, has a high potential to identify the selected tree species a nalyzed and to distinguish among many species. An important aspect of this analysis is using a c rown pixel cluster vs. per pixel measure of reflectivity. The c rown pixel cluster data converge on an accurate reflectance measure for each band (Cho et al. 2008) The separability in canopy reflectance spectra implies that several additional tree species can be discriminated using WorldView 3 imagery.

PAGE 105

90 In addition, the utilization of the proper bands and Vegetation Indices can theoretically yield a classification accuracy for tropical tree species. The results of this analysis provide some confidence that the combination of direct band information, plus the addition of select vegetation indices, could significantly contribute to any classification schema used for accurate vegetation species detection. A classification schema that could successfully perform individual tree taxonomy identification would be of great benefit to a variety of applications (M. van Leeuwen and Nieuwenhuis 2010) including the realization of an improved biomass estimates (Wulder et al. 2004) and a better understanding of the relationship between forest processes and climate variability. The output from this study defines a starting point for a large area canopy based classification in a tropical rainforest environment.

PAGE 106

91 References 2 and the Evolution of the DigitalGlobe Remote Sensing Satellite Const ellation: Introductory Paper for the Special Session on WorldView Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery XVIII, SPIE 8390 (83900L 9): 1 15. doi:10.1117/12.919756. TM 136. Remote Sensing of Environment 64 (3): 234 253. doi:10.1016/S0034 4257(98)00 014 5. Remote Sensing Reviews 13 (1 2): 95 120. doi:10.1080/02757259509532298. Abovegro International Journal of Forestry Research 2014: 1 14. doi:10.1155/2014/715796. BRDF Measurements and Model for Spe cular and Diffuse Components Remote Sensing of Environment 98 (2 3): 201 211. doi:10.1016/j.rse.2005.07.005. Breunig, Fabio M., Lenio Soares Galvao, Yhasmin Mendes Moura, and Rafaelo Balbinout. ence of a Subtropical Semideciduous Anais XVI Simposio Brasileiro de Sensoriamento Remoto SBSR, Foz Do Iguau, PR, Brasil, 13 a 18 de Abril de 2013, INPE no. 1986: 6917 6922. Campbell, James, B., and Randolph H. Wayne. 2011. Introduction to Remote Sensing 5th ed. New York, NY: The Guilford Press. Photogrammetric Engineering Remote Sensing 70 (1 ): 135 140. doi:10.14358/PERS.70.1.135. Organization for Tropical Studies http://sura.ots.ac.cr/florula4/index.php. for Canadian Journal of Remote Sensing 22: 229 242. The International Archives of Photogrammet ry, Remote Sensing and Spatial Information Science 37 (Part B7): 1 8. Cifuentes Workshop on Tree Volume and Biomass Allometric Equat ions in South and Central UN REDD Programme 92.

PAGE 107

92 Cifuentes Jara, M, Miguel Morales, and D Henry. 2013. Inventory of Volume and Biomass Tree Allometric Equations for Central and South America UN REDD MRV Report 11, CATIE, Turialba, Costa Rica, F ood & Agriculture Organization of the United Nations Remote Sensing of Environment 96 (3 4): 375 398. doi:10.1016/j.rse.2005.03.009. Clark, Roger N, G A Swayze, K Eric Livo, Raymond F Kokaly, Trude V V King, J Brad Dalton, J Sam Vance, Barnaby W Rockwell, Todd Hoefen, and Robert R Mcdougal. USGS Spectroscopy Lab 1 21. Condit, Richard, Rolando Perez, and Nefertaris Daguerre. 2011. Trees of Panama and Costa Rica Princeton University Press. Cross, Matthew D., Ted Scambos, Fabio Pacifici, and Wes Marsh Use of Metre Scale Multi Spectral Satellite Image Data for Identifying Tropical Forest International Journal of Remote Sensing 25 (11): 3723 3752 iomass and Spectral International Journal of Engineering Research & Technology (IJERT) 1 (5): 1 13. doi:ISSN: 2278 0181. Ed ge Spectral Index for Remote Sensing Estimation of Green LAI over European Journal of Agronomy 46. Elsevier B.V.: 42 52. doi:10.1016/j.eja.2012.12.001. DigitalGlobe, DS WV3 09/14 2. Elvidge, C Band and Narrow Band Red and near Remote Sensing of Environment 54 (1): 38 48. doi:10.1016/0034 4257(95)00132 K. Foody, Giles M., Doreen S. Boyd, and Mark E J Cutler. Tropical Forest Biomass from Landsat TM Data and Their Transferability between Remote Sensing of Environment 85 (4): 463 474. doi:10.1016/S0034 4257(03)00039 7. Gao, Xiang, Alfredo R. Huete, Wenge Ni, and Tomoaki M Biophysical Remote Sensing of Environment 74 (3): 609 620. doi:10.1016/S0034 4257(00)00150 4. Gastellu Etchegorry, J.P., P. Guillevic, F. Zagolksi, V. Demarez, V. T richon, D. Deering, and Remote Sensing of Environment 68 (July 1998): 281. Quantific Journal of Plant Physiology 161 (2): 165 173. doi:10.1078/0176 1617 01176.

PAGE 108

93 Channel in Remote Sensing of Global Vegetati on from EOS Remote Sensing of Environment 58 (3): 289 298. doi:10.1016/S0034 4257(96)00072 7. Advances in Space Research 22 (5): 689 692. doi:10.1016/S0273 1177(97)01133 2. Ieee Geoscience and Remote Sensing Letters 5 (2): 133 13 7. doi:10.1109/Lgrs.2008.915598. Haboudane, Driss, John R. Miller, Elizabeth Pattey, Pablo J. Zarco Tejada, and Ian B. Green LAI of Crop Canopies: Modeling and Validation in the Context of Precision Remote Sensing of Environment 90 (3): 337 352. doi:10.1016/j.rse.2003.12.013. Hair, Joseph F. Jr., William C. Black, Barry J. Babin, and Rolph E. Anderson. 2010. Multivariate Data Analysis Seventh Ed. Pearson Pre ntice Hall. Analytial Spectral Devices, Inc. (ASD) no. 4th Ed.: 144. to Quantify Agricultural Crop Characteristic s at Different Growth Stages under Varying Remote Sensing 2 (2): 562 578. doi:10.3390/rs2020562. Classification with Random Forest Using Very High Spatial R esolution 8 Band worldView Remote Sensing 4 (9): 2661 2693. doi:10.3390/rs4092661. Journal of Forest Research 9 (1): 7 14. doi: 10.1007/s10310 003 0045 z. 3 Imagery Technical Note 1 WorldView 3 Instrument WorldView 12. Li, Guiying, Dengsheng Lu, Emilio Moran, and Scott Hetrick. 201 Cover Classification in a Moist Tropical Region of Brazil with Landsat Thematic Mapper International Journal of Remote Sensing 32 (23): 8207 8230. doi:10.1080/01431161.2010.532831. Based Feature Extraction for I mproved Endmember Abundance IEEE Transactions on Geoscience and Remote Sensing 42 (3): 644 649. doi:10.1109/TGRS.2003.822750. Lillesand, Thomas M., Ralph W. Kiefer, and Jonathan W. Chipman. 2008. Rem ote Sensing and Image Interpretation 6th Editio. John Wiley and Sons Ltd.

PAGE 109

94 Manakos, Ioannis, Kiril Manevski, Chariton Kalaitzidis, and Dennis Edler. 2009. Comparison Between FLAASH and ATCOR Atmospheric Correction Modules on the Baisis of WorldView 2 Image ry and in Situ Spectroradiometric Measurements Mediterranean Agronomic Institute of Chania Value of the New Spectral Bands in the WorldWiew 2010 IEEE Int ernational Geoscience and Remote Sensing Symposium 2723 2726. doi:10.1109/IGARSS.2010.5649771. 1 Satellite Sensor for Multi Procee dings of the 1 10. http://www.eoc.csiro.au/hswww/jpl_99.htm. Journal of Geophysical Research Atmospheres 97 (D17): 18935 1 8946. doi:10.1029/92JD00879. Proceedings of the Annual Conference of the Remote Sensing and Photogrammetry Society 1 11. Miura, Tomoaki, and Alfredo R. Huete. 2009. Sensors 9 (2): 794 813. doi:10.3390/s90200794. Moura, Yhasmin Mendes, Lnio Soares Galvo, Joo Roberto dos Santos, Dar A. Roberts, and Fbio Marcelo Breun and Inter Remote Sensing of Environment 127. Elsevier Inc.: 260 270. doi:10.1016/j.rse.2012.09.013. Myint, S W, J Franklin, M Buenemann, W K Kim, Photogrammetric Engineering and Remote Sensing 80 (10): 983 993. doi:10.14358/PERS.80.10.983. ithm for Very High Proc. JACIE 1 43. In Geoscience and Remote Sensing Symposium (IGARSS) 2016 IEEE Inte rnational 1973 1975. Physical Quantities for the Analysis of Multitemporal and Multiangular Optical Very IEEE Transactions on Geoscience an d Remote Sensing 52 (10): 6241 6256. doi:10.1109/TGRS.2013.2295819.

PAGE 110

95 Hyperspectral Satellite Imagery for Regional Inventories: A Test with Tropical Emergent Trees in the Journal of Vegetation Science 21 (2): 342 354. doi:10.1111/j.1654 1103.2009.01147.x. Band and Broad Band Indices for Assessing Nitrogen Availability and Water Stress in an A Agronomy Journal 100 (4): 1211 1219. doi:10.2134/agronj2007.0306. Ieee Geoscience and Remote Sensing 31 (3): 727 734. ld Spectroscopy Facility Post Post Processing Spectral Data in MATLAB Review of Remote Sensing Technology in Support of the Kyoto Protoc Environmental Science and Policy 6 (5): 441 455. doi:10.1016/S1462 9011(03)00070 4. Remote Sensing of Environment 51 (3): 375 384. doi:10.1016/0034 4257(94)00114 3. Roy, D. P., H. K. Zhang, J. Ju, J. L. Gomez Dans, P. E. Lewis, C. B. Schaaf, Q. Sun, J. Li, H. Data to Nadir BRDF Adjusted Remote Sensing of Environment 176: 255 271. doi:10.1016/j.rse.2016.01.023. and Spectral Reflectance across a Wide Range of Species, Leaf Structures and Dev Remote Sensing of Environment 81 (2 3): 337 354. doi:10.1016/S0034 4257(02)00010 X. STRI Website www.si.edu. Thenkabail, Prasad S., Eden A. Enclona, Mark S Ashton, Christopher Legg, and Minko Jean Remote Sensing of Environment 90 (1): 23 43. doi:10.1016/j.rse.2003.11.018. UN ative Programme on Reducing Emissions UN REDD Website www.un redd.org/. Vaiphasa, Chaichoke, Suwit Ongsomwang, Tanasak Vaiphasa, and Andrew K. Skidmore. iscrimination Using Hyperspectral Data: A Estuarine, Coastal and Shelf Science 65 (1 2): 371 379. doi:10.1016/j.ecss.2005.06.014.

PAGE 111

96 Vaiphasa, Chaichoke, Andrew K. Skidmore, Willem F. de Boer, and Tanasak Vaiphasa. ISPRS Journal of Photogrammetry and Remote Sensing 62 (3): 225 235. doi:10.1016/j.isprsjprs.2007.05.006. Parameters Using Li European Journal of Forest Research 129 (4): 749 770. doi:10.1007/s10342 010 0381 4. IEEE Transactions on Geoscience and Remote Sensing 44 (7): 1931 1933. doi:10.1109/TGRS.2006.873688. Vescovo, Loris, Georg Wohlfahrt, Manuela Balzarolo, Sebastian Pilloni, Matteo Vegetation Indices Ba sed on the near Infrared Shoulder Wavelengths for Remote International Journal of Remote Sensing 33 (7): 37 41. doi:10.1080/01431161.2011.607195. Weyermann, Jrg, Alexander Damm, Mathias Kneubhle r, and Michael E Schaepma n. 2014 IEEE Transactions on Geoscience and Remote Sensing 52 ( 1 ): 616 627 2 Vis NIR Multispectral I magery to Support Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery XVIII, SPIE 8390 (83900N): 1 8. doi:10.1117/12.917717. Wulder, Michael Bioscience 54 (6): 511 521. doi:10.1641/0006 3568(2004)054[0511:HSRRSD]2.0.CO;2. Xue, Jinru, and Baofeng Su. Zhang, Jinkai, Benoit Rivard, Arturo Snchez Azofeifa, and Karen Castro Esau. 2006. and Inter Class Spectral Variabil ity of Tropical Tree Species at La Selva, Costa Remote Sensing of Environment 105 (2): 129 141. doi:10.1016/j.rse.2006.06.010.

PAGE 112

97 CHAPTER IV CLASSIFICATION OF TR OPICAL FOREST TREE S PECIES USING METER SCALE IMAGE DATA Abstract Accurate classification of tropical tree species is critical for studies of forest habitat, forest composition, biomass, and ultimately a better understanding of the role trees play in climate variability throug h carbon uptake. The aim of this study was to derive an accurate classification method for a tropical tree species at a research forest site in Costa Rica (La Selva Biological Station) using high resolution imagery (WorldView 3) validated with field gathe red spectrometer data An object oriented classification was the basis of this study, using multi pixel scale full tree canopies extracted through a segmentation process in ENVI and a series of pre processed WV 3 bands and band combinations. We evaluate d various combinations of image bands and multi band indices to determine their feasibility in differentiating tree species ; we then tested the classification on a managed arboretum site within La Selva. The results suggest that WorldView 3 bands in the G reen, Red, Red Edge, Near IR2 portions of the EM spectrum, with the addition of two specialized vegetation indexes derived from the WorldView 3 imagery, are excellent data inputs for segmentation and classification of the complex tropical forest. Classifi cation results for the six tree species analyzed yielded an accuracy of 85.4%, with minimal errors of commission and omission. Although the focus was emergent species, shadows played a role in the classification accuracy through their impact on segmentati on accuracy. The process provides a path to better characterization of tropical forest composition and species distribution for improving biomass studies in a tropical environment.

PAGE 113

98 Introduction The aim of this study was to determine if a rule based objec t oriented classification schema could provide an accurate and objective classification and assessment of tropical canopy tree species within a tropical study area. An important step in the study of forest habitat, biomass and composition would be the id entification of the proper data preparation processes, scientific approaches and analytical tools necessary for such an endeavor. Accurately measuring species type, would support a more accurate measure of above ground biomass (Wulder et al. 2004) which could improve current carbon offset programs (Cifuentes Jara and Henry 2014) Currently large uncertainties exist in e stimates of these parameters for tropical forests (Barbosa, Broadbent, and Bitencourt 2014) The use of remote sensing imagery for forest analysis has a long history, from the utilization of the Landsat legacy series imagery (Li et al. 2011) and the current Landsat 8 satellite (Roy et al. 2016) to high resolution imaging systems from commercial systems (Katoh 2004; Wulder et al. 2004) Re garding forest analysis, many studies have attempted to use remotely sensed imagery for tree type identification in complex forest assemblages, including tropical forests (Li et al. 2011; Carleer and Wolff 2004) Many pixel based classification studies have used several different classification schema (maximum likelihood, spectral angle mapper, support vector mach ine, random forest etc.) to determine which works best in different forests, but with minimal success. Most studies generated accuracies from 42 74% depending on the forest assemblage and environments studied (Shafri, Suhaili, and Mansor 2007; Yu et al. 2006; Plourde and Congalton 2 003) While pixel base approaches have been the standard for remote sensing classification, Geospatial Object Based Image Analysis (GEOBIA) has made significant advances in the

PAGE 114

99 last decade and has proven to be superior to pixel based approaches, especia lly when utilizing very high resolution imagery (Immitzer, Atzberger, and Koukal 2012; Blaschke 2010) Ut ilizing the entire canopy structure as the bas is of species characterization, which includes all of the nuanced variances in the canopy rather than individual pixels within the canopy or even individual leaves, has proven to be the more accurate method f or a complex classification procedure (Myint et al. 2011) Very high spatial resolution imagery provides greater spatial a nd spectral detail for any feature, including vegetation. Because of this increase in spatial information, a single pixel will likely not capture the general characteristics of a classification target, as the fine resolution of newer space borne collectio n systems provide multiple pixels that characterize any feature. Some features, such as tree canopies, are typically not homogeneous, and a certain amount of variability occurs within the canopy, which can lead to a reduction of separability between other features (Yu et al. 2006) This can lead to poor definition of features and low class ification accuracies when using pixel based methods (Zhang et al. 2006) Object based classification methods provide a good alterna tive to overcome these problems They add additional whole obj ect feature information beyond spectral content such as shape, size pixel variability and proximity to other objects (Blaschke 2010; Yu et al. 2 006) that can provide critical information to identify features even with non homogeneity within a defined object. With the introduction of high resolution imagery the application of a segmentation process to identify image objects has been successfu lly extended to vegetation studies, specifically the identification of emergent tree types within a forest (Blaschke 2010) Many studies have evaluated various types of object based classification approaches, such as Random Forest (RF), Decision trees, Support Vector Machines (SVM) and Artific ial Neural

PAGE 115

100 Network (ANN) schemas. Also more traditional statistical techniques such as Multinomial Logistics Regression and Linear Discriminant Analysis (Fagan et al. 2015; Prospere, McLaren, and Wilson 2014; Feret and Asner 2013; Clark and Roberts 2012; Tuia et al. 2009) have also been used with some success. Clark and Roberts (2012) achieved an 87.4% cl assification accuracy in their study of a tropical forest regime in Costa Rica when using an object based approach through a Random Forest classifier. Others have directly compared the pixel based approach to the objects based approach, with the object ba sed approach achieving superior results in a complex forest setting (Feret and Asner 2013; Clark and Rob erts 2012; Myint et al. 2011) A study by Immitzer, Atzberger, and Koukal (2012) used a random forest classifier to compare pixel based v ersus object based classification in an Austrian mixed forest. Worldview 2 image data provided the bas is for the classification. Manually extracted tree crown data provided ground truth from known tree locations in the study area. Through the analysis of several classification procedures, an object based random forest approach provided the highest classi fication accuracy at 82.4% and the pixel based classification was on average 10 percentage points lower in accuracy Another promising classification option is a rule set object based classification approach which provides some ability of input from the u ser on variable choices orders of variable importance (Myint et al. 2014; van der Sande, de Jong, and de Roo 2003) Myint et al. ( 2014) used a rule set object based classification to quantify mangrove extent in Bangladesh using Landsat imagery. The r ule set approach produced an overall ac curacy of 84.1%, but the authors noted that the settings for segmentation and rules applied to the particular study might not be suitable for other study areas. The research also discovered that more bands included in the classification process did not eq uate to higher accuracies but

PAGE 116

101 instead led to lower accuracies due to signature confusion related to high correlation between certain Landsat bands. Additionally, Ke, Quackenbush, and Im (2010) used a rule based classification to map forests in New York State by using both QuickBird imagery (p roviding reflectivity information) and LiDAR data (tree height information) for both the segmentation and rule set data input. Classification accuracies of near 90% were achieved in classifying 4 specific evergreen species and a broad deciduous species gr oup. Including the additional height information increased classification accuracy by approximately 10%, supporting the premise that more and diverse data sets of a study area increases the classification accuracy (Ke, Quackenbush, and Im 2010) In this study, our goal was to move the current research forward by utilizing very high resolution re motely sensed image derived data products for use in a rule set object based classification schema for accurate delineation of tropical forest species. This approach utilizes the information from several image bands and two image derived spectral vegetati on indices in an object based rule set schema to differentiate six different species in a tropical forest setting. Viewing and illumination corrections, and the utilization of an atmospheric compensation procedure, assist ed in creating an accurate image d ata set representative of the forest canopy conditions in the area of study The use of an object based classification approach is not unique (Myint et al. 2014) and has promise, as discussed above, but heavily depend s on the data input types and the segm entation procedure to accurately define the objects being classified ( Myint et al 2014) In this study, only imagery from the WorldView 3 sensor allowed an assessment of the ability of the imagery in classifying tree species. Guidance on the makeup of th e segmentation settings ( Myint et al. 2014; Latif and Ibrahim

PAGE 117

102 2014; Yu et al. 2006; Wang et al. 2004) allowed this study to accurately define crowns within the Arboretum. Overall Approach Our aim was to determine if WorldView 3 imagery with the proper pre processing, band selection, and data products (vegetation indices), could accurately identif y emergent canopy tree species in a tropical forest assemblage through the use of an object based rule set classification procedure La Selva Biological Stati was chosen as the study site for this research due to its advantageous location for easy tropical forest access, up to date infrastructure and an excellent support staff. The specific study location used was the H 1968, the Arboretum (Figure 1) is a 3.5 ha managed research area within La Selva. It contains approximately 929 cataloged plants, with the most recent census occurring in August 2016 to March 2017 Of these are 727 trees that represent 185 native species (Vargas and Castro 2017) The Arboretum is the focus of many research efforts at La Selva and provides a baseline for taxo nomic studies of tropical tree species within Central America. The first step in this research effort was to establish appropriate ground control of tree species by finding well exposed examples of canopy species in the field outs ide of the Arboretum stu dy area and accurately measuring their locations via GPS. WorldView 3 imagery provided by DigitalGlobe was geocorrected, atmospherically compensated and converted to surface reflectance for the study area (Cross et al. 2018) C rown wide pixel informatio n of those select trees extracted from the WorldView 3 imagery provided the ground truth data for this study. Tree species identical to the ground truth within the La

PAGE 118

103 Selva Arboretum provided a direct comparison to an independent measure of each species o utside of the Arboretum. An object based rule set classification schema, utilizing multiple pixels per tree crown provided a good c rown representation, including the intra c rown variations (Clark, Roberts, and Clark 2005) A segmentation process identified specific tree crowns using the WorldView 3 multi spectral bands. The most effective image data and derived vegetation indices (identified through an o bjective statistical procedure) were the input to the rule se t process. A series of error matrices defined the accuracy of the classification procedure. The processes defined in this study will help future classification efforts in the complex tropical forest, and the results will assist in determining the most ef fective types of data input for classification analysis of tropical tree species. The use of a multi spectral imaging system to achieve these goals is a traditional choice, but not necessarily the most effective imaging syst em for species identification H yperspectral systems ( Papes et al. 2010; M L. Clark, Roberts and Clark 2005) and LiDAR systems ( Bergen et al. 2009; Evans et al. 2009) have proven to be excellent choices for species discrimination/identification, especially in complex forest areas ( van Leeuwen & Nieuwenhuis 2010; Zhang et al. 2006) The purpose in using the WorldView 3 imaging system for this study is to determine if a more cost effective image data set is effective for use in a tropical forest. The advantages of using the WorldView 3 sensor are large area collection good revisit times over broad areas of interest, and overall cost advantages over specialized aircraft based sensor systems, such as hyperspectral or LiDAR. Even though the WorldView 3 sensor does not match the spatial r esolution of LiDAR or the spectral resolution of a hyperspectral sensor, there is a high enough spatial resolution to collect many

PAGE 119

104 inter crown pixels within a particular tree, and a sufficient number of image bands for spectral characterization of a partic ular species (Cross et al. 2018). WorldView 3 Imagery and Ground Truth Data Preparation Imagery Acquisition and Preparation The imagery used was a nearly cloud free, dry season, DigitalGlobe DigitalGlobe provided the imagery with AComp already applied. AComp is a proprietary atmospheric compensation process developed by Figure 1. A true color composite WorldView 3 image of the Holdridge Arboretum. The outline of the Arboretum boundary is in yellow. Trails into and through the Arboretum are in white. Tree canopies are evident in the multisp ectral imagery. Please refer to Table 1 for imagery specifications.

PAGE 120

105 DigitalGlobe for their fleet of satellites and associated sensors. It is a physics based compensation schema that uses observed in scene pixel spectra for its procedure (Pacifici 2013) The derived pixel based Aerosol Optical Depth (AOD) information is applied into a radiative transfer schema for the imagery bands collected (Pacifici 2016) The AComp product output is an atmospherically compensated image of surface reflectance values per pixel. AComp also accounts for water vapor in the atmospheric column. Due to tree canopies being non Lambertian, and with the imagery collected at a non nadir imaging view angle (Table 1), a Bidirectional Reflectance Factor (BRF) correction was required to convert data to an on nadir view (Moura et al. 2012; Bousquet et al. 2005; Middleton 1992) Reflectance anisotropy can vary by up to 30% within a closed forest canopy ( Weyermann et al. 2014 ) and high spatial resolution imagery with variable view angles (WorldView 3) can be especially affe cted (Pacifici, Longbotham, and Emery 2014) BRF correction values (Table 2) for different wavelength regions were derived from Breunig Bands Spectral Range (nm) Resolution Panchromatic 450 800 Panchromatic 0.31m Costal 400 450 Multispectral 1.24m Blue 450 510 Dynamic Range 11 bits/pixel Green 510 580 Specific Image Information for Study Yellow 585 625 Date/time 11/11/2014, 15:52:28Z Red 630 690 Zenith View Angle 26.2 Red Edge 705 745 Azimuth View Angle 108.8 Near IR1 770 895 Cloud Cover 0.5% Near IR2 860 1040 Data Extent NW Corner 10.48 N, 84.14 W SE Corner 10.25 N, 83.99 W Table 1. WorldView 3 imagery s pecifications (DigitalGlobe 2014) and specific image information extracted from the metadata file from the imager y used in this study.

PAGE 121

106 et al. (2013) for a mixed tropical tree canopy and supporting information from Gastellu Etchegorry et al. (1999) provided the correction parameters needed for the imagery. Image Rectification Because of the off nadir image collection and the inh erent error of displacement due to the height of the overall tree canopy, it was necessary to geo rectify the imagery to maximize the locational accuracy of tree species within the study area. GPS ground control locations, collected in May of 2015 from op en areas within the La Selva Biological Station area were the basis of the correction procedure. A Trimble GeoExplorer 2008 series unit with Trimble Zephyr 2 external backpack antenna, utilizing differential GPS through the Satellite Based Augmentation S ystems, Wide Area Augmentation System, and multipath correction, provided the positional accuracy necessary in collecting GPS points for an accurate rectification procedure. Most points collected for the correction had a positional accuracy better than 1 meter. An image to m ap rectification procedure within ENVI, using a nearest neighbor process to ensure the pixel values remained fixed, geometrically corrected the imagery to an accurate geolocation position. WorldView 3 Imagery Collection Backscatter Condition Band Locations Wavelength Range (nm) BRF Factor 26.2 OZA Coastal 4 00 450 1.01 Blue 450 510 1.08 Green 510 580 1.15 Yellow 585 625 1.17 Red 630 690 1.20 Red Edge 705 745 1.25 Near IR1 770 895 1.33 Near IR2 860 1040 1.34 Table 2. BRF Correction factors for the WorldView 3 Imagery. Values were derived from (Breunig et al. 2013) and represent backscatter viewing conditions for the imagery. Dividing the observed reflectance by th e BRF correction factors above derives the on nadir surface reflectance.

PAGE 122

10 7 Ground Truth Data Collection and Processing A n acquisition of tree species collections within La Selva Biological Reserve during May 2017 provided the ground truth necessary for this study. All tree samples for ground truth were from outside of the Arboretum (within the confines of La Selva), and th us independent examples from those used for the evaluation. In addition, tree crown samples were of sufficient size to be emergent within the forest canopy structure and visible in the WorldView 3 image (Table 3). Trees were chosen based on their availab ility within the Arboretum and the ability of the tree crown to be clearly identified within the imagery when selecting ground truth samples. All of the c rown samples are from at least two trees from various locations within La Selva, facilitating a prope r and accurate species ground truth representation for the classification. GPS positions of the selected tree observations from various locations within La Selva provided data for matching the field identified tree to the proper canopy vi sible in the Tree Species Family Approximate Location C rown Total Total pixels Collection Size (m 2 ) Castilla elastica Moraceae Lab area and west La Selva 3 81 100.44 Cedrela odorata Me liaceae Lab area and east La Selva 2 81 100.44 Cordia alliodora Boraginaceae Arboretum trail from lab 2 39 48.36 Pentaclethra macroloba Fabaceae Station entrance and west La Selva 9 217 269.08 Pterocarpus sp. A Fabaceae Lab area and east La Selva 2 128 158.72 Stryphnodendron microstachyum Fabaceae Station entrance and lab area 3 87 107.88 Table 3. Six tree species extracted for the classification ground truth in May 2017 from various locations within La Selva Biological Station, Costa Rica.

PAGE 123

108 WorldView 3 imagery. A conversion of the GPS coordinates of each tree crown used for ground truth to an ESRI Shapefile facilitated a direct comparison with the imagery. Additional spatial data provided by the staff at La Selva (trail location s, trail signs, streams and river Shapefiles, etc.) assisted in correctly locating the selected tree species within the image. An extraction of the average of full crown pixel clusters yielded mean image derived spectra providing the ground truth needed f or the analysis. Figure 2 shows examples of typical canopy data extractions from the WorldView 3 imagery. Figure 2. D ata extraction s for two tree crowns in La Selva. In the image above, the purple polygon represents a c rown extraction from S. microstachyum and the red polygon is a P. macroloba extraction. Upper left is the extraction areas superimposed on the WorldView 3 panchromati c image. Left is a closer view of the c rown extraction for P. macroloba showing the individual multispectral pixels within the defined c rown sample.

PAGE 124

109 Figure 3 illustrates the average reflectance per tree crown for each for each WorldView 3 band of the six tree species studied. T he variance in the reflectivity values between is species is evident. In addition to the spectral reflectivity for each species, an extracted canopy average of the Arboretum represented the collective response of emergent vegetation (Figure 3). Variation s in tree health and seasonality can affect variations in inter crown variability and ultimately change the mean crown value for a particular species collection. Inter genus and inter family variability can be extreme, as trees that look similar both vis ually and spectrally (Cross et al. 2018) can be associated with completely different family or genus types (e.g. C. odorata and S. microstachyum ) This is in contras t to some trees within the same family or genus type (e.g. P. macroloba and Z. longifolia ) which can have radically different leaf and branch construction (Condit et al. 2011) and ultimately have significantly different spectral reflectance characteristics (C ross et al. 201 8 ). All of the possible variations above will affect the overall clas sification, and problems can happen when over generalizing tree crown reflectivity assuming that the values collected are applicable to any environmental study in any tropical setting. The average data values used represent trees within the end of the dry season and their specific reflectivity characteristics that reflect that portion of the growing season. Variance can also exist between trees of the same species, but taking an overall crown average to represent a tree species incorporates inter crown va riability, allowing reflectivity values between crowns of the same species to be very close in reflectivity across the available WorldView 3 bands (Cross et al. 2018) Avoiding shadow areas was a priority during the extraction of the canopy average to en sure that the value calculated included mostly tree canopies. A sampling of most of the

PAGE 125

110 non shadow area of the Arboretum yielded a collection of 9,750 pixels. In addition, a vegetation index used in this study requires the input of the canopy average in its calculation. Arboretum Classification Analysis A series of preliminary statistical evaluation procedures helped determine which image data would be most effective in the classification process. We then performed a segmentation process on the image ry from the Arboretum to define tree c rowns and then Figure 3. Average reflectance the six tree species used as ground control for this study. Canopy average values for the Arboretum area also included. Lines between points are for clarity to show the trend in variability between bands.

PAGE 126

111 applied a rule set object based classification approach, with the appropriate data input guided by the statistical analysis. An error matrix determined the accuracy of the object oriented classificatio n procedure performed. Data Selection The ground truth data collected outside of the Arboretum (Table 3 ) determined which canopy trees would be a part of the classification analysis within the Arboretum. The Arboretum data file includes precise location s of canopy and emergent tree species (Vargas and Castro 2017) and were used for determining the accuracy of the classification in this study. A detailed catalog for the Arboretum d escribes all trees according to species, size measured in Diameter at Breast Height (DBH), and precise geolocation referenced from a permanent 25m x 25m grid (azimuth and distance from each grid post). The staff at La Selva produced an ESRI Shapefile (UTM Zone16N grid, meters) of the Arboretum catalog for easy integration with the imagery data (Vargas and Castro 2017) A discriminant analysis (DA) was performed to determine which of t he eight multispectral WorldView 3 imagery bands (independent values) had the most discriminatory power for the tree types (specific classes) studied (Hair et al. 2010) was performed within the DA to evaluate the discriminatory power of the independent variab les (Immitz providing objective clarity as to which WorldView 3 bands and Vegetation Indices are most important for characterizing and differentiating tropical tree species. This test can indicate an initial assess ment of group membership from the independent samples and insight into the importance of each independent variable. small, a higher discriminatory ability is realized and corresponds to a statistically greate r

PAGE 127

112 separability between the classified groups (Thenkabail et al. 2004) This is a common use of (Immitzer, Atzbe rger, and Koukal 2012; Papes et al. 2010). An analysis of all eight bands of WorldView 3 imagery for all seven trees analyzed determined which bands have the highest discriminatory power when differentiating the selected tree species. The placement of the imagery bands within the electromagnetic spectrum maximizes the differentiation of earth surface features including vegetation. T he most important bands for vegetation are in the mid to long wavelength visible range showing the variation in chlorophyll activity and near infrared, showing the variation in structure and water content (Campbell, James and Wayne 2011) WorldView 3 has additional image bands in the blue yellow red edge and near infrared providing a greater potential for differentiat ing veg etation (Wolf 2012) Chlorophyll absorption is pronounced in the B lue and R ed bands (Lillesand, Kiefer, and Chipman 2008; Gitels on and Merzlyak 1998) and the Yellow band holds promise in providing additional information for species identification (Asner 1998) The Red Edge band characterizes plant health and is sensitive to Leaf Area Index (Delegido et al. 2013; Merton et al ., 1999; Gitelson et al., 1998) An additional near infrared band allows more information about vegetation water content and overall structure. In addition to the sensor band information from WorldView 3, we analyzed two Spectral Vegetation Indices (SV Is) to determine their ability to classify vegetation. The use of SVIs has been well documented for determining a variety of vegetation parameters, including chlorophyll production, gross primary productivity, leaf area index vegetation type and biomas s estimates (Xue and Su 2017; Gitelson et al. 2008; van Leeuwen and Orr

PAGE 128

113 2006; Gao et al. 2000; Price 1993) In Chapter 3 above 14 specific spectral vegetation indices (SVIs) were analyzed for their discriminatory power for differentiating tropical tree species. Of the 14 SVIs analyzed, two developed specifically for the WorldView 3 sensor possessed a discrimin atory power that matched or exceeded the discriminatory power of the imagery bands. The WorldView Average Canopy Reference Index (WV ACRI) uses a canopy average value for its calculation. The index is a measure of the differentiation of a specific tree s pecies from the overall complex forest canopy. The overall infrared response is calculated using the average of the near infrared values in Bands 7 and 8, and the visible response is a combination of the Green, Yellow and Red Edge band reflective measure ments as they compare to the canopy average for those bands. In the equations below, ACnir1 refers to average canopy Near IR1 band, ACgrn refers to average canopy Green band, ACre refers to the average canopy Red Edge Band, etc. Cross et al. (2018) used data from an ASD spectrometer to determine canopy averages for each band. In thi s study, and extraction of canopy average values for each WorldView 3 band provided the data input necessary for the calculation of WV ACRI for the Arboretum. where; The second specialized SVI is the WorldView Red Edge Slope Weighted Index (WV RESWI) that looks specifically at the Red Edg e variability measured from the imagery. This index provides a measure of Red Edge band intensity and a measure of the reflectivity slope.

PAGE 129

114 Table 4. test significance values for discriminating the six tree = 0.05. Variation in red absorption by chlorophyll and infrared reflectivity by plant structure are important reflective ch aracteristics that define specific tree species (Xue and Su 2017; Lillesand, Kiefer, and Chipman 2008) WorldView 3 imagery and the two suggested SVIs for the six species studied. riminatory power in the traditional remote sensing bands typically used for vegetation analysis (Campbell, James and Wayne 2011) specialized SVIs constructed specifically for the WorldView 3 sensor performed as well or better than individual imagery bands in their discriminatory power. A correlation p erformed on the bands /indices listed in Table 4 provided additional information of the importance of Band Locations F test Costal 0.862 20.090 Bl ue 0.662 64.082 Green 0.717 49.514 Yellow 0.710 51.138 Red 0.626 74.983 Red Edge 0.560 98.605 Near IR1 0.500 125.583 Near IR2 0.529 111.681 WV ACRI 0.529 111.568 WV RESWI 0.485 133.042

PAGE 130

115 Table 5. Correlation matrix of the independent variables from Table 4. each band/index studied. The optimum independent data grouping for a classification would be the values with the highest discriminatory and smallest cor relation between the variables (Immitzer, Atzberger, and Koukal 2012; Thenkabail et al. 2004) In addition, Figures 4 and 5 show the WorldView 3 band reflectivity and SVI values for each tree species used in this stud y. Figure 4 shows the visible portion of the spectrum, from the Costal Band to the Red Band, and Figure 5 shows range of values from the Red Edge Band through the two SVIs analyzed for this study. The separation of the average reflectance values for the seven trees studied divided into two graphs better visualizes the variance between values, as the visible portion of the spectrum (Figure 4) is approximately one order of magnitude smaller than the other reflectivity values in Figure 5. WV 3 Bands and Indices Coastal Blue Green Yellow Red Red Edge Near IR1 Near IR2 WV ACRI WV RESWI Costal 1.000 .321 .212 .257 .326 .081 .017 .002 .058 .014 Blue 1.000 .522 .529 .621 .236 .041 .081 .202 .109 Green 1.000 .683 .586 .587 .369 .381 .580 .454 Yellow 1.000 .647 .446 .137 .183 .387 .257 Red 1.000 .153 .054 .017 .123 .004 Red Edge 1.000 .757 .817 943 .903 Near IR1 1.000 .766 .891 .942 Near IR2 1.000 .914 .837 WV ACRI 1.000 .958 WV RESWI 1.000

PAGE 131

116 Figure 5. Av erage WorldView 3 band reflectivity in the near infrared EM spectrum and SVI values for the seven tree species in this study. Figure 4. Average WorldView 3 band reflectivity in the visible EM spectrum for the seven tree species in this study.

PAGE 132

117 The necessity of the additional SVIs is evident in their ability to separate tropical forest species that are extremely close in spectral response through the original WorldView 3 bands. Values for some of these species are nearly identical in the visible and near infrared, with some minor separation in the Red Edge band and the Near IR2 band. Both SVIs performing slightly better than the original WorldView 3 bands, providing an additional set of information for species differentiation. Object ba sed Classification Within this study, a full crown measure for each species from Table 3 is the basis for the classification within the Arboretum. Each of the tree c rowns in the study area were organized into specific object segments comprising similar d igital pixel values, and the groupings will have defined edges that separate each group from other distinct groupings (van der Sande, de Jong, and de Roo 2003) These objects we re the basis of each crown in the study area, and multispectral information from each of these objects define d each as belonging to a tree canopy of a specific species. Therefore, these groupings or objects are spectrally homogeneous within their individu al regions, ideally have a distinct boundary, and the shape and size should be mostly representative of tree crowns or continual species stands within the study area. The mean spectral response for each image band or image product used of each object defi nes the canopy of each tree species. A comparison of this result to tree response data collected for each species in question determines the tree type per object, and the ground truth crown locations for particular species in the Arboretum defines the cla ssification accuracy.

PAGE 133

118 Image Segmentation The object oriented classification process requires an image segmentation to define polygons of pixels (tree crowns) with common data characteristics. The process of defining objects from pixel information is not new, as it has its roots in medical and industrial image processing, and has been applied to remotely sen sed imagery since the late 1990 s (Blaschke 2010) Several different types of segmentation processes exist but fall into two major groups: boundary based algorithms and region based algorit hms (Carleer, Debeir, and Wolff 2005) Boundary based algorithms use an edge filter combined with a threshold setting to close polygons, whereas region based algorithms are based on spect ral similarity (scale setting) and spatial similarity (merge setting). The latter is a typical process used in image processing software, such as eCognition and ENVI, and is less sensitive to slight variations in texture, which is a significant advantage when using high resolution imagery (Carleer, Debeir, and Wolff 2005) and is overall a more accurate process when characterizing tree crowns (Singh et al. 2015; Ferreira et al. 2014) The region based process within ENVI is the basis of the segmentation approach in this study. When performing a segmentation process to define objects within an image, settings for scale level and merge level must be defined appropriately based on the image resolut ion and object complexity in the landscape (Blaschke 2010) Scale level defines how many segments are created (a smaller scale level equates to more segments) and the size of each segment and the merge level setting defines the strength of merging adjacent segments. User defined thresholds f or the scale level and merge level define the segments and the success of the segmentation is dependent on the complexity of the landscape features (Carleer, Debeir, and Wolff 2005) The di scretion of the investigator through a visual

PAGE 134

119 Figure 6. False color composite of the Arboretum study area showing individual tree canopies This composite image is for compa rison of features to the segmentation provided in Figure 7. determination of the segmentation defines the settings for scale and merge level, as there is no perfectly objective segmentation approach or process (Myint et al. 2011) If either setting is too extreme the objects are not defined correctly, leading to either segments that are too complex, which subdivide known features, or assi milation of features into large polygons (Blaschke 2010)

PAGE 135

120 Figure 7 The result of the segmentation process for the Arboretum based on a setting of 11 for scale and 86 for merge within the ENVI segmentation process. The false color composite ( Figure 6 ) is for comparison of features to the segmentation shown. Each defined segmented feature represents an object with like qualities. The values of the two SVIs used, which maximize the differentiation in the imagery between features (Table 4), defined the segments within this study. Typically, a good segmentation is created through a low scale setting (creating more segments) and a higher merge setting (merging several segments together), ensuring a good characterization of features within an image. For both settings, standard values range from 0 100. Several studies working with imagery from tropical areas have defined optimal

PAGE 136

121 setting for these parameters as between 10 40 for the scale level and between 70 90 for the merge level (ENVI 2017; Myint et al. 2014; Latif and Ibrahim 2014; Yu et al. 2006; Wang et al. 2004) These optimal settings set the best possible object for tree crown definition. With initial testing of several differ ent combinations of values for scale and merge settings of 11 for the scale level and 86 for the merge level ensured good initial object differentiation and further consolidation of like objects in the Arboretum study area. Figure 6 is a false color comp osite to help in the visualization of the individual tree c rowns in the Arboretum from the result of the segmentation, shown in Figure 7 The segmentation process for this study used the SVIs from the multispectral imagery because each of the SVIs maximiz e characteristics of specific b and responses (Figures 4 and 5) in delineating vegetation (DigitalGlobe 2014; Latif and Ibrahim 2014) Other ancillary SVI information was included in the classification process to assist in differentiating tree species. Classification Settings A rule set object based classification procedure (ENVI 2017) identified the tree species based o nly on their distinctive mean crown response from the imagery bands and SVIs from segmentation output (Figure 6). No other values were included in the classification procedure (texture, shape metrics of the segments, other data) in order to g et a direct m easure of the ability of the imagery to classify tree species. After testing various data ranges to determine which would be an appropriate representation of a tree crown at a specific spectral location, a value range of +/ 5% from each mean value were s et for each WorldView 3 band and SVI. If values for a defined canopy segment fall into the +/ 5% range for any data within the rule set, then the segment classified as a particular species. The range chosen encompasses the variability that can exist wit hin a particular species ( Cross et

PAGE 137

122 al. 2018). This data range allowed for any inter species variation while still maintaining a species based consistency in mean spectral response within the defined object. The classification was applied only to those Ar boretum trees with a DBH >50 cm, considered sufficient to ensure that the tree was a part of the emergent Arboretum canopy (Chave et al. 2014) An accuracy analysis utilizing an Error Matrix compared the classification r esults to known locations of the same canopy tree species within the Arboretum. Arboretum Classification Results The first classification analysis performed assessed all bands and SVIs available for the process, excluding the Costal band and Blue band, as both of these bands are susceptible to severe atmospheric attenuation in a humid tropical environment (Lillesand, K iefer, and Chipman 2008) and provide questionable data that are not representative of the actual reflectivity of those bands on the surface. In this first classification (Classification 1), we chose the top three image bands and the top SVI (based on the Correlation values) as the data input. Within the rule set process, an assigned weight can be given to any data input, allowing specific data inputs to have more importance in the classification procedure. We assigned weights to Lambda score. This ensured that the application of each band and SVI was at a level commensurate with their ability to differentiate tree species. Table 6 shows the results from the object based classification. Cla ssification 2 (Table 7) started with the same data inputs and weights as Classification 1, but additional bands and SVIs were included based on information from Tables 4 and 5, and the visual separation shown in Figures 4 and 5. For each tree species, the

PAGE 138

123 defined that particular species in the imagery. Bands/Indices Used: Red Edge, Near IR1, Near IR2, WV RESWI Overall Accuracy = 75.61%, Kappa = 0 .685 Field Reference Classification C. elastica C. odorata C. alliodora P. macroloba P. sp. A S. micro. Total Users Acc. C. elastica 1 1 100% C. odorata 1 2 1 4 50.00% C. alliodora 1 10 11 90.91% P. macroloba 13 13 100% P. sp A 1 4 5 80.00% S. micro. 1 2 1 4 25.00% Unknown 1 1 1 3 Total 3 3 12 15 7 1 41 Prod. Acc. 33.33% 66.67% 83.33% 86.67% 57.14% 100% 75.61% Bands/Index Used: Dependent on Rule set for each Species Overall Accura cy = 85.37%, Kappa = 0.808 Field Reference Classification C. elastica C. odorata C. alliodora P. macroloba P. sp. A S. micro. Total Users Acc. C. elastica 2 2 100% C. odorata 3 1 4 75.00% C. alliodora 1 11 12 91.67% P. macroloba 13 13 100% P. sp. A 1 5 6 83.33% S. micro. 1 1 100% Unknown 1 1 1 3 Total 3 3 12 15 7 1 41 Prod. Acc. 66.67% 100% 91.67% 86.67% 71.43% 100% 85.37% Table 7. Error Matrix Classification 2. The classification results for all rule sets specified by tree species. The abbreviation S. micro. represents the tree species S. microstachyum Table 6. Error Matrix Results Classification 1. The classification results for all rule set inputs equally weighted across all bands/SVIs for all tree species. The abbreviation S. micro. represents the tree species S. microstachyum

PAGE 139

124 Figure 8 The output of Classification 2 usi ng custom rule sets for each species. Legend: C. elastica green, C. odorata blue, C. alliodora red, P. macroloba yellow, P. sp. A teal, S. microstachyum magenta. Table 7 defines the classification accuracy regarding this specific band selection in defining the tree species cor rectly for this study. Figure 8 specifies the geospatial distribution of the classification results for all emergent objects of the six tree species studied within the Arboretum and surrounding area.

PAGE 140

125 As is evident f rom the error matrix results, the WorldView 3 imagery captures the spectral variations between the tree species studied, including the intra crown variability that exists within each canopy. This canopy wide response is the key to identification of specie s in a complex forest assemblage. In comparing the two classification outputs, a rule set process defined for individual tree species based on their spectral characteristics will produce a high association of segmented tree species crowns to their verifie d GPS locations within the Arboretum. Improvement in analysis from Classification 1 to Classification 2 occurred by customizing band and SVI selection for on the additional information from Tables 4 and 5 and Figures 4 and 5. This information also deter mined the specific weights to apply to each band and SVI in the rule set. Having some known relation to other species sampled allows the rule set process to maximize the differentiation between species, improving the classification output and the ability to identify species within a complex forest environment. This high association (low errors of commission and omission) with the segmented tree species crowns and the verified GPS locations of those trees within the Arboretum in Classification 2 are encou raging. The low miss classification between species (Table 7) suggests that the segmentation parameters and rule sets chosen maximized the unique species reflectivity response from the WorldView 3 bands. Unknown tree classifications are likely due to the significant shadowing that still exists within the Arboretum, as is evident by the low Canopy Average values derived from the Arboretum (Figure 3), even when avoiding obvious shadow areas in creating the average. This inherent shadowing was a result of t he off nadir WorldView 3 image acquisition, at 26.2 zenith view angle (Table 1 ) allowing foreground tree crowns to obscure other possible emergent crowns in the background.

PAGE 141

126 The error matrix only measures classification accuracy when comparing the output of the classification to known tree species locations. Over classifying individual crowns to a specific species is possible, as the rule set process attempts to fit a tree type to a defined segment that represents all or part of a tree crown. Figure 8 il lustrates this effect, where C. alliodora and C. odorata dominate the spatial distribution of trees in and around the Arboretum. It is evident that these two species are spectrally close to other species in the study area that are not included in the clas sification. Adding more emergent species to the classification would assist in defining these over classified areas more accurately. From the analysis above it is evident that both the SVIs significantly improved classification accuracy and the segmentat ion process, with the WV RESWI outperforming WV ACRI and all other bands available for the classification. The WV RESWI performs well in showing variability between the chlorophyll production and plant structure because of its focus on the Red Edge variab ility between species (Lillesand, Kiefer, and Chipman 2008) The construction of the WV ACRI is highly dependent on a canopy average measurement. Since the Canopy Average values were low (Figure 3) due to the difficulty in removing shadows, these values likely had an effect on the overall performance of WV ACRI in this study. Conclusions and Recommendations We sought to establish a process of differentiating selected crown emergent tree species within the tropical forest regime through the implementation of an object based rule set classification schema. This process incorporated WorldView 3 high resolution image dat a appropriately processed for this task, and a series of ground truth data to verify the accuracy of the developed schema. The complexity of a tropical forest assemblage, both in

PAGE 142

127 species diversity and inter species variability, poses a great challenge to identifying individual tree species (Barbosa, Broadbent, and Bitencourt 2014) Successfully achieving a deeper understanding of the species assemblage within the tropical forest could improve our overall understanding of the role of forests in climate (Wulder et al. 2004) The results of this study show that if the proper imagery is used (WorldView 3), the imagery is corrected for illumination and atmospheric attenuation and absorption e ffects, and the appropriate segmentation and rule parameters are set, an object based rule set classification can yield extremely accurate results in a complex forest. An important aspect of this analysis is using a crown wide measure of reflectivity for the tree species in question (Cho et al. 2008) providing an object based approach that encompasses intra crown variability. This approach provides a disti nct identific ation of a species and affords an advantage to classifying species in the field compared to pixel based approaches (Immitzer, Atzberger, and Koukal 2012) We are encouraged by the results in this study, with future research including more species and a broader study area. In addition, continuing to ref ine the imagery correction processes, and improving the object based rule set classification parameters will be important steps in future analyses, which move towards accurately defining a more spatially broad tropical rainforest environment.

PAGE 143

128 Referenc es Remote Sensing of Environment 64 (3): 234 253. doi:10.1016/S0034 4257(98)00014 5. Barbosa, J. M., E. N. Br International Journal of Forestry Research 2014: 1 14. doi:10.1155/2014/715796. Bergen, K. M., S. J. Goetz, R. O. Dubayah, G. M. Henebry, C. T. Hunsaker, M. L. Imhoff, R. D Structure for Biodiversity and Habitat: Review and Implications for Lidar and Radar Journal of Geophysical Research: Biogeosciences 114 (4): 1 13. doi:10.1029/2008JG000883 ISPRS Journal of Photogrammetry and Remote Sensing 65 (1). Elsevier B.V.: 2 16. doi:10.1016/j.isprsjprs.2009.06.004. Bousquet, Lau BRDF Measurements and Model for Specular and Diffuse Components Remote Sensing of Environment 98 (2 3): 201 211. doi:10.1016/j.rse.2005.07.005. Breunig, Fabio M., L enio Soares Galvao, Yhasmin Mendes Moura, and Rafaelo Balbinout. Anais XVI Simposio Brasileiro de Sensoriamento Remoto SBSR, Foz D o Iguau, PR, Brasil, 13 a 18 de Abril de 2013, INPE no. 1986: 6917 6922. Campbell, James, B., and Randolph H. Wayne. 2011. Introduction to Remote Sensing 5th ed. New York, NY: The Guilford Press. ery High Spatial Resolution 71 (11): 1285 1294. doi:10.14358/PERS.71.11.1285. Species Identificat Photogrammetric Engineering Remote Sensing 70 (1): 135 140. doi:10.14358/PERS.70.1.135. Chave, Jerome, Maxime Rejou Mechain, Alberto Burquez, Emmanuel Chidumayo, Matthew c Global Change Biology 20 (10): 3177 3190. doi:10.1111/gcb.12629. Using Hyperspectral Indices At Leaf and Ca The International Archives of Photogrammetry, Remote Sensing and Spatial Information Science 37 (Part B7): 1 8.

PAGE 144

129 Level Differences in Hyperspectral Metrics among Tropical Rainforest Trees a s Determined by a Tree Remote Sensing 4 (6): 1820 1855. doi:10.3390/rs4061820. Remo te Sensing of Environment 96 (3 4): 375 398. doi:10.1016/j.rse.2005.03.009. Condit, Richard, Rolando Perez, and Nefertaris Daguerre. 2011. Trees of Panama and Costa Rica Princeton University Press. Cross, Matthew D., Ted Scambos, Fabio Pacifici, and Wes M Use of Metre Scale Multi Spectral Satellite Image Data for Identifying Tropical Forest International Journal of Remote Sensing 25 (11): 3723 3752 Delegido, J., J. Verrelst, C. M. Meza, J. P. Rivera, L. Alonso, Edge Spectral Index for Remote Sensing Estimation of Green LAI over European Journal of Agronomy 46. Elsevier B.V.: 42 52. doi:10.1016/j.eja.2012.12.001. DigitalGlob e, DS WV3 09/14 2. Excelis Visual Information Solutions Return Lidar in Natural Resources: R ecommendations for Project Planning, Data Remote Sensing 1 (4): 776 794. doi:10.3390/rs1040776. Fagan, Matthew E., Ruth S. DeFries, Steven E. Sesnie, J. Pablo Arroyo Mora, Carlomagno Soto, Aditya Singh, Philip A. Townsend, an Species Composition of Forests and Tree Plantations in Northeastern Costa Rica with an Remote Sensing 7 (5): 5660 5696. doi:10.3390/rs70505660. Feret, Jean Baptiste Jean IEEE Transactions on Geoscience and Remote Sensing 51 (1): 73 84. Ferreira, M.P., D.C. Zanotta, M. Zortea, T.S. Kortin g, L.M.G. Fonseca, Y.E. Shimabukuro, International Geoscience and Remote Sensing Symposium (IGARSS) 784 787. doi:10.1109/IGARSS.2014.6946541. Gao, Xian Biophysical Remote Sensing of Environment 74 (3): 609 620. doi:10.1016/S0034 4257(00)00150 4. Gastellu Etchegorry, J.P., P. Guillevic, F. Zagolksi, V. Demarez, V. Trichon, D. Deering, and Remote Sensing of Environment 68 (July 1998): 281.

PAGE 145

130 Gitelson, Anatoly A., and Mark N. Merzly Advances in Space Research 22 (5): 689 692. doi:10.1016/S0273 1177(97)01133 2. Monitoring of Ieee Geoscience and Remote Sensing Letters 5 (2): 133 137. doi:10.1109/Lgrs.2008.915598. Hair, Joseph F. Jr., William C. Black, Barry J. Babin, and Rolph E. Anderson. 2010. Multivariate Data Analysi s Seventh Ed. Pearson Prentice Hall. Classification with Random Forest Using Very High Spatial Resolution 8 Band worldView Remote Sensing 4 (9): 2661 2693. doi :10.3390/rs4092661. Journal of Forest Research 9 (1): 7 14. doi:10.1007/s10310 003 0045 z. Ke, Yinghai, Lindi J. Quackenbush, and Jungho Im. 2010. Multispectral Imagery and LIDAR Data for Object Based Forest Species Remote Sensing of Environment 114 (6). Elsevier Inc.: 1141 1154. doi:10.1016/j.rse.2010.01.002. Latif, Zulkiflee A B D, and Nurhidayah Ibra High Resolution Remotely Sensed Data Tree Species Identification Using High Resolution Remotely FIG Congress 2014 16 21. Cover Classification in a Moist Tropical Region of Brazil with Landsat Thematic Mapper International Journal of Remote Sensing 32 (23): 8207 8230. doi:10.1080/01431161.2010.532831. Lillesand, Thomas M., Ralph W. Kiefer, and Jonathan W. Chipman. 2008. Remote Sensing and Image Interpretation 6th Editio. John Wiley and Sons Ltd. 1 Satellite Sensor for Multi Proceedings of the 1 10. http://www.eoc.csiro.au/hswww/jpl_99.htm. Journal of Geophysical Research Atmospheres 97 (D17): 18935 18946. doi:10 .1029/92JD00879. Moura, Yhasmin Mendes, Lnio Soares Galvo, Joo Roberto dos Santos, Dar A. Roberts, and Inter Remote Sen sing of Environment 127. Elsevier Inc.: 260 270. doi:10.1016/j.rse.2012.09.013.

PAGE 146

131 Photogrammetric Engineering and Rem ote Sensing 80 (10): 983 993. doi:10.14358/PERS.80.10.983. Myint, S W, Patricia Gober, Anthony Brazel, Susanne Grossman Clarke, and Qihao Weng. Pixel vs. Object Based Classification of Urban Land Cover Extraction Using High Spatial Resolution Im Remote Sensing of Environment 115 (5). Elsevier Inc.: 1145 1161. doi:10.1016/j.rse.2010.12.017. Proc. JACIE 1 43. In Geoscience and Remote Sensing Symposium (IGARSS) 2016 IEEE International 1973 1975. Pacifici, Fabio, Nathan Longbotham, and William J. Emery. 2014. Physical Quantities for the Analysis of Multitemporal and Multiangular Optical Very IEEE Transactions on Geoscience and Remote Sensing 52 (10): 6241 6256. doi:10.1109/TGRS.2013.2295819. Hyperspectral Satellite Imagery for Regional Inventories: A Test with Tropical Journal of Vegetation Science 21 (2): 342 354. doi:10. 1111/j.1654 1103.2009.01147.x. Photogrammetric Engineering & Remote Sensing 69 (3): 289 297. doi:10.14358/pers.69.3.289. Ieee Geoscience and Remote Sensing 31 (3): 727 734. Remote Sensing 6 (9): 8494 8523. doi:10. 3390/rs6098494. Roy, D. P., H. K. Zhang, J. Ju, J. L. Gomez Dans, P. E. Lewis, C. B. Schaaf, Q. Sun, J. Li, H. Remote Sensing of Env ironment 176: 255 271. doi:10.1016/j.rse.2016.01.023. of Maximum Likelihood, Spectral Angle Mapper, Neural Network and Decision Tree Classifiers in Hyperspectral Image Journal of Computer Science 3 (6): 419 423. doi:10.3844/jcssp.2007.419.423.

PAGE 147

132 Characterizing Selected Canopy Tree Species at the Angkor World Heritage Site in C PLoS ONE 10 (4): 1 26. doi:10.1371/journal.pone.0121558. Thenkabail, Prasad S., Eden A. Enclona, Mark S. Ashton, Christopher Legg, and Minko Jean Ra Remote Sensing of Environment 90 (1): 23 43. doi:10.1016/j.rse.2003.11.018. IEEE Transactions on Geoscience and Remote Sensing 47 (7): 2218 2232. doi:10.1109/TGRS.2008.2010404. Classification Approach of IKONOS 2 Imagery for Land Cover Mapping to Assist Flood Risk and Flood Damag International Journal of Applied Earth Observation and Geoinformation 4 (3): 217 229. doi:10.1016/S0303 2434(03)00003 5. Parameters Using LiDAR Remote Sensin European Journal of Forest Research 129 (4): 749 770. doi:10.1007/s10342 010 0381 4. IEEE Transactions on Geoscience an d Remote Sensing 44 (7): 1931 1933. doi:10.1109/TGRS.2006.873688. Vargas, Orlando, and Enrique Castro. 2017. Species List of the Leslie R. Holdridge Arboretum Organization for Tropical Studies (OTS), Scientific Department La Selva Bilogical Station. https ://sura.ots.ac.cr/species/arboleda.php. IKONOS and QuickBird Images for Mapping Mangrove Species on the Caribbean Remote Sensing of Environment 91 (3 4): 432 440. doi:10.1016/j.rse.2004.04.005. Weyermann, Jrg, Alexander Damm, Mathias Kneubhle r, and Michael E Schaepman. 2014 IEEE Transactions on G eoscience and Remote Sensing 52 ( 1 ): 616 627 2 Vis NIR Multispectral Imagery to Support Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery XVIII, SPIE 8390 (83900N): 1 8. doi:10.1117/12.917717. Spatial Resolution Remotely Sensed Data for Ecosystem Characterizatio Bioscience 54 (6): 511 521. doi:10.1641/0006 3568(2004)054[0511:HSRRSD]2.0.CO;2.

PAGE 148

133 Yu, Qian, Peng Gong, Nick Clinton, Greg Biging, Maggi Kelly, and Dave Schirokauer. sed Detailed Vegetation Classification with Airborne High Spatial Photogrammetric Engineering and Remote Sensing 72 (7): 799 811. doi:10.14358/PERS.72.7.799. Zhang, Jinkai, Benoit Rivard, Arturo Snchez Azofeifa, and Kar en Castro Esau. 2006. and Inter Class Spectral Variability of Tropical Tree Species at La Selva, Costa Remote Sensing of Environment 105 (2): 129 141. doi:10.1016/j.rse.2006.06.01 0.

PAGE 149

134 CHAPTER V CONCLUSION AND FUTUR E RESEARCH Conclusion The dissertation results provide a significant step forward in understanding the complex tropical forest and its connection to the environment. Overall, the results highlighted the potenti al for high resolution spatial and radiometric imagery for differentiating very complex and subtle surface reflectivity responses in a complex study area. Being able to draw out specific reflective responses from the image data, if properly corrected, cou ld lead to major improvements in assessments of tree species diversity, calculations of above ground biomass, and an improved understanding of the anthropogenic effects on complex tree assemblages. A key element in the research was to envision the tree a s a complete, whole crown This broke from traditional identification and classification remote sensing techniques, which use a pixel based procedure for analysis. This object based approach provided a more realistic framework by which to delineate speci es within a complex forest. The developed process has the potential to differentiate more species beyond the focus of this dissertation, which has far reaching implications regarding accurate measures of forest diversity and biomass within a tropical fore st environment. This can ultimately lead to a better understanding of the role forest have in the overall carbon cycle, providing a more accurate insight into global climate change scenarios. The three papers that comprise this dissertation strive to es tablish a schema for proper image processing to provide accurate information for tree species identification, establish the best data grouping to use for such a purpose, and apply those findings to determine their

PAGE 150

135 effectiveness in tree species differentiat ion through an object based classification. When moving forward on developing an approach to determine if high resolution multispectral data could delineate tree species, it was evident early on that achieving a true surface reflectance from the image da ta would be essential. This required the corrections defined in this dissertation, including the determination of the best atmospheric correction possible for the tropical environment, and addressing the bi directional aspects of tree crown reflectance. Regarding atmospheric compensation procedures, other studies have not compensated for the above effects, and the humid tropical environment compromises the image data used, where atmospheric attenuation is problematic. This is especially true in short wav elengths in the visible portion of the spectrum, which are important data inputs to tree species differentiation. Bi directional reflectance with varying surface types significantly reduce the accuracy of any classification procedure performed. We used a general tree canopy measure from a previous study of a South American tropical forest, and the results showed that the inclusion of this correction is critical to accurately differentiating species. It is possible to achieve an improvement in this corre ction by determining the bi directional variances in a specific study area, or possibly using the crowns of individual tree species. The research has far reaching possibilities in using the defined process for other forest ecosystems beyond a tropical fo rest assemblage. Forest diversity issues and biomass estimate errors are not restricted to tropical environments. In addition, the appropriate data preparation, the application of appropriate image products, and applying an object based classification ap proach could assist in other forest mitigation issues, such as widespread disease, drought, and assessments of the condition and make up of urban forests. Providing

PAGE 151

136 insight into ecosystem processes assists planners, policy makers, and scientists in acquir ing a greater understanding of the importance of forests at the local, regional, and global scale. Future Research Efforts Determination of Forest Biomass The scientific community has used a variety of approaches for the identification of trees in complex forests through remote sensing techniques, with the ultimate goal of achieving an accurate biomass calculation for trees worldwide. Over the past two decades, researchers have attempted to quantify carbon storage resulting from the measured biomass in a study area. Several traditional techniques exist to measure biomass, such as small scale ground surveys of forest density (Da Silva Scaranello et al. 2012; P. Clark and Gregg 2011) specific tree species inventories with in small stands (Ebuy et al. 2011; Nvar Chidez 2010; Cordero and Kanninen 2003) and estimates of tree density over regions (Lefsky et al. 2005) Unfortunately, limitations in biomass estimates are manifest by the complexity of the forests stands and the inability to differentiate specific tree species, which have different biomass. Estimations of above ground biomass from traditional destructive t echniques typically have had errors of +/ 20%, mostly because forests have been viewed in broad groups, such as conifer and deciduous (Barbosa, Broadbent, and Bitencourt 2014) The development of allometric equations has been an important advancement in the abilit y of producing biomass estimates. The acquisition of allometrics, like the calculation of biomass, has traditionally been a through a destructive process (Cole and Ewel 2006; Segura and Kanninen 2005; Cordero and Kanninen 2003) More recently, through non destructive processes, biomass has been estimated using regression equations utilizing measured living tree crown diameter and total height (Ba naticla, Sales, and Lasco 2007) GlobAllomeTree,

PAGE 152

137 an international online repository for allometric equations supported by the Food and Agricultural Organization of the United Nations, currently has 6,300 allometric equations established for forests world wide, and 1,659 exist for South America alone (GloballomeTree 2015) Allometric equations are also available in Costa Rica for plantations (Cordero and Kanninen 2003) and trees in pasture ecosystems (Andrade, Brook, and Ibrahim 2008) Use of LiDAR A promising contemporary approach in determining above ground biomass from complex forest stands is using LiDAR systems to generate an above ground forest canopy height (Lim et al. 2003) The LiDAR height information is then used to estimate total tree volume based on known allometric relationships (Drake et al. 2003) Some LiDAR based techniques are as simple as u sing basic tree type (conifer, deciduous) and tree height (Drake et al. 2003) Others approaches include a variety of variables as part of the allometric above ground biomass estimate (Zhao et al. 2009) including the diameter at breast height (DBH), above ground canopy height, and crown diameter. The acquisition of detailed LiD AR data over a regional area to acquire forest canopy heights is cost prohibitive. A method based on a spaceborne LiDAR sensor (low density laser altimetry) would facilitate a more cost effective path to a regional to global characterization of above groun d forest canopy heights and ultimately above ground forest biomass (Wulder et al. 2012) In addition several studies have identified t he advantage of including a tree species delineation process with the LiDAR data for accurately calculating above ground biomass (Z hao, Popescu, and Nelson 2009) Unfortunately, many tree segmentation algorithms using LiDAR alone fall short of this capability. Van Leeuwen and Nieuwenhuis (2010) identified the potential of a complimentary data paring of LiDAR, fused with passive

PAGE 153

138 optical remote sensing imagery for id entification of species, to advance the accuracy of biomass models within forests stands. This suggested approach is the focus of the following proposed research effort. A Proposed Approach for Calculating Biomass A proposed three step process provides t he framework for measuring the viability of calculating a biomass in a complex tropical environment. An AOI that consists of a series of homogeneous tree stands, such as a plantation, provides the first study area. The verification of the AOI will be thr ough precise ground truth acquired at the study site and through the application of the process developed in this dissertation using WorldView 3 image data ( Zhang et al. 2006; M. L. Clark, Roberts, and Clark 2005) A LiDAR collection of the plantation will provide a tree crown height measurement (Bergen et al. 2009; Evans et al. 2009) which will be verified on the ground using a range height instrument. The species type, height measurements, and other estimated physical measurements of the tree will be input into the proper allometri c equations for calculating biomass of the plantation site (Banaticla, Sales, and Lasco 2007; Cole and Ewel 2006; Jenkins et al. 2004; Cordero and K anninen 2003) The next step will be to create a tree species inventory of varying canopy emergent species within the confines of Holdridge Arboretum at La Selva Biological Station. It is envisioned that this will be performed using the information desc ribed above in an image based segmentation/classification process. After the completion of an accurate tree inventory, an estimate of total above ground biomass will be calculated using the generated tree inventory, above ground tree heights from LiDAR, a nd established tree allometrics.

PAGE 154

139 Once the process proposed is tested and verified in a controlled study area, the application of this schema to other areas within the La Selva Biological Station assesses the accuracy of the process in a more natural fore st setting. In achieving the above tasks, an improved above ground biomass measurement of the tropical forests is expected, which could lead to an improved understanding of the role of tropical forests on the carbon cycle. Forest Diversity Assessments V arious factors, such as climate, groupings of species, predators, water, and nutrient availability affect species diversity. Because of this complexity, diversity has been a difficult measure to quantify objectively (Ghazoul 2010) so a more theoretical approach has been applied to the problem with mixed results. Through the theory of Competitive Exclusion, superior species thrive and eventually force inferior species to extinction (Ghazoul 2010) Niche Differentiation describes the process by which species, which normally compete for similar environmental resources, actively use common resources differently to coexist within an ecosystem (Adler et al. 2007) The Neutrality Theory suggest that forest develop through equal resource availability and species dispersal within an ecosystem are mostly due to f luctuations (Mikkelson 2005) Others theorize that that forest diversity will vary depending on the species under consideration, the overall ecosystem composition, and environmental conditions (Gravel et al. 2006) The likely explanation is that combinations of these theories is the appropriate explanation for most forest growt h and species distributions within the tropical landscape (Adler et al. 2007) Achieving a better understanding of the processes by which species diversify and exist within a complex forest assemblage would benefit a variety of environmental and economic efforts. A direct measure of tree species distributions, utilizing the processes

PAGE 155

140 outlined in this dissertation, would be directly applicable to achieving a better understanding of species diversity, moving from the theoretical to a grounded understanding of the processes within a complex tropical forest. Spatial Variability and Quantity Spatial variability and species quantity within the tropics is quite var iable between sites and regions, with a complex interaction of various forces driving this variability. It addition, the scale of observation can have a dramatic effect, as comparison of locations at the plot level can be quite different from the regional level (Ghazoul 2010) Gaining a complete understanding of where species are distributed and quantity helps scientists and policy makers address issues such as environmental degradation and efficient us e of tropical resources while minimizing impacts on the environment. Reasons for varied patterns of diversity include endemism, climate variations, landscape variations, disease, water availability, temperature variations, soil variations, nutrient avail ability, predation, and resistance to those (Loreau et al. 2001) Factors such as water availability, soil moisture, soil nutrient availability, and light availability have a significant e ffect on growth and mortality (Gaston 2000) Other factors, such as c rown structure and canopy openings (gaps), may have a sign ificant role in the success of species growth and propagation (Denslow 1987) When regarding environmental preservation, there is a growing realization and und erstanding of the human impacts on the environment. Changing economies, national security issues, and human health are elements that can both initiate, and be affected by environmental change (Lubchenco 1998) The research presented in this dissertation has the potential of providing a greater insight into ecosystem processes. Having an accurate measure of the spatial extent of tree

PAGE 156

141 specie s, both at the local and regional scale, could help refine our overall understanding of the changing dynamics within a complex tropical forest. This would provide the framework for a better understanding of the interactions between species and the effect of outside environmental pressures, leading to better assessments of forest health and longevity.

PAGE 157

142 References e for Ecology Letters 10 (2): 95 104. doi:10.1111/j.1461 0248.2006.00996.x. Carbon Sequestration of Silvopastoral Systems with Native Timber Species in the D ry Plant and Soil 308 (1 2): 11 22. doi:10.1007/s11104 008 9600 x. for Tropical Tree Plantation Species Using Secondary Data From the Philippines Annals of Tropical Research 29 (3): 73 90. http://www.worldagroforestrycentre.org/sea/Publications/files/paper/PP0186 06.PDF. Aboveground Biomass in Tropical Secondary Fores International Journal of Forestry Research 2014: 1 14. doi:10.1155/2014/715796. Bergen, K. M., S. J. Goetz, R. O. Dubayah, G. M. Henebry, C. T. Hunsaker, M. L. Imhoff, R. Vegetation 3 D Structure for Biodiversity and Habitat: Review and Implications for Lidar and Radar Journal of Geophysical Research: Biogeosciences 114 (4): 1 13. doi:10.1029/2008JG000883. Clark, Matthew L., Dar A. Roberts, and David B Remote Sensing of Environment 96 (3 4): 375 398. doi:10.1016/j.rse.2005.03.009. Security no. April. doi:July 01, 2016. Forest Ecology and Management 229 (1 3): 351 360. doi:10.1016/j.foreco.2006.04.017. Cordero, L.D. Perez, and M. Kannine Journal of Tropical Forest Science Da Silva Scaranello, M A, L F Alves, S A Viera, P B De Camargo, C A Joly, and L A Diameter Relationships of Tropical Atlantic Moist Forest Scientia Agricola 69 (1): 26 37. doi:10.1590/S0103 90162012000100005. Annual Review of Ecology and Systematics 18: 431 451.

PAGE 158

143 Drake, Jason B., Robert G. Knox, Ralph O. Dubayah, David B. Clark, Richard Condit, J. Ground Biomass Estimation in Closed Canopy Neotropical Forests Using Lidar Remote Sensing: Factors Affecting the Generality Global Ecology and Biogeography 12 (2): 147 159. doi:10.1046/j.1466 822X.2003.00010.x. J ournal of Tropical Forest Science 23 (2): 125 132. Return Lidar in Natural Resources: Recommendations for Project Planning, Data Remote Se nsing 1 (4): 776 794. doi:10.3390/rs1040776. Nature 405 (6783): 220 227. doi:10.1038/35012228. Ghazoul, Jaboury. 2010. Tropical Rain Forest Ecology, Diversity, and Conservation No. 577.34. Oxford U niversity Press. GloballomeTree Website www.globallometree.org. Gravel, Dominique, Charles D. Canham, Marilou Beaudet, and Christian Messier. 2006. Ecolo gy Letters 9 (4): 399 409. doi:10.1111/j.1461 0248.2006.00884.x. Jenkins, Jennifer C., David C. Chojnacky, Linda S. Heath, and Richard A. Birdsey. 2004. Based Biomass Regressions for North American doi:10.2737/NE GTR 319. Lefsky, Michael A., David J. Harding, Michael Keller, Warren B. Cohen, Claudia C. Carabajal, Fernando Del Bom Espirito Santo, Maria O. Hunter, and Raimundo De ss Using Geophysical Research Letters 32 (22): 1 4. doi:10.1029/2005GL023971. Lim, K., P. Treitz, M. A. Wulder, B. St Progress in Physical Geography 27 (1): 88 106. doi:10.1191/ 0309133303pp360ra. Loreau, M, S Naeem, P Inchausti, J Bengtsson, JP Grime, A Hector, DU Hooper, et al. 2001. Biodiversity and Ecosystem Functioning: Current Knowledge and Future Science 294 (5543): 804 808. http://apps.webofknowledg e.com/full_record.do?product=UA&search_mode=GeneralS earch&qid=1&SID=1FCcGbGcDabNdG4een5&page=1&doc=4. 16. Mikkelson, Gregory Based vs. Neutral Models of Ecological Biology & Philosophy 20 (2 3): 557 566. doi:10.1007/s10539 005 5583 7.

PAGE 159

144 Navar Tropical and Subtro pical Agroecosystems 12 (3): 507 519. http://redalyc.uaemex.mx/src/inicio/ArtPdfRed.jsp?iCve=93915170011. Biotropica 37 (1): 2 8. doi:10.1111/j.1744 7429.2005.02027.x. European Journal of Forest Research 129 (4): 749 770. doi:10.1007/s10 342 010 0381 4. Wulder, Michael A., Joanne C. White, Ross F. Nelson, Erik Naesset, Hans Ole Orka, Nicholas C. Coops, Thomas Hilker, Christopher W. Bater, and Terje Gobakken. 2012. Remote Se nsing of Environment 121. Elsevier B.V.: 196 209. doi:10.1016/j.rse.2012.02.001. Zhang, Jinkai, Benoit Rivard, Arturo Snchez Azofeifa, and Karen Castro Esau. 2006. and Inter Class Spectral Variability of Tropical Tree Species at La Selva, Costa Ri Remote Sensing of Environment 105 (2): 129 141. doi:10.1016/j.rse.2006.06.010. Biomass: A Scale Invaria Remote Sensing of Environment 113 (1). Elsevier Inc.: 182 196. doi:10.1016/j.rse.2008.09.009.

PAGE 160

145 APPENDIX MATLAB Program for Convolving ASD Data to WorldView 3 Image Bands clc clear % %Selecting the s atellite for the data correction sat = input('Enter 2 for WV2 or Enter 3 for WV3: '); disp (' ') % %Selecting the proper radiance response table (WV 2 or WV 3) for correction if sat == 2 rad = 'radiance_response_WV02.txt'; n = 9; bandname = ch ar('Pan', 'Costal', 'Blue', 'Green', 'Yellow', 'Red', ... 'Red Edge', 'Near IR', 'Near IR2'); elseif sat == 3 rad = 'radiance_response_WV03.txt'; n = 17; bandname = char('Pan', 'Costal', 'Blue', 'Green', 'Yellow', 'Red', ... 'Re d Edge', 'Near IR', 'Near IR2', 'SWave IR1', 'SWave IR2', ... 'SWave IR3','SWave IR4','SWave IR5','SWave IR6','SWave IR7', ... 'SWave IR8'); end % %Selecting the ASD tree data to be converted to WV 2 or WV 3 values treeref = input ('Enter t he ASD data [example Zygia_WV2.txt]: ', 's'); disp (' ') % %Loading the tree reflectance data into Matlab in1 = importdata(treeref); in2 = importdata(rad); % %Selecting the name of the output file treeout = input ('Enter name of output file [example text.txt]: ', 's'); disp (' ')

PAGE 161

146 %Calculate the correction % if sat == 2 %Opening up a file storing the calculated values fileID = fopen(treeout, 'w'); fprintf (fileID,'Calculated ASD Values for Worldview 2 \ n'); disp('Calculated ASD V alues for Worldview 2') for i = 1:n S = sum(in1(:,2).*in2(:,i+1)); T = sum (in2(:,i+1)); out1 = S./T; name = bandname (i,:); %Values below used for dlmwrite %calc(i) = out1; % %Printing out the result in the Command Window (15 sig digits) fprintf (name); fprintf (' = %1.15f \ n', out1); % % Printing the output to a file (15 sig digits) fprintf ( fileID, name); fprintf (fileID,' \ t %.15f \ n', out1); end fclose(fileID); elseif sat == 3 %Opening up a file storing the calculated values fileID = fopen(treeout, 'w'); fprintf (fileID,'Calculated ASD Values for Worldv iew 3 \ n'); disp('Calculated ASD Values for Worldview 3') for i = 1:n S = sum(in1(:,2).*in2(:,i+1)); T = sum (in2(:,i+1)); out1 = S./T; name = bandname (i,:); %Values below used for dl mwrite %calc(i) = out1; % %Printing out the result in the Command Window (15 sig digits) fprintf (name); fprintf (' = %1.15f \ n', out1); %

PAGE 162

147 % Printing the output to a file ( 15 sig digits) fprintf (fileID, name); fprintf (fileID,' \ t %.15f \ n', out1); end fclose(fileID); end % %Save to an ascii txt file with 15 significant digits %calc2 = calc' %treeout = input ('Enter name of output file [ example test.txt]:: ', 's'); %dlmwrite(treeout, calc2, 'delimiter', \ t', 'precision', '%.15f',... % 'newline', 'pc')