APPICATION OF REMOTE SENSING TECHNIQUES AT DIFFERENT SCALES OF OBSERVATION ON WETLAND EVAPOTRANSPIRATION By CHUNG-HSIN JUAN A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2001 Dedicated to my parents Chun-Chang and Hua-Yin ACKNOWLEDGMENTS First, I would like to express my deepest gratitude to my former advisor, the late Dr. Sun-Fu Shih, and current advisor Jonathan D. Jordan for their guidance, support, and patience. Dr. Shih passed away suddenly on June 16, 2000, which was a tragic loss for his loving family and his students. I feel mournful that Dr. Shih couldn’t share my joy at the culmination of my Ph.D. study journey. I also greatly appreciate Dr. Allen Overman, Dr. Byron Ruth, Dr. Mark T. Brown, and Dr. Bon Dewitt for serving on my graduate committee, providing knowledge and precious suggestions. Many people helped and contributed to this Ph.D. research. I am thankful to everyone at the Center for Remote Sensing who ever worked with me. The lab manager, Orlando Lanni, was generous with his time and experience; former graduate students, Chih-Hung Tan, Lei-Wen Chen, and current graduate student, John Craig, offered great assistance in many field trips. Current graduate students, Assefa Melessa, Kai-Jen Tien, and Keng-Liang Huang, encouraged and inspired me in research. I would also like to thank the St. John’s River Water Management District for installation of the lysimeter system, collecting some data, and partially sponsoring the research, especially the great assistance from Dr. Maria Mao and Ken Snyder. I also want to express my great appreciation to the Institute of Technology Development, Stennis Space Center, NASA for providing access to the hyperspectral imager. Moreover, I would like to acknowledge a very important person during these years of graduate study in United States, my best friend in Gainesville, Phillip M. Whisler. We share the sincerest friendship with all my joys and sorrows. With his friendship I was able to get through many frustrating moments. He also helped me with editing and proofreading this dissertation. Finally, I express my deepest gratitude to my parents for their many years of love, support, and encouragement. IV TABLE OF CONTENTS page ACKNOWLEDGMENTS iii LIST OF TABLES viii LIST OF FIGURES x ABSTRACT xiii CHAPTERS 1 INTRODUCTION 1 1 . 1 Importance of Wetland Evapotranspiration 1 1 .2 Applicability of Conventional ET Methods to Wetland ET 1 1.3 Limited Field Study for Wetland ET 2 1 .4 Conventional Remote Sensing Application on ET Estimation 3 1 .5 Important Vegetation Parameters for ET Estimation 4 1.5 Problems in Wetland Vegetation Mapping 7 1.6 Goals and Objectives 8 2 LITERATURE REVIEW 11 2. 1 Evapotranspiration Theories 11 2.1.1 Energy Balance Approach 12 2.1.2 Aerodynamic Vapor Transport Approach 14 2. 1 .4 Penman-Monteith Equation and Vegetation Influence on ET 1 8 2.1.5 Water Budget and Lysimeter 21 2.1.6 Wetland ET Studies 22 2.2 Remote Sensing Techniques 24 3 MATERIALS AND METHODOLOGY 33 3.1 Study Site Description 33 3.2 Fundamental Study of Wetland ET 36 3.2.1 Vegetation Parameters Measurements 36 3.2.2 Estimation of Canopy Resistance 42 3.2.2. 1 Revised procedure of measuring stomatal resistance 43 3. 2.2. 2 Revised procedure of measuring LAI 43 v 3.2.4 Methods for Evapotranspiration Estimation 43 3.2.4. 1 Modified FAO Penman method 44 3. 2. 4.2 Priestley-Taylor method 46 3. 2. 4.3 Penman-Monteith combination method 47 3.3 Spectroradiometry on the Responses of Vegetation Parameters 48 3.3.1 Monitoring of Spectral Responses 48 3.3.2 Spectral Analysis of Stomatal Resistance 5 1 3.3.3 Spectral Analysis of LAI 52 3.4 Aerial Hyperspectral Imaging 53 3.4. 1 Preparations Before Aerial Imaging 55 3.4. 1 . 1 Preparation of geolocation targets 55 3.4. 1.2 Preparation of calibration panels 56 3.4.2 Setup for Aerial Hyperspectral Imaging 57 3.4.3 Ground Truthing 57 3.4.4 Hyperspectral Image Processing and Calibration 60 3.4.4. 1 Geometric rectification 60 3.4.4.2 Radiometric calibration 61 3.4.5 Vegetation Mapping Using Aerial Hyperspectral Image 63 3.4.5. 1 Test of contingency for the selection of decision rules 64 3. 4. 5. 2 Test of separability for the selection of the most effective wavebands ... 64 3.5 Application of Satellite Images 66 3.5.1 Spectral Calibration of ETM+ Images 67 3. 5. 1.1 Calculation of spectral radiance 67 3. 5. 1.2 Calculation of at-satellite planetary reflectance 68 3.5. 1 .3 Calculation of at-satellite temperature 69 3.5.2 Vegetation Mapping Using ETM+ Images 70 3.5.2. 1 Spectral analysis of different vegetation types on the ETM+ image 70 3. 5.2. 2 Knowledge based classification 71 3.6 Accuracy Assessment of Vegetation Maps 72 3.7 Estimation of ET over the Fort Drum Marsh 74 4 RESULTS AND DISCUSSION 75 4.1 Fundamental Study of Wetland ET 75 4.1.1 Conditions of the Lysimeters 75 4. 1 .2 Vegetation Parameter Measurements 76 4. 1 .3 Estimation of Canopy Resistance 80 4. 1 .4 Methods for Evapotranspiration Estimation 86 4. 1 .4. 1 Correlations between weather parameters and ET 89 4. 1 .4.2 Evaluation of the ET Methods for sawgrass 90 4. 1.4. 3 Evaluation of the ET methods for cattail 94 4. 1 .4.4 Overall evaluation of different ET methods 98 4. 1.4.4 Verification of different ET methods 102 4.2 Spectral Radiometric Analysis 103 4.2.1 Spectral Response of Different Vegetation Types 103 4.2.2 Spectral Responses of Different Stomatal Resistance 1 10 4.2.3 Spectral Responses of Different LAI Values 1 16 vi 4.3 Hyperspectral Imaging 119 4.3. 1 Geometric Rectification 119 4.3.2 Radiometric Calibration 119 4.3.3 Vegetation Mapping Using the Aerial Hyperspectral Image 125 4.4 Application of Satellite Images 131 4.5 Accuracy Assessment of Vegetation Maps 136 4.6 Estimation of ET over the Fort Drum Marsh 139 5 CONCLUSION AND RECOMMENDATION 141 5.1 Conclusion 141 5.2 Recommendation for Future Research 143 APPENDICES A WEATHER DATA 145 B ALBUM OF FIELD WORK 157 LIST OF REFERENCES 161 BIOGRAPHICAL SKETCH 166 vii LIST OF TABLES Table Page 3.1 Spectral regions of different wavebands in the Landsat-7 ETM+ scanner 51 3.2 Typical error matrix display for accuracy assessment 73 4. 1 Vegetation parameters for sawgrass in seven field trips 77 4.2 Vegetation parameters for cattail in seven field trips 78 4.3 Correlation coefficients of weather parameters and ET values 89 4.4 Correlation coefficients between the weather parameters 89 4.5 Annual mean ET and RMSE of different estimation methods for sawgrass in 1997 94 4.6 Annual mean ET and RMSE of different estimation methods for cattail in 1997 98 4.7 Monthly estimated and lysimeter ET in 1997 99 4.8 Annual mean ET and RMSE of different ET methods in 1999 102 4.9 Computed NDVI of sawgrass and cattail from spectral measurements in each field trip 110 4.10 Different vegetation indices at different stomatal resistance values (sawgrass) 115 4. 1 1 Different vegetation indices at different stomatal resistance values (cattail) 115 4. 12 Vegetation indices of sawgrass at different LAI values 1 16 4.13 Vegetation indices of cattail at different LAI values 116 4. 14 Knowledge rules for classification of wetland vegetation using the ETM+ images 132 viii 4. 1 5 Matrix of accuracy of the classification using the hyperspectral image 137 4. 16 Matrix of accuracy of the classification using the ETM+ images 138 IX LIST OF FIGURES Figure Page 1 . 1 Illustration of the scaling-up concept 9 2.1 Resistance model within canopy (adapted from Monteith & Unsworth, 1990) 19 2.2 Typical spectral reflectance curves for vegetation, soil, and water (adapted from Lillesand & Kiefer, 1994) 27 2.3 Reflected spectra for different combination of chlorophyll, anthocyanin. (adapted from Swain & Davis, 1978) 28 2.4 Average course of reflectance, absorption, and transmittance of a green healthy plant leaf (adapted from de Boer, 1993) 29 2.5 Spectral reflectance curves of four agricultural crops (adapted from de Boer, 1993) 31 2.6 Changes in the spectral reflectance of oak leaves during the growing season (adapted from de Boer, 1993) 1 1 3.1 Location of the experimental site, (adapted from Mao & Berman, 1999) 34 3.2 Location of lysimeters and Fort Drum marsh (Mao & Berman, 1999) 37 3.3 Desired aerial hyperspectral imaging area (the yellow square) 54 3.4 Spectral reflectance of the calibration panels 58 3.5 Possible classes in the aerial hyperspectral imaging area according to the 1999 aerial image of the same fly over path 59 3.6 Atmospheric absorption effects especially at some particular wavelengths (adapted from Erdas, 1995) 62 4. 1 Stomatal resistance of sawgrass versus height 8 1 4.2 Stomatal resistance of cattail versus height 82 x 4.3 Measured and estimated LAI of sawgrass versus height 83 4.4 Measured and estimated LAI of cattail versus height 84 4.5 Computed ET of sawgrass versus different canopy resistance. The x-axis was located at the actual lysimeter ET 88 4.6 Computed ET of cattail versus different canopy resistance. The x-axis was located at the actual lysimeter ET 87 4.7 Estimated ET using Penman equation versus lysimeter ET for sawgrass in 1997 91 4.8 Estimated ET using Priestley-Taylor equation versus lysimeter ET for sawgrass in 1997 92 4.9 Estimated ET using Penman-Monteith equation versus lysimeter ET for sawgrass in 1997 93 4. 10 Estimated ET using Penman equation versus lysimeter ET for cattail in 1997 95 4. 1 1 Estimated ET using Penman equation versus lysimeter ET for cattail in 1997 96 4.12 Estimated ET using Penman-Monteith equation versus lysimeter ET for cattail in 1997 97 4.13 Reflectance curves of cattail and sawgrass measured on 19 October 1996 104 4. 14 Reflectance curves of cattail and sawgrass measured on 23 December 1996. ..105 4. 1 5 Reflectance curves of cattail and sawgrass measured on 28 March 1997 1 06 4. 17 Reflectance curves of cattail and sawgrass measured on 1 1 August 1997 108 4. 19 Spectral reflected radiance of cattail at the different values of stomatal resistance 1 1 1 4.20 Spectral reflectance of cattail at the different values of stomatal resistance. .112 4.2 1 Spectral reflected radiance of sawgrass at the different values of stomatal resistance 113 4.22 Spectral reflectance of sawgrass at the different values of stomatal resistance 114 4.23 Spectral reflectance of sawgrass at different LAI values 1 17 xi 4.24 Spectral reflectance of cattail at different LAI values 118 4.25 Hyperspectral image with the observed flight direction change 121 4.26 Aerial hyperspectral image after piecewise polynomial rectification procedure 122 4.27 Spectral reflectance of different wetland vegetation species measured by a hand-held spectroradiometer 123 4.28 Spectral reflectance of different wetland vegetation species from the spectrally calibrated hyperspectral image 124 4.29 Vegetation map of Ft. Drum marsh generated from the aerial hyperspectral image 127 4.30 Vegetation map of the Fort Drum marsh generated from the hyperspectral image using 506.8 nm, 672.3 nm, and 813.0 nm wavebands 128 4.3 1 Vegetation map of the Fort Drum marsh generated from the hyperspectral image using 515.0 nm, 672.3 nm, 721.9 nm, 837.8 nm wavebands 129 4.32 Vegetation map of the Fort Drum marsh generated from the hyperspectral image using 515.0 nm, 672.3 nm, 697.1 nm, 746.8 nm, 862.6 nm wavebands 130 4.33 Spectral reflectance of different vegetation types extracted from Feb 05, 2000 ETM+ image 133 4.34 Spectral reflectance of different vegetation types extracted from May 11, 2000 ETM+ image 134 4.35 Vegetation map of Fort Drum marsh generated from the ETM+ images of Feb 5, 2000 and May 11,2000 135 4.36 Estimated ET distribution in the Fort Drum marsh on May 1 1, 2000. The whiter color represents higher ET value 140 B. 1 Cattail lysimeter in the Fort Drum marsh 157 B.2 Weather station in the Fort Drum marsh 158 B.3. Geolocation target sitting on the water and vegetation surface 1 59 B.4. Calibration panels set up along the dike 160 xii Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy APPLICATION OF REMOTE SENSING TECHNIQUES AT DIFFERENT SCALES OF OBSERVATION ON WETLAND EVAPOTRANSPIRATION By Chung-hsin Juan May 2001 Chairman: Jonathan D. Jordan Major Department: Agricultural and Biological Engineering The establishment and maintenance of the structure and functions in wetland ecosystems is greatly influenced by hydrologic conditions. Evapotranspiration (ET) is the major output component in the hydrologic water budget. Therefore, in order to provide efficient information for water resources management and the conservation of wetland ecosystems, research on ET is urgently needed. Moreover, to overcome the variable spatial vegetation distribution and the temporal change of wetlands, appropriate remote sensing techniques are also greatly needed. The goal of this research was to study fundamental wetland ET and then with the aid of remote sensing techniques from the micro scale to the macro scale to develop useful wetland ET estimation methods. The study site was located in the Ft. Drum Marsh, Upper St. John’s River Basin in Indian River County, Florida. The site is a freshwater marsh with southern cattail ( Typha domingensis Pers.) and sawgrass ( Cladium xiii jamaicense Crantz) as the dominant vegetation species. There were four stages of the study: 1) a fundamental ET study with a lysimeter system, 2) ground measurements and analyses of spectral responses of wetland vegetation using a field spectroradiometer , 3) wetland vegetation mapping using aerial hyperspectral images, and 4) application of satellite images to delineate wetland vegetation and estimate marsh-wide ET. The results of the fundamental ET study showed the various important vegetation parameters of sawgrass and cattail. A more appropriate estimation method of canopy resistance for sawgrass and cattail was proposed. Among the various ET estimation methods, the Priestley-Taylor method was found to be most applicable. The ground spectral response measurements of sawgrass and cattail demonstrated a distinguishable difference in red wavebands and normalized difference vegetation index (NDVT), which indicated the spectral separability of the two wetland species. Leaf area index and stomatal resistance displayed a high correlation to spectral reflectance. Aerial hyperspectral imaging proved very successful in the identification of wetland vegetation species. Among all 64 wavebands, the separability tests revealed that the wavebands in the blue-green, red edge, and near-infrared spectral regions are the most important contributors for the identification of wetland vegetation species. The satellite image was applied to map wetland vegetation using the knowledge based classification method. Integrating the results from the four stages of study, the marsh-wide ET was estimated. The results of this research can have extensive application to wetland ET, wetland delineation, and remote sensing techniques. xiv CHAPTER 1 INTRODUCTION 1 . 1 Importance of W etland Evapotranspiration Evapotranspiration (ET), which is the conversion of water to vapor and the transport of that vapor away from the earth's surface into the atmosphere, accounts for a large portion of the hydrological water budget. Around 73% of Florida precipitation is returned to the atmosphere through ET (Femald and Patton, 1984). Wetland systems are the transition zones between upland systems and aquatic systems. The hydrology of wetlands has an extremely great influence on the establishment and maintenance of the structure and functions in wetland ecosystems. Even a slight change in hydrology may cause a change in the wetland ecosystem or even the degradation of the wetlands. Therefore, in order to provide efficient information for water resources management and the conservation of wetland ecosystems, research on the major hydrologic parameter, ET, is urgently needed. 1 .2 Applicability of Conventional ET Methods to Wetland ET Among the conventional methods for the calculation of reference crop ET, the most common methods include the pan evaporation (Doorenbos & Pruitt, 1977; Smith, 1992), Blaney-Criddle (Blaney & Criddle), Priestley-Taylor (Jensen et al., 1990), and Penman-Monteith (Monteith, 1990) techniques. These are the common methods for crop 1 2 ET estimation. Except for the Penman-Monteith method, they were developed empirically. ET estimate is formulated by the pan evaporation from a reference evaporation pan, from the temperature by the Blaney-Criddle method, from radiation by the Priestley-Taylor method, and from the Shih method (Shih, 1981) by both temperature and solar radiation. The Penman-Monteith method differs from the other empirical methods in that it was derived physically by utilizing micrometeorology methods. Thus, when the ET function of certain vegetation types are considered, the Penman-Monteith method is usually adopted to characterize a detailed ET process. 1 .3 Limited Field Study for Wetland ET Even though voluminous research has contributed to crop ET estimation, the applicability of the parameters concluded from crop ET research to the parameters for wetland ET may need further verification. Because wetland vegetation between terrestrial and open- water aquatic ecosystems is transitional, the characteristics of ET in wetlands should be different from those in the terrestrial systems and especially different from agricultural systems. Although, in recent decades, the importance of wetland ET has been gradually gaining recognition, the quantity of wetland ET research has been comparatively sparse. In addition, lack of consensus between researchers has lead to considerable debate regarding ET estimation (Abtew & Obeysekera, 1995). Anderson and Idso (1987) suggested that the ratio of wetland plant ET to open water evaporation was linearly related to the ratio of vegetation area to open surface area. Allen et al. (1994) found the ratio of cattail ET to potential ET to be 1 . 1 5 and their earlier research (Allen et al., 1992) showed the ratios of cattail ( Typha spp.) ET and bulrush ( Scirpus spp.) ET to 3 potential was 1.6 and 1.8, respectively. Bematowicz et al. (1976) demonstrated the ratios of cattail ET of two different species ( Typha angustifolia and Typha latifolia L .) to potential ET were 3.2 and 3.4, respectively. Having reviewed the experimental data from 25 sources, Allen et al. (1997) concluded that the wide range in ratios of wetland ET to potential ET (0.5 to 5.3) resulted from the “clothesline effect” or “oasis effect” caused by improper experimental design and then suggested that well-watered, fully vegetated surfaces in similar climatic conditions should have similar ET rates. Several different reasons could explain the divergent opinions on wetland ET proposed by these researchers, such as the different wetland types, different wetland vegetation species, the misconducts of wetland ET mechanisms, or even improper experiment design. Thus, in order to estimate wetland ET with as little bias as possible, a fundamental study with a properly-designed lysimeter system is necessary to clarify the characteristics of wetland ET and to ensure reliable data collection. 1 .4 Conventional Remote Sensing Application on ET Estimation Generally, these previously depicted ET estimation methods are based either on the concept of energy balance or empirical formulas. Regardless of which ET method is used for estimating regional ET, a sound network of weather stations or observation stations is necessary in order to collect climatic data in cooperation with vegetation distribution information. Unfortunately, both weather station and vegetation distribution information for most Florida wetlands regions is not available at fine detail. Thus, using the sparse existing information available, it is difficult to estimate a region’s wetland ET condition. The development of an alternative, therefore, is urgently needed. Moreover, 4 the accuracy and applicability limited by the spatial and temporal variability in hydrological modeling is of great concern. Measuring the characteristics of an area rather than a point, remote sensing is a unique tool that enables hydrologists to visualize the hydrological processes over different spatial locations and time periods (Engman & Gumey, 1991). Raymond and Owen-Joyce (1985) analyzed Landsat images to identify the landuse types of Palo Verde Valley, California and then estimated the total ET by combining the ET values of the different landuse types. Heimburg (1982) used the thermal band of satellite images to estimate regional ET. ET is a hydrologic phenomenon involving several climatic and land-surface parameters which usually change with time and space. Therefore, utilization of remote sensing techniques for collection of the needed spatial and temporal information has proven valuable. 1.5 Important Vegetation Parameters for ET Estimation In recent years, several ET studies have been done through the application of remote sensing. Caselles et al. (1998) utilized the combination of Landsat TM and NOAA-AVHRR images to estimate the actual ET in Spain using a semi-empirical temperature equation. Choudhury (1997) using satellite and assimilated data estimated the global pattern of potential ET utilizing the Penman-Monteith equation. Those studies determined the vegetation parameters in the equations, either by some assumptions or by the semi-empirical estimations. Therefore, their remote sensing applications to ET estimations could not fully demonstrate the differences of regional ET values where the different vegetations were mixed. Thus, when using a remote sensing application, the exploration of the deterministic relations between spectral reflectance and vegetation 5 parameters is necessary for the accurate estimation of regional ET. Therefore, in order to accurately determine vegetation parameters from spectral reflectance, a technique of ground-based remote sensing measurements associated with measurements of vegetation parameters is essential for the development of ET remote sensing methods on a micro scale and application of these methods on medium and regional scale. Three important vegetation parameters (that is, vegetation species, leaf area index, and stomatal resistance) affecting ET estimation methodology can be grouped into two categories. The first category is the vegetation parameters which are the empirical coefficients calculated from the regression of experimental data used to identify the specific species. Once the vegetation species are identified, the vegetation parameters in the first category may be determined for the specific species. The vegetation parameters in ET estimation methods, except the Penman-Monteith method, are in the first category. The ET method in the second category, the Penman-Monteith method in addition to the identification of vegetation species, requires the vegetation parameters such as leaf area index (LAI) and stomatal resistance. Because the Penman-Monteith method is a physical-based approach and is usually considered to be important to the understanding of the characteristics of the ET processes in different vegetation types, research on the estimation of the LAI and the stomatal resistance is emphasized. Therefore, the three important vegetation parameters will be studied with efforts using remote sensing techniques. Generally, different ranges of wavelengths of the reflected radiance observed using remote sensing techniques can identify different vegetation types and their growth stages (de Boer, 1993). To identify the vegetation species by using remote sensing 6 requires advance knowledge of their spectral signatures which can then be used as references. Thus, the spectral signatures of wetland vegetations need to be investigated. The normalized difference vegetation index (NDVI), defined as the combination of reflectance on red and infrared wavebands has been widely used to characterize the vegetation parameters. The relationship between NDVI and LAI has also been studied by several researchers (Curran, 1983; Liu and Huete, 1995). A linear relation for the value of NDVI lower than 3-4, has been observed (Curran, 1983; Nemani and Running, 1989). However, Curran (1983) pointed out an asymptotic regime in which NDVI increases very slowly with increasing LAI for NDVI higher than a certain threshold of 3-4. Carlson and Ripley (1997) suggested the linear relation of LAI and NDVI resulted from the variation in the fractional vegetation cover and was only available when the fractional vegetation cover was less than 100%. Thus, using NDVI to estimate LAI needs further confirmation. In addition, developing suitable vegetation indices other than NDVI may be a good alternative to estimate LAI. Very few studies have been done on the relationship between spectral reflectance and canopy resistance. Carter (1998) studied the reflectance pattern and indices for stomatal conductance of CO2 assimilation rate in pine canopies. He displayed the significant relations for the net CO2 assimilation at wavelength of 700.2 nm. Given the low availability of relevant research in this area, an experiment with measurements on the canopy resistance of H20 and the spectral reflectance is necessary. 7 1.5 Problems in Wetland Vegetation Mapping Due to the differences in the physiological structures of different species, the amount of ET changes from species to species even though the outer climatic conditions are similar. Therefore, the main key to applying remote sensing tools to wetland ET would be to accurately identify wetland vegetation species. Wetland vegetation species are very sensitive to environmental changes and wetland vegetation species may be a good indicator of the environmental changes. In a freshwater marsh, both cattail {Typha spp.) and sawgrass ( Cladium jamaicense Crantz) can grow in similar geomorphological and geographical locations and may mix together but sawgrass tends to be found in lower nutrient level conditions (Kadlec & Knight, 1996). Therefore, government agencies routinely monitor the spatial distribution of the cattail and sawgrass in freshwater marshes. However, both of the species, from a distance, look very similar and are difficult to identify without close observation. Thus, wetland delineation to the specie level usually requires more intensive tasks in the form of field surveys. In order to save the time and labor in field work needed to identify different species, remote sensing was employed as an ideal and convenient tool to serve this purpose (Doren, 1999). The most frequently used conventional remote sensed data were either satellite images or airborne color infrared (CIR) photos and images (Madden, 1999; Welch, 1996). Most satellite data, however, do not have fine enough spatial resolution to differentiate detailed ground information. Although the airborne CIR photos and images may have fine enough spatial resolution, their spectral wavebands may be too broad to identify two 8 similar looking species. Therefore, in order to properly delineate wetlands for study purposes, a new well-developed remote sensing technique is desirable. 1.6 Goals and Objectives To complete a study of remote sensing application to wetlands ET, fundamental wetland ET, spectral responses to wetland vegetation, and proper remote sensing techniques are the main keys. Therefore, the purpose of this dissertation is to integrate all aspects of fundamental research to estimate wetland ET by remote sensing methods. First, fundamental research on wetland ET with a sound lysimeter system needs to be achieved. Then, to apply remote sensing techniques, the observation, from the micro, meso, and macro scales of spectral responses of wetland vegetation is also necessary. The different scales in this research are defined by the positions of the remote sensors. If a remote sensor is operated at ground surface, then it is considered as micro scale. If remote sensed data is acquired by a low-altitude airplane it is defined as meso scale. If remote sensing is executed by a satellite sensor it is regarded as macro scale. In terms of spatial resolution used in this research, remotely sensed data of micro, meso, and macro scales is smaller than 1 meter, equal to 1 meter, and larger than 1 meter. A scaling-up concept as illustrated in Figure 1.1 was employed in this research. The concept is to study the spectral responses of wetland vegetation at ground micro-scale level, then to analyze the spectral responses using airborne images at mesoscale level based on the observation at microscale and eventually, to apply the results of the previous two scales to satellite images at the macroscale level. This concept is an integrated remote sensing research procedure and is expected to create solid linkage between ground truth situation 9 Macroscale observation Mesoscale observation Microscale observation l jjii/iiiitjiit mm iik Figure 1 . 1 Illustration of the scaling-up concept. 10 and the remotely sensed data. In summary, the following four goals were involved: 1 ) Fundamental research on the functions of wetland ET, 2) Fundamental research on the relations between the spectral reflectance of wetland vegetations and the vegetation parameters in the ET estimation equations, 3) Wetland vegetation mapping using remote sensing techniques, 4) Applications of the above three fundamental studies to the wetland ET with remotely sensed data. In order to complete the proposed research goals, the objectives for different research stages were set as follows: 1) Collect the relevant vegetation parameters, 2) Evaluate, identify and modify existing conventional ET estimation methods, 3) Implement ground-based remote sensing tools to collect spectral information for wetland vegetation species, 4) Develop proper methods to differentiate the vegetation species and to analyze the spectral characteristics of LAI and canopy resistance from the ground-based remotely sensed data, 5) Design and execute an airborne hyperspectral imaging mission, 6) Use satellite images to estimate the regional wetland ET. CHAPTER 2 LITERATURE REVIEW The study of multi-scale remote sensing application to wetland ET involves two different disciplines: 1) ET process and estimation methods; 2) remote sensing theory and methods. Therefore, the literature review first started with the concept of ET process and the different estimation methods. Some previous wetland ET studies were also included. Optical characteristics of different ground vegetation objects in different conditions were then reviewed. 2.1 Evapotranspiration Theories ET research includes the study of both ET from soil free water surfaces and transpiration from stomata of live vegetation. The complex ET process involves radiation exchanges, vapor transport, and the physiological structure and growth status of vegetation. There have been many ET studies, however most of them were focused on crop ET. Conventional ET estimation methods were derived basically from an energy balance perspective, aerodynamic vapor transport perspective, or a combination of the two. The amount of ET is usually expressed as the rate of water volume evopotranspired per unit ground surface area. The unit for this expression is Length/Time. Another common ET expression is the rate of latent heat of evapotranspiration per unit ground surface area. The unit for this expression is 11 12 Energy/Time. To avoid confusion, the terms related to ET as used in this study are defined below. Evaporation, noted as E, is the physical process by which a liquid or solid is transferred to the gaseous rate. Potential evaporation, noted as Eq, is the evaporation from a surface when all surface-atmospheric interfaces are wet so that there is no restriction on the rate of evaporation from the surface. Evapotranspiration, noted as ET, is the combined processes of evaporation and transpiration by which water is transferred from the earth’s surface to the atmosphere. Potential evapotranspiration, noted as ETo, is the rate at which water, if available, would be removed from wet soil and plant surfaces. Reference crop evapotranspiration, noted as ETr, is the rate at which water, if readily available, would be removed from the soil and plant surfaces for a given crop. When mentioning reference evapotranspiration, also noted as ETr, it means the reference evapotranspiration for grass or alfalfa of a given height. 2.1.1 Energy Balance Approach The process of ET requires a large amount of energy in order to transform water from liquid (or solid) form to gaseous form. The primary energy source is from solar radiation. Solar radiation usually supplies 80 to 100 percent of needed energy and is often the limiting factor of ET (Saxton & McGuinness, 1982). In cold, humid climates only 50 to 60 percent of the net radiation may be converted to ET, where in a hot, arid climate latent heat may exceed net radiation by 10 to 50 percent with sensible heat derived from the air and converted to ET (Jensen et al., 1990). Therefore, an energy 13 balance approach was widely developed by early researchers. A typical energy balance equation may be expressed as Rn=H + XE + G + X (2.1) and R=Rs-aRs+R,-Rlr (2.2) where Rs = incoming solar radiation (short wave), aRs = solar radiation reflected, Ri = incoming radiation (long wave), Rir = emitted long wave radiation, Rn = net radiation, H = sensible heat of air, XE = latent heat of water vapor, X = latent heat of vaporization, E = depth of evaporative water, G = soil heat, and X = miscellaneous heat sinks, like plant and air heat storage. Because the contribution of miscellaneous heat sinks is usually much less, the other three components it is often neglected and the energy balance equation is expressed as: XE-R-H-G (2.3) 14 2. 1 .2 Aerodynamic Vapor Transport Approach The measurement of water vapor, as it is transported away from an evaporating surface, offers the potential for the most direct measurement of ET. The approach usually involves measuring the vapor pressure of the air at two or more heights above the evaporating vegetation and a profile of wind velocities in order to define moisture gradients and wind transport or fluctuations of vertical velocity and humidity at a single height. All of the measurements are quite sensitive and the amount of required data is voluminous. For adiabatic profiles ET can be determined by using the eddy diffusion equation (Thomthwaite and Holzman, 1939) as «T_ ~Pak\u2 -ux){q2 -qx) E~ nvot (24) where k = von Karman constant, pa = density of air, q i = specific humidity at the position i , Ui = wind speed measured at the position i , and Zi = height at the position i . 2.1.3 Combination Approach The actual ET process is driven by both energy and vapor deficit. Therefore, an ideal ET estimation should combine an energy balance approach and aerodynamic vapor transport approach. Because the atmospheric transport mechanisms of sensible heat are similar to those of water vapor, Bowen (1926) assumed that the sensible heat 15 flux and the latent heat flux are proportional and the proportionality constant is called the Bowen ratio. (2.5) From the energy balance equation, sensible heat flux H is equivalent to (R- G) - XE. Then Equation 2.5 can be rewritten as, AE = ( R-G ) (1 -P) (2.6) If the vapor transport process and heat transfer process are assumed to be involved only in the diffusive process and only the vertical diffusive process is considered, then the latent heat flux and sensible heat flux can be expressed by the following diffusive equation, E = -p.K. (2.7) dz H = -p.CrK„ ^ (2.8) where pa - density of air, Kw = vapor eddy diffusivity, qv = specific humidity, T - temperature, Cp = specific heat at constant pressure, and Kh = heat diffusivity. When considering the gradient at near surface, the latent heat and sensible heat can be expressed in the form of two different equations as 16 (2.9) (2.10) where qa = air specific humidity near above the surface, qs = specific humidity at the surface, Ta = air temperature near above the surface, and Ts - temperature at the surface. Thus, the Bowen ratio can be rewritten by substituting E and H with Equations and the ratio Kh and Kw of the heat and vapor diffusivities are commonly taken to be 1 (Priestley and Taylor, 1972). In order to determine the Bowen ratio, measurements of temperature and vapor pressure are required at two different heights. Under the circumstance of saturated water vapor pressure, the relationship between temperature and saturated vapor pressure at that temperature is certain. Thus, the temperature parameters in Equation 2. 1 1 can be defined as the gradient of the saturated vapor pressure curve (Penman, 1948), 2.9 and 2.10 as (2.11) where y is the psychrometric constant and expressed as 0.622^ w 17 T-T A (2.12) where ea* is the corresponding saturated vapor pressure at the air temperature and e* is the corresponding saturated vapor pressure at the surface temperature. Because the surface is considered as wet and very near saturated in the combination approach, es is assumed to be equal to e* . Thus, the Bowen ratio can be depicted as another form P = Y e —e | a a e-e (2.13) a J Substitution of Equation 2.13 into Equation 2.6 yields R„-G= l + £ l A , e-L ( * v e-e. \es ~ea) (2.14) The second term of E can be considered as the result of wind effect and water vapor diffusion from surface to air. Thus, it can be described as E = f(u)(es -ea) (2.15) where f(u) is called wind function. Therefore, substituting Equation 2.15 into Equation 2.14, the Penman equation is shown as = VMe.-e.) (2.16) A+y A+y In Equation 2.16, the term before the plus sign is the part of ET contributed from energy balance and the term after the plus sign is the part of ET contributed from aerodynamic vapor transport. In the Penman combination method, the assumption is made that the humidity at surface is wet and near saturated. This assumption makes the usage of the Penman equation more convenient by reducing the measurements of temperature and vapor pressure to only one height. Thus, the Penman combination 18 method can be widely used wherever a complete weather station is installed. Because the Penman equation combines both energy balance and aerodynamic vapor transport approaches, it usually gives higher accuracy where these two approaches are both important in the real situations. Due to the assumption of a vapor-saturated surface, the Penman equation is more applicable in well-irrigated and water-sufficient conditions that are also similar to the definition of reference ET. Therefore, the Penman equation is often used to calculate reference ET. 2.1.4 Penman-Monteith Equation and Vegetation Influence to ET Penman’s combination equation assumes that water is evaporated from an open water surface or well-irrigated short grass surface. In order to introduce the saturated vapor pressure curve and the gradient A in Equation 2.12 into the derivation of the Penman equation, a main assumption of the method is that vapor pressure at the surface is saturated. Nevertheless, a vegetated surface is not usually saturated in humidity except after rainfall or dew formation (Brutsaert 1986). Thus, the assumption of saturated vapor pressure is generally not true for well-grown, tall vegetation. Monteith (1990) modified the Penman equation with the adjustment of resistance within a canopy. In Monteith’ s approach, instead of using saturated surface vapor pressure e* , the saturation value in open stomata on plant leaf surface was introduced. Monteith considered the bulk effects of open stomata at different layers within the canopy. Canopy resistance was introduced to represent the accumulated effects of each individual stomatal resistance at different layers. It can be related to the concept of bulk resistance of a set of parallel electronic resistances (Figure 2.1). When the vegetation is not actually saturated, the surface vapor pressure es is not equal to e* . Monteith adjusted saturated surface vapor pressure as 19 Figure 2.1 Resistance model within canopy (adapted from Monteith & Unsworth, 1990). 20 (2.17) where ra is the diffusion resistance of air layers and rc is the diffusion resistance within vegetation canopy. Applying the relation of Equation 2.17 instead of Equation 2.12 to substitute into the Bowen ratio (Equation 2. 1 1), the Penman-Monteith can be derived as Even though Monteith modified the Penman equation to better fit the real mass transport from vegetation and soil to atmosphere in the conditions of non-saturated surface vapor pressure and tall vegetation, how to measure and estimate the canopy resistance still took the diligent efforts of many researchers. The stomatal resistance values for a number of crops have been studied and can be found throughout the literature. However, research on the stomatal resistance for wetland vegetation is still sparse. Stomatal resistance can be also considered the external display of a vegetation’s internal conditions, such as its physiological structures, growth stages, and stress conditions. Physiological structures may vary from one vegetation species to another. Growth stages for the same vegetation species may differ from one year to another. Stress conditions may be influenced by the environmental factors such as flood, drought, freeze, heat, and et al. Therefore, the practical application of the Penman-Monteith equation may be limited to those major crops for which research has been most common, growth environmental factors are well defined, and species are monocultured. Although (2.18) where y' -y 1 + — r„ (2.20) 21 the practical application of the Penman-Monteith equation is limited, in many studies it is still considered better for revealing the process and structure of ET for different plant species. 2.1.5 Water Budget and Lvsimeter The value of ET is not directly measurable. Water budget is the main method for calculating ET. To date, the most accurate ET values were usually obtained through water budget calculation and accurately measuring other hydrological components. The water budget equation for wetlands can be expressed as (Mitsch & Gosselink. 1993) = P + Si+Gi-S0-Ga+ET (2.21) dt where S(l) = storage of water at the water level of /, Si = surface inflow, Gi = recharge from groundwater, S0 = surface outflow, and G0 = discharge into groundwater. Basinwide water budget calculation is complicated by many spatial and temporal uncertainties in the measurement of each hydrological component, so usually, its accuracy is not very high accuracy. Therefore, the use of tanks with proper design and operation is an alternative way to measure actual ET. Specific tanks for ET measurement are called lysimeters. Lysimeters have been used for over 300 years to determine water use by vegetation via the ET process. Precise lysimetry for measuring ET has been developed mainly within the past 60 years (Howell et al., 1991). Designs of lysimeters 22 vary with the type of vegetation, the surrounding environment, and the designer’s focus and style. Data recording systems were gradually improved and today, fully automatic data recording systems are common. Allen et al. (1991) specifically pointed out several important considerations in lysimeter design, such as evaporative area and vegetative area inside lysimeters, clothesline effects, and oasis effects. The evaporative area and the vegetative area inside lysimeters should represent the ratio of these two areas in the surrounding environments. If the plant heights inside and outside lysimeters are different it may cause a so-called clothesline effect and result in some inaccuracy. Moreover, if the vegetation of the lysimeters are surrounded by a large expanse of drier vegetation, then the oasis effect occurs. Therefore, a proper design of the lysimeter system is very important in obtaining actual ET values. 2.1.6 Wetland ET Studies Compared with the numerous ET studies for crops, fundamental wetland ET studies are rare. Earliest literature related to wetland vegetation lysimeter studies can be traced back to the water hyacinth ( Eichhornia crassipes (Mart.) Solms ) and cattail ( Typha latifolia L.) lysimeter study by Otis (1914). However, reported ratios of wetland ET to open- water evaporation (potential ET) of lysimeter measurements ranges from 12 to 0.87 (Allen et al., 1997). Allen et al. further pointed out that the wide range of the ratios between ET and potential ET resulted from the oasis effect and clothesline effect caused by inadequate design of the lysimeters. In the literature cited by Allen et al. (1997), some researchers evaporation measurements were obtained from standard Class A pans as the potential ET and some used measurements obtained from open water lysimeters as the potential ET. Therefore, not only the adequate design of lysimeters 23 influences the validity of wetland ET study, but also the selection of a standard potential ET for wetland ET studies effects the overall comparison of wetland ET. Conventional ET estimation methods are still widely used for wetland ET estimation. A particular direct measurement method for wetland ET is to use the diurnal cycles of groundwater or surface water in wetlands (Mitsch & Gosselink, 1993). This method can be described as ET = Sy(24h±s ) (2.22) where Sy = specific yield of aquifer (unitless), = 1.0 for standing water wetlands, < 1.0 for groundwater wetlands, h = hourly rise in water level from midnight to 4:00 A.M., 5 = net fall (+) or rise (-) of water table or water surface in one day. The method assumes that ET nears zero during midnight to 4:00 A.M. and the recharge water from the aquifer is a constant rate. A few Florida wetland studies exist. In central Florida, Dolan et al. (1984) studied the daily ET in a marshland along the Palatlakaha river near the city of Clermont. They used water-level records to calculate the daily ET from May 1977 to May 1978. They also developed a simple ET model using biomass and saturated deficiency. The highest daily ET was recorded as 10 mm day'1 in September 1977 and in February 1978, the lowest daily ET was recorded as 0.5 mm'1. The Center for Wetlands at the University of Florida has completed a series of pond cypress ( Taxodium ascendens) studies in Alachua County in north-central Florida. 24 Heimburg (1984) computed the ET of pond cypress domes using water-level records and water budget methods. Heimburg also pointed out that ET is about 80% of pan evaporation in spring and autumn dry periods, falls as low as 60% of pan evaporation during the summer wet season, and is minimal in winter. In the Everglades Nutrient Removal (ENR) project area in south Florida, Abtew et al. (1995) studied the ET of southern cattail ( Typha domingensis Pers.) in a constructed wetland for reducing phosphorus concentrations in agricultural runoff. They installed an automatic lysimeter system in the study area with careful consideration in lysimeter design to avoid oasis effects. The lysimeter was a circular polyethylene tank 3.5 m in diameter and 90 cm in depth with a 4.8 mm thickness. The lysimeter was installed about 80 m from the major internal levee. The plants and soil were transplanted from the surrounding environment. The average measurement ET rate was 3.9 mm day'1. Among the estimation methods, they found that the Penman-Monteith method showed the least error of estimation. 2.2 Remote Sensing Techniques Remote sensing is the science of obtaining information about an object area or phenomenon through the analysis of data acquired by a device that is not in contact with the object, area, or phenomenon under investigation (Lillesand & Kiefer, 1994). Therefore, remote sensing techniques involve sensors, relations between objects and sensed measurements, data interpretation, photogrammetry, data processing and analysis. When remote sensing is applied to natural environments, spectroradiometry and some 25 vegetation indices are widely used. Moreover, different resolutions in spectrum, space, and time have different applications and influence on the results. Spectroradiometry is the science that studies the spectral responses of an object, area, or phenomenon. Field spectroradiometric study allows the fundamental understanding of relations between spectral responses and studied objects or phenomena. This understanding aids in the interpretation of the remote sensed data and the design of proper sensors for desired research goals. People identify different objects by their color, size and shape. Color of an object is the spectral reflectance in visible light. The size and shape of objects may also change the spectral reflectance at a given scale of observation. Spectral reflectance is the fraction of electromagnetic energy at a given spectral waveband which is reflected by the observed objects. When electromagnetic energy projects onto an object several different interactions can happen. The interactions can be reflectance, absorption, or transmittance. By applying the principle of conservation of energy, we can state the interrelationship between these three energy interactions as (Lillesand & Kiefer, 1994) E, (A) = Er (A) + Ea (A) + Et (A) (2.23) where E/(A) = incident energy, Er(X) = reflected energy, Ea(X) = absorbed energy, and Ej{X) = transmitted energy. The type of energy that the human eye or most remote sensing sensors pick up is reflected energy or, in some cases, emitted energy. Some light or energy sources emit high amounts of electromagnetic energy but most vegetation objects do not emit 26 significant electromagnetic energy. Therefore, in vegetation remote sensing, reflected energy is still the main focus. Because the reflected energy intensity may change with the incident energy intensity spectral reflectance is used mostly to represent the reflectance characteristics of an object. Its mathematical definition is (Jenson, 2000) EM) EM) (2.24) where p>. is the spectral reflectance and is often expressed as a percentage. Spectral reflectance curves for three basic types of earth features: healthy green vegetation, dry bare soil and clear lake water are demonstrated in Figure 2.2. Water absorbs most incident energy and reflection is low. Green vegetation reflects some electromagnetic energy at the green wavelengths but reflection is high at near infrared wavelengths. Electromagnetic energy reflectance by dry bare soil is high, especially at mid-infrared wavelengths. The difference in the spectral characteristics of these three objects is significantly distinguishable. The main contributions of vegetation pigments are chlorophyll (green), xanthophylls (yellow), carotene (orange), and anthocyanins (red). Different combinations and types of these pigments readily identify different vegetation species (Figure 2.3). Especially two major types of chlorophylls which exist in most plants and have different reflected patterns for their energy needs in powering photosynthesis. Chlorophylls absorb light mainly in the blue and red regions of the spectrum. The general spectral reflectance pattern of a green healthy plant leaf is illustrated in Figure 2.4. Green-yellow chlorophyll a is present in all living plants and plays a dominant role. Higher-level plants and some green algae contain small quantities of blue-green chlorophyll b. Chlorophylls absorb most light in the visible region except that absorption 27 Figure 2.2 Typical spectral reflectance curves for vegetation, soil, and water (adapted from Lillesand & Kiefer, 1994) Reflectance (%) 28 Anthocyanin, no chlorophyll Chlorophyll — — Anthocyanin and chlorophyll Wavelength (pm) Figure 2.3 Reflected spectra for different combination of chlorophyll, anthocyanin. (adapted from Swain & Davis, 1978). 29 0/ /© Figure 2.4 Average course of reflectance, absorption, and transmittance of a green healthy plant leaf (adapted from de Boer, 1993) 30 in the green region may be slightly less than other visible regions. As a result, the absorption peaks of chlorophylls are approximately 450 nm and 650 nm, where the reflectance peak is around 540 nm. Most near-infrared light is either reflected or transmitted through leaves. The amount of reflectance in the near infrared region is mainly dependent on leaf mesophyll structure and plant health conditions (Lillesand & Kiefer, 1994). Mid-infrared light (1300-2500 nm) is mainly absorbed by the water inside plant leaves, so mid-infrared reflectance is a good indicator for the liquid water content of leaves. The water absorption peaks of mid-infrared light falls in the 1400 to 1900 nm range. Figure 2.5 is an example of how different vegetation species reflect differently. Even though the general reflectance patterns of vegetation are similar, there is still some slight but distinguishable differences between various vegetation species. The spectral characteristics of different vegetation species are usually called spectral signatures. By using spectral signatures of vegetation species one can recognize different vegetation species. Even for the same plant, spectral reflectance at different growing stages is different. Figure 2.6 shows the spectral reflectance of oak leaves from premature stage to mature stage. Vegetation spectroradiometry is essential for remote sensing in natural and agricultural environment, and it involves plant physiology. Therefore, for more precise remote sensing application, intensive studies associating vegetation conditions and the relevant spectral responses are desirable. 31 reflectance (%) ^ ^ ^ ^ Visible Near IR Mid-IR Figure 2.5 Spectral reflectance curves of four agricultural crops (adapted from de Boer, 1993) 32 reflectance (%) Figure 2.6 Changes in the spectral reflectance of oak leaves during the growing season (adapted from de Boer, 1993). CHAPTER 3 MATERIALS AND METHODOLOGY 3.1 Study Site Description The experimental site is located in the Fort Drum marsh, part of the Fort Drum Marsh Conservation Area (FDMCA), in Indian River county, Florida (27°35'12"N, 80°41'17"W, Figure 3.1). The FDMCA is the southern-most headwater area of the St. John’s River. It is bordered on the north by State Road 60, on the south by the Florida Turnpike, on the east by a levee adjacent to a private citrus grove, and on the west by the county border between Indian River and Okeechobee counties. The total area of the FDMCA is about 8,300 hectares (Mao & Rao, 1997). In the central part of the FDMCA is the Fort Drum creek floodplain swamp. The Fort Drum creek floodplain swamp is about 1,200 hectares. The dominant vegetation in the Fort Drum creek floodplain swamp is bald cypress ( Taxodium distichum ) and Florida elm ( Ulmus americana floridana (Chapm.) Little). Most of the land on the west side and southwest side of the Fort Drum creek is dry. The Fort Drum marsh is located on the east side of the Fort Drum creek. The dominant vegetation is sawgrass ( Cladium jamaicense Crantz), southern cattail ( Typha domingensis Pers.), and ludwigia ( Ludwigia spp.) The area of the Fort Drum marsh is around 2,200 hectares Soils of the Fort Drum marsh are poorly drained hydric organic soils, primarily Gator and Terra Ceia muck (Ponzio, 1995). Under natural conditions, the water table is 33 34 Figure 3.1 Location of the experimental site (adapted from Mao & Berman, 1999). 35 exposed above the surface for most of the year. The soils support vegetation that is typical of freshwater marshes. The sawgrass marsh community is the historical and desired vegetation community in this area (Mao & Bergman, 1999). Climate of the study area is subtropical. Summers are long, warm and relatively humid. Winters are generally mild because of the southern latitude and proximity to the Atlantic Ocean. The mean temperature is 22.2 C (Mao & Bergman, 1999). July and August temperatures are the warmest. December and January are the coolest months. The normal annual rainfall (1960-1990) measured at Fort Drum is 125.5 mm (Rao et al., 1995). Generally, more than 60 percent of the annual rainfall occurs during a five-month period, June through October (Mao & Bergman, 1999). A lysimeter system including three lysimeters and a weather station was installed in the Fort Drum marsh (Figure 3.2, Mao & Bergman, 1999) in 1996. The three lysimeters installed at the experimental site contain cattail (27°34'51"N, 80°41T9"W; Figure B.l), sawgrass (27°35T2"N, 80°41T8"W), and open water (27°35'23"N, 80°41T4"W). The vegetation with soil within a given lysimeter was transplanted from the surrounding area so that the lysimeter represents common environmental features. Each lysimeter, made of polyethylene having a thickness of 5 mm, has a surface area of 9.8 m and is 1.0 m deep. The bases of all three lysimeters were filled with 65 cm of subsoil from the surrounding area. Mature cattail and sawgrass plants were transplanted to the lysimeters. Marsh water was added to within 7 cm of the tank top and fluctuations in water level were controlled within a range of 3.8 cm using an automated inflow-outflow pump. Volume of inflow and outflow was recorded by the flow meters and water level was also monitored using electronic sensors. Data were recorded by data 36 loggers (Model CR10, Campbell Scientific, Inc., Logan, Utah) and transmitted via radio transmission. A weather station (27°35'21"N, 80°41'16"W) was placed adjacent to the three lysimeters (Figure B.2). Solar radiation, net radiation, photosynthetic photon flux density, air temperature, marsh water temperature, relative humidity, atmospheric pressure, wind speed, wind direction, and rainfall were continuously recorded. Wind speed at a height of 10 m above ground level was measured every 15 minute. All other parameters were measured every 5 minutes, with average values being recorded every 15 minutes. All recorded data was transmitted via radio transmission. 3.2 Fundamental Study of Wetland ET The fundamental wetland ET study involved several interests and included several procedures. First, a sound lysimeter system was installed and continuously operated. Then, proper sampling procedures and routine field trips were executed in order to measure the related vegetation parameters. The collected lysimeter data and field measurements were then analyzed statistically and different ET estimation methods were also tested for validity. 3.2.1 Vegetation Parameters Measurements Vegetative parameters related to ET estimation were measured seasonally. A sampling procedure was implemented so that the measurements could represent the total plant community within the lysimeters. Four 59 x 59 cm wood frames were placed in four quarters of each lysimeter. The vegetative parameters were sampled and measured 37 Figure 3.2 Location of lysimeters and Fort Drum marsh (Mao & Berman, 1999). 38 within each frame. This procedure not only reduces the sample size but can also maintain a nearly random sample pattern. The total sampled area is about 1/5 of the total surface area of lysimeter. The measuring procedure and definition of each parameter was described as follows: 1 . Plant canopy height The plant canopy height (in units of cm) is defined as the distance between the canopy component of interest and the water surface. Plant heights were measured by placing a measuring stick vertically adjacent to selected plants and recording the distance from the canopy component to water surface. Three measurements were recorded for each frame. The mean height was the average of the 12 measurements. 2. Plant density Plant density (in units of stem m'2) is defined as the number of plant stems per unit area. The number of plants inside each frame was counted and recorded for each of four frames. Mean plant density was calculated as the average stem count dividing by the area of the frame. 3. Air temperature and leaf surface temperature Air temperature is the temperature of air surrounding the measuring instrument, usually 1 .5 m above the ground surface. Leaf surface temperature is the temperature of the very top layer at the leaf surface. Leaf surface temperature was measured with an infrared thermometer (Model 210, Everest Interscience Inc., Tucson, Arizona). Surface temperature can be read from a thermometer LCD panel by pointing it to a leaf and 39 triggering it. The difference of surface temperature and air temperature can also be read from the thermometer LCD pane. 4. Leaf transpiration, stomatal diffusive resistance, and photosynthetic photon flux density (PPFD) Leaf transpiration (Tr) is defined as the mass rate of water vapor from the leaf per unit area (in unit of pg cm2 s'1). The LI-COR steady state porometer was designed to calculate transpiration by the following equation (LI-COR, 1984): T _ m»i A“p (3.1) where mwt is the mass rate of water vapor from the leaf and Aap is the area of the aperture in use on the cuvette of the porometer. Stomatal resistance has several definitions. The following definition was used by the LI-COR steady state porometer to calculate the stomatal resistance: „ . . Water vapor density gradient Stomatal resistance (3.2) Transpiration Photosynthetic Photon Flux Density (PPFD) is the measurement of the photosynthetically active radiation (PAR) ranged within the wavelengths of 400 to 700 nm. The unit of PPFD is micromoles per square meter per second (pmol m'2 s'1). 1 pmol m'2 s'1 = 6.023 x 1023 photons. The leaf transpiration, stomatal resistance, and Photosynthetic Photon Flux Density (PPFD) were measured with the LI-COR steady state porometer (Model 1600, LI-COR, Inc., Lincoln, Nebraska). A plant was selected randomly within each sampling frame. Six different locations of the selected plant were measured for their leaf transpiration, stomatal resistance, and PPFD. The plant was first separated 40 into three parts: inner, middle, and outer. Then, each part was divided into upper and lower portions. Therefore, measurements were made for each portion in order to describe the spatial difference of measurements. 5. Weighted Stomatal Resistance In order to compare the stomatal resistance at the same basis of mean PPFD value of several measurements, the weighted value of stomatal resistance was weighted by mean PPFD value and PPFD value at the time of measurement. Stomatal resistance was weighted by mean PPFD value and PPFD value at the time of measurement by the following equation (Abtew et al., 1995): PPFD mean X Rs, Rsw , = (3.3) PPFD, where Rswi = weighted stomatal resistance of the z-th measurement, PPFDmean = mean of all PPFD values, RSi - the z-th measurement of stomatal resistance, and PPFDi ~ the z'-th measurement of PPFD. 6. Canopy resistance The resistance for the water vapor exchange of a canopy which is in terms of integral resistance of leaf stomatal resistance and boundary resistance in each leaf. The canopy resistance can be calculated by the following equation (Abtew and Obeysekera, 1995): R c 0.5 LAI (3.4) where 41 Rc = canopy resistance, Ri = stomatal resistance, and LAI = Leaf Area Index. 7. Leaf Area Index (LAI) Leaf area index (LAI) is defined as total area of green leaves (one side only) from plants within a given area divided by the ground surface area from which these plants are grown (in units of percent or dimensionless). Leaf area index was calculated as the average leaf area per plant multiplied by the average number of plants per frame and further divided by the area of the frame. Three cattail and two sawgrass plants (entire plant) were cut and brought back from the Fort Drum marsh area to the lab for leaf area measurement. All leaves of the plant were separated and measured individually with the LI-COR portable area meter (model LI-3000, LI-COR, Inc., Lincoln, Nebraska). The LAI was calculated by the following equation: LAI = (mean leaf area per plant) * (plant density) (3.5) 8. Plant fresh biomass Plant fresh biomass is defined as the mass of fresh plant material within a defined area divided by the area size, and measured in units of gm'2. In order to retain the integrity of the lysimeter, fresh biomass was determined with plants from outside the tank. The aforementioned frame (see 3.2.1) was placed in the cattail and sawgrass marshes outside the lysimeter. All the vegetation within the frame was harvested and weight of all the vegetation was measured using a Mettler PC 440 scale. 42 9. Dry biomass The dry biomass is defined as the mass of dry plant material within a defined area divided by the area size, and measured in units of g m'2. The dry plant was obtained by oven-drying with a temperature setting at 70 degree centigrade for a 72 hour period. After the weight of fresh biomass was measured, the vegetation was then placed into an oven for 72 hours with the temperature setting at 70 °C. The weight of all the vegetation after oven-drying was measured using a Mettler PC 440 scale. The dry biomass was the vegetation weight at the end of the 72 hours dry period. 3.2.2 Estimation of Canopy Resistance During the two year monitoring period of this project, a large variability of stomatal resistance was observed. This anomaly has also been observed by other researchers (Saguaro, 1990; Abtew et al., 1995). Abtew et al. (1995) suggested using the combination of stomatal resistance in upper and lower portions of inner, middle, outer parts of leaves to represent the canopy resistance. However, problems still remain because there is still a large variability of stomatal resistance values within each portion of the leaves. Unless the leaves are separated into sufficiently small portions where the variability of stomatal resistance values can be neglected, the suggested method is not practical. Therefore, to reduce the variability of stomatal resistance, an alternative approach is needed. Some measurements of cattail stomatal resistance in Gainesville, Florida were made in April 1998. The results showed that the stomatal resistance of cattail is very likely to be linearly related to plant height. Therefore, an additional field trip (Aug. 29, 43 1998) was made using a different procedure. In this additional field trip, stomatal resistance was measured along the plant height to study the stomatal resistance of cattail and sawgrass in the Fort Drum study area. The revised procedures are described in the following paragraphs, 3.2.2. 1 Revised procedure of measuring stomatal resistance The definition of stomatal resistance and the measuring instrument is the same. However, instead of measuring stomatal resistance in the upper and lower portions of the inner, middle, and outer parts of leaves, the stomatal resistance values were measured along the leaves at locations of different heights. 3.2.2. 2 Revised procedure of measuring LAI The definition and the measuring procedure of LAI is the same LAI. However, instead of measuring the total leaf area above water surface, plant leaves were divided into several sections along different height locations and were measured. To represent the LAI contributed by the leaves above the relevant height location, LAI was calculated as the leaf area above different height locations divided by the sampling ground surface area. 3.2.4 Methods for Evapotranspiration Estimation Among the conventional methods for the calculation of reference crop ET (including pan evaporation techniques), the following methods are most common (Doorenbos and Pruitt, 1977; Smith, 1992), Priestley-Taylor (Jensen et al., 1990), and Penman-Monteith (Monteith, 1990). Because there are various forms for these methods, the forms used in this research are described in the following paragraphs. 44 3.2.4. 1 Modified FAQ Penman method In the late 1990s, The original FAO-24 Penman combination method (Doorenbos and Pruitt, 1977) was revised by several authors (Jensen et al., 1990; Smith 1992). The modified Penman combination equation has the form: *ET = -£-{Rn-G) + A + / ~~ — 6A3W f (e° - e ) A + y (3.6) Wf = aw + bwu 2 where (3.7) X = latent heat of evaporization (MJ kg'1), A = slope of vapor pressure curve (kPa°C''), y = psychrometric constant (kPa0C_1), R„ = net radiation (MJ m'2 d'1), G = ground heat flux (MJ nf2 d'1), Wf = wind function, u2 = wind speed at 2-m height (m s'1), and aw, bw = wind function coefficients. Latent heat of vaporization Latent heat of vaporization was estimated by an equation provided by Harrison (1963) as: where X = 2.501 - 2.361 x 10‘3r (3.8) T = air temperature (°C). 45 Saturation vapor pressure The saturation vapor pressure was estimated by an expression (Murray, 1967): 0 f 16.787-116.9) e = exp V T + 237.3 ) ^ where T = air temperature (°C). Slope of the saturation vapor pressure curve The slope of the saturation vapor pressure curve was calculated by differentiating the equation for saturation vapor pressure: A = 4098e° {T + 237. 3)2 (3.10) where e = saturation vapor pressure (kPa), and T = air temperature (°C). Psychrometric constant The psychrometric constant, y, represents a balance between the sensible heat gained from air and the sensible heat transformed into latent heat (Brunt, 1952) and is calculated as: where CPP 0.622A (3.11) P = atmospheric pressure (kPa); cp = specific heat of moist air at constant pressure, 1.013 kJ kg'1 °C, and X - latent heat of vaporization (kJ kg'1). 46 Penman wind function The wind function used in the Penman combination equation was determined by Wright (1982) as: Wf=aw+bwu2 (3.12) (D-mV flw =0.4 + 1. 4e ^ 58 J (3.13) (D- 243 V bw =0.605 + 0.345e ^ 80 ^ (3.14) where U2 = wind speed at 2 meter height (m s'1), and D = day of the year. 3.2.4.2 Priestlev-Tavlor method The Priestley-Taylor method is an empirical radiation-based equation with the following general form (Jensen et al., 1990): lET = a-±-(R,-G) (3.15) A + y where X = Latent heat of evaporization (MJ kg'1), a = Priestley-Taylor constant with a locally calibrated value of 1 . 1 8 for cattail (Abtew and Obeysekera, 1995), A = slope of vapor pressure curve (kPa°C‘1), y = psychrometric constant (kPa°C'1), R„ = net radiation (MJ m'2 d'1), and 47 G = ground heat flux (MJ m'2 d'1). 3.2. 4,3 Penman-Monteith Combination method The Penman-Monteith combination method included the addition of a surface and aerodynamic resistance function to the Penman equation: X ET= * A + y where and (z - d) i N ^ 3 1 In \ w J In Zom 1 > N° i k 2 uz (3.16) (3.17) where r*=r( i + -) r„ (3.18) p = atmospheric density (kg m'3), cp = specific heat of moist air (kJ kg’1 “C'1), ra = aerodynamic resistance (s m'1), rc = canopy resistance (s m'1), zw = height of the wind speed measurement (m), zp = height of the humidity and temperature measurement (m), Zom roughness length for momentum transfer (m), Zov roughness length for vapor transfer (m). 48 u2 = wind speed at height zw (m s'1), and k = von Karman’s constant, 0.41 (dimensionless). Some of the parameters can be computed as: Zom 0. 123 he, he is mean height of crop canopy in mm, zov =0.1zom, and d = (2/3) he. The above are the common methods for crop ET estimation. The Priestley-Taylor method focuses mainly on energy balance where the Penman and Penman-Monteith methods are combination methods. 3.3 Spectroradiometry on the Responses of Vegetation Parameters The spectral reflectance of different vegetation parameters was analyzed in this section. In addition to the spectral response of different vegetation types, two vegetation parameters related to ET estimation, stomatal resistance and LAI, were also considered. 3.3.1 Monitoring of Spectral Responses The plant spectral response curves were measured with a GER-1500 hand-held spectroradiometer with 512 different wavelength bands ranging from 350 nm to 1050 nm (Model GER-1500, Geophysical & Environmental Research Corp., Millbrook, New York). A Spectralon diffuse white standard plate (SRT-99-50) calibrated by Labsphere (Labsphere, Inc., North Sutton, New Hampshire) was used as a standard to calibrate the reflectance curve. The Spectralon diffuse white standard plate is made of a polytetrafluoroethylene (PTFE) compound which maintains strong spatial and spectral 49 uniformity and stability over time (Bruegge et al., 1995). It reflects 99% of incident light uniformly from ultraviolet, visible to near infrared wavelengths. The measuring process with the standard white plate is to record a measurement of the standard white plate as a reference reading and then to record a measurement of the target object with the same measuring angle. In this manner, one can compare the reflected radiance of the target object using that of the reference white plate. The spectral readings can be stored and downloaded using a RS-232 serial port. The spectroradiometer can also be operated in a laptop PC with a RS-232 serial port connection. After the spectral readings were downloaded, the spectral reflectance was calculated (in an expression of percentage) by dividing the plant spectral reading with the spectral reading of the white standard plate. Rad,. R... = Rad -x 100% (3.19) ref where R,ar = spectral reflectance of a target object (%), Rad tar = reflected radiance form a target object (W cm'2 nm'1 sr'1 x 10'10), and Radref = reflected radiance form a target object (W cm'2 nm'1 sr'1 x 10'10). From 1996 to 1997, during the field trips for measuring the vegetation parameters, the spectral reflectance of sawgrass and cattail were also measured for further analysis Normalized difference vegetation index (NDVI), Green NDVI, and different band ratios were applied to examine the spectral reflectance of cattail and sawgrass. NDVI is widely used for identifying the spectral characteristics of vegetation. The definition of NDVI is NpviJNm-R) (NIR + R) 50 (3.20) where NIR = the spectral reflectance in the near infrared region, and R = the spectral reflectance in the red region. Gitelson et al. (1996) analyzed the spectral reflectance of different concentration of algae, and found that by using a green waveband instead of a red waveband in the NDVI definition, it could better distinguish different algae concentration. Therefore, they suggested the green NDVI as Green NDVI = (NIR~G>> (3 21) (NIR + G) K J where G is the spectral reflectance in the green region. Band ratios are also commonly used to observe the features of different objects. The definition of band ratio is Band ratio = Band Band 1 2 where Band 1 and Band 2 are the wavebands selected by users. The regions of blue, green, red, and near infrared regions wavebands were selected to match the spectral regions of the wavebands in the Landsat-7 ETM+ scanner (Table 3.1; Landsat Project Science Office, 2001 ). 51 Table 3.1 Spectral regions of different wavebands in the Landsat-7 ETM+ scanner. Spectral waveband Spectral region Blue 450-520 nm Green 520-600 nm Red 630-690 nm Near infrared 760-900 nm 3.3.2 Spectral Analysis of Stomatal Resistance On May 28, 1999 a field trip was made to measure the stomatal resistance and spectral responses of cattail and sawgrass. The measuring equipment was the same as described in the previous section. In order to show significant spectral responses at different stomatal resistance, the measuring process started at 7:00 AM. Due to the increasing available solar radiation stomatal resistance decreased with the rising sun. A mature leaf was selected from each sawgrass and cattail lysimeter. In each leaf, around 15 cm from the leaf top, a point was marked and selected to be the measuring point. The dew, if present, on the measuring points was wiped off by tissues. Then, stomatal resistance was measured first, followed by the measurement of spectral reflectance. Measuring first started in the cattail lysimeter. After the measurements in the cattail lysimeter were completed, all equipment was moved to the sawgrass lysimeter and the measuring process resumed. The measuring procedures were switched and repeated back and forth between the measuring points in the sawgrass and cattail 52 lysimeters until the stomata! resistance did not show significant change. In other words, the measuring ceased when the stomatal resistance neared the lowest values. NDVI, Green NDVI, and different band ratios were also applied to examine the spectral reflectance of cattail and sawgrass. As in the previous section, the spectral regions of different wavebands were selected to match the spectral regions of the wavebands in the Landsat-7 ETM+ scanner 3.3.3 Spectral Analysis of LAI LAI is another important vegetation parameter for ET estimation. Higher LAI tends to result in higher transpiration. A field trip was made on May 26, 1999 to measure the LAI of cattail and sawgrass and the corresponding spectral reflectance. The assistance of an airboat was not available at that time and without an airboat it was difficult to measure LAI inside the marsh. Therefore, the measuring points were acquired along the dike and this limited the number of measuring points. Along the dike, as many measuring points as possible were selected. At each measuring point the spectral reflectance was measured using the GER-1500 spectroradiometer. Then, the dead leaves were cleared and the LAI of live leaves was measured. In addition, to display the spectral reflectance curves, some indices frequently used in remote sensing were entered into the analyses. As in the previous section, those indices were NDVI, green NDVI, and band ratios with different band combinations. The spectral regions of different wavebands were selected. 53 3.4 Aerial Hyperspectral Imaging In this research, aerial imaging as the meso-scale remote sensing technique is the key linkage between the ground microscale remote sensing study and the satellite macroscale remote sensing study. Therefore, the design and execution of the aerial imaging mission was important in this research. There are many different types of aerial images. Traditionally, the available remote sensed data from an airplane were aerial photos. As technology has developed, the available remote sensed data can now be obtained from different types of digital imagers. Moreover, many different new types of sensors for aerial imaging are being developed and under experiment. With the courteous assistance of the Institute of Technology Development, Stennis Space Center, NASA, flights with an experimental hyperspectral imager were employed. The hyperspectral imager is band-adjustable and can have a maximum of 128 wavebands. Due to the limitation of the transfer rate and the storage capacity, the number of available wavebands is directly related to the pixel size and therefore inversely to the imaged area. In this research, 64 wavebands with the pixel size of 1 meter and a wavelength range from 399.2 nm to 920.5 nm were selected. Moreover, because of the storage capacity of the hyperspectral imager systems, the imaging area was limited to within 1500 x 2000 meters. As a result of the limitation, the aerial imaging area was limited to the parts of the Fort Drum marsh around the lysimeters (Figure 3.3). ZJp 54 Figure 3.3 Desired aerial hyperspectral imaging area (the yellow square). 55 3.4. 1 Preparations Before Aerial Imaging Utilizing aerial hyperspectral images usually generates two main concerns. First, due to the current data transferring speed and storage capacity of storage devices, many hyperspectral imagers are not fixed-frame imaging systems. In other words, hyperspectral imagers may scan line by line along a track or scan back and forth across a track. Because of the flying instability of the airplane, keeping the imager scanning ground areas exactly along the designated route is difficult, so it usually results in more distortion than fixed-frame imagers (Lee et al., 2000). Under this circumstance, a set of dense ground control points is required. Secondly, the hyperspectral imager measures reflected radiances of ground objects which may change with respect to incoming radiances and may be compared with other images of the same objects taken under different light conditions. The more useful measuring unit to represent the signatures of measured objects is the ratio of the reflected radiance of an object to the incoming radiance. In order to convert the values of spectral radiance to spectral reflectance, a set of reference calibration panels is essential. Therefore, to complete an aerial imaging mission, a set of bright white geolocation targets and a set of calibration panels with different gray levels, from visible wavebands and infrared wavebands, were made and placed in the target imaging area. 3 .4. 1 . 1 Preparation of geolocation targets Because the imaging area was a marsh, the geolocation targets needed to be water-proof, floatable, and anchorable in the marsh. One meter by one meter white foam boards coated with plastic films were used as geolocation targets. They were used because they are water-proof, floatable and, with plastic films, are more resistant to 56 bending stress. Two wooden sticks crossing each other were affixed on the back of each white foam board to increase its resistance to wind and waves. A fishing line with 1 00 lbs resistance was tied onto the wood frame and linked to a concrete block serving as an anchor (Figure B.3). If the geolocation targets were placed adjacent to plants, then the targets were taped to the plants to ensure higher fixation. Before the imaging day, an airboat trip was taken in the Ft. Drum Marsh toplace the geolocation targets and record their coordinates using a code differential GPS unit (Model Pathfinder Pro XR, Trimble Navigation, California). 3.4. 1 .2 Preparation of calibration panels There are few companies that make large size calibration panels with an assortment of standard reflectance gray shades. Such calibration panels are ideal for this type of research but sometimes beyond the budgets of research centers at university and local government levels. Therefore, other alternative materials were used. Spectral reflectance of cloths in local textile stores were examined using the GER- 1500 spectroradiometer. An ideal fabric is expected to be able to reflect a similar percentage of incident light at each wavelength through the ultra violet, visible and near- infrared spectral regions. Most white, gray, or black cloths were found to be unsuitable and did not uniformly reflect in the near-infrared spectral region. After an extensive search through local fabric stores, some black, white and white-meshy fabrics were found to reflect uniformly from the blue waveband through near infrared wavebands. Those fabrics were purchased, sewed, and properly layered to make five calibration panels with a size of 3.7 x 3.7 meters (Figure B.4). To prevent the spectral quality from being influenced by moisture and dirt, each calibration panel was attached to a piece of black 57 plastic tarpaulin. The spectral reflectance of the calibration panels is 4%, 10%, 25%, 45%, and 56%, respectively (Figure 3.4). 3.4.2 Setup for Aerial Hyperspectral Imaging In order to obtain the most significant information from the results of the airborne hyperspectral images and their comparison to the satellite images, the ideal imaging data would be calculated on or nearest to the days when the Landsat 7 would pass over the study area. In Florida, spring is usually the least cloudy season. Thus, the imaging days around April 25 and May 11, 2000, which were Landsat 7 satellite flyover days, were considered. Before aerial imaging on the imaging day the calibration panels and extra geolocation targets were placed along the dike of the Ft. Drum Marsh. In order to avoid shadowing from the airplane and tall vegetation, the aerial imaging was taken from two different directions, north to south and west to east. 3.4.3 Ground Truthing After the aerial imaging was completed, ground truthing was performed via airboat. Based upon the results of an unsupervised classification performed in a 1999 aerial image (Figure 3.5), there were some classes of vegetation, based upon the spectral characteristics, which were identifiable. The same basic flyover pattern used in the collection of the 1999 images was followed for this second serial image collection. Each geographical location point was first measured by the GPS unit. Then, the plants as well as their growth stages were identified. The spectral reflectance was also measured using a spectroradiometer. 58 Spectral Reflectance of Calibration Panels Pend A Pend B Panel C — Pend D PenelE Figure 3.4 Spectral reflectance of the calibration panels. 59 Legend Class Names Dike Water Bush & Tree Sawgrass Cattail Unknown Unknown Unknown Figure 3.5 Possible classes in the aerial hyperspectral imaging area according to the 1999 aerial image of the same fly over path. 60 3.4,4 Hyperspectral Image Processing and Calibration Raw images usually contain some spatial distortion and the digital number in each pixel only represents the relative reflected radiance. Therefore, for further application, geometric rectification and radiometric calibration is required. 3.4,4. 1 Geometric rectification Raw digital images usually are not suitable for use as maps because they contain significant geometric distortion. These distortions may be a result of platform instability and airplane motions such as roll, pitch, and yaw, when using an along-track type hyperspectral imager. This distortion may be corrected by analyzing a set of well distributed ground control points (GCPs) in the imaging area. The actual coordinates of the GCPs need to be known and the positions (column and row) of these GCPs must be identifiable from the images. The relationship between the actual coordinates and the positions on the image can be tested using different test functions. The most desirable function is the one containing the minimum errors. Once the function is selected and the relevant coefficients are obtained, the raw image can be transferred into the map coordinate system by applying the function and coefficients. This process can be explained by the following mathematic expression (Lillesand & Kiefer, 1994) x = ft(X,Y) y = f2(XJ ) (3.20) where (x, y) = map coordinates, (X, Y) = image position (column, row), and //>/? = transformation functions. 61 Typical transformation functions can be linear or of various order polynomial functions. However, depending on the type of distortion, other functions may be applicable. 3.4.4.2 Radiometric calibration In most digital images, the digital number in each pixel may only represent relative reflected radiance. The actual reflectance in each pixel requires further calibration. The task of the radiometric calibration in hyperspectral imaging contains two steps. The first step is to calculate the wavelength of each waveband. Because the employed hyperspectral imager is tunable, the actual wavelength of each waveband may vary at every adjustment. At certain wavelengths, when solar radiation permeates the atmosphere, radiation is absorbed by some gases such as carbon dioxide, water vapor, and ozone. At those wavelengths, the actual radiance is minute (Figure 3.6). Therefore, these atmosphere-absorbed wavelengths can be easily identified from the hyperspectral images and the wavelength of each hyperspectral waveband can be calibrated using these known atmosphere-absorbed wavelengths. The employed imager uses a liquid crystal filter to spread the incoming radiance into different spectra. Because the imager is still under experiment, and of different adjustments, the spectral spreading of liquid crystal may vary. Therefore, a piecewise linear function was used for the calibration function. Because the spectral reflectance percentage from the calibration panels was known and used as references, digital numbers at each waveband in the uncalibrated image were linearly interpolated. This method can be expressed mathematically as 62 Figure 3.6 Atmospheric absorption effects especially at some particular wavelengths (adapted from Erdas, 1995). 63 IF DN e (DNi , DNm ) , then R- — — — — — x(DN-DNi)+R (3-21) dnm-dn , '' where ZW = digital number of a pixel /? = spectral reflectance of a pixel DNi = digital number of the /th reference, DNj+i = digital number of the i+lth reference, Ri = spectral reflectance of the zth reference, and Ri+i = spectral reflectance of the i+lih reference. 3.4.5 Vegetation Mapping Using Aerial Hyperspectral Image Without knowing the species, the vegetation parameters of ET estimation methods cannot be applied. Therefore, the essential task for further ET estimation is to clearly delineate sawgrass, cattail, and other wetland vegetation species After the raw image was geometrically rectified and radiometrically calibrated, the areas with known vegetation species at the ground truth points were selected from the hyperspectral image. Those selected areas were later used as training data for supervised classification. Because the classification of this study was expected to be by species and the training data were available, the supervised training method was exploited. The desired classes for identification were sawgrass, cattail, water lily ( Nymphaea spp.) and other emergent species, ludwigia (Ludwigia spp.) and other deciduous shrubs, wax myrtle ( Myrica cerifera, an evergreen shrub), and bald cypress ( Taxodium distichum). Among the emergent species in the Fort Drum marsh, the water lily was 64 dominant, so the water lily was specified in the class name. Also ludwigia was the dominant species among the deciduous shrubs in the Fort Drum marsh, so, in order to represent its dominance, it was shown in the class name. 3.4.5. 1 Test of contingency for the selection of decision rules Because the pixels of the training data are not always so homogeneous that every pixel in a training group will actually be classified to its corresponding class, selection of decision rules for the classification was based on the contingency test. The mis- classification of these distinct points may be caused by the improper utilization of decision rules or by the inherency of divergence in the data. If it is caused by the latter, systematic improvements are needed in the whole remote sensing approach, such as obtaining images with proper spectral and spatial resolution in the proper time frame. If the mis-classification is caused by the former reason, then the proper decision rule will improve the results. A contingency test is a quick classification of the training data using different decision rules. By using the selected decision rule, each test will generate a matrix of percentage of pixels that were classified as expected. Then, the decision rule with highest contingency will be selected. The test of contingency can represent the validity of the results of classification before further assessment of accuracy. 3 .4. 5 .2 Test of separability for the selection of the most effective wavebands The separability test is a statistical measurement of the distance between two classes in training data. The higher value in separability means a higher possibility of separating these two classes. The separability test can be calculated for any combination of bands used in the classification. Therefore, the separability test can test the separability of two classes as well as test separability of chosen bands. 65 Even though the hyperspectral image has the advantage of numerous narrow wavebands, the efficiency of using all those numerous wavebands may be doubted. Some wavebands may provide a crucial contribution to the classification, but some may not. Some wavebands may be redundant because the pixel values at these wavebands are very similar to those at their adjacent wavebands. Therefore, it is very possible to use fewer but more crucial wavebands to get similar results of classification. The Jeffries-Matusita (JM) distance was used to calculate the measurement of separability. The JM distance is expressed mathematically as (Swain & Davis, 1978) (C,+C ' U -/0+^in fc + Cj)/: c, + c, JM , = JA\-e-a) (3.22) (3.23) where i, j = the two classes being compared, C, = the covariance matrix of class i, Hi = the mean vector of class i, and | C, | = the determinant of Q. The value of JM distance is between 0 and 1414. The higher value means higher separability. If a value of JM distance reaches the upper boundary, then it means that the two compared classes are totally separable. Conversely, if a value of JM distance is equal to 0, then the two compared classes are inseparable. There are three steps of executing the test of separability. For example, if the selection of the three most effective wavebands is desired, the first step is to find out the possible combination of the three bands out of all 64 wavebands. The second step is to 66 calculate the JM distance of all selections of two compared classes and get the average JM distance of this waveband combination. The third step is to compare all the average JM distance of all the waveband combinations and choose the highest waveband combination. This process is tedious and the computation time increases exponentially as the number of total calculated wavebands increases. 3.5 Application of Satellite Images Application of satellite images to this study is a challenge because the study area is comparatively small. The current commercial satellite providing images of smallest spatial resolution is IKONOS which has a spatial resolution of 4 meters in its four available multispectral wavebands of blue, green, red, and near-infrared. However, comparatively, it is very expensive, the available spectral wavebands are limited, and the historical images are not available. The available satellite images of the second smallest spatial resolution are Landsat-7 ETM+, SPOT, and IRS images. They all have multispectral images with spatial resolutions between 20 to 30 meters. However, the Landsat-7 ETM+ images have additional wavebands in thermal and mid infrared regions. Considering the desired spectral information and availability of data, the Landsat-7 ETM+ images were chosen. Two Landsat-7 ETM+ (Enhanced Thematic Mapper Plus) images were acquired. One was taken on May 1 1 , 2000 which was a day before the aerial hyperspectral imaging day. Therefore, it could represent the ground situation of Fort Drum marsh for essentially the same time as the aerial hyperspectral images. Another ETM+ image was taken on Feb 05, 2000. It was kindly offered by SJRWMD. 67 3.5.1 Spectral Calibration of ETM+ Images The pixel values of raw ETM+ images are shown as the relative degree of reflected radiance from 0 to 255. Those same pixel values in different images may represent different radiance. Thus, the comparison of different images will be difficult due to differences in the original pixel values. In order to compare the different images and take advantage of spectral analyses at ground and aerial levels, the radiometric calibration of images is necessary. The task of radiometric calibration involves two steps. The first step is to calculate spectral radiance, and the second step is to calculate at-satellite planetary reflectance or at-satellite temperature for TM band 6 (thermal band). 3 . 5 . 1 . 1 Calculation of spectral radiance The MSS, TM, and ETM+ sensors and the data systems were designed to produce a linear response to incident spectral radiance (Markman & Barker, 1984; Landsat Project Science Office, 2001). Each satellite sensor has its own response functions for each waveband. The ETM+ sensor has onboard calibration lamps and temperature references. The parameters in the response functions can be calibrated using these onboard calibration lamps and temperature references. Those parameters can be later used for post-processing radiometric calibration of images. The conversion of digital numbers of each image pixel to the absolute spectral radiance can be expressed mathematically as follows (Landsat Project Science Office, 2001): L = ^ max ^rmn x DN + L- min V 255 (3.24) 68 where L = spectral radiance (mW cm'2 ster'1 pm'1), DN = digital number, Lmax max radiance when DN=255 (mW cm'2 ster'1 pm'1), and Lmin = min radiance when DN=0 (mW cm'2 ster'1 pm"1). 3.5. 1 .2 Calculation of at-satellite planetary reflectance Because the sun can be considered to constantly emit the same amount of irradiance, the solar irradiance at a given location on the earth’s surface is influenced by sun angle, sun-earth distance, and atmospheric effects. If atmospheric effects are ignored, the solar irradiance can be determined by the known sun angle and sun-earth distance. The sun-earth distance is usually expressed in astronomical units. An astronomical unit is the mean distance between the earth and the sun, approximately 149.6 x 106 km (Lillesand & Kiefer, 1994). The solar irradiance on the earth’s surface can be expressed mathematically as (Landsat Project Science Office, 2001) Sn cos 0 S = (3.25) where S = solar irradiance (mW cm'2 pm'1), So = solar irradiance at mean earth-sun distance (mW cm'2 pm'1), do = solar zenith angle, and d = earth sun distance in astronomical units. 69 If the emission from the earth objects is also neglected, the at-satellite planetary reflectance is the ratio of spectral radiance to solar irradiance, and can be expressed as (Landsat Project Science Office, 2001) 7rLd2 R = S0 cos0s (3.26) where R = at-satellite planetary reflectance. 3.5. 1.3 Calculation of at-satellite temperature After the calculation of spectral radiance of each waveband, the calculation of at- satellite temperature is different from that of at-satellite planetary reflectance. The at- satellite temperature can be calculated as the following function (Landsat Project Science Office, 2001) (3.27) where Trad = at-satellite radiant temperature (K), K2 = calibration constant 2, Ki = calibration constant 1 , and L = spectral radiance (mW cm'2 ster'1 pm'1). The temperature computed by Equation 3.27 is radiant temperature. Radiant temperature is the measurement from a thermal sensor. The actual temperature has to be adjusted by the emissivity of the measured object. The adjustment can be expressed as (Lillesand & Kiefer, 1994) T =■ 1 rad In i+i 70 where T = actual temperature (K), and £ = broadband thermal emissivity. The emissivity for freshwater marsh can be considered to be 0.99 (Tan, 1998). 3.5.2 Vegetation Mapping Using ETM+ Images Vegetation identification to species level using satellite images is much more difficult than using aerial hyperspectral images. The latter provides much better spatial and spectral resolutions than the former. The method of vegetation mapping using Landsat-7 ETM+ images in this research is not to directly use any classification scheme but to develop a feasible method based on the vegetation mapping results of hyperspectral imaging. 3.5.2. 1 Spectral analysis of different vegetation types on the ETM+ image To determine the best strategy of vegetation mapping using Landsat-7 ETM+ images, one must first understand the spectral characteristics of different vegetation types in ETM+ images. Because the vegetation types were mapped through the analysis of the aerial hyperspectral image, the satellite image within the boundary of the hyperspectral imaging area could be recognized. Thus, spectral responses of different vegetation types shown on the satellite image could be observed by extracting the spectral values of the pixels within the vegetation map generated from the hyperspectral image. 71 3. 5.2. 2 Knowledge based classification Even though the spatial and spectral resolutions of ETM+ images are much larger than those of hyperspectral images, there were some small but noticeable differences between different vegetation types from the results of the spectral analysis of different vegetation types in ETM+ images. For instance, cattail and sawgrass have close values in visible and near-IR bands but cattail has higher values in mid-IR bands. Another good example is that the ludwigia may be confused with water lily or wax myrtle in spring time (the image on May 1 1 , 2000) but leaves of ludwigia mostly drop out in winter (the image on February 05, 2000). To form a reliable set of training data, typical supervised classification requires many ground truth points. As the study site is relatively small, the appropriate number of training data points is not available for ETM+ images. However, the knowledge obtained from the spectral analysis of different vegetation types can be used in another type of classification, knowledge based classification. Knowledge based classification utilizes the knowledge of the characteristics of different classes to perform classification rules. In essence, the knowledge of classification system is a hierarchy of rules, or a decision tree, that describes the conditions under which a set of low level constituent information gets abstracted into a set of high level informational classes. The constituent information consists of user- defined variables and includes raster imagery, vector coverages, spatial models, external programs, and other data sources. Each rule is a conditional statement, or list of conditional statements, about the variable’s data values and/or attributes that determine an informational component or hypotheses. Multiple rules and hypotheses can be linked 72 together into a hierarchy that ultimately describes a final set of target informational classes or terminal hypotheses. Based on the knowledge of the results for the spectral analysis in the different vegetation types obtained in the previous section, the rules for different vegetation types were formed. Each rule utilized the spectral characteristics of each vegetation type in the spring and winter images and set up the upper and lower bound of those key wavebands. For example, if the red and mid-infrared values are both high in the winter and spring images, then it will be classified as cattail. Thus, the vegetation types over the Fort Drum Marsh were identified using the knowledge-based classification method. 3.6 Accuracy Assessment of Vegetation Maps After classification is performed, the accuracy of the classification results need to be further assessed. To assess the accuracy, another set of ground truth points is required. Ideally, the ground truth points are randomly selected and include every class of vegetation. Because the water table was very low in 2000 and 2001, an airboat was not able to penetrate some very high, dense vegetation. With the assistance of SJRWMD, two more field trips were performed using a special vehicle, a marshmaster. A marshmaster looks like a bulldozer without a blade but can float on water. Usually, a marshmaster destroys the vegetation in its path so can only be used after imaging. One field trip for collecting ground truth points in the hyperspectral imaging area was completed on April 3, 2001. The other field trip of the whole marsh was completed on April 13, 2001. 73 The accuracy assessment is usually displayed as an error matrix of reference (true) classes and mapped classes (Table 3.2). Each number in the error matrix indicates the percentage (or number of points) that are in one reference class and mapped into another class. Therefore, the diagonal of the matrix reveals the accuracy of each class. The overall accuracy can be obtained by dividing the total accurate points with the total reference points. Table 3.2 Typical error matrix display for accuracy assessment. Reference Class 1 2 q 1 P.i P12 Plq Pl+ 2 P21 P22 P2q P2+ Map q Pql Pq2 Pqq Pq+ P+1 P +2 P+q However, even if the classification is done by random assignment of pixels, there is still a random chance of accuracy. Therefore, to display the accuracy without the random chance, the Kappa coefficient is considered. The Kappa coefficient can be calculated by the following equation (Stehman, 1999), 74 p,-Tp.*p>, K = 7 (3.29) i where Pc = overall accuracy p^ - i>. , i=l P+I- = 2>fe,and *=i P,y = the percentage of points which are in the j class and mapped as the i class. 3.7 Estimation of ET over the Fort Drum Marsh The available spatial information of the Fort Drum marsh was obtained from the vegetation map and the temperature distribution came from the thermal band of ETM+ images. Therefore, the marsh-wide ET in the Fort Drum marsh was computed using this information. Since the Priestley-Taylor equation performed better than the other methods (see the results of 4.1.4), the marsh-wide ET in the Fort Drum marsh was estimated using the Priestley-Taylor equation. The temperature information was plugged into the calculation. In addition to temperature, the Priestley-Taylor equation also requires net radiation information. Unfortunately, the 2000 weather data were not available. Therefore, the required weather parameters using the historical mean values of May were used. By this approach, the ET of the Fort Drum marsh on May 1 1 , 2000 was computed. CHAPTER 4 RESULTS AND DISCUSSION 4. 1 Fundamental Study of Wetland ET 4.1.1 Conditions of the Lvsimeters Seven field visits were scheduled for vegetation monitoring and measuring in Oct. 19, 1996, Dec. 23, 1996, Mar. 28, 1997, May 30, 1997, Aug. 11, 1997, Nov. 11, 1997, Apr. 23, 1998, respectively. In each field trip, vegetation parameters were carefully measured. The growth conditions of cattail and sawgrass inside the lysimeters were similar as those outside the lysimeters. The conditions of lysimeters in each field trip were depicted as follow: On Mar. 28, 1997 70 % of cattail was found dry, but only 20 % sawgrass was dry. The cattail was flowering. On May 30, 1997 after some of the cattail died off, young cattail was growing up and had almost reached mature status. However, 50 % of the cattail was still dry, but 70 % of the leaves of the dry cattail had dropped off. The sawgrass inside the lysimeter was growing but had not reached the mature status. On Aug. 1 1, 1997 both the cattail and sawgrass grew well and reached mature status. Due to July’s abundant rainfall, the water level of the marsh was higher than the lysimeters. 75 76 On Nov. 1 1, 1997 approximately half of the cattail was dead and dry inside the lysimeter. The dead and dry situation also occurred outside the lysimeter. The sawgrass grew well and only about 20 % of sawgrass was dead. On Apr.23, 1998 about half of the cattail was dead. To date, the cattail outside the lysimeter grew to its greatest height. It was about 30 cm higher than that inside lysimeter. The reason for higher growth of the cattail outside the lysimeter may have been caused by the flood in Feb and Mar, 1998. Sawgrass grew fine. About 25% of the sawgrass was dead. Generally, areas of sawgrass and cattail died off during the winter, but the cattail tended to have a greater percentage of die off. When cattail died off, the dead leaves would form a layer of dead biomass. Because the cattail leaves have many inside pores, layers of dead cattail biomass would float and cover the water surface. 4. 1 .2 Vegetation Parameter Measurements Each vegetation parameter was properly measured and recorded. The vegetation parameters were calculated and are summarized in Tables 4. 1 and 4.2 The summarized results provide information on the characteristics of wetland vegetation. Mean canopy height of the cattail was 183.9 cm with a standard deviation of 9.3 cm and that of the sawgrass was 139.5 cm with a standard deviation of 12.6 cm. Mean density of cattail was 21.9 plants m'2 with a standard deviation of 13.9 plants m'2 and that of the sawgrass was 24.3 plants m'2 with a standard deviation of 1 .2 plants m‘2 . Mean canopy temperature of the cattail was 27.9 C with a standard deviation of 3.8 C, and that of the sawgrass was 28.0 C with a standard deviation of 5.5 C. Mean leaf area index of the cattail was 3.74 with a standard deviation of 2.14 and that of the Table 4. 1 Vegetation parameters for sawgrass in seven field trips. 77 3 !3 -a c on e o '■«-< > u Q CN VO 1/1 oo oo On o r- 00 o CO oo m vo CN CN S o S p co 00 00 NO CN p co ON (< n p in t^ iri p ro z x; (N in CN »— i m «— « H a- 'C H 0JQ S' r- ID ■rf p 1— < o «— i 00 t p vn P 2 3 ON G ON CN CN r— I CN (N CN o < CN n CN co rb mi < ON ro CN CN CN CN vo VO ON O ON < O 2 <<<<<<<<<< ^zzzzzzzzz (73 L (D 4— • % Gh 3 CO CO o G U 1-. fl G (D 00 E o o o c/3 % a, o c 3 u T3 17 4— » -G GO 'S £ X 17 T3 C CO 17 E 03 <4-1 co 17 H-l GO C/3 C/3 03 E o -C (73 u Q 2 03 > ON vo VO rn m 00 ON oo in ON rn CN d CN 00 rn CN CN On ON ON ON in rn vo o r-~ in r-; rn 00 cn ON in rn CN 00 CN CN o CN 00 CN ON d VO CN CN m CN V- 00 O, ON < ON CN CN m oo m CN 00 in CN (N oo < z oo CN CN < Z m rn rn O -'S' CN c n "C E— 2 2 2 Nov 1997 in in rn o in m oo rn vd r-~ 00 CN 00 CN CN rn in CN m in d Aug 1997 12.9 68.8 30.9 32.0 6VZ 2.48 148.5 168.1 2.95 LYZ o m >s Er cO On on I- t"- CS ON ^ ON o ON 1 o NO o o t'' <— i in On vq d f— H 1—4* 1—4* CN 00 CN rn CN ON m m r- i— i »— 1 CN in 00 o t-" CN in o ON vo 00 ON vo ON O rn CN d rn CN vd d rn 1 00 m rn t"- CN m i) On (N Q On ^ VO ON o On ~ o Os 00 o in ON 00 r-^ rn m ON CN CN 44.5 184 24.1 N/A m < z o o < Z 00 vd o < Z o rn O < z r- on r-~ n- r-^ cn < < Z Z o -4— > V E 73 E C/2 3 3 55 c ■*-* 4b .s? ‘53 £ o Q c/2 •4-* .2? *53 £ x 0.08 >0.05 >0.05 >0.25 < 0.045 <0.24 < 0.225 <0.22 <0.25 <0.2 <0.25 Near Infrared >0.18 >0.145 >0.15 >0.1 >0.165 >0.22 < 0.045 <0.24 <0.165 <0.11 <0.15 <0.16 <0.17 Mid- infrared >0.15 >0.145 >0.06 >0.14 >0.13 Criteria used in the May 1 1 , 2000 image Near- Infrared <0.25 >0.23 >0.22 >0.168 <0.17 Mid- infrared >0.14 133 Wavelength (nm) Figure 4.33 Spectral reflectance of different vegetation types extracted from Feb 05, 2000 ETM+ image. Spectral reflectance (%) 134 40% 35% 30% 25% 20% 15% 10% 5% 0% Whter Sly & other emerged species Ludwigia & other shrubs Wax Mertle Cypress Saw grass Cattail 560 660 830 1650 2215 Wavelength (nm) Figure 4.34 Spectral reflectance of different vegetation types extracted from May 1 1, 2000 ETM+ image. 135 Figure 4.35 Vegetation map of Fort Drum marsh generated from the ETM+ images of Feb 5, 2000 and May 1 1, 2000. 136 4.5 Accuracy Assessment of Vegetation Maps In the field trip on April 3, 2001, 35 ground truth points were collected for the hyperspectral image coverage area. The ground truth results for the hyperspectral vegetation map are shown in Table 4.15. The few mis-classified points were located in the mixture area, so the spectral reflectance of these points was not within any classes. In the case, it was classified to the most statistically similar class. The overall accuracy is 0.914, and the Kappa value is 0.889. Therefore, overall performance of the classification using the hyperspectral image was very accurate. In the field trip on April 13, 2001, 41 ground truth points were collected for the whole study area. The ground truth results for the Landsat vegetation map are displayed in Table 4.16. The ground truth results of cattail, sawgrass, and shrub had high accuracy of classification. Emergent species had fair classification accuracy. However, wax myrtle and cypress were not well classified. Because the spatial resolution of ETM+ images is 30 meters, pixels with one dominant species tend to be successfully identified. In the Fort Drum marsh, the dominant vegetation types are cattail, sawgrass, and shrub, so these vegetation types are more effectively classified. However, wax myrtle and cypress are present as lone specimens or sparse clumps in the Fort Drum marsh. Therefore, these two vegetation types had higher errors in classification. The overall accuracy was 0.803, and the Kappa value was 0.73 1 . Though the accuracy and the Kappa coefficient were lower than the those of using the hyperspectral image, the performance of classification using ETM+ images was still satisfactory. 137 Table 4. 15 Matrix of accuracy of the classification using the hyperspectral image. Ground truth points in each class Cattail Sawgrass Emergent Wax Cypress Shrub Total species myrtle points C/0 C/0 a o X o C c/o •4—* g ‘3 Dh