|UFDC Home||myUFDC Home | Help|
This item has the following downloads:
1 FUSION OF LANDSAT-7, IRS-1D AND RADARSAT-1 DATA FOR FLOOD DELINEATION By SIRIPON KAMONTUM A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2009
2 2009 Siripon Kamontum
3 To my family
4 ACKNOWLEDGMENTS My dissertation would not have been possibl e without kindly guidan ce and supervision of my advisor, Dr. Scot Smith and the r est of my advisory committee members: Dr. Bon Dewitt, Dr. Grenville Barnes, Dr. Amr Abd-Elrahman and Dr. Kirk Hatfield who always dedicated their precious time to educate me and gave me cr itical comments. I am truly grateful. My continued gratitude to organizations in Thailand who supplied data for my study: Geo-informatics and Space Technology Developmen t Agency (GISTDA), Computer Center of Khon Kaen University, Royal Irrig ation Department, and Department of Environmental Quality Promotion. Thanks to the Royal Thai Governme nt and GISTDA for giving me an opportunity to pursue my PhD degree. Special thanks to my friends and family fo r all supports. Huge thanks to my husband, Ped, for his unceasing love and support.
5 TABLE OF CONTENTS page ACKNOWLEDGMENTS...............................................................................................................4 LIST OF TABLES................................................................................................................. ..........7 LIST OF FIGURES.........................................................................................................................8 ABSTRACT...................................................................................................................................10 CHAP TER 1 INTRODUCTION..................................................................................................................12 Problem Statement.............................................................................................................. ....12 Objectives...............................................................................................................................15 Study Area..............................................................................................................................15 2 LITERATURE REVIEW.......................................................................................................18 Image Fusion Techniques....................................................................................................... 18 Principal Component Analysis........................................................................................ 18 Intensity-Hue-Saturation.................................................................................................19 Flood Assessment Using Remote Sensing............................................................................. 21 Landsat............................................................................................................................21 IRS...................................................................................................................................27 Radarsat...........................................................................................................................30 Flood Models..........................................................................................................................36 Flood Models and Remote Sensing........................................................................................ 41 3 MATERIALS AND METHODS........................................................................................... 53 Spatial Data................................................................................................................... ..........53 Satellite Imageries........................................................................................................... 53 GIS Data..........................................................................................................................54 Hydrological Data and HEC-RAS Model..............................................................................55 Data Preprocessing............................................................................................................. ....55 Speckle Noise Suppression..............................................................................................56 Geometric Co rrection ...................................................................................................... 57 HEC-RAS Data Preparation............................................................................................ 58 Data Analysis..........................................................................................................................59 Calculation of Water Surface Elevation Using HEC-RAS ............................................. 59 Radarsat Image Interpretation......................................................................................... 61 PCA.................................................................................................................................62 Pan-sharpening................................................................................................................62 Classification of Fused Image.........................................................................................63
6 4 RESULTS...............................................................................................................................72 HEC-RAS and HEC-GeoRAS Results................................................................................... 72 Flood Maps Interpreted from Radarsat Images...................................................................... 72 Comparison of Radarsat Flood Maps Referencing HEC-RAS Flood Map ............................ 72 PCA Results............................................................................................................................74 Fused Image and Its Classification......................................................................................... 75 Comparison of Flood Maps from Fused Im age and Best Radarsat Image.............................77 Underwater Feature Detection................................................................................................ 79 5 DISCUSSION.........................................................................................................................94 Flood Prediction Using HEC-RAS......................................................................................... 94 Radarsat W1 versus Radarsat S7............................................................................................ 95 The Fused Image.....................................................................................................................96 6 CONCLUSIONS.................................................................................................................. 100 REFERENCES............................................................................................................................103 BIOGRAPHICAL SKETCH.......................................................................................................112
7 LIST OF TABLES Table page 2-1 Sensors and bands of Landsat-1 to 7.................................................................................. 52 2-2 Description of Rada rsat im aging modes............................................................................ 52 3-1 RMS residual errors of the GCPs.......................................................................................71 3-2 Estimated Mannings n values........................................................................................... 71 4-1 Error matrix of classification of Radarsat W1 image........................................................ 91 4-2 Accuracy totals of classification of Radarsat W1 image................................................... 91 4-3 Kappa statistics of classification of Radarsat W1 image................................................... 91 4-4 Error matrix of classification of Radarsat S7 image.......................................................... 91 4-5 Accuracy totals of classification of Radarsat S7 image.....................................................91 4-6 Kappa statistics of classification of Radarsat S7 image.....................................................91 4-7 Eigen values and eigen vectors.......................................................................................... 92 4-8 Percentage of total scene variance..................................................................................... 92 4-9 Error matrix of flood map interp reted from Radarsat S7 image........................................ 93 4-10 Accuracy totals of flood map inte rpreted from Radarsat S7 image................................... 93 4-11 Kappa statistics of flood map inte rpreted from Radarsat S7 image................................... 93 4-12 Error matrix of flood map in terpreted from fused image.................................................. 93 4-13 Accuracy totals of flood map interpreted from fused image............................................. 93 4-14 Kappa statistics of flood map interpreted from fused image............................................. 93
8 LIST OF FIGURES Figure page 1-1 Study area................................................................................................................. ..........17 2-1 Radarsat imaging modes.................................................................................................... 51 3-1 Radarsat W1 image.......................................................................................................... ..65 3-2 IRS-1D image............................................................................................................... .....65 3-3 Radarsat S7 image.......................................................................................................... ....66 3-4 Landsat-7 image............................................................................................................ .....66 3-5 Topographic maps........................................................................................................... ...67 3-6 Hydrological stations...................................................................................................... ...67 3-7 Flowchart of my study...................................................................................................... .68 3-8 Geometry data in HEC-RAS.............................................................................................. 69 3-9 Histogram of Radarsat W1................................................................................................. 70 3-10 Histogram of Radarsat S7.................................................................................................. 70 3-11 Spectral bands of Landsat -7, IRS-1D and Radarsat-1 ....................................................... 70 4-1 Flooded areas and water surface elevation simulated by HEC-RAS................................. 81 4-2 Radarsat images............................................................................................................ .....81 4-3 Overlaid flooded areas from HEC-RAS, Radarsat W 1 and Radarsat S7 images.............. 82 4-4 Inputs of PCA....................................................................................................................82 4-5 Principal component images and per centages of the tota l scene variance......................... 83 4-6 Fused image of PC4 (red), Radarsat S7 (green), and PC1 (blue) im ages.......................... 84 4-7 Fused image classification................................................................................................. 85 4-8 Water surface elevation measur ed from hydrological stations..........................................86 4-9 Dead vegetation (area X), built up water body (area Y), and a row of trees (area Z). ...... 86 4-10 Flooded areas of each land cover type............................................................................... 87
9 4-11 Differences caused by changing in water surface elevation.. ............................................88 4-12 Overlaid flooded areas from the reduced flood map of Radarsat S7 and the fused im age flood map................................................................................................................ 89 4-13 Inspection of underwater features...................................................................................... 90 5-1 Radarsat S7 images......................................................................................................... ...99
10 Abstract of Dissertation Pres ented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy FUSION OF LANDSAT-7, IRS-1D AND RADARSAT-1 DATA FOR FLOOD DELINEATION By Siripon Kamontum August 2009 Chair: Scot E. Smith Major: Forest Resources and Conservation In my study, the extent of areas inundated by flood water was delineated. Landsat-7, IRS-1D and Radarsat-1 images were used to create a fused image containing essential information of a flood occurring in 2002 in Thai land. This is the first study to address flood analysis using Landsat-7, IRS-1D and Radarsat-1data acquired during the same flood event. Two flood maps interpreted from two Radarsat images acquired in W1 and S7 modes were compared against a flood map simulated by HEC-RAS, a hydr aulic model that calculates water surface elevation from hydrological and t opographic data. It was found that classification accuracies of the flood map interpreted from the Radarsat S7 im age were higher than those from the Radarsat W1 image with an overall Kappa value of 88.0 pe rcent for S7 compared to 75.2 percent for W1. Therefore, the Radarsat S7 image was chosen to be the input of a pan-sharpening process. Next, principal component analysis wa s applied to Landsat and IRS da ta to reduce data redundancy. First principal component (PC1) image that c ontained 63.5 percent of th e total scene variance and clearly portrayed flood boundaries and four th principal component (PC4) image that contained 4.9 percent of the to tal scene variance and showed details in flooded areas were selected to be two additional inputs of the pan-sharpening process. A combination of the
11 Radarsat S7, PC1 and PC4 images was sharpe ned to a five meter ground sampled distance by fusion with the IRS panchromatic image. Th e boundary of the 2002 flood was successfully delineated from the fused image. Classification a ccuracies of flooded area s interpreted from the fused image were slightly higher than those inte rpreted from the Radarsat S7 image alone. The overall Kappa value was 93.2 percent versus 90.8 pe rcent when visual inte rpretation was used as a reference. Damaged areas due to the flood a nd flooded transportation ro utes were identified from the fused image. Underwater features c ould not be detected using the fused image.
12 CHAPTER 1 INTRODUCTION Problem Statement Flooding is both a natural and a m an-made phe nomenon that can damage the environment, property and destroy life. As more people decide to live in floodprone areas such as deltas and river basins, it is important to accurately a ssess the location and boundari es of flooded land so that relief efforts can be coordinated. To facilit ate this process, remote sensing data have been utilized, such as Landsat (Green et al. 1983; Townsend and Walsh, 1998; Frazier and Page, 2000; Ryu et al. 2002; Wang et al. 2002; Hudson and Colditz, 2003; Wang, 2004; Neuenschwander et al. 2005), SPOT (Toyra and Pietroniro, 2005), and IRS (Kar, 1994; Sharma et al. 1996; Siegel and Gerth, 2000). These examples used passive remote sensi ng. Passive sensors require clear atmospheric conditions which is almost neve r the case during the rainy seas on. Consequently, active remote sensing systems, which have the ability to penetr ate clouds and other attenuating factors, play an important role in the delineation of flooded lands because the data can be acquired regardless of the weather. A most common example of an active remote sensing system is Radio Detection and Ranging (radar). Radar data have been applie d to many flood mapping applications (Townsend and Walsh, 1998; Horritt et al. 2001; Liu et al. 2002; Toyra and Pietroniro, 2005). However, radar has significant drawbacks such as the fact th at it is expensive and the spatial quality of the imagery is relatively poor in comparison to passive remote sensing imagery. Furthermore, radar images are not useful in mapping of urban flooding due to large amounts of corner reflection from buildings (Kiage et al. 2005). Therefore, it is best to co mbine imagery sets of both passive and active systems for flood delineation.
13 Fusion of radar and optical data acquired dur ing a rainy season to map inundated areas has rarely been studied, especially in monsoonal zone, because there is little opportunity for (1) clear atmospheric conditions dur ing flooding period and (2) both rada r and optical satellites acquiring data of the same location during that period. During the typhoon season of 2002, a flood occurred in the Nort heast region of Thailand but the sky remained clear. These conditions enab led the IRS-1D satellite to acquire data on October 14 and Landsat-7 acquired data on Octo ber 25. Radarsat-1 was programmed to acquire data of the same location on October 11 and Oct ober 15. For this reason, fusion of optical data (IRS-1D and Landsat-7) and radar data (Radarsa t-1) for flood delineation in monsoonal area could be performed. These images were acq uired during October 1125 (14-day duration). During this period, water surface elevation dropped gradually causing different flood extents in these satellite images. For the Radarsat wide mode 1 (W1) acquired on October 11, IRS-1D acquired on October 14 and Radars at standard mode 7 (S7) acquired on October 15 images, surface water extents were not greatly different, but for the Landsat image acquired on October 25, the surface water extent was noticeably smal ler. Therefore, when IRS-1D, Landsat-7and Radarsat-1data were fused, the influence of diffe rent acquisition dates was included in the fused image. Thus, the temporal factor must be carefully considered when the fused image is interpreted (Pohl and Van Genderen, 1998). Radarsat imageries available for my study we re taken in different modes. The image acquired on October 11, 2002 is the W1 with a 30 m spatial resolution and a 25 incidence angle. Another image acquired on October 15, 2002 is the S7 with a 25 m spatial resolution and a 47 incidence angle. Because spatial data in my study are based on 1:50,000 topographic maps, these Radarsat images with 25 and 30 m pixel size ar e deemed appropriate. Since the images were
14 taken at different incidence angl es, it was considered necessary to determine which image is better for flood delineation. Flooded areas interprete d from both images were compared to a flood map calculated from Hydrologic Engineering Centers Ri ver Analysis System (HEC-RAS), a hydraulic model developed by the U. S. Army Corps of Engineers. Inputs for HEC-RAS include hydrological data collected in the field and topographic data. There are many studies using only radar data to delineate floods (Dwivedi et al. 1999; Horritt et al. 2001; Freeman et al. 2002; Liu et al. 2002; Rosenqvist and Birkett, 2002; Rosenqvist et al. 2002; Kiage et al. 2005; Henry et al. 2006). In my study, it is hypothesized that classification accuracy of flooded areas derived from a fused image of IRS-1D, Landsat-7 and Radarsat-1 data is higher than that derive d from Radarsat-1 data al one. Thus flooded areas interpreted from the fused image were compared to that interpreted from the Radarsat-1 image. Data fusion techniques have been commonly used to improve interpretability of images, increase spatial resolution, enhan ce certain features not visible in either of the single data alone, complement data sets for improved classificati on and substitute missing information (Pohl and Van Genderen, 1998). The techniques were applied in many fields, such as underground coal fires detection (Zhang et al. 1999), land use/land cover mapping (Haack and Bechdol, 2000; Zhu and Tateishi, 2006; Santos and Messina, 2008), and landslide invent ory (Nichol and Wong, 2005). In my study, there are Landsat-7 data taken in the blue wavelength that can penetrate clear water and detect underwater features, and Radarsat-1 data that can detect change of water surface roughness caused by underwater feat ures and/or wind. If fusion tec hniques are applied to these data, underwater features such as submerged aqua tic vegetation may be identified and the fusion techniques may prove to be useful in pr oviding more information of flooded areas.
15 Objectives Objectives of m y study are as follows: 1. Delineate the flood extent in the study area occurring between October 11-25, 2002 using a combination of IRS-1D, Landsat-7 and Radarsat-1 data 2. Compare flooded areas derived from (a) Radarsat wide mode 1 at 25 incidence angle (b) Radarsat standard mode 7 at 47 in cidence angle and (c) HEC-RAS calculation 3. Determine if classification accuracy of flooded areas derived from the fused image of IRS-1D, Landsat-7 and Radarsat-1 is higher than that derived from Radarsat-1 alone 4. Determine if underwater features such as s ubmerged aquatic vegetation can be detected using these data Study Area In m y research, the study ar ea is in the Northeast of Thailand (also known as Korat Plateau), as shown in Figure 1-1, A. This region is 170,226 sq.km., situated around 90 1,400 m above mean sea level. The plateau is bounded by Phetchabun Range and flat-topped mountains (Phu Luang, Phu Kra Dung, and Phu Khieo) in the west, bounded by Dong Phayayen and Dangrek Ranges in the south, and the Mekong River in the north and east. Major rivers of this area are Mun and Chi rivers, which flow eastward. Area inside the plateau is flat with low hills. Average slope of the whol e area is less than 0.5%. The Northeastern region of Thailand has a popul ation of approximately 21 million, one-third of the countrys total population (National Sta tistical Office, 2000). Most of the land is used for agriculture with 70% rice, 17% field crop, and 13% co mbination of forest, pasture, natural rubber and others (Na tional Statistical Office, 2003). Yi eld rates per unit area of this region however, are relatively low compared to other regions of the country, leading to low average income (Kermel-Torres, 2004). Only 10% of agricultural land in the area is irrigated (National Economic and Social Advisory Council 2007) and the rest of the agricultural land depends solely on rainfall. Agricultural areas are also seasonally inundated by floodwater.
16 The rainy season is from May to Octobe r when the Southwest Monsoon wind conveys humid air from the Gulf of Thaila nd to the region (Figure 1-1, A), however the wind strength is reduced by mountain ranges located in the we st (Phetchabun Range) and the south (Dong Phayayen and Dangrek Ranges) of the region. Th is wind brings small amounts of rain. Larger amounts of rain occur between August and Octobe r due to the influence of storms from the South China Sea (Figure 1-1, A) and flooding usually occurs du ring this period. The floods damage residential and agricultural areas along the floodplains of the Mun and Chi rivers, the most important rivers in the Northeast. The 2002 floods in the Mun and Chi Basins appear in dark tone in the southeast as shown in Figure 1-1, B. The dry season of the Northeast region runs from November to April and water shortages usually occur between February and April. At this time, a large proportion of land cannot be cultivated. The study site is the overlapping area of the IRS-1D, Landsat-7 and Radarsat-1 W1and S7 images as highlighted by a yellow frame in Figur e 1-1, B. Latitude and Longitude of upper left corner are 15 46 47.76 N, 103 45 42.34 E, upper right are 15 43 45.90 N, 104 03 08.48 E, lower right are 15 08 28.65 N, 103 55 56.00 E, and lower left are 15 11 41.30 N, 103 38 46.44 E. The area is approximately 30 65 sq.km.
17 Storms Radarsat-1(W1) Landsat-7 Chi River IRS-1D Study area Mun River Radarsat-1 (S7) A B Figure 1-1. Study area. A) Nort heast of Thailand and rainfall influences. B) The 2002 flood shown in IRS-1D, Radarsat-1 (W1), La ndsat-7, and Radarasat-1 (S7) images. Korat Plateau
18 CHAPTER 2 LITERATURE REVIEW Image Fusion Techniques Re mote sensing data have been acquired in various wavelengths with different spatial, temporal, and spectral resolution. In my study, La ndsat, IRS and Radarsat were used. Due to various input data, image fusion techniques were needed to process the data and create an image containing information of the 2002 flood. Goshtasby and Nikolov (2007) defined image fusion as the process of combining information from two or more images of a scen e into a single composite image that is more informative and is more suitable for visual perception or computer pr ocessing. Research on image fusion began in the mid-eighties. It was focused on fusing visible and infrared images for surveillance purposes. Many more image fusion te chniques were developed in the 1990s and beyond. Image fusion has been successfully applie d in medical, industrial, military and civilian remote sensing applications (Goshtasby and Nikolov, 2007). According to Pohl and Van Genderen (1998), im age fusion techniques can be categorized into two groups: (1) statistical/numerical met hods, which include Principal Component Analysis (PCA) and (2) color-related techniques, which include Intensity-Hue-Saturation (IHS) color spaces (Pohl and Van Genderen, 1998). Principal Component Analysis Redundancy commonly occurs in remote sensing data. For exam ple, images taken in green and red wavelength usually appear similar. They are highly correlated and contain much of the same information. PCA is used to reduce redundancy in remote sensing data by generating fewer bands of non-redundant data from many bands of original data (Lillesand et al., 2007). The first principal component image (PC1) re presents the largest percentage of scene variance, while PC2,
19 PC3, PCn represent successively smaller pe rcentages of the variance. The successive components are chosen to be perpendicular to all previous components to make them uncorrelated. As a result, unnoticed features in PC1 could be identifi ed in PC2 and PC3, for example. Principal component images can be individually analyzed as black and white images or combined to make a color image and also can be used as input for image classification. Lillesand et al. (2007) mentioned that the intrinsic dimensio nality of Landsat TM or ETM+ was three. Thus, just PC1, PC2 and PC3 of TM or ETM+ were often used in image classification. PCA has been successfully applied in land use/land co ver mapping (Henderson et al., 1998; Saindranath et al. 2000; Saroglu et al. 2004; Santos and Messina, 2008), flood delineation (Hudson and Colditz, 2003), vegetation change detection (Lu et al. 2008), and surface fuel mapping (Mutlu et al. 2008). In my study, multi-spectral data of Landsat and IRS were inputs. There is redundancy in each sensor data and between both sensor data Thus, PCA is critical in redundancy reduction. Intensity-Hue-Saturation IHS is a color m odel used for merging images with different spatial resolutions. In an IHS system, color is described in te rm of intensity, hue and saturation. Intensity characterizes the brightness of a color (lower intensity = darker color; higher intensity = brighter color). Hue defines color. For example 0 or 360 (0 or 255 in 8-bit system) represents red, and 60 (43 in 8-bit system) represents yellow. Saturation indicates purity of a color (Lillesand et al. 2007). By using an IHS color model, spatial and spectral information of an image can be separated. Intensity represents spatial informati on, whereas hue and satura tion represent spectral information (Pohl and Van Genderen, 1998).
20 To increase the spatial resolution of a sate llite image, the low resolution multispectral image originally in RGB is tran sformed to the IHS system and then the intensity component is replaced by values from a higher resolution image such as a panchromatic image. This is called pan-sharpening. The higher resolution IHS image is transformed back to the RGB color system. In order to produce the best fused imag e in terms of color bala nce, the histograms of intensity and the high resolution image must be matched prior to the substitution (Lillesand et al., 2007). A constraint of this method is that number of input bands is limited to only three from a lower resolution image and one band of a higher resolution image. IHS transformation was successfully applied in geomorphology (Singhroy, 1995; Singhroy et al. 1998; Schetselaar, 2001), we tland delineation (Dwivedi et al. 1999), urban structure classification (Netzband et al. 1999), burned land mapping (Koutsias et al. 2000), flood extent delineation (Srivastava et al. 2000; Spruce and McKellip, 2006; National Emergency System of Uruguay, 2007; Akar et al ., 2008; International Charter, 200 8), and land use/land cover mapping (Saroglu et al. ,2004). Most of these studies used a pansharpening, except for studies of Singhroy (1995), Singhroy et al. (1998), Dwivedi et al. (1999) and Saroglu et al. (2004), which C-band radar data were used to replace inte nsity component of multispectral data. In my study, a panchromatic image (five m sp atial resolution) of IRS was one of input data. Through a pan-sharpening, spatial resolution of the fused image can be increased. Flooded roads can be identified from the pan-sharpened im age and this information is necessary for relief effort planning. PCA and pan-sharpening techniques applied in my study were performed at the pixel level of input images. These pixel level fusion algori thms assume correspondence between pixels in input images; therefore it is critical to accurate ly co-register all input images to avoid false
21 classification later on (Pohl and Van Gendere n, 1998; Goshtasby and Nikolov, 2007). Pohl and Van Genderen (1998) also commented that fusion of multi-sensor data mostly included temporal factor due to differences in ac quisition date/time and users have to be aware of the physical characteristics of the input data in order to be able to judge the resulting data. Flood Assessment Using Remote Sensing Flood boundary is im portant information for flood situation assessment. It is difficult to produce flood extent data through field survey es pecially for a large ar ea because it is a time consuming process and flood boundary is dynamic. In some cases, a flooded area is inaccessible due to transportation limitation or epidemic in the area. Remote sensing is a powerful technology to acquire flood information of a specific area in a certain time. Furt hermore, high temporal resolution remote sensing data can be used to monitor a flood movement. Data from remote sensing satellites, such as Landsat, IRS and Radarsat, are widely used in flood-related applications. Landsat The Landsat system has continuously collect ed data since 1972. Its sensor suite has evolved through several generations. Multispectral Scanner (MSS) on Landsat-1 (launched in 1972), Landsat-2 (launched in 1975), Landsat-3 (launched in 1978), Lands at-4 (launched in 1982) and Landsat-5 (launched in 1984) had four spectral bands (green, red and two bands in near infrared) with 80 m resolution (Lillesand et al., 2007). The thematic Mapper (TM) on Landsat-4 and Landsat-5 had seve n spectral bands (blue, green, re d, near infrared, two bands in mid infrared, and thermal infrared). The spatia l resolution of the thermal band is 60 m while those of the others are 30 m. Landsat-7 (la unched in 1999) has an Enhanced Thematic Mapper Plus (ETM+) sensor with eight spectral bands, seven bands like the TM and a new panchromatic band with a 15 m spatial resoluti on. Details of the Landsat sensor suit are shown in Table 2-1.
22 The long term record of Landsat data has been particularly useful in flood studies. Green, et al. (1983) used a Landsat-1 MSS image (standard color-enhanced combination of green, red and near infrared bands) to define spatial pattern of 1974 flooding in the Darling River floodplain in New South Wales, Australia. The image was acquired four days after the flood peak. Floodwaters (turbid water = lig ht blue tone in the image) we re differentiated from stagnant areas (clear water = black tone in the Landsat image). The flood da ta could be used in planning of roads, levees, and channel construction. Townsend and Walsh (1998) assessed abiliti es of Landsat TM, JERS-1 (L-band, HH) Synthetic Aperture Radar (SAR) and ERS-1 (C-b and, VV) SAR sensors for detecting flooding in bottomland forests of the lower Roanoke River floodplain, North Carolina. For the classification accuracy assessment, 1126 random sample points were selected. From ANOVA statistics, JERS-1 classification had higher F-value than ER S-1 classification. Thus, JERS-1 had a better ability to identify flooded areas. ERS-1 was less effective than JERS-1 because the shorter wavelength of C-band was scattered by forest canopy. However, for this study area, ERS-1 sensor could detect flooded forest This might have resulted from open canopies of forests in the study area. For Landsat, the authors recommende d that it was suitable for flood detection only during leaf-off periods. The authors also commente d that radar data provided independent data that could be used for hydrological model va lidation, which reduced cost for field-based measurements. Frazier and Page (2000) compared TM bands 1-7 (not thermal) of Landsat-5 in water body detection. Density slicing classi fication method was applied to the six bands of Landsat. In classification accuracy assessment, color aerial photographs taken on the same day as the Landsat data were used as reference. The st udy found that band 5 yielded the highest overall
23 classification accuracy (96.9%). Band 7 yielde d a good result (94.4% overall classification accuracy), but some irrigated crops were miscla ssified as water. Band 4 could be used to identify most of major water bodies (93.6% overall classificati on accuracy), but some urban areas, hill shadows and paddocks were miscla ssified as water. Band 1, band 2, and band 3 yielded 7.7%, 6.4%, and 8.0% overall classifica tion accuracy, respectively. Band 5 density slicing classification was compared to a 6-band maximum likelihood classification, and found that using only Band 5 to classify water bodies was as successful as using 6 bands of Landsat (96.9% versus 97.4% overall classification accuracy). Ryu et al. (2002) extracted waterline in a tidal flat in Gomso Bay, Korea using Landsat ETM+ data. Density slicing method was used to cl assify the images. Classification outputs were compared to ground leveling data and waterl ine tracking using differential GPS observed simultaneously with Landsat-7 acquisition time. It was found that TM band 4, band 5, and band 6 were reliable for waterline extraction during the flood tide. On an ebb tide, band 6 was the most accurate, but it has lower spatial reso lution. Thus, the authors recommended using combination of different bands rather than a single band for the waterline classification. Furthermore, the authors found that the Norma lized Difference Vegetation Index (NDVI), (TM4 TM3) / (TM4 + TM3), could be successfully applied to distinguish between turbid water and the tidal flat because reflectance of band 4 was high for both tidal flat and turbid water (reflectance increased as suspended sediment in creased) while reflectance of band 3 was low for the tidal flat but high for the turbid water. This band discriminated between clay and silt. Wang et al. (2002) mapped flood extent in a flat terrain along th e Tar River, Pitt County, North Carolina using Landsat TM data. TM band 4 was found useful in delineating boundary of land and water, but it was not us eful to differentiate between water and asphalt areas because
24 both areas had low reflectance. In TM band 7 image, water has low reflectance, while asphalt pavement has intermediate reflectance. Thus, the sum of TM band 4 and TM band 7 was used to identify water and non-water areas. A histogram of TM band 4 + TM band 7 was produced and showed that two distinct distri butions of wet and dry areas were evident, and a threshold value for water and non-water classification was iden tified. Cloud shadows, which were initially misclassified as water, were later recode d to be non-flooded areas. In the study, a Digital Elevation Model (DEM) was integrat ed in the classification to iden tified floodwater under forest canopies, which was an alternative to using ra dar data. The DEM used in the study was a 30 m USGS DEM with 1 m vertical accuracy. The DEM was filled using flood water elevation recorded at a river gauge. In the study, DEM filling was re-run at 1 m of the river gauge reading, but the pattern of the flood did not significantly change. The authors commented that using data from a river gauge to fill DEM only worked for a short distance. Hudson and Colditz (2003) deli neated flooding in a large a lluvial valley of the lower Panuco basin, Mexico using Landsat-5 TM acquired eight days afte r the peak of a large flood in 1993, and Landsat-7 ETM+ acquired in the dry season in 2000. PCA was applied to the Landsat-5 data and input data were bands 1, 2, 3, 4, 5, and 7. It was found that mid-range PCs (PC2 and PC3) were useful for flood mapping. For multi-temporal analysis, band 1, 3, 4, and 7 of Landsat-5 and Landsat-7 were inputs of PCA (band 2 and 5 were excluded because of similarity to other visible and in frared bands) and the re sult showed that PC2 was the most useful for flood mapping. The authors r ecommend mid-range PCs for flood delineation, whether using single scene or multi-temporal scenes. Wang (2004) defined extent of the 1999 flood on a coastal floodplain along the Tar/Pamlico River, North Carolina using Landsat-7 images. The flood peak was on September
25 21, 1999. Three date data of Landsat were acquired on July 28, 1999 (pre-flood), September 23, 1999 (two days after the peak), and September 30, 1999 (nine days after the peak). The pre-flood image was used for change detection. To delineate flood, various band and band ratio combinations were explored and found that TM band 5 + TM band 7 and TM band 4 + TM band 8 were suitable for wet and dry area disc rimination. For the image taken on September 30, flooded areas defined by TM band 4 + TM band 8 were larger than those defined by TM band 5 + TM band 7 because damaged crops still had low reflectance in bands 4 and 8. Therefore, the TM band 4 + TM band 8 of the image acquired da ys after the peak illustrated the maximum extent of the flood and this method could reduce the requirement of getting simultaneous optical remote sensing data to define maximum exte nt of a flood. For classification accuracy of TM4+TM8, data from 85 flooded and non-flooded sites were use as reference, and the classification accuracies were calculated for open fields, developed areas, and forested areas. Overall accuracy was highest for classification in open fields: 99.3% for the September 23 image and 96.1% for the September 30 image. For deve loped areas, the overall accuracy was 89.7% for the September 23 image and 82.5% for the September 30 image. For forested areas, overall accuracy was lower than other areas because the wavelengths used in Landsat cannot penetrate forest canopies (overall accuracy was 84.6% fo r the September 23 image, and 87.0% for the September 30 image). Neuenschwander et al. (2005) compared ability of La ndsat-7 ETM+ sensor to the Advanced Land Imager (ALI) sensor (on the NAS A EO-1 satellite) to map flood in the lower Okavango Delta, Botswana. Bayesian Pairwise Cla ssification (a supervised classification that uses a class-dependent band sele ction technique) was applied to both data, and found that, ALI yielded 100% in both producers and users accuracies of wate r class, whereas ETM+ yielded
26 99.63% producers accuracy and 99.26% users accuracy. In the ETM+ data, swamp was often misclassified as water causing overestimation of total flooded area in the delta. For other land cover classes, the ALI sensor also yielded hi gher classification accuracies than ETM+ sensor. According to this research, Landsat data alone provide satisf actory results in flood delineation. Using only one band or summation of two bands of Landsat (TM4, TM5, TM6, TM5+TM7,or TM4+TM8), land and water were su ccessfully distinguished and histogram threshold (also known as density slicing) classifi cation method was often used (Frazier and Page, 2000; Ryu et al. 2002; Wang et al. 2002; Wang, 2004). PCA was also applied to Landsat data, and mid-range PCs (PC2 and PC3) were recommended for flood delineation (Hudson and Colditz, 2003) Like other optical sensor data, cloud c ontamination caused a discrepancy in flood delineation using Landsat data (Wang et al., 2002). To avoid cloud c over, a Landsat image acquired days after a flood peak could be used to delineate maximum flood extent by using TM4+TM8 (Wang, 2004). To detect flood under forest canopies, L-ba nd and C-band radar data were found to be superior to Landsat data (Townsend and Walsh, 1998) but the cost of radar data is higher. Thus, Wang et al. (2002) used stream gauge data and DEM to estimate flood in a forest and they found that using the stream gauge data to fill the DE M could be used for a short distance. Therefore, using a flood model probably yields a better result and thus can be an alternative to costly radar data to detect flood underneath canopies. Landsat-7 ETM+ sensor had a limitation to differentiate betw een water and swamp (Neuenschwander et al., 2005). If Landsat data are fused to othe r sensor data, certain features not visible in either of the individual data sets ca n be enhanced. There have been many studies that
27 fused Landsat and other sensor data to detect features that were ha rd to discriminate using single sensor data alone. They include fusion of Landsat and airborne SAR data for landslide, erosion and deposition assessment (Singhroy, 1995), fusi on of Landsat and SPOT, Landsat and IRS, Landsat and Radarsat, and Landsat and ERS data to differentiate urbanized area, bare ground, and estuarine emergent (Henderson et al ., 1998), fusion of Landsat and Radarsat for landslide identification (Singhroy et al. 1998), fusion of spaceborne images (L andsat, SPOT, NOAA-AVHRR, and ERS-1), airborne images (optical and thermal infrared), and GIS data to detected underground coal fires (Zhang et al. 1999), fusion of Landsat, and C-band and L-band images of Shuttle Imaging Radar (SIR-C) to differentiate mud-brick houses, scattered agriculture, and natural vegetati on in savanna environment (Haack and Bechdol, 2000), fusion of Landsat and SPOT images for ur ban growth detection (Gluch 2002), fusion of Landsat and ERS-1 data for land cover mapping (Zhu and Tateishi, 2006), fusion of Landsat and Radarsat for land cover mapping (Huang et al. 2007; Santos and Messina, 2008), and fusion of Landsat and SPOT for land cover ch ange detection (Lu et al., 2008). IRS The Indian Re mote Sensing (IR S) satellite was first launc hed in 1988. IRS-1A, the first satellite of the series, carried tw o types of Linear Imaging Self-scanning Sensor (LISS), LISS-I and LISS-II. Spatial resolution of LISS-I was 72.5 m while that of LISS-II was 36.25 m. Both LISS-I and II had four spectral bands (blue, gr een, red and near infrared). In 1991 IRS-1B was launched and equipped with the same sensors as IRS-1A. Later, in 1995, IRS-1C was launched and equipped with LISS-III sensor with 23 m spatial resolution and four spectral bands (green, red, near infrared and mid infrar ed). IRS-1C also carried panchr omatic sensor (5.8 m resolution) and Wide Field Sensor (WiFS) with two spectra l bands (red and near infrared) at 188 m spatial resolution. In 1997, IRS-1D was launched and e quipped with the same sensors as IRS-1C.
28 IRS data have been applied in many flood a pplications. Kar (1994) investigated the 1990 flood in Luni-Jawai plains of the Thar Desert in India using false colo r composites of IRS-1A LISS-I and LISS-II images acquired during pre-flood (October 25, 1988) and post-flood (November 12, 1990) periods. The study found that spatial distribution of flood damages was controlled by lineaments in the study area. Sharma, et al. (1996) delineated the 1993 flood in P unjab, India using IRS-1A LISS-I and ERS-1 SAR (C-band, VV) images. The flood was caused by a heavy rain during July 10-12, 1993 (58% of the annual rainfall) Flooded areas were visually interpreted from false color composite images of IRS-1A (acquired on Ju ly 27-29, 1993), and ERS-1 images (acquired on July 19, 1993). The 1993 flood map was then superi mposed with the 1988 flood map to generate a general flood prone area map. Siegel and Gerth (2000) studied 1997 floodwater distribution patterns of the Oder River, Germany. The flood patterns were observed in te rms of concentration of suspended matter and sea surface temperature (SST). Data from WiFS sensor (188 m spatial resolution, 774 km swath width) equipped on IRS-P3 and IRS-1C were used to map distribution of suspended matter while infrared data of NOAA-AVHRR were used to cr eate SST maps. The maps were then compared to ship-borne measurements, and both data sets were well matched. Srivastava et al. (2000) interpreted 1998 flood extent in Marigoan district in Assam State, India from two bands (red and near infrared) of WiFS sensor of IRS-1C acquired in September 1998. To assess the damage, the flood map was intersected with a land cover map derived from a pan-sharpened image of IRS-1C LISS-III and pa nchromatic data acquired before the flood. Dwivedi and Kandrika (2005) delineated aqua culture areas using fused image of LISS-III and panchromatic sensor data from IRS-1C and IRS-1D. In the fusion, PCA was applied to
29 multispectral data, and then first principa l component was replaced by a high resolution panchromatic image. Next, inverse PCA was app lied to the data. In the fused image, bunds of aquaculture ponds were clear and sharp. The imag e was used to digitiz e the boundaries of ponds. From field survey, shapes of prawn ponds were narrow and long comparing to fish ponds. To segregate prawn and fish ponds, shape factor (polygon area divided by polygon perimeter) was calculated, and a threshold value of 18.0 was found to be optimal for the segregation. Classified prawn and fish ponds were compared to field su rvey data. Overall classification accuracy was 82.6%. The data derived from the fused image were important for environmental health studies. Boruah et al. (2008) used IRS LISS-I and LISS-III sensor data acquired in 13 years to detect change in channel planform and physical habitat on the braided Brahmaputra River in Assam, India. ENVIs K-means unsupervised classi fication was used to map water, sand bar, and riparian vegetation. The authors found that GROW facility in ENVI software was more suitable than the K-means method to classify river wate r. GROW allowed users to set up a training area and define a specific standard deviation (in this study, standard deviation was 2), and then connected river channel and its branches were re liably mapped. River water in the study area was classified into three sub-classes: deep water, medium water, and shallow water. Overall classification accuracy was 85%, when aerial p hotographs, local knowledge of the area, and scrutiny of the images were used as references. Data from three sensors of IRS (WiFS, LI SS, and panchromatic) were used in many studies. WiFS sensor data are suitable for flood de tection in large areas (Siegel and Gerth, 2000; Srivastava et al. 2000; Bhan and Flood Team, 2001) because of its large coverage and 3-day revisit time (Van der Sanden et al ., 2001). LISS sensor data, wh ich has higher spatial and
30 spectral resolution than WiFS, were also used to differentiate land and water (Kar, 1994; Sharma, et al. 1996; Boruah et al. 2008). High spatial resolution panchromatic image of IRS plays an important role in image fusion. Fusion of IRS LISS and panchromatic sens or data was proved to be successful in many applications: land cover mapping (Saraf, 1999; Srivastava et al., 2000; Saroglu et al. 2004), settlement structure identification (Netzband et al., 1999), salt-affected soil delineation (Dwivedi et al. 2001) and segregation of prawn and fish ponds (Dwivedi and Kandrika, 2005). IRS and other sensor data were fused in many studies, such as fusion of IRS-1B LISS-II and ERS-1 data (ERS-1 data replaced intens ity component of the IRS image) for wetland mapping (Dwivedi et al ., 1999), fusion of IRS-1D LISS-III and ERS-1 data for land use/land cover mapping (Saroglu et al ., 2004), and fusion of IRS-P6 and Landsat-5 data for flood mapping (National Emergency System of Uruguay, 2007). Like other optical sensor data, cloud cover cau ses missing data in IRS images. In a study of Dwivedi et al (1999), IRS and ERS-1 images were acquired at a concurrent time, and there was cloud contamination in the IR S image. After the two images were fused, wetlands of the Sundaban Delta (West Bengal, India) were successf ully mapped. This indicated that missing data in the IRS image were substituted by the ERS-1 data through the fusion process. Therefore, fusion of optical and radar data is not only improve interpretab ility of the image, but also substitute missing data due to cloud cover. Radarsat Radarsat-1 is a Canadian earth observation satellite launched in Novem ber 1995. It is equipped with a SAR sensor using a single frequency (C-band at 5.3 GHz frequency), and single polarization (HH). Radarsat-1 ca n acquire data at various inci dence angles and swath widths
31 (bigger swath width, lower spatial resolution), as shown in Figure 2-1 an d Table 2-2 (Canadian Space Agency, 2008). Radarsat data were used in many flood studies. Yang et al. (1999) used Radarsat ScanSAR (50 m spatial resolution) acquired on July 28, 1998 to determine flood extent in Qianshan county, Anhui province, China. Land and water were classified using a histogram threshold method. The study area is mountainous, and mountain shadows appeared in the Radarsat image. These shadows caused confusion in the classificati on. Furthermore, parts of water in a large lake were misclassified as dry area due to rough water surface (caused by wind) that returned large amount of backscatter to the ra dar antenna. To reduce the miscla ssification, Landsat TM data (acquired on December 7, 1995, pre-flood period) were used to extract hill shade and water bodies. Histogram threshold classification method was applied to Landsat band 2 to extract hill shade. If digital number (DN) of a pixel was less than 23, that pixel was then classified as hill shade. Next, flooded areas interpreted from the Radarsat image was subtracted by hill shade. Landsat data were also used to classify water bodies during pre-flood period. A pixel was classified as water body, if TM2 >= 10 and (TM2 + TM3) > (TM4 + TM5) 10. The Landsat derived water bodies were included in the flooded ar eas interpreted from the Radarsat image. In classification accuracy assessment, visual interpretation was used as reference. Overall accuracy for mountainous areas was 85%, and that of flat terrain was 90%. This study shows that merging of information from Landsat can improve classi fication accuracy of Radarsat data. Information from Landsat reduced misclassification due to hill shadows and rough water surface, which are disadvantages of radar sensor. Zhou et al. (2000) monitored the 1998 flood in the Nenjiang and Songhua River Basins, China using multi-temporal NOAA AVHRR and Radarsat images. Pe ak discharge of the flood
32 was on August 12, 1998. Flooded areas were cl assified from cloudfree AVHRR images acquired on July 30, August 14 and 24. AVHRR images were suitable for monitoring dynamic of flooding due to its high temporal resolution, but the authors required higher spatial resolution images to evaluate flood damages. Landsat TM data acquired during that period could not be used due to serious cloud contamination. Theref ore, Radarsat images (ScanSAR, 50 m spatial resolution) acquired on August 15, 23, and 29 were used. To identify flooded areas, a histogram threshold classification method was used. The au thors concluded that multi-temporal images from both optical and radar sensors were succes sfully applied to monitor the dynamics of flooding. Townsend (2001) mapped seasonal flooding in fo rested wetlands of the Roanoke River floodplain in North Carolina using multi-temporal Ra darsat S1, S2 and S6 images taken at 23.1, 27.5 and 44.0 incidence angles, respectively. Elev en Radarsat images were acquired in 1996 1998 during both leaf-on and leaf-o ff periods. Threshold values we re used to classify flooded and non-flooded forest. Overall classi fication accuracy of leaf-off images was higher than that of leaf-on images (98.11% versus 89.09%), when data from 13 U.S. Geological Survey wells were used to validate the resu lts. This study showed that Radarsat data could be used to map flooded forests in temperate regions re gardless of season and water stage. The author did not comment influence of different incidence angles on the outputs. Toyra et al. (2001) mapped extent of standing water in the Peace-Athabasca Delta, Canada from Radarsat S2 (27.5 incidence angle) and SPOT multispectral images. Mahalanobis distance supervised classification was app lied to the data. To validate the classification output, data from field measurements, aerial photos, and maps were reference data, and then Kappa coefficients were calculated. Kappa coefficient of combinati on of Radarsat and SPOT, Radarsat alone, and
33 SPOT alone were 0.92, 0.76, and 0.80 respectively. At 95% confident level, Kappa coefficient of combination of Radarsat and SPOT was significantly different from that of Radarsat and SPOT alone, but Kappa coefficient of Radarsat and SPOT were not significan tly different. The study found that the Radarsat image taken at small in cidence angle (20-31) could penetrate canopies of willow, grasses or sedges and detect standing water under th e vegetation canopies. Furthermore, this study examined the effect of incidence angle on flood mapping by comparing combination of Radarsat S1 (23.5 incidence a ngle) and SPOT to combination of S7 (at 47.0 incidence angle) and SPOT. Kappa coefficient of Radarsat S1 and SPOT was higher than that of Radarsat S7 and SPOT (0.83 versus 0.74), and the Kappa coefficients were significantly different at 95% confident level. The Radarsat S7 taken at bigger incidence angle had longer path length through the canopies; theref ore its signal was more attenuated. By using the same data set, a study of Toyra and Pietroniro (2005) found that si gnal of Radarsat S1 and S2 (small incidence angle) were very sensitive to waves on wate r surfaces, and many pixels in Lake Claire was misclassified as dry vegetation. To eliminate this confusion, SPOT data were combined with the Radarsat data, but for areas outside covera ge of the SPOT image, the misclassified areas were manually recoded as open water. The study also found that Radarsat S7 (big incidence angle) image was not sensitive to wave action. Kiage et al. (2005) used Radarsat ScanSAR narrow (50 m spatial resolution) images to assess flooding in coastal Louisiana due to Hurri cane Lili causing flood peak during October 3-4, 2002. Three Radarsat images used in the study were acquired on September 23 and 28 (pre-flood), and October 3 ( during the flood). The study found a new method to identify flooded areas. First of all, mean values of pixels from the two pre-flood im ages were calculated to create a mean image, and then subtracted the mean image from the Radarsat image acquired during the
34 flood period. Flooded areas appeared as darker regions and were easily identified. The study confirmed that Radarsat ScanSAR data were suitable for delineate flood in coastal marshes following hurricanes. However, the ScanSAR im ages were not useful in mapping of urban flooding in New Orleans after Hurricane Ka trina, August 2005, due to large amounts of backscatter from the buildings, and combinati on of Radarsat and SPOT data to map the flood after Hurricane Katrina was in progress. Lang et al. (2008) examined influence of different incidence angles (23.5, 27.5, 33.5, 39.0, 43.5, and 47.0) of Radarsat images to de tect flood underneath forest canopies during leaf-on and leaf-off periods. The study area was the lower Roanoke and Cashie River floodplain, North Carolina. To discriminate flooded and nonflooded areas in each image, a threshold value was used. Flood extents interpreted from the Radars at images were verified using flood extents estimated by a flood simulation model of Roa noke River floodplain developed by Townsend and Foster (2002), and the data were used only when flood extents from both sources were in agreement (no comparison detail was described in the literature). A hypot hesis of the study was that an image taken at a larger incidence angle was less sens itive to flooding. The study found that the sensitivity slightly decreased with increas ing incidence angle, but the sensitivity did not significantly decrease as found in the study of Toyra et al. (2001). According to these studies, Radarsat data we re very useful for flood mapping due to cloud penetration ability. To differentiate between flooded and non-flooded areas, a threshold value was often used (Yang et al. 1999; Zhou et al. 2000; Townsend, 2001; Liu et al., 2002; Lang et al. 2008). Flooded areas interpre ted from Radarsat images were so reliable that they were used to validate ability of MODIS to detect flood in Cambodia and Vietnamese Mekong Delta (Sakamoto et al. 2007)
35 However, using Radarsat data alone to delineate flood was not suitable for some study areas, where mountain shadows, waves on water surf ace, and large amount of corner reflections caused confusions in Radarsat images (Yang et al. 1999; Toyra and Pietroniro, 2005; Kiage et al. 2005; Song et al. 2007). To eliminate thes e confusions, Radarsat data were combined with local slope information (Song et al. 2007) or data from optical sensors, such as Landsat (Yang et al. 1999) and SPOT (Toyra and Pietroniro, 2005; Kiage et al. 2005). Radarsat can acquire data at va rious incidence angles. Staples et al. (1998) suggested that a Radarsat image taken at large incidence angle (more than 40) was appropriate for land and water differentiation. To detect flood water under vegetation canopies, Toyra et al. (2001) found that a Radarsat image taken at a small incidenc e image was significantly better than one taken at a big incidence angle, but Lang et al. (2008) found that different in cidence angles of Radarsat images did not significantly affect ability to de tect flood under forest canopies. This may result from different study areas. A Radarsat image taken at a big incident angle was used to detect submerged reef in a study of Staples et al. (1998). Submerged reef caused turbul ent water, and large amounts of backscatter were returned to the antenna, but its surroundings had lower backscatter due to smoother water surface. Thus, the reef was detected in the Radarsat image. Radarsat data have been combined with ot her sensor data in many applications. They include fusion of Landsat and Radarsat data to differentiate urbanize d area, bare ground, and estuarine emergent (Henderson et al ., 1998), combination of Rada rsat, airborne SAR, and Landsat TM data to identif y landslide features (Singhroy et al. 1998), fusion of Radarsat and ERS-2 data for land cover mapping (Saindranath et al. 2000), combination of Radarsat and JERS-1 data for land cover mapping (Souza Filho et al. 2005), combination of Radarsat and
36 Landsat data for land cover mapping (Huang et al. 2007), and combination of Radarsat texture information, Landsat, and digital video data for land cover mapping (Santos and Messina, 2008). For Thailand, a country in monsoonal zone, Rada rsat plays a crucial ro le in flood detection because of cloud penetration ability. Flood extent a nd marooned villages can be identified from a Radarsat image. However, decisi on makers require more informati on to evaluate flood situation. To provide a range of helps to flood victims, up-to-date information of roads that may be damaged by the flood is needed, and this informati on can be derived from panchromatic data of IRS. Most Thai people do not have flood insura nce; therefore, the gove rnment has to provide flood compensation. Land cover types affected by the flood can be extract ed from multi-spectral data of Landsat and IRS. This information ca n be used as an input for flood compensation calculation. For these reasons, da ta of Landsat, IRS, and Rada rsat acquired during the flood period have to be fused to create an image c ontaining necessary information for decision makers. According to literature review, there is no research that fuses data of Landsat, IRS, and Radarsat acquired in the same flood period to create an im age illustrating the flood situation. Therefore, fusion of these data needed to be done. Flood Models To prevent f lood damage, information of flood risk areas should be provided to people settling in those areas. A histor ical record of past floods allo ws these people to understand the level of risk they face. To prepare this information, a flood model can be used to simulate different flood scenarios. Also, flood models are important in flood management analysis because they are essential to simulate flow befo re a flood control structure is built. Flood models are critical in historic flood studies as well. Flood forecasting methods have been continuously developed. For the Nile, Hurst (1952) concluded that two successful methods included: (1) railway time-table method where if the
37 water surface elevation at an ups tream is measured, the time that this volume of water would reach a downstream point could be calculated and (2) forecasting based on the concept that after the end of a rainy season at headstream area, th e discharge of one month influences the discharge of the following month. Then correlations of the discharge can be calculated and were used in the forecasting. This method worked well during N ovember May before the beginning of the next rainy season. For the Mississippi River, the U.S. Army Co rps of Engineers (USACE) has applied several hydraulic models to calculate water surface elevation. According to Haestad Methods (2007), during 1950s 1960s calculation of the Mississi ppi Basin Model (MBM) was initially done by hand until the mid to late 1960s when computers began to be used, and Hydrologic Engineering Center (HEC) was formed in 1964 (U.S. Army Corps of Engineers, 2007). In 1970s, flow parameters were statistically an alyzed to predict 5 500 year return period of floods and cross sections along the 300 mi. (483 km.) reach of the Mississippi were survey ed. In 1980s, the cross section data were used to develop HEC family software: HEC-1 (watershed hydrology), HEC-2 (river hydraulics), HEC-3 (reser voir analysis for conservation), and HEC-4 (stochastic streamflow generation program) (Haestad Methods, 2007; U.S. Army Corps of Engineers, 2007). In 1995, HEC-2 was superseded by HEC-RAS that can calculate water surface elevation based on one-dimensional steady and unsteady flow regi mes (U.S. Army Corps of Engineers, 2002). For my study, the focus is on depth and boundary of flood water, not in water quality or sediment transportation. Therefor e, models for flow simulation were considered. According to U.S. Geological Survey (2008), fl ow simulation models could be categorized into three groups: one dimensional (1D), two dimensional (2D), and three dimensional (3D). The 1D model considers flow parameters (such as velocity) only in the same direction of the flow (longitudinal
38 direction). The 2D model considers flow para meters in longitudinal direction and another direction, vertical or horizont al direction. The 3D model considers flow parameter in longitudinal, vertical, and hor izontal directions. Higher dime nsional models can thoroughly simulate flow characteristics, but th ey need more input data. Correia et al. (1998) suggested that more complex models were not necessarily bette r for some specific uses. On the other hand, a simpler model requiring less input data might be more suitable for a study area that lacks input data. Di Baldassarre et al. (2009) also commented that 1D models were suitable for flood simulation in natural rivers when hydraulic si tuation was not dominated by 2D phenomena, such as dam or levee failure. A study of Horritt and Bates (2002) confirmed that the ability to predict river flood inundation of a 1D model was equal to a 2D model. In the study, HEC-RAS (1D model), LISFLOOD-FP (2D model), and TELEMAC-2D (2D model) were compared on a 60 km reach of the Severn river, UK with 19 ground surveyed cross sections and a LiDAR derived DEM. Test flood events were the 1998 (435 m3/s peak flow) and 2000 (391 m3/s peak flow) floods. Flooded areas were interpreted from a 1998 Radarsat-1 (C-band, HH) image and a 2000 ERS-2 (C-band, VV) image. There were six calibration/validation combina tions for the three models: (1) calibrated on the 1998 flood wave travel time and validated by the 1998 Radarsat image, (2) calibrated on the 1998 flood wave travel time and validated by the 2000 ERS-2 image, (3) calibrated on the 1998 Radars at image and validated by the 2000 ERS-2 image, (4) calibrated on the 2000 flood wave travel time and validated by the 2000 ERS-2 image, (5) calibrated on the 2000 flood wave travel time and validated by the 1998 Radarsat image, and (6) calibrated on the 2000 ERS-2 image and validated by the 1998 Rada rsat image. After the calibration, optimum friction coefficients of main channel and floodplain were designed for each model and each flood
39 event. The study found that HEC-RAS and TELEM AC-2D yielded equally good predictions, and were better than LISFLOOD-FP. The authors commented that perf ormances of the models were different because they were differently sens itive to changes in friction coefficients. Although 1D models are widely used to simu lated river flood inunda tion, they are not appropriate for all kind s of study areas. Mark et al. (2004) found that 1D models had limitations for predicting floods in an urban area, where bot h overland flow and flow in underground pipe systems existed and interacted to each other. In my study, the study area was a rural area dominated by non-irrigated agricultural land. Flooding in the study area occurred along natural rivers, Mun and Lam Sieo Yai rivers. Available hydrological data were daily discharge and daily wate r surface elevation collected from ground-based stations situated far away from each other, and some river reaches did not have stations. Due to physical characteristic s of the study area and limitation of available hydrological data, 1D model was deemed an ap propriate model for my study. One dimensional flood simulation models widely applied to st udy areas in Thailand were HEC-RAS (developed by the U.S. Army Corps of Engineers) and MIKE 11 (developed by Danish Hydraulic Institute). Weesakul (1995) set up a real time flood for ecasting model of Klong Utaphoa River Basin in Songkhla province. In the study a rainfall-runo ff model, NAM, was used to calculate runoff from rainfall data, and MIKE11 was used to simula te flow in the river. A limitation of the study was insufficient input data. Only daily hydrol ogical data were available, while 6-hour hydrological data were recommended, and more rain fall stations in the study area were needed. Petchprayoon (2001) predicted flood wave characteristics and estimated damaged areas caused by failure of Tha Dan dam in Nayok province using MIKE11. In the study, damaged areas from dam failure and overflow through spillway were predicted in four different severities.
40 To estimate flood damage of each land cover type flooded areas were overl aid to land cover data interpreted from a Landsat-5 image. Taragsa et al. (2004) used HEC-RAS to pred icted flood along Chi rive r floodplain in Maha Sarakham, Kalasin and Roiet provinces. A HE C-RAS result was compared to a flood map interpreted from a Radarsat image, and found that both flooded areas were in agreement by 52.26%. Minimum, normal, and maximum Mannings n values of each land cover type were tested, and found that normal values yielded a better result. The aut hors commented that the model needed more input data and more calibration. Kidson et al. (2005) used HEC-RAS to estimate discharge of two palaeofloods (84-year and 49-year floods) of the Mae Chaem River, Northern Thailand, based on palaeo-stage indicators (sand and silt terraces located independently, wood debris deposited in three separate caves). The authors commented that HEC-RAS, a 1D model, was suitable for the calculation because a greater dimensional model requires more input data, which were difficult to identify from palaeoflood evidence. Sansena and Bhaktikul (2007) created flood risk maps of 5, 10, 25, 50,100 year return periods of Mae Klong River in Ratchaburi province. In the st udy, hydrological data collected during 1985-2002 were used for frequency analysis, and 2-parameter log normal distribution was found to be the best fit for the data set. HEC-GeoRAS and HEC-RAS were used to create flood risk maps. The 1996 flooded areas predicted by HEC-RAS was compared to 1996 flood map of the Royal Irrigation Department, and found that both flooded areas were in agreement by 60.52%. The authors also commented that the DE M was the most important factor for flood simulations, and a higher resolution DEM was required to obtain more accurate results.
41 According to these studies, HEC-RAS and MIKE 11 were successfully applied in Thailand for flow simulation in Northern region (Kidson et al., 2005), Northestern region (Taragsa et al., 2004), Central region (Petchpra yoon, 2001; Sansena and Bhaktikul, 2007), and Southern region (Weesakul, 1995). In my study, HEC-RAS was chos en because it yielded a good result and the software (HEC-RAS and HEC-GeoRAS) are free for download, while MIKE11 is a commercial software. Advantages of HEC-RAS are not only its ability to produce an accurate result, but also its availability and accessibility for al l users to download it through the internet. As a result, HEC-RAS was widely used to simulate water flow in many study sites, such as the Livramento catchment in Setubal, Portugal (Correia et al. 1998), Ste. Agathe in Manitoba, Canada (Ahmad and Simonovic, 2001), the river Mhne in Germ any (Dose et al., 2002), the Tagus Basin in Central Spain (Benito et al. 2003), the Areias creek cat chment in Brazil (Diniz et al., 2003), San Antonio River Basin (Knebl et al. 2005), Tabunganen Scheme in South Kalimantan, Indonesia (Wignyosukarto, 2006), the South Nati on watershed in Canada (Yang et al. 2006), Manoa Valley (El-Kadi and Yamashita, 2007), Shitalakhya River in Bangladesh (Islam, 2007), and Atrato River in northwest of Colo mbia (Mosquera-Machado and Ahmad, 2007). Flood Models and Remote Sensing Re motely sensed data have been used in hydr ological models as a direct input (such as LiDAR DEM), a source of information for parameter setting, and reference data for model calibration/validation. Van de Griend and Engman (1985) commented that remote sensing data enhanced ability of hydrological models based on partial area concept, a method dividing watershed into contributing areas Temporal and spatial characte ristics of these contributing areas (such as vegetation types, temperature and soil moisture) could be identified using remote sensing data instead of laborious field survey.
42 Schultz (1988) concluded that information from remote sensing have been used as input for hydrological models or used to estimat e parameters in hydrological models. Model parameters based on watershed characteristics, such as drainage density, la nd use/land cover, and soil types could be interpreted from Landsat and SPOT images, and slope could be obtained from a stereo pair of SPOT images. The author also used NOAA infrared data with a long time record to calculate historic runoff va lues of catchments in Southern France. One-year period of concurrent data from both field survey and NOA A were used to calibrate a model. By using input from NOAA data alone, the model could estimate historic r unoffs that were useful for planning purposes. Remote sensing data from ground-based C-band weather radar were used as an input for real time flood forecasting of Gunz River catchment, Germany. Blyth (1993) concluded that remo te sensing data were used to aid hydrological models in three ways: site selection for instrument se tting, illustrating physical characteristics of catchments, and hydrological variable esti mation. Location of ground-based measurement devices could be designed prope rly based on information from remote sensing that showed variation and homogeny of a study area. Physical characteristics of catchments (such as stream length, surface water extent, drainage ditches, slope and land cover) could be extracted from Landsat and SPOT data, and surface roughness informa tion could be derived from a radar image. Hydrological variables (such as precipitation, vege tation moisture and soil moisture) could not be directly measured using remote sensing data but the variables could be estimated using a transfer function relating ground-ba sed data to remote sensing da ta. By applying the transfer function to remote sensing data, spatial variati on of hydrological variable s could be illustrated. Sharma et al (1996) simulated rainfall-runoff proce ss of the Divisadero Largo basin in Argentina using a 1D raster-based hydrological model, SWAMREG. Input data of the model
43 were soil type, slope, vegetation (interpreted from a Landsat TM image), and rainfall data. Thirteen rainfall-runoff events were used to calibra te the model, and other 13 events were used to validate the model. The model divi ded inflow (direct rainfall and inflow from the upstream) into infiltration and outflow, and simulated rainfall excess and outflow hydrographs. Outputs of the model were overall hydrograph, peak discharge, runoff volume, and fl ow duration. Relative square errors of the 13 validati on events were between 3.7% 13.2%. Profeti and Macintosh (1997) used Landsat TM images (acquired on January 27, August 7, and October 26, 1991) to estimate soil water cont ent of the Fucecchio Marsh in Tuscany, Italy. Band5/band7 (correlating to soil water content) was compared to band1/band2 (decorrelateing to soil water content). Validation data were calc ulated from the Fucecchio Marsh hydrological model, developed by Civil Engineering Department, Florence University. The model is raster-based using cumulative daily rainfall and soil moisture data measured in the field as inputs. Correlation coefficient between the Landsat band ratios and model outputs were calculated based on each soil type. The study found that band1/band2 had higher average correlation coefficient than that of band5/band7 (0.727 versus 0.639). Highest correlation coefficients obtained from soil types those had low vegetation c over. This study also monitored the 1992 flood in the study ar ea using ERS-1and ERS-2 data acquired on October 16, 1992 (pre-flood) and November 4, 1992 (f our days after the flood). Both images were inputs of a supervised classification to id entify flooded and non-flooded areas. Bates and De Roo (2000) developed a ne w flood simulation model, LISFLOOD-FP. The model is raster-based, and was designed to ope rate with high resolution and high accuracy DEMs (such as LiDAR DEM) that represent physical ch aracteristics of channel and other elements (such as dykes, embankments, depressions, and former channels). The test site was the River
44 Meuse in The Netherland, where a large flood oc curred in January 1995 and validation data from aerial photos and ERS-1were available. The model yielded percentage correct at 81.9%, when it was compared to flooded/non-flooded areas interp reted from the aerial photos and the ERS-1 image. Fox and Collier (2000) presented a method to calculate probability of a flood caused by heavy convective rain. The prediction could be done at a lead-time of up to seven days. The test site was the River Irwell catchment in the No rthwest of England. Inputs of the model were sensible heat flux data estimated from NOAA-AVHRR, precipitable water content of the atmosphere, and velocity of ascending air. The model was tested using data acquired in the second half of July 1996, and found that predicted convective rainfa ll had some correlation with observed convective rainfall. However, the prop osed method required a longer trial period to evaluate reliability of the model. Biftu and Gan (2001) developed a physically based hydrological model, named DPHM-RS, of the Paddle River Basin, Central Alberta in Canada. DPHM-RS was designed to incorporate remote sensing data in the model. Input data of the model were obtained from both field measurement and remote sensing (Lands at TM/MSS, NOAA-AVHRR, and Radarsat). The remotely sensed data yielded information of land cover, surface albedo, and vegetation index. The model could simulate runoff, surface temperat ure, net radiation, and soil moisture of the Paddle River Basin. Simulated and observed runoffs were in agreement of 60-85%. Simulated surface temperatures were close to those derived from Landsat TM and NOAA-AVHRR (discrepancies were less than 2 K), but for cl oudy day discrepancies were 3-7 K because the surface temperature derived from the images represented temperature of clouds, which was lower than temperature of the ground surface. Simu lated net radiation agreed well with observed
45 except for rainy days due to cloud cover. Daily mean soil moisture data estimated from the model agreed well with those es timated from Radarsat images. Bjerklie et al (2005) developed a method to estimate river discharge using remotely sensed data. The test sites were 17 single channel rivers (in New Hampshire, Vermont, Connecticut, Illinois, Maryland, South Dakota, Colorado, Montana, Kansas, Oregon, Delaware and Washington) and three braided rivers (in Alaska and British Columbia). Inputs data were water surface width and maximum channel width (measured from Digital Orthophoto Quadrangles and airborne SAR images), and channel slope data (obtained from 1:24,000 topographic maps). To improve accuracy of the es timated discharges, observed discharges were used to create a correction function for singl e channels and another function for braided channels. After applying the correc tion function to the estimated discharges, larges t log residuals were at two reaches, where flow activity was managed (the Mississippi River at Thebes in Illinois and the Sacramento River near Red Bl uff in California). Furt hermore, the authors proposed another method to estimate discharge by using width, slope and velocity. In this method, water surface velocity (converted to mean velocity) was used as a correction function instead of creating a new correction function. A te st site was the Missour i River in Sioux City, Iowa. Water surface velocity was derived from interferometry airborne SAR. Estimation accuracy was improved as the estimated value was within ten percent of the observed value. The methods using remote sensing data to estimate di scharge could be useful for a study area lacking ground-based data. Knebl et al. (2005) developed a flood modeling fr amework including NEXRAD (weather radar) data, GIS and hydrological models (HEC-HMS and HEC-RAS) for San Antonio River Basin in Central Texas. Study period was June 30 July 9, 2002 covering summer storm season
46 of the area. HEC-HMS, a rainfall -runoff model, was used to es timate overland flow and channel runoff from excess precipitation. HEC-HMS made a prediction based on 4 4 sq.km. spatial resolution according to input from NEXRAD. To calibrate HE C-HMS, predicted runoffs and predicted hydrographs were compared to observe d data, and then, watershed parameters were manually adjusted for each subbasin. The calibrated model was used to estimate hydrological data, which later on were inputs of HEC-RAS to predict floode d areas. The predicted flooded areas were validated against flooded areas interp reted from a Landsat TM image, and then, a map illustrating matched, overestimated and underestimated areas was produced. From the Landsat data, it could be implied that actual infiltration and disper sal of runoff was faster than that represented in the model. Bates et al (2006) observed dynamic movement of flooded areas in the River Severn, west-central England during Nove mber 8-17, 2000 using 1.2 m spa tial resolution airborne radar images. By combining series of inundated area data with topographic data (obtained from LiDAR), discharge, flooded area, reach storage and rate of re ach dewatering since last image could be calculated. Comparison of these calculat ed data to those observed in the field yielded more insight into understanding of factors cont rolling flood pattern (suc h as local drainage networks, embankments and culverts). Furthermore, the flooded areas interpreted from the airborne radar images were used to validat e ability of LISFLOOD-FP model to perform flood prediction. The study found that th e model performed well during th e peak flow (November 8), but performance decreased during low flow (N ovember 17). This might be resulted from increasing influence of small-scale features (not represented at the model grid scale) during the low flow period.
47 Paz and Collischonn (2007) pres ented a new method to automa tically extract length and slope of river reaches from a 90 m DEM pr oduced by the Shuttle Radar Topography Mission (SRTM). The extracted data could be used as i nputs of a large-scale raster-based hydrological model. In the study, the test si te was the Uruguay River basin a nd the extracted outputs were validated against vector data of the river netw ork. Quality of the extracted river length was improved using a stream burning method that cond itioned the extracted river length by vector data of the river. Quality of calculated slope was not validated because no reliable slope data were available for comparison. Schumann et al. (2007) used remote sensing data to estimate roughness value of each cross section in HEC-RAS, based on a concept that roughness values were spatially heterogeneous. The study site was River Alzette in Luxembourg. LiDAR derived DEM was used in the study. HEC-RAS was calibrated using data from the 2003 flood event (70.5 m3/s peak discharge). Water levels were identified for 55 cross sections from integration of flood boundaries (interpreted from ENVISAT acquired near the time of peak discharge) and the DEM. By assessing model error at each cross section, local roughness and correlation between cross sections were analyzed. Roughness values at each cross section were adjusted, and this adjustment compensated for all kinds of erro rs in the prediction. The calibrated model was validated using flooded boundaries interpreted fr om ERS data acquired ten hours prior to the peak of the 2003 flood, and flood boundaries interp reted from aerial photo s acquired during the 1995 flood that had a much higher peak discharge (95.6 m3/s). The study found that the calibrated model performed very well for both validation events. Th is indicated that the model calibration was robust because it was independent from different flood magnitudes and different
48 sources of remote sensing data. This study showed potential of remote sensing in flood model configuration. Di Baldassarre et al (2009) used a near real time Envisat image (C-band, VV, approximately 100 m resolution) to verify a flood simulation model, UNET (a package for unsteady flow calculation in HEC-RAS), and then made a prediction of the 2008 flood in a time shorter than the flood water travel time. The stu dy site was a 98 km reach of the River Po in Italy. Topography data were deri ved from LiDAR and sonar. The flood simulation model was previously calibrated using data of the 2000 hi gh magnitude flood, and ab solute relative error between simulated and observed water surface did not exceed four percent. When a flood occurred in 2008, the authors rece ived the near real time En visat image, interpreted the inundation width, and made a fl ood prediction using the calib rated model. However, the prediction was not correct, which confirmed that a well calibrated model based on a flood event (the 2000 high magnitude flood, approximately 60 year return period ) might give an incorrect prediction for another flood event (the 2008 low magnitude flood, three four year return period). According to these studies, remote sensing data have been used as direct inputs of flood models. Those data were DEM obtained fr om LiDAR (Bates and De Roo, 2000; Bates et al ., 2006; Schumann et al., 2007; Di Baldassarre et al ., 2009), and channel topography from sonar (Di Baldassarre et al ., 2009). Remote sensing data also provided physical data of a basin, which were river network (Blyth, 1993; Paz and Collischonn, 2007), Slope (Schultz, 1988; Blyth, 1993; Paz and Collischonn, 2007), and land cove r (Schultz, 1988; Blyth, 1993; Sharma et al ., 1996; Biftu and Gan, 2001). For some study areas lacking ground-based data, remote sensing data were utilized to estimate hydrological variable s: precipitation (Schultz, 1988; Sharma et al ., 1996;
49 Knebl et al. 2005), river discharge (Schultz, 1988; Sharma et al ., 1996; Bjerklie et al ., 2005; Knebl et al. 2005), runoff (Sharma et al ., 1996; Biftu and Gan, 2001; Knebl et al. 2005), soil moisture (Profeti and Macintosh, 1997; Biftu and Gan, 2001), and sensible heat flux (Fox and Collier, 2000). Remote sensing has proved to be useful for hydrological variable estimation. However, using optical sensor data to estim ate surface temperatures and net radiation was reliable only on cloud-free days, while accuracy of soil moisture estimated from radar data was more consistent owing to atmosphe ric independence (Biftu and Gan, 2001). Remote sensing has played a significant role in flood model calibration and validation. Both optical sensor (Knebl et al. 2005; Schumann et al. 2007) and radar data (Bates and De Roo, 2000; Bates et al ., 2006; Schumann et al. 2007; Di Baldassarre et al ., 2009) have been used. Because flooding usually occurs during storm seasons, radar data were more popular due to cloud penetration ability. Furthermore, Townse nd and Walsh (1998) and Toyra and Pietroniro (2005) recommended using radar images in flood model validation. However, when interpreting flooded areas from a radar image; ambiguity may happen due to waves on water surface, which usually occurs in the middle of rivers or large lakes. Horritt and Bates (2002) suggested that uncertainty areas in the middle of a river could be corrected using outpu t from a flood model. Thus, flooded areas predicted by a flood model can be used to complement radar data (uncertainty is caused by waves) and optical sensor data (uncer tainty is caused by clouds). In a study of Profeti and Macintosh (1997), out puts from a hydrological model were used to validate soil water content calculated from diffe rent band ratios of Landsat TM data. This is an example that a model can be used to determin e a suitable method to obtain information from remote sensing. Also, Lang et al. (2008) used flood extents calculated by a flood simulation model of Roanoke River floodplain to verify flood extents interpreted from Radarsat images.
50 Only Radarsat flood extents in agreement with the model outputs were used to detect flood underneath forest canopies. Likewise, in my study, flooded areas simulated by HEC-RAS were used to validate flooded areas interpreted from two Radarsat im ages. An image yielded higher classification accuracy of flooded and non-flooded ar eas was deemed the best Radarsat image in my study, and was then selected to be an input in the image fusion. According to previous studies, there is no research that fuses Landsat, IRS and Radarsat data acquired during the same flood event. Therefor e, fusion of these data to create an image illustrating the flood situation is necessary. When a flood occurs, information of flood boundaries, flooded transportation routes, and damaged areas are nece ssary to provide a range of helps and compensation. This information could be obtained from a fused image of Landsat, IRS and Radarsat through PCA and pansharpening techniques. A flood ma p calculated by HEC-RAS was used to validate fl ooded areas interpreted from Ra darsat W1 and Radarsat S7 images.
51 Figure 2-1. Radarsat im aging modes. (Source: http://www.asc-csa.gc .ca/eng/satellites/ radarsat1 /components.asp. Last date accessed November 2008).
52 Table 2-1. Sensors and bands of Landsat-1 to 7 Sensor Satellite Band number Spectral band Wavelength (m.) Spatial resolution (m.) MSS Landsat-1 to 5 4 Green 0.5-0.6 80 5 Red 0.6-0.7 80 6 Near infrared 0.7-0.8 80 7 Near infrared 0.8-1.1 80 TM Landsat-4 to 5 1 Blue 0.45-0.52 30 2 Green 0.52-0.60 30 3 Red 0.63-0.69 30 4 Near infrared 0.76-0.90 30 5 Mid infrared 1.55-1.75 30 6 Thermal infrared 10.4-12.5 120 7 Mid infrared 2.08-2.35 30 ETM+ Landsat-7 Same as TM Same as TM Same as TM 30, 60 for thermal band Panchromatic 0.52-0.90 15 Table 2-2. Description of Radarsat imaging modes Mode Nominal resolution (m.) Number of positions/beams Swath width (km.) Incidence angles (degrees) Fine 8 15 45 37-47 Standard 25 7 100 20-49 Wide 30 3 150 20-45 ScanSAR narrow 50 2 300 20-49 ScanSAR wide 100 2 500 20-49 Extended high 18-27 3 75 52-58 Extended low 30 1 170 10-22
53 CHAPTER 3 MATERIALS AND METHODS Spatial Data Satellite Imageries Satellite im ageries used in my study were s ponsored by GISTDA, a major data provider of Landsat, SPOT, IRS, and Radarsat in Thailand. Ideally, Landsat, SPOT, IRS, and Radarsat images covering a flooded area in the Northeast of Thailand acquired on the same date were needed. However, the probability of achieving this ideal is low. This is largely due to cloud cover, which typically accompanies a flood event. Therefore, an attempt was made to find the narrowest range of image dates in order to minimize the effect s the temporal variation. Also, processing techniques were used in the analysis to mitigate these effects. Among the available sensors, Radarsat plays the most important role in flood detection in mo nsoonal areas. Therefore, the availability of Radarsat data was checked fi rst. Unlike those optical sensors that continuously acquire data, Radarsat acquires data only when it is programmed in advance. As a result, archived Radarsat data were available for certain areas in certain times according to requests in the past. Fortunately, the archive had several imag es with different imaging modes and incidence angles illustrating flooding in the study area since 2002. Next, the availability of Landsat, SPOT and IR S data was investigated for images taken on the same dates as the available Radarsat imag es. In the investigation, it was found that cloud contamination was a serious problem for the optic al sensor data taken during rainy seasons, and there were no optical data taken on the same date of the Radarsat that has acceptable cloud cover. However, on October 14, 2002 the sky was clear and IRS-1D obtained data over the study area, and on October 25, 2002, the sky was partly clear, and the flooded area could be seen from Landsat-7 image. No SPOT image, however, was available during this period. As a result, the
54 images used in my study were Radarsat-1 (W 1), IRS-1D, Radarsat-1 (S7) and Landsat-7 acquired during October 11-25, 2002. On October 11, 2002, Radarsat-1 acquired data of the study area in W1 mode at 25 incidence angle with 30 m spa tial resolution using C-band and HH polarization. The original data were stored in 16-bit unsigned format. A fu ll scene image of the Radarsat W1 is shown in Figure 3-1. On October 14, 2002, atmospheric conditions over the study area were clear enough for IRS-1D to acquire data. A full scene image of path 125, row 62 of IRS-1D in a false color composite is illustrated in Figure 3-2. In my st udy, all bands of LISS-III sensor (green, red, near infrared, and mid infrared) with 23 m spatial resolution, and a panchromatic image with 5.8 m resolution were used. The original data were stored in 8-bit unsigned format. On October 15, 2002, Radarsat-1 obtained data in S7 mode at 47 incidence angle, 25 m spatial resolution, and 16-bit unsigned pixel depth. A full scene image of the Radarsat S7 (C-band, HH polarization) is shown in Figure 3-3. On October 25, 2002, Landsat-7 acquired data of the study area. At that time, a storm hit the Northeast of Thailand, but sky over the st udy area was clear enough to identify flooded areas. A full scene image of Landsat (path 127, row 49) covering the study area in a false color composite is shown in Figure 3-4. All bands of ETM+ sensor were used in my study. The original data were stored in 8-bit unsigned format. GIS Data A land cover layer used in m y study was supported by Department of Environmental Quality Promotion, Ministry of Natural Resources and Environment, Thailand. Scale of this layer is 1:50,000 referencing UTM pr ojection, and Indian 1975 datum.
55 Other GIS layers such as administration bounda ry, transportation, river network, and river basin boundary were supported by Computer Ce nter, Khon Kaen University, Thailand. These data were created based on 1:50,000 map scal e, UTM projection, and Indian 1975 datum. A DEM with metric vertical units and 30 m pixel size referencing UTM projection and Indian 1975 datum was supported by GISTDA. A set of 1:50,000 digital topographic maps (UTM, WGS 1984) covering the study area were also supported by GISTDA (Figure 3-5). Hydrological Data and HEC-RAS Model For m y study, the Royal Irrigation Department of Thailand provided river cross section and daily hydrological data: discharge (m3/s) and water surface elev ation (m above mean sea level). Data stored in text f iles were collected at hydrological stations along the Mun and Lam Sieo Yai Rivers (Figure 3-6). HEC software used in my study were HE C-GeoRAS version 4.1.1 for ArcGIS 9.1 and HEC-RAS version 4.0 Beta, downloaded from U. S. Army Corps of Engineers website ( http://www.hec.usace.army.mil/software/hec-ras/). Th e webs ite also provided a set of practical manuals, such as HEC-GeoRAS users manual, HEC-RAS users manual, Hydraulic reference manual and HEC-RAS application guide. Data Preprocessing Data were manipulated in rater-based envir onm ent of Erdas Imagine software and vector-based-environment of ArcGIS software. In raster-based environment, all satellite imageries and the DEM were co-registered (Fig ure 3-7, upper part), but for Radarsat images, speckle noise was suppressed before the registration. In vector-b ased environment, geometry data of HEC-RAS were prepared using ArcGIS software and exported to HEC-RAS for water surface elevation calculation, and then, a HE C-RAS flood map was produced (Figure 3-7, upper right portion). The HEC-RAS flood map was used to validate two flood maps interpreted from
56 Radarsat W1 and S7 images to determine best Radarsat image in my study, which was objective 2 (Figure 3-7, middle portion). The best Radars at image was an input of a pan-sharpening process, while three other input images were IRS panchromatic image and two images from PCA, which input data were Landsat and IRS mu ltispectral (Figure 3-7, middle left portion). The fused image was then created. By using unsuperv ised classification, a flood map was interpreted from the fused image, which was objective 1 (Fig ure 3-7, lower left portion). The fused image flood map and the best Radarsat flood map were validated agai nst my visual interpretation, which was objective 3 (Figure 3-7, bottom right portion). To detect under-water features (objective 4), areas in flooded zone were compar ed to land cover layer to determine relationship between colors in the fused image and land c over types (Figure 3-7, bottom left portion). Data preprocessing was divided into three parts: speckle noise suppression, geometric correction and HEC-RAS data prep aration. The first two processes were done in raster-based environment, while the last was done in vector-based environment. Speckle Noise Suppression Like other radar im ages, the Radarsat images used also have speckle noise. By using a radar speckle suppression module in Erdas Imag ine software, this noise was reduced. Erdas Imagine provides seven speckle noise filters: mean filter, median filter, Lee-sigma filter, local region filter, Lee filter, Frost filter and Gamma map filter. A study of Yongwei and Zong-Guo (1996) eval uated performance of the seven speckle noise filters in Erdas Imagine using five crit eria: speckle suppression index, edge enhancing index, feature preserving index, im age detail preserving coefficient, and speckle image statistical analysis. They found that Frost and Lee filters were good in preserving edges and features, but the Frost filter had lowest ability in speckle s uppression. Therefore, the Le e filter was chosen to
57 be applied because it can suppress speckle and pr eserve edges of wet and dry areas, which is most important for flood delineation. Geometric Correction To avoid dis crepancy in the fused image, all images used must be tightly fitted to each other. Thus geometric correction is a very important step. First of all, the panchromatic image of IRS was geometrically corrected referencing 1:50,000 digital topographi c maps (UTM, WGS84), as shown in top left portion of Figure 3-7. In th e rectified panchromatic image, linear features (such as roads and man-made channels) were more up to date than in topographic maps. These obvious features made finding ground control point s (GCPs) easier. As a result, the rectified panchromatic image was used as a reference to register IRS multispectral, Landsat and Radarsat images, and DEM. First order polynomial coordinate transformation was applied to all images and the DEM due to the flat topography in the study area. Bicubic spline resampling method was used to avoid a blocky appearance in rectified images. Table 3-1 describes root mean square (RMS) residual errors of the GCPs in the rectification of IRS panchromatic and multispectral, Landsat, Radarsat W1 and S7 images, and DEM. Total RMS error of each rectification was less than half pixel of input data. Rectified IRS multispectral and Landsat images were input in PCA later on (middle left portion of Figure 3-7). Because PCA requires eq ual pixel size for all inputs, pixel size of rectified IRS multispectral (originally 23 m) and Landsat (originally 30 m) was changed to five m during the registration process. For the radar im ages, the pixel size of the Radarsat W1 (30 m) and Radarsat S7 (25 m) did not change because th ese images were later on classified to produce flood maps (middle portion of Figure 3-7). The or iginal spatial resoluti on of these Radarsat images was one factor considered when their ab ility to illustrate flooded areas was evaluated. However, the pixel depth of the Radarsat images was changed from 16 bits to eight bits to reduce
58 data volume. Moreover, one of these Radarsat im ages was an input in a pan-sharpening process (middle left portion of Figure 3-7) requiring same pixel depth for all inputs, and pixel depth of other inputs (IRS panchromatic and principal component images) was eight bits. After the geometric correction, all satellite images (IRS panc hromatic and multispectral, Landsat, and both Radarsat W1 and S7 images) and DEM were tightly fit to each other. HEC-RAS Data Preparation Input data for HEC-RAS are geom etry and stead y flow data. Geometry data were prepared in ArcGIS environment, but steady flow data (discharge and water su rface elevation) were directly input in HEC-RAS softwa re (top right portion of Figure 3-7). HEC-GeoRAS, an ArcGIS extension, was used to prepare geometry data. These included stream cen ter lines, bank lines, flow path center lines, cross s ection cut lines, and roughness coeffi cients according to land cover categories. Stream center lines were digitized from ups tream to downstream according to a topology rule of HEC-RAS. A junction was formed wher e endpoints of three or more rivers/reaches connected. Next, the stream center lines were copied as a flow path of the main channel. Left and right bank lines were also di gitized, but direction of thes e lines was not important. Cross section cut lines were generated using an XS Cut Line tool in HEC-GeoRAS. These lines started from left overbank to right overbank when facing downstream, and were perpendicular to the flow path lines. Next, le ngth of the cross section cut lines was manually adjusted to cover the floodplain. In this step, all rectified sate llite imageries were used as backgrounds to estimate the flood boundary. Some cross section cut lines were modified to comply with rules of HEC-RAS (cut lines must cross the main channel only once, and cross section cannot intersect each other), as shown in Figure 3-8 A. Elevations of cross section surfaces were generally extracted from the DE M, except for cross sections at hydrological
59 stations, where elevation data were available. Fi gure 3-8 B shows a profile of a cross section in a circle of Figure 3-8 A. Roughness coefficients used in HEC-RAS are values of Mannings n estimated using land cover data. Mannings n of each land cover type applied was estimated based on values proposed by Chow (1959). For example, Mun River was co mparable to main channel (clean, winding, some pools and shoals), thus Mannings n wa s estimated to be 0.040. Estimated Mannings n value for each land cover type applied in my st udy is shown in Table 3-2. After geometry data were ready, these data were exported from ArcGIS, and th en imported to HEC-RAS for calculation of water su rface elevation later on. Data Analysis Calculation of Water Surface Elevation Using HEC-RAS HEC-RAS calculates water su rface elevation using geometry and steady flow data. Geometry data included stream cen ter lines, bank lines, flow path center line s, cross section cut lines and roughness coefficients (Mannings n). After using HECGeoRAS to create geometry data in the ArcGIS environment, data were then imported to HEC-RAS (top right portion of Figure 3-7). If number of elevation points at a cross section was more than 500, exceeding the HEC-RAS limitation, those elevation points were filte red using cross section point filter, a tool in geometric data section. In addition, if number of values of Mannings n at a cross section was more than 20, then values covering a small area farthest from the main channel were manually deleted. In other word, small land cover areas were merged into larger land cover type dominating that area. Steady flow data for HEC-RAS were (1) aver age daily discharge collected on October 11 and 15, 2002, and (2) average daily water surfac e elevation collected on October 11 and 15, 2002. The average values were representatives of hydrological data of October 11 and 15, 2002,
60 when Radarsat W1 and S7 images were acq uired consecutively. A flood map calculated by HEC-RAS was later compared to flood maps interpreted from Radarsat W1 and S7 images (objective 2 of my study), as shown in middle right portion of Figure 3-7). In HEC-RAS, each reach must have at leas t one discharge data point, but in my study, some reaches did not have hydrological station, and consequently no discharge data. As a result, discharge values were estimated using addition of upstream discharges because conservation of mass was assumed and the flow regime was assume d steady (no temporal change in velocity). After a discharge value was ente red for each reach, HEC-RAS assumed that discharge remained constant for a reach. Next, water surface elevatio n at the downstream ends of the river system was entered. Additional data, su ch as known water surface elevati ons at hydrological stations, were also entered. Computational starting point was a cross section at the downstream end of the river system. HEC-RAS computed wate r surface elevation from one cr oss section to the next one upstream using the standard step method. Duri ng computation, HEC-RAS did not compute water surface elevation of cross sections where wate r surface were previously entered. The program used the known water surface to calculate water surface elevation of the next upstream cross section. The calculation process continued until water surface elev ation of every cross section was computed. After the calculation was finished, the result was exported from HEC-RAS to HEC-GeoRAS. The data were stored in an Ar cGIS geodatabase. The calculated water surface elevation was stored as an attribute at each cross section cut line. Next, water surface Triangulated Irregular Network (TIN) was generated using the data from each cross section. To create a flood map, the water surface TIN was overlaid with the 30 m DEM. The flood map was
61 in raster format with the same ground cell si ze of DEM. Afterward, the flood map calculated by HEC-RAS was used as a reference in a comparis on of flood maps interpreted from Radarsat W1 and S7 images (middle ri ght portion of Figure 3-7). Radarsat Image Interpretation To identify wet and dry pixels in Radarsat W 1 and S7 images, a threshold value of each image was determined using histogram investigati on. If the DN of a pixel was less than or equal to the threshold, that pixel was classified as a wet (flooded), and otherw ise a dry (non-flooded) pixel when its DN was bigger than the threshold. For the Radarsat W1 image, the histogram di stribution is nearly bimodal, as shown in Figure 3-9. The first mode with a lower DN repr esents wet areas, where water surface was a specular reflector reflecting radar signal away from the antenna. The second mode with brighter tone (higher DN) represents dry areas, which were diffuse reflector s returning larg e proportion of microwave energy to the antenna. In the classi fication, threshold values from 16 to 25 were tested because from the histogram, these values were potential thresholds separating between wet and dry areas. After these thresholds were used to classify wet and dry areas, visual evaluation was used to identify an appropr iate threshold. The tested thres holds smaller than 20 were not suitable because many wet pixels in water bodies were misclassified as dry pixels. The tested threshold bigger than 20 were not suitable because plenty of pixels representing shadow of forest canopy in high land areas were misclassified as we t pixels. As a result, 20 was deemed the most appropriate threshold to discrimi nate between wet and dry areas in the Radarsat W1 image. For the Radarsat S7 image, the histogram di stribution is bimodal representing wet and dry areas (Figure 3-10). Threshold values from 18 to 27 were tested, and visual evaluation was used to identify an appropriate threshold. It was f ound that 24 was the most suitable threshold to separate between wet and dry area s in the Radarsat S7 image.
62 After two flood maps were inte rpreted from Radarsat W1 and S7 images, they were compared to the HEC-RAS flood map (reference da ta), and error matrix and Kappa statistics were calculated. This comparison of flood maps derived from Radarsat W1 and S7 images to the HEC-RAS result constitutes objective 2 of my study. In the classification accuracy assessment, a st ratified random sampling method was used to place 250 sample points in each class (wet and dr y). A Radarsat image that yielded higher classification accuracy of flooded areas was deemed the best Radarsat image in my study. Then, it was selected as an input in a pan-shar pening process (middle portion of Figure 3-7). PCA Due to the high spectral redundancy in Lands at and IRS data (Fi gure 3-11), PCA was applied to com press data into fewer bands. Input s of PCA were the IRS multispectral acquired on October 14, 2002 and Landsat acquired on October 25, 2002 data (upper left portion of Figure 3-7). These images were acquired on differe nt dates. Thus outputs from PCA did not only show compressed data, but also showed changes detected during the twel ve-day period. Output images from the PCA were separately analy zed as black and white images. A principal component image that showed the clearest boundary of flood extent was selected to be an input in the pan-sharpening. Another prin cipal component image that show ed the greatest variation in gray levels over flooded areas was chosen to be another input for pan-sharpening. Pan-sharpening Inputs of the pan-sharpening process were (1) the best Radarsat im age that yielded the highest classification accuracy of the flooded areas comparing to HEC-RAS results, (2) the principal component image that clearly show ed flood boundaries, (3) th e principal component image that showed various gray levels in fl ooded areas and (4) the IRS panchromatic image (middle left part of Figure 3-7). A color imag e combining the best Radarsat image and two
63 principal component images was transformed fr om RGB to IHS color model. The intensity component was replaced by the IRS panchromatic image, and then the image in IHS color model was transformed back to RGB color model. The result was a fused image containing data of IRS, Landsat, and Radarsat with five m spatial resolution. Classification of Fused Image To classify the fused im age, unsupervised cl assification was used to obtain 250 classes. These classes were manually grouped into eleven land cover classes comparing to the IRS, Landsat and Radarsat images, and QuickBird images acquire during January February 2006 in GoogleEarth. A 7 7 median filter was applied to the classified image, and any area smaller than 30 30 m (Landsat pixel size that was the largest pixel size of the input data) was eliminated from the classified image. Finally, classification of the fused image was obtained. Of the eleven land cover classes was water body and flooded area. This class represented flood water extent delineated from the fused image, which was objectiv e 1 in my study (lower left portion of Figure 3-7). Next, flooded areas interpreted from the fu sed image were compared to flooded areas interpreted from the best Radarsat image, and my visual interpretation ba sed on the IRS, Landsat and Radarsat images was used as a referen ce. In this step, a HEC-RAS flood map was not suitable to be used as a reference because the fused image included data acquired during October 14-25 (twelve-day period), whereas the best Radarsat image (later found to be Radarsat S7) was acquired on October 15, and using a HEC-RAS flood map of October 19 or 20 (mid-point of the twelve-day period), or October 15 (acquisition date of the Rada rsat S7 image) was not well represent the flood during that period. On the other hand, vi sual interpretation and flood movement knowledge was more suitable to be used as the reference.
64 In the classification of the best Radarsat image, stratified random sampling method was used to place 250 sample points in each class (wet and dry). By using coordinates of these sample points, their locations co uld be identified in IRS, Landsat, and Radarsat S7 images, and then visual interpretation was used to identify if the sample points fell in wet or dry areas. Next, an error matrix, classification accuracies and Ka ppa statistics were calc ulated. This method was also used to assess classification accuracy of th e fused image. Next, clas sification accuracies and Kappa statistics of the fused image were compared to those of the best Radarsat image to determine if classification accuracy of flooded areas derived from the fused image was higher than that from the best Radarsat image alone which was objective 3 of my study (lower right portion of Figure 3-7). According to objective 4 of my study, identifyi ng underwater features were expected to be a derivative of the fused image. Areas within fl ood boundaries were scrutinized comparing to the land cover layer to identify underw ater features (bottom left portion of Figure 3-7). The principal component image that showed various gray levels in flooded areas and the best Radarsat image showing water surface roughness were anticipated to provide information of underwater features.
65 Figure 3-1. Radarsat W1 image taken on Oc tober 11, 2002 at 25 incidence angle, 30 m resolution. Figure 3-2. IRS-1D image take n on October 14, 2002, 23m resolution.
66 Figure 3-3. Radarsat S7 image taken on Oc tober 15, 2002 at 47 in cidence angle, 25 m resolution. Figure 3-4. Landsat-7 image take n on October 25, 2002, 30 m resolution.
67 Figure 3-5. Topographic maps. Figure 3-6. Hydrological stations.
68 Figure 3-7. Flowchart of my study. Erdas Imagine ArcGIS, HEC-GeoRAS Radarsat S7 IRS multispectral Landsat Remove speckle noise Register Topographic maps Hydrological data Create geometry data HEC-RAS Calculate water surface elevation ArcGIS, HEC-GeoRAS Create flood ma p HEC-RAS flood map Radarsat W1 Rectified Radarsat S7 Rectified Radarsat W1 S7 flood map W1 flood map Rectified IRS multispectral Rectified Landsat PCA Compare (Objective 2) Best Radarsat image PC image PC image Flood map from fused image Pan-sharpening Rectified IRS pan Unsupervised classification ( Ob j ective 1 ) GIS layers DEM 30 m IRS pan Register Rectified DEM Flood map from best Radarsat image Compare (Objective 3) Best image for flood delineation Fused image Detect underwater features (Objective 4) Under-water feature map Visual interpretation Land cover layer Histogram threshold classification Analysis each PC image
69 B Figure 3-8. Geometry data in HEC-RAS. A) Rive r channels and cross sections. B) Profile for cross section shown in a circle in A. Mannings n A
70 Figure 3-9. Histogram of Radars at W1. Threshold of 20 between wet and dry pixels is shown. Figure 3-10. Histogram of Rada rsat S7. Threshold of 24 between wet and dry pixels is shown. Figure 3-11. Spectral bands of Landsat-7, IRS-1D and Radarsat -1 showing high redundancy in Landsat and IRS spectral bands.
71 Table 3-1. RMS residual errors of the GCPs Data Input cell Number of RMS re sidual errors (m) Output cell size (m) GCPs X Y Total size (m) IRS panchromatic 5.8 419 1.72 1.65 2.48 5 IRS multispectral 23 413 3.89 3.99 5.57 5 Landsat 30 378 6.04 5.27 8.01 5 Radarsat W1 30 175 4.49 4.20 6.15 30 Radarsat S7 25 158 2.44 2.49 3.48 25 DEM 30 17 4.02 3.93 5.62 30 Table 3-2. Estimated Mannings n va lues (adapted from Chow, 1959) Land cover Mannings nComparable to Built-up area 0.017 Unfinished concrete Grass land 0.030 Short grass Non-forested wetland 0.030 Short grass Paddy field 0.035 High grass Horticulture 0.035 Mature row crops Field crops 0.040 Mature field crops Water body 0.040 Natural streams, clean, winding, some pools and shoals Orchard 0.080 Heavy stand of timber, a few down trees, little undergrowth, flood stage below branches (minimum Mannings n value) Forest plantation 0.080 Heavy stand of timber, a few down trees, little undergrowth, flood stage below branches (minimum Mannings n value) Riverine trees 0.100 Heavy stand of timber, a few down trees, little undergrowth, flood stage below branches (normal Mannings n value) Deciduous forest 0.100 Heavy stand of timber, a few down trees, little undergrowth, flood stage below branches (normal Mannings n value) Forested wetland 0.100 Heavy stand of timber, a few down trees, little undergrowth, flood stage below branches (normal Mannings n value)
72 CHAPTER 4 RESULTS HEC-RAS and HEC-GeoRAS Results A result fro m HEC-RAS was a calculated water surface elevation. This output was stored as an attribute of each cross section line. To create a flood map, the calculated water surface elevation was exported from HEC-RAS, and then imported to HEC-GeoRAS in ArcGIS environment. Next, water surface elevation data at each cross section line was used to create a water surface TIN, which was then overlaid with the 30 m DEM and finally the HEC-RAS flood map was created. Flooded areas and water surfac e elevation simulated by HEC-RAS were shown in Figure 4-1. The flooded areas were with in coverage of cross section lines. Flood Maps Interpreted from Radarsat Images Radarsat W 1 and S7 images (Figure 4-2, A and B respectively) were classified using threshold values of 20 and 24 respectively, a nd flood maps were produced. The ground cell size of the flood map interpreted from Radarsat W1 image was 30 m, while that interpreted from Radarsat S7 was 25 m according to pixel size th e original images. Figure 4-3 shows overlaid flooded areas interpreted from the Radarsat W1 and S7 images and those simulated by HEC-RAS. Comparison of Radarsat Flood Ma ps Referen cing HEC-RAS Flood Map To compare the flood maps interp reted from the Radarsat W1 and S7 images, the HEC-RAS flood map was used as a reference. In the Radarsat W1 and S7 flood maps, flooded areas outside the floodplain were masked out be cause coverage of the HEC-RAS flood map was the floodplain area (within the covera ge of cross section lines). In the Radarsat W1 flood map, 250 sample point s were randomly placed in areas classified as wet, and other 250 sample points were randomly placed in areas classified as dry. Locations of
73 these 500 sample points were projected on the HEC-RAS flood map (the reference data) to evaluate classification accuracy of the Radarsat W1, and then, an error matrix, classification accuracies and Kappa sta tistics were calculated. This method was also applied to evaluate classification accuracy of the Radarsat S7 image. For the classification of the Radarsat W1 image, overall classification accuracy was 87.60%. Because the primary purpose of my study was to map flooded areas, wet category was in focus. Producers accuracy of the wet cat egory was 95.63% that was 95.63% of flooded areas was correctly identified as wet areas. Users accuracy of the wet category was 78.80% that was only 78.80% of pixels classified as wet were rea lly wet areas (Table 4-1 and 4-2). Kappa value of the wet class was 0.6395 (Table 4-3) that was classification of flooded areas was 63.95% better than randomly classify pixels as wet pi xels. The overall Kappa statistic was 0.7520 (this classification was 75.20% better than a random classification). For the classification of the Radarsat S7 image, overall classification accuracy was 94.00%. For the wet category, producers accu racy was 97.01%, and users accuracy was 90.80%, which were 97.01% of wet areas were correctly identified, a nd 90.80% of pixels classified as wet were truly wet (Table 4-4 and 4-5). Kappa value of the wet class was 0.8271 (Table 4-6) that was classification of flooded areas was 82.71% better than a random classification. The overall Kappa statistic was 0.8800 (this classificati on was 88.00% better than a random classification). Considering only the wet class (flooded), producers and users accuracies of the Radarsat S7 classification were higher th an those of the Radarsat W1 classification. Moreover, kappa value of the wet class and the overa ll kappa statistic of the Radarsat S7 classification were higher than those of the Radarsat W1 classification. Thus, the Radarsat S7 image yielded higher
74 classification accuracies of flooded areas than th e Radarsat W1 image referencing the HEC-RAS flood map, and objective 2 of my st udy was completed. The Radarsat S7 image was then chosen to be an input of pan-sharpening process. PCA Results Outputs of the PCA wer e 12 black and white images because inputs of PCA were four bands of IRS and eight bands of Landsat data. Fa lse color composite images of IRS and Landsat, input data of the PCA, are shown in Figure 44, A and B respectively. Eigen values and eigen vectors were shown in Table 4-7. Eigen values we re used to calculate pe rcentage of the total scene variance of each principal component, sh own in Table 4-8. PC1 image contained 63.53% of the total scene variance, while PC2, PC3, PC12 contained successively smaller percentages. In PC1 image (Figure 4-5, A), flooded areas a ppeared as dark areas while dry areas and clouds appeared as brighter tone. This imag e clearly showed flood boundaries. For PC2 image that contained 21.32% of the total scene varian ce (Figure 4-5, B), flooded areas and clouds appeared as dark areas causing confusion in flood delineation. PC3 image contained 6.22% of the total scene variance (Figured 4-5, C). Both flooded areas and dry areas appeared as bright tone and flood boundaries were not obvious. Road ne tworks were obvious in this image. For PC4 image that contained 4.91% of th e total scene variance (Figure 45, D), details in flooded areas were illustrated. There were various gray levels in the floodplain and th e meandering river (dark tone) was differentiated from its surroundings. The PC1, PC2, PC3 and PC4 explained 95.98% of the variance in the Landsat and IRS images. PC5 to PC12 images contained only 4.02% of the total scene vari ance. Although PC5 to PC12 images were individually examined, no image showed flood boundaries or details in flooded areas.
75 Among the principal component images, the PC1 image showed the most obvious flood boundary, while the PC4 showed details in the flooded areas. As a result, the PC1 and PC4 images were selected to be inputs of the pa n-sharpening. Although the PC 3 showed obvious road networks, it was not chosen because the IRS panchromatic image illustrated road networks better. Fused Image and Its Classification In order to create a fused im age containing all necessary information about flooding, the Radarsat S7, PC1 and PC4 images were select ed, and then were sharpened by five m IRS panchromatic image. The fused image of PC4 (in red), Radarsat S7 (in green), and PC1 (in blue) is shown in Figure 4-6. The fused image was classified into 250 classes and finally grouped in to eleven land cover classes including water body and flooded area, paddy field, field crops, orchard, horticulture, forest plantation, deciduous forest, forested wetland, non-forested wetland, built-up area, and dead vegetation, as shown in Figure 4-7. The water body and flooded area class represented flooded areas delineated from the fused im age, which was objective 1 of my study. Dead vegetation appears in the fused image because during October 11-25, 2002 (when Radarsat, IRS, and Landsat acquired the data ) water surface elevation of the Mun River decreased gradually (Figure 48). On October 14, 2002 when IRS acquired the data, water surface elevation was approximately one m higher than that on October 25, 2002 when Landsat acquired the data. As a result, some flooded areas in the IRS imag e were exposed dead vegetation in the Landsat image. In the IRS im age (Figure 4-9, A) roads and villages are in bright tone, paddy field app ears as red, and flooded area appears in darker tone. After flood water receded, dead vegetation exposed. An exampl e of dead vegetation area is area X in the Landsat image (Figure 4-9, B). Also, a built up water body (area Y) can be seen from the
76 Landsat image. After applied PCA to the IRS and Landsat data, dead vegetation showed as a very bright tone in PC4 image (area X in Figure 4-9, C). In Radarsat S7 and PC1 images (Figure 4-9, D and E respectively), dead vegetation (area X) is not different from its surroundings, but the built up water body (area Y) is obvious. After the PC4, Radarsat S7, and PC1 images were assigned in red, green, and blue channels respec tively, dead vegetation became visible as red areas in the fused images (area X in Figure 4-9, F, for example). The built up water body is visible in the fused image (area Y) A row of trees that appeared as a bright line in the Radarsat S7 image (area Z in Figure 4-9, D) showed in the fused image as a green line (area Z in Figure 4-9, F). The row of trees appeared in gree n because this information only showed in the Radarsat S7 image (in green channel of the fused image). Another important category in my study is water body and flooded areas. To get truly flooded areas from the water body and flooded area cla ss, areas in this class were subtracted by water body polygons from the land cover layer. Next, the truly flooded areas and the dead vegetation areas were combined and total damage d areas were obtained. To assess damage of the 2002 flood, the total damaged areas were overlaid with the land cover la yer. Finally, damaged areas of each land cover category were derive d, as shown in Figure 4-10. The 2002 flood mainly inundated paddy field (73.31 sq.km.) and horticult ure (23.99 sq.km.), whic h were 63% and 20% of the flooded areas, respectivel y. Although 7% and 8% of flooded areas were forested and non-forested wetland, these areas were not consid ered damaged because they were seasonally flooded. A total flooded area of other land cover categories (orcha rd, grass land, riverine tree, and built up area) was 1.80 sq.km (2% of the flooded areas).
77 Comparison of Flood Maps from Fu sed Image and Best Radarsat Image Because the Radarsat S7 im age was considered the best Radarsat image in my study, the flood map interpreted from the Radarsat S7 was compared to that interpreted from the fused image according to objective 3 of my study. However, flooding condition on October 15, 2002 when Radarsat acquired data in S7 mode was different from the condition on Octobe r 25, 2002 when Landsat acquired data in that the water surface elevation when the Landsat data were taken was lower than that when the Radarsat S7 data were taken. A ccordingly, some flooded areas in the Radarsat S7 image (area X in Figure 4-11, A) appear as dry areas in the La ndsat image (area X in Figure 4-11, B), and also in the fused image (area X in Figure 4-11, C) These dry areas were dead vegetation. The classification of the Radarsat S7 and fused images also showed that some parts of Radarsat S7 flooded areas were dead vegetation in the fused image (area X in Figure 4-11, D and E). Thus, before flooded areas interpreted from the Radarsat S7 image were compared to those interpreted from the fused image, the Radarsat S7 flooded areas were needed to be reduced by the dead vegetation areas. A purpose of this reduction was to eliminate influence of difference in water surface elevation on October 15 and 25, 2002. Figur e 4-12 shows overlaid flooded areas of the reduced flood map of Radarsat S7 and the flood map from the fused image. The two flood maps were compared referenci ng my visual interpre tation of the IRS, Landsat, and Radarsat S7 images. In the redu ced flood map of Radarsat S7, 250 sample points were randomly placed in areas classified as we t, and other 250 sample points were randomly placed in areas classified as dr y. Locations of these 500 sample points were projected on overlaid IRS, Landsat, and Radarsat S7 images, and visual interpretation was used to design if the sample points fell in wet or dry areas. Then, an e rror matrix, classification accuracies and Kappa
78 statistics were calculated. This classification accu racy assessment method was also applied to the fused image. For the flood map interpreted from Radarsat S7 image, overall classification accuracy was 95.40%. Because the primary purpose of my study wa s to map flooded area, wet category was in focus. Producers accuracy of the wet cate gory was 98.30% meaning that 98.30% of flooded areas were correctly identified as wet areas. Users accuracy of the wet category was 92.40%, which meant that 92.40% of pixels classified as wet were really wet areas (Table 4-9 and 4-10). Kappa value of the wet class was 0.8566 (Table 4-11) meaning classification of flooded areas was 85.66% better than randomly assigned pixels as flooded pixels. The overall Kappa statistic was 0.9080 (this classification was 90.80% better than a random classification). For the flood map interpreted from the fuse d image, overall classi fication accuracy was 96.60%. For the wet category, producers accu racy was 99.16%, and users accuracy was 94.00%, which meant that 99.16% of wet areas were correctly identifie d, and 94.00% of pixels classified as wet were truly wet (Table 4-12 and 4-13). Kappa value of the wet class was 0.8859 (Table 4-14), which meant that classification of flooded areas was 88.59% better than a random classification of flooded areas. The overall Kappa statistic was 0.9320 (this classification was 93.20% better than a random classification). Comparing between two flood ma ps derived from the Radarsat S7 and fused images, overall classification accuracy, producers accuracy of the wet class, and users accuracy of the wet class of that interpreted from the fused image were slightly higher than that interpreted from the Radarsat S7. Moreover, the overall Kappa value and the Kappa value of the wet class also confirmed that the classification of the fused imag e was better. Therefore, classification accuracy
79 of flooded areas derived from the fused image was higher than that derived from the Radarsat S7 alone and object 3 of my study was completed. Underwater Feature Detection To detect underwater features in the f use d image, non-flooded areas were masked out and only areas classified as water body and flooded ar eas were inspected and compared to the land cover layer. In the land cover layer and the fused image (Figure 4-13, A and B), flooded areas including paddy field, horticulture, forested we tland and non-forested wetl and appears in three major colors, black, green and orange. The black areas were completely flooded zones because these areas appeared dark in PC4 (red), Radarsat S7 (green) and PC1 (blue). The green areas were high moisture content vegetations or dikes surrounding paddy fiel ds that had high backscatter. These objects appeared only in th e Radarsat S7 image (Figure 4-13, D) that was displayed through green channel in the fused image. The orange tone areas were wet areas (wet vegetation and wet soil) appeari ng in both Radarsat S7 image (i n green channel) and PC4 image (in red channel). For the completely flooded areas, which appe ared black in the fused image, it provided only flood boundary data, but did not provide info rmation about underwater features. For green areas, which were high moisture vegetations or dikes surrounding paddy fields, they were not submerged features. Therefore, they did not provide information of underwater features. However, pattern of green lines (dikes) showed that the flooded areas were paddy fields (area X in Figure 4-13, B). Wet vegetation and wet soil appeared orange in the fused image (Figure 4-13, B). When the fused image was compared to th e land cover data (Figure 4-13, A), the orange areas could be flooded paddy field, horticulture, forested wetland or non-forested wetland. They did not exclusively relate to any land cover category. A relationship be tween land cover types
80 and the three colors (black, green and orange) in flooded areas of the fused image could not be found. Therefore, underwater features could not be detected using the fused image. At first, data of Landsat blue band (Figur e 4-13, C) was expected to yield underwater information. Also, the Radarsat data (Figure 413, D) was supposed to show change in water surface roughness caused by different kinds of unde rwater features. However, the expected information did not show in the Landsat blue band, Radarsat and fused images. Therefore, we cannot detect underwater features from the fused image, and objective 4 of my study was completed.
81 Figure 4-1. Flooded areas and water su rface elevation simulated by HEC-RAS. A B Figure 4-2. Radarsat images. A) Radarsat W1 acquired on Oc tober 11, 2002 at 25 incidence angle. B) Radarsat S7 acquired on Oc tober 15, 2002 at 47 incidence angle.
82 Figure 4-3. Overlaid flooded ar eas from HEC-RAS, Radarsat W1 and Radarsat S7 images. A B Figure 4-4. Inputs of PCA. A) IRS multisp ectral acquired on Oct ober 14, 2002. B) Landsat acquired on October 25, 2002.
83 Figure 4-5. Principal component images and percentages of the total scene variance. A) PC1 contained 63.53%. B) PC2 contained 21.32%. C) PC3 contained 6.22%. C) PC4 contained 4.91%. A B C D
84 Figure 4-6. Fused image of PC4 (red), Radarsat S7 (green), and PC1 (blue) images.
85 Figure 4-7. Fused image classification.
86 Figure 4-8. Water surface elev ation measured from hydrological stations during October 11-25, 2002, dates of data acquisition. A B C D E F Figure 4-9. Dead vegetation (area X), built up wa ter body (area Y), and a row of trees (area Z). A) IRS image. B) Landsat image. C) PC4 image. D) Radarsat S7 image. E) PC1 image. F) Fused image (PC4 in red, Rada rsat S7 in green, and PC1 in blue). X Y Z Z X X X X Y Y Y Y
87 Figure 4-10. Flooded areas of each land cover type.
88 D E Figure 4-11. Differences caused by changing in water surface elev ation. A) Larger flooded areas in Radarsat S7image. B) Smaller floode d areas and exposed dead vegetation in Landsat image. C) Flooded areas (in black) and dead vegetation (in red) in fused image. D) Flooded areas classi fied from Radarsat S7 image. E) Classification of fused image. B C A X X X X X X X X X X
89 Figure 4-12. Overlaid flooded areas from the re duced flood map of Radarsat S7 and the fused image flood map.
90 A B C D Figure 4-13. Inspection of underwater features. A) Land cover laye r. B) Fused image. C) Blue band image of Landsat ETM+. D) Radarsat S7. X
91 Table 4-1. Error matrix of cla ssification of Radarsat W1 image Classified Data Wet Dry Row Total Wet 197 53 250 Dry 9 241 250 Column Total 206 294 500 Table 4-2. Accuracy totals of cl assification of Radarsat W1 image Class name Reference totals Classified totals Number corrected Producers accuracy (%) Users accuracy (%) Wet 206 250 197 95.63 78.80 Dry 294 250 241 81.97 96.40 Totals 500 500 438 Overall Classificati on Accuracy = 87.60% Table 4-3. Kappa statistics of classification of Radarsat W1 image Class name Kappa Wet 0.6395 Dry 0.9126 Overall Kappa Statistics = 0.7520 Table 4-4. Error matrix of cla ssification of Radarsat S7 image Classified Data Wet Dry Row Total Wet 227 23 250 Dry 7 243 250 Column Total 234 266 500 Table 4-5. Accuracy totals of cl assification of Radarsat S7 image Class name Reference totals Classified totals Number corrected Producers accuracy (%) Users accuracy (%) Wet 234 250 227 97.01 90.80 Dry 266 250 243 91.35 97.20 Totals 500 500 470 Overall Classificati on Accuracy = 94.00% Table 4-6. Kappa statistics of classification of Radarsat S7 image Class name Kappa Wet 0.8271 Dry 0.9402 Overall Kappa Statistics = 0.8800
92 Table 4-7. Eigen values and eigen vectors. PC Eigen Eigen vector value ETM+1 ETM+2 ETM+3ETM+4ETM+5ETM+6ETM+7 ETM+8IRS2IRS3IRS4IRS5 PC1 1914.87 0.30012 0.31707 0.35545 0.32802 0.47208 -0.04567 0.32914 0.28538 0.03955 0.03913 0.38836 -0.08247 PC2 642.52 -0.34110 -0.27820 -0.39885 0.24251 0.05198 0.05308 -0.13202 0.01483 -0.06612 -0.11552 0.73369 0.09575 PC3 187.59 -0.14832 -0.11258 0.10416 -0.34432 0.13347 0.12218 0.15520 -0.18420 0.50815 0.63452 0.18908 0.22040 PC4 147.98 -0.26336 -0.21986 -0.17080 0.11979 0.59482 0.19343 0.32655 -0.02851 -0.19037 -0.10341 -0.41403 0.34893 PC5 46.92 -0.10471 0.20230 -0.01916 0.54643 -0.20329 0.20902 -0.32874 0.37189 0.29223 0.20289 -0.22133 0.37711 PC6 46.78 0.23642 0.15828 0.21148 -0.27097 -0.22402 0.36298 0.06901 -0.08033 -0.21243 -0.28664 0.22842 0.65444 PC7 8.01 0.17105 0.18653 -0.00237 0.36894 0.15679 0.01533 -0.20160 -0.82342 0.20218 -0.12084 -0.00928 0.02792 PC8 6.20 0.31925 -0.13931 -0.19203 0.30003 -0.27351 0.02122 0.27901 -0.15505 -0.49857 0.56978 0.00312 0.03742 PC9 5.77 -0.45093 -0.00620 0.62090 0.04592 0.09109 -0.01397 -0.35157 -0.11967 -0.46161 0.21753 0.03990 -0.02440 PC10 4.58 0.46972 -0.09819 -0.20716 -0.23751 0.44399 0.03778 -0.62301 0.14491 -0.15136 0.18732 0.00473 0.06843 PC11 2.70 -0.28320 0.79448 -0.40515 -0.20979 0.08052 -0.01388 0.02124 -0.03827 -0.20891 0.16379 0.02861 -0.02351 PC12 0.21 0.00049 -0.00062 0.00006 0.00020 -0.00030 -0.87453 0.00039 0.00008 0.00008 -0.00003 0.00002 0.08497 Table 4-8. Percentage of total scene variance Principal component image Percen tage of total scene variance PC1 63.53 PC2 21.32 PC3 6.22 PC4 4.91 PC5 1.56 PC6 1.55 PC7 0.27 PC8 0.21 PC9 0.19 PC10 0.15 PC11 0.09 PC12 0.01
93 Table 4-9. Error matrix of flood map interpreted from Radarsat S7 image Classified Data Wet Dry Row Total Wet 231 19 250 Dry 4 246 250 Column Total 235 265 500 Table 4-10. Accuracy totals of flood ma p interpreted from Radarsat S7 image Class name Reference totals Classified totals Number corrected Producers accuracy (%) Users accuracy (%) Wet 235 250 231 98.30 92.40 Dry 265 250 246 92.83 98.40 Totals 500 500 477 Overall Classificati on Accuracy = 95.40% Table 4-11. Kappa statistics of flood map interpreted from Radarsat S7 image Class name Kappa Wet 0.8566 Dry 0.9660 Overall Kappa Statistics = 0.9080 Table 4-12. Error matrix of flood map interpreted from fused image Classified Data Wet Dry Row Total Wet 235 15 250 Dry 2 248 250 Column Total 237 263 500 Table 4-13. Accuracy totals of fl ood map interpreted from fused image Class name Reference totals Classified totals Number corrected Producers accuracy (%) Users accuracy (%) Wet 237 250 235 99.16 94.00 Dry 263 250 248 94.30 99.20 Totals 500 500 483 Overall Classificati on Accuracy = 96.60% Table 4-14. Kappa statistics of flood map interpreted from fused image Class name Kappa Wet 0.8859 Dry 0.9831 Overall Kappa Statistics = 0.9320
94 CHAPTER 5 DISCUSSION Discussion chapter is divided into three parts. The first part is fl ood prediction using HEC-RAS, which explains lim itation s of my study due to lack of input data, and using remotely sensed data to properly adjust geometry data of HEC-RAS. The second part is Radarsat W1 versus Radarsat S7 discussing influence of incide nce angle and spatial resolution of the Radarsat images on classification accuracy. The last part of this chapter is the fused image discussion about its information and limitation. Flood Prediction Using HEC-RAS An i mportant data requirement of HEC-GeoRAS is a high resolution DEM because ground cross section data are mainly extracted from the DEM, and the DEM is also used to overlay with a water surface TIN (simulated by HEC-RAS) to produce a flood map. In my study, a 30 m DEM with elevation values to the whole meter was used. However, spatial pattern of flooded areas in the HEC-RAS result was still similar to the pattern of flooded areas in the Radarsat images because the DEM was good enough to represent topography of the study area, a rural area with flat topography (slope < 0.5%). If the sa me DEM was to be used to predict flood in an urban area, the result likely would be unacceptabl e because the DEM, at this resolution, does not well represent roads, buildings, or other objects obstructing the flow. Other inputs of HEC-RAS are hydrological data. The model requi res at least one discharge for a reach, but in my study area, some reaches did not have discharge data. These missing data were estimated from addition of upstream discharg es, and the estimated input discharges affected the output of HEC-RAS. Furthermore, hydrological stations in my study ar ea are too far apart. As a result, spatial variation of hydrological data may not be measured well. Due to these
95 limitations, HEC-RAS might not yi eld its best output in my study. However, flooded areas simulated by HEC-RAS still had spatial patterns similar to those of the Radarsat images. HEC-RAS delineates inundated areas only within coverage of cross sections. At a location, a cross section should long enough to cover the entire floodplain. If cross secti ons are too short, predicted flooded areas may be narrower than what they should be. On the other hand, if cross sections are too long, the number of elevation points of a cros s section may exceed 500 and/or number of Mannings n values may exceed 20, a nd then data modification (such as applying cross section point filter to th e data, and grouping land cover cla ss of a smaller area into a bigger area) is required. If cross section lines are too long, this will lead to a larger area of interest and longer calculation time. Also, false flooded areas may appear as many small polygons due to depressions or sinks in DEM da ta. In my study, satellite imager y was used as a background in cross section length adjustment. Therefore, leng th of cross section lines could be properly adjusted to cover the flooded area s, and remotely sensed data c ould facilitate users in defining boundary of HEC-RAS calculation. Radarsat W1 versus Radarsat S7 It has been shown that a Radars at image taken at large incid e nce angle (more than 40) is useful for land and water differentiation. For flooded forests, a small incidence angle (20 to 31) of Radarsat signal could penetr ate canopies of willow, grasses or sedges, and detect standing water under the vegetation. However, it has also been shown that increasing incidence angle of Radarsat signal did not substantially decr ease distinction between flooded and non-flooded forest. According to my study, the Radarsat S7 image (47 incidence angle and 25 m pixel size) was found to give higher classification accura cy of flooded and non-flooded areas than the Radarsat W1 image (25 incidence angle and 30 m pixel size). This may have resulted from two
96 factors: (1) an advantage of the Radarsat S7 im age that it has higher spatial resolution than the Radarsat W1 image (25 m versus 30 m) and (2) an influence of land cover types in the flooded zones. In my study area, 63% of flooded areas was paddy field, 20% was horticulture, 8% was non-forested wetland, 7% was forested wetland, and 2% was other land cover types (orchard, grass land, riverine tree, and built up area). If the major land c over type in flooded areas is forest, the Radarsat S7 image taken at a large incide nce angle may give lowe r classification accuracy, and the Radarsat W1 taken at a small incidence angle may give higher classification accuracy. Although classification accuracy derived from th e Radarsat W1 image was lower than that from the Radarsat S7 image, the Radarsat W1 data were still valuable for flood monitoring. In a study of flooding, high temporal resolution data are important. Therefore, various incidence angles and/or various spatial resolutions of radar images may unavoidable because a radar satellite has to take images at di fferent incidence angles in order to obtain data of the same study site as frequently as possible. The Fused Image The prim ary purpose of the fused image was to include all necessary information about flooding into one image, and a more obvious flood boundary (compared to a Radarsat image alone) was expected. When considering only a pansharpened image of the Radarsat S7 (one of three bands in the fused image), the sharpened image yielded more distinct flood boundaries (area X in Figure 5.1, A) compared to the original 25 m Radarsat S7 image (area X in Figure 5.1, B). Furthermore, the sharpened image showed a more obvious bridge across the river and clearer floodwater over the ends of the bridge (area X in Figure 5.1, A and B). Boundary of a man-made pond was more apparent in the sharpened imag e (area Y in Figure 5.1, A and B). Roads and a man-made channel could be seen from the sharpened image, while these features could not be identified in the original Radarsat S7 image (area Z in Figure 5.1, A and B).
97 Classification accuracies and kappa statistics also confirmed that the classification accuracy of flooded areas derived from the fuse d image of IRS-1D, Landsat-7 and Radarsat-1 was higher than that derived from Radarsat-1 alone. In classification accuracy assessment of the fused image and the Radarsat S7 image, my vi sual interpretation of the IRS, Landsat and Radarsat S7 images and flood movement knowle dge were used as the reference. Overall classification accuracy of the fused image (96.60%) wa s slightly higher than that of the Radarsat S7 (95.40%). For the flooded area category, prod ucers accuracy of the fused image was 99.16% while that of the Radarsat S7 was 98.30% (0.86% difference). Users accuracy of the fused image and the Radarsat S7 image were 94.00% and 92.40%, respectivel y (1.60% difference). Kappa value of the flooded area category derive d from the fused image was 0.8859 while that obtained from the Radarsat S7 image was 0.8566 (0.0293 difference). Therefore, if the flood boundary is the only point of intere st, use of the Radarsat image alone is a fast and convenient way to produce the result. However, from point of view of decision makers, more information relating to the flood situation is required, and the fused image fro m my study, for instance, is not only to provide a slightly bette r flood boundary, but also give in formation about damaged areas that were classified as dead vegetation in my study. The IRS panchromatic image (a component of the fused image) also provided information of flooded transportation ro utes (for example, a flooded bridge showed in area X of Figure 5.1, A). Although the fused image contains various in formation of flooding, some other land cover information were neglected because second and third principal components (containing 21.32% and 6.22% of scene variance respectively) of IR S and Landsat data were left out during the fusion process. As a result, some land cover type s could not be identified in the fused image. Comparing to land cover layer, gr ass land and riverine trees were two land cover categories that
98 were not identified in the fused image. Informati on that can be used to differentiate these two land cover classes may appear in the omitted prin cipal component images. Thus, the fused image is suitable for a flood study, but is not appropriate for a land cover classification. In my study, underwater features were not illustrated in the fused image. This may be because flood water was too turbid for the bl ue band of Landsat to penetrate and detect underwater features. Besides, flooded areas we re mainly agricultural land (paddy field and horticulture). This vegetation is not a large object as are coral r eefs or sand bars, which cause breaking waves or changes in water surface roughne ss, which could be detected by radar signal. Moreover, the study area is not a windy area and water surface appeared smooth in the Radarsat image. Information of underwater features did not show in the Radarsat image. Therefore, underwater features in my study area could not be detected us ing the fused image of IRS-1D, Landsat-7 and Radarsat-1.
99 Figure 5-1. Radarsat S7 imag es. A) Pan-sharpened image. B) Original 25 m image. A B X Y Z X Y Z
100 CHAPTER 6 CONCLUSIONS It is im portant to accurately delineate flooded areas because these data can be used as inputs for flood compensation calcula tion. Optical remote sensing data, such as Landsat and IRS, have been used to delineate flood, but optical sensor data are limited because flooded areas usually have extensive cloud cover, especially at the time of the flood. Radar data are suitable for flood delineation due to cloud penetra tion ability. However, it has been shown that radar data are not useful to detect flood in urban areas. Theref ore, in my study, optical sensor data of Landsat and IRS were fused with radar data (Radarsat) to create a fused image, and flooded areas of the 2002 flood in the Mun River Basin were successfully delineated. Flooded and non-flooded areas illustrating the 2002 inundation in the Mun River Basin were interpreted from two Radarsat images: Rada rsat W1 (25 incidence angle, 30 m pixel size) acquired on October 11, 2002 and Rada rsat S7 (47 incidence angl e, 25 m pixel size) acquired on October 15, 2002. Classification accuracy of these Radarsat images was validated against the HEC-RAS flood map, for which requ ired input data included hydrol ogical data (average water surface elevation and average disc harge of October 11 and 15, 2002) and topographic data of the study area. The Radarsat S7 yielde d higher classification accuracy than the Radarsat W1 (overall Kappa value 88.0% versus 75.2%). This has resulte d from the higher spa tial resolution of the Radarsat S7 image and land cover types in the flood zones, dominated by paddy field and horticulture. Therefore, the Radarsat S7 image was deemed the best Radarsat image for flood detection in my study, and was chosen to be an input for image fusion. Another input of the image fusion was the IR S panchromatic image with five m spatial resolution, acquired on October 14, 2002. This image illustrated the road networks in the study area and how the flood situation a ffected transportation routes. Two other inputs of the image
101 fusion, derived from PCA of Landsat (acquire d on October 25, 2002) and IRS multispectral (acquired on October 14, 2002) data, were (1) PC1 (the image that showed the most obvious flood boundary) and (2) PC4 (the image that show ed details in the flooded areas). The color composite image of PC4 (red), Radarsat S7 (gre en) and PC1 (blue) was sharpened by the IRS panchromatic image to create the fused image containing information of the 2002 flood in the Mun River Basin. The fused image was classified using unsupervised classification to obtain eleven land cover classes including water body and floode d area, paddy field, field crops, orchard, horticulture, forest plantation, de ciduous forest, forested wetl and, non-forested wetland, built-up area, and dead vegetation. The fused image contained a temporal factor due to different acquisition dates of the input data (twelve-day period durin g October 14 25, 2002). Exposed dead vegetation due to decreasing of water surf ace elevation during the twelve-day period was identified in the fused image. The water body and flooded area class was used to create the flood map from the fused image, and was then compared to the flood map fr om the Radarsat S7 image (the best Radarsat image in my study). In the comparison, my visual interpretation based on the IRS, Landsat and Radarsat S7 images was used as the reference. It was found that cla ssification accuracy of flooded and non-flooded areas derived from the fused image was slightly higher than that of the Radarsat S7 image alone (overall Kappa value 93.2% versus 90.8%). Therefore, if the flood boundary is the only point of intere st, use of the Radarsat image alone is a fast and convenient way to produce the result. Nevertheless, th e fused image provided a slightly better flood boundary, damaged area information (dead vege tation), and flooded transportation routes.
102 Although the fused image illust rated information relating to the flood, it did not show underwater features because flood water was too turb id for the blue band of Landsat to penetrate and detect underwater features. Also, s ubmerged vegetation (flooded paddy field and horticulture) did not cause change in water su rface roughness. Thus, Radarsat signal could not detect turbulent water caused by the submerged vegetation in the study area. In my study, components of the fused imag e were selected based on a flood study. The same data set of Landsat, IRS and Radarsat can be used for other purposes, such as a road network study, for which the most useful information was shown in the PC3 image. For future studies, fusion of optical sensor a nd radar data to detect flooding in a monsoonal zone will be continued because flood information is necessary for providing relief efforts. Flood boundary data is an important input for flood compensation calculation, and fusion of multispectral and radar data with higher spatial resolution panchromatic imagery will be needed. As more optical sensor satellit es become available, the opportunities to gather data on flooded areas under clear atmospheric c onditions during a rainy season in crease. For Thailand, its first optical sensor satellite, Theos, was launched in October 2008. It provides 2 m spatial resolution panchromatic data and 15 m spatial resolution of multispectral data in blue, green, red and near infrared. With a steerable mirror like that of th e SPOT satellite, Theos can acquire data of a location every three days. The high temporal resolution makes a flood movement study possible. These data can be fused with radar data acqui red regardless of weather conditions. In addition, fusion of optical sensor data with various polar izations of radar data should be studied to determine information content of ea ch polarized image for a flood study.
103 REFERENCES Ahm ad, S., and S. P. Simonovic, 2001. Integration of heuristic knowledge with analytical tools for the selection of flood damage reduction measures, Canadian Journal of Civil Engineering 28(2):208-221. Akar, I., C. Uysal, and D. Maktav, 2008. Determin ation of natural disaster by integration of remote sensing and GIS: the Yeniiftlik stream basin model in Istanbul, Turkey, Proceeding of the the XXIth ISPRS Congress Beijin, China. Bates, P. D., and A. P. J. De Roo, 2000. A simple raster-based model for flood inundation simulation, Journal of Hydrology 236(1-2):54-77. Bates, P. D., M. D. Wilson, M. S. Horritt, D. C. Mason, N. Holden, and A. Currie, 2006. Reach scale floodplain inundation dynamics observed us ing airborne synthetic aperture radar imagery: Data analysis and modelling, Journal of Hydrology, 328(1-2):306-318. Benito, G., A. Diez-Herrero, and M. F. de Villalta, 2003. Magnitude and frequency of flooding in the Tagus Basin (central Spain) over the last millennium, Climatic Change, 58(1/2):171192. Bhan, S. K., and Flood Team, 2001. Study of fl oods in West Bengal during September, 2000 using Indian Remote sensing satellite data, Journal of the Indian Society of Remote Sensing 29(1-2):1-3. Biftu, G. F., and T. Y. Gan, 2001. Semi-distribut ed, physically based, hydrologic modeling of the Paddle River Basin, Alberta, using remotely sensed data, Journal of Hydrology 244(34):137-156. Bjerklie, D. M., D. Moller, L. C. Smith, and S. L. Dingman, 2005. Estimating discharge in rivers using remotely sensed hydraulic information, Journal of Hydrology, 309(1-4):191-209. Blyth, K., 1993. The use of microwave remote sens ing to improve spatia l parameterization of hydrological models, Journal of Hydrology 152(1-4):103-129. Boruah, S., D. Gilvear, P. Hunter, and N. Sharma, 2008. Quantifying channel planform and physical habitat dynamics on a la rge braided river us ing satellite data the Brahmaputra, India, River Research and Applications 24(5):650-660. Canadian Space Agency, 2006. Radarsat-1. URL: http://www.asccsa.gc.ca/eng/satellites/radarsat1/components. asp (last date accessed: 26 November 2008) Chow, V. T., 1959. Open-channel hydraulics McGraw-Hill, New York. Correia, F. N., F. C. Rego, M. D. G. Saraiva, and I. Ramos, 1998. Coupling GIS with hydrologic and hydraulic flood modeling, Water Resources Management 12(3):229-249.
104 Di Baldassarre, G., G. Schumann, and P. D. Bate s, 2009. Near real time satellite imagery to support and verify timely flood modelling, Hydrological Processes 23(5):799-803. Diniz, M. G. M., N. d. O. Nascimento, and M. B. Baptista, 2003. Risk assessment of flood control failure in a storm water drainage system incorporating multiple detention basins, Proceeding of the Discharged urban waters : Ressource or Risk ? 2nd WWW-YES Vitry sur Seine, France. Dose, T., G. Morgenschweis, and T. Sc hlurmann, 2002. Extrapolating stage-discharge relationships by numerical modeling. (last date accessed: 7 May 2009) Dwivedi, R. S., and S. Kandrika, 2005. Delineati on and monitoring of aquaculture areas using multi-temporal spaceborne multispectral data, Current Science, 89(8):1414-1421. Dwivedi, R. S., K. V. Ramana, S. S. Thamma ppa, and A. N. Singh, 2001. The utility of IRS-1C LISS-III and Pan-merged data for mapping salt-affected soils, Photogrammetric Engineering & Remote Sensing 67(10):1167-1175. Dwivedi, R. S., B. R. M. Rao, and S. Bha ttacharya, 1999. Mapping wetlands of the Sundaban Delta and its environs using ERS-1 SAR data, International Journal of Remote Sensing 20(11):2235-2247. El-Kadi, A. I., and E. Yamashita, 2007. Modelin g stream flows and flood delineation of the 2004 flood disaster, Manoa, O'ahu, Hawai'i, Pacific Science 61(2):235(21). Fox, N. I., and C. G. Collier, 2000. Estimating medium-range catchment flood potential, Journal of Hydrology 237(1-2):1-16. Frazier, P. S., and K. J. Page, 2000. Water body detection and delineation w ith Landsat TM data, Photogrammetric Engineering & Remote Sensing 66(12):1461-1467. Freeman, A., B. Chapman, and P. Siqueira, 2002. The JERS-1 Amazon Multi-season Mapping Study (JAMMS): Science objectives and implications for future missions, International Journal of Remote Sensing 23(7):1447-1460. Gluch, R., 2002. Urban growth detection using te xture analysis on merged Landsat TM and SPOT-P data, Photogrammetric Engineer ing & Remote Sensing 68(12):1283-1288. Goshtasby, A. A., and S. Nikolov, 2007. Image fusi on: Advances in the state of the art, Information Fusion, 8(2):114-118. Green, A. A., G. Whitehouse, and D. Outhet, 1983. Causes of flood streamlines observed on Landsat images and their use as indicators of floodways, International Journal of Remote Sensing 4(1):5-16. Haack, B., and M. Bechdol, 2000. Integrating multisensor data and radar texture measures for land cover mapping, Computers & Geosciences 26(4):411-421.
106 Kiage, L. M., N. D. Walker, S. Balasubramania n, A. Babin, and J. Barras, 2005. Applications of Radarsat-1 synthetic aperture radar imagery to assess hurricane-related flooding of coastal Louisiana, International Journal of Remote Sensing 26(24):5359-5380. Kidson, R., K. S. Richards, and P. A. Carling, 2005. Reconstructing th e ca. 100-year flood in Northern Thailand, Geomorphology 70(3-4):279-295. Knebl, M. R., Z. L. Yang, K. Hutchison, and D. R. Maidment, 2005. Regional scale flood modeling using NEXRAD rainfall, GIS, and HEC-HMS/RAS: A case study for the San Antonio River Basin summer 2002 storm event, Journal of Environmental Management 75(4):325-336. Koutsias, N., M. Karteris, and E. Chuvic o, 2000. The use of intensity-hue-saturation transformation of Landsat-5 thematic mapper data for burned land mapping, Photogrammetric Engineering & Remote Sensing 66(7):829-839. Lang, M. W., P. A. Townsend, and E. S. Ka sischke, 2008. Influence of incidence angle on detecting flooded forests using CHH synthetic aperture radar data, Remote Sensing of Environment 112(10): 3898-3907. Lillesand, T. M., R. W. Kief er, and J. W. Chipman, 2007. Remote Sensing and Image Interpretation 6th Ed, John Wiley & Sons, New York. Liu, Z., F. Huang, L. Li, and E. Wan, 2002. Dyna mic monitoring and damage evaluation of flood in north-west Jilin with remote sensing, International Journal of Remote Sensing 23(18):3669-3679. Lu, D., M. Batistella, and E. Moran, 2008. Inte gration of Landsat TM and SPOT HRG images for vegetation change detec tion in the Brazilian Amazon, Photogrammetric Engineering & Remote Sensing 74(4):421-430. Mark, O., S. Weesakul, C. Apirumanekul, S. B. Aroonnet, and S. Djordjevic, 2004. Potential and limitations of 1D modelling of urban flooding, Journal of Hydrology 299(3-4):284-299. Mosquera-Machado, S., and S. Ahmad, 2007. Flood hazard assessment of Atrato River in Colombia, Water Resources Management, 21(3):591-609. Mutlu, M., S. C. Popescu, C. Stripling, and T. Spencer, 2008. Mapping surface fuel models using lidar and multispectral data fusion for fire behavior, Remote Sensing of Environment 112(1):274-285. National Economic and Social Advisory Council 2007. Royal irrigation department prepares 1,000 irrigation projects. URL: http://www2.nesac.go.th/document/show12.php?did=07070024 (last date accessed: 17 September 2007)
107 National Emergency System of Ur uguay, 2007. Flooding in Uruguay. URL: http://www.disasterscharter.org/disasters /CALLID_159_e.html (last date accessed: 29 September 2007) National Statistical Office, 2000. 1960-2000 Populati on and Housing Census, National Statistical Office, Bangkok. National Statistical Office, 2003. 2003 Agricultur al Census: Northeastern Region, National Statistical Office, Bangkok. Netzband, M., G. Meinel, and R. Lippold, 1999. Cla ssification of settlement structures using morphological and spectral features in fuse d high resolution satellite images (IRS-1C). URL: http://www.data-fusion.org/ps/sig/m eeting/Spain99ps/netzband.pdf (last date accessed: 1 October 2007) Neuenschwander, A. L., M. M. Crawford, and S. Ringrose, 2005. Results from the EO-1 experiment A comparative st udy of Earth Observing-1 Advanced Land Imager (ALI) and Landsat ETM+ data for land cover ma pping in the Okavango Delta, Botswana, International Journal of Remote Sensing 26(19):4321-4337. Nichol, J., and M. S. Wong, 2005. Satellite remo te sensing for detaile d landslide inventories using change detection and image fusion, International Journal of Remote Sensing 26(9):1913-1926. Paz, A. R., and W. Collischonn, 2007. River reach length and slope estimates for large-scale hydrological models based on a relatively highresolution digital elevation model, Journal of Hydrology 343(3-4):127-139. Petchprayoon, P., 2001. The prediction of flash flood caused by dam failure: a case study at the Tha Dan dam, Thailand, Proceeding of the 22nd Asian Conference on Remote Sensing Singapore. Pohl, C., and J. L. Van Genderen, 1998. Review article Multisensor image fusion in remote sensing: Concepts, methods and applications, International Journal of Remote Sensing 19(5):823-854. Profeti, G., and H. Macintosh, 1997. Flood management through Landsat TM and ERS SAR data: a case study, Hydrological Processes 11(10):1397-1408. Rosenqvist, A., and C. M. Birkett, 2002. Evalua tion of JERS-1 SAR mosaics for hydrological applications in the Congo River Basin, International Journal of Remote Sensing 23(7):1283-1302. Rosenqvist, A., B. R. Forsberg, T. Pimentel, Y. A. Rauste, and J. E. Richey, 2002. The use of spaceborne radar data to mode l inundation patterns and trace gas emissions in the central Amazon floodplain, International Journal of Remote Sensing 23(7):1303-1328.
108 Ryu, J. H., J. S. Won, and K. D. Min, 2002. Waterlin e extraction from Landsat TM data in a tidal flat: A case study in Gomso Bay, Korea, Remote Sensing of Environment 83(3):442-456. Saindranath, J., P. V. N. Rao, and S. Thiruve ngadachari, 2000. Radarsat data analysis for monitoring and evaluation of irri gation projects in the monsoon, International Journal of Remote Sensing 21(17):3219-3226. Sakamoto, T., N. Van Nguyen, A. Kotera, H. Ohno, N. Ishitsuka, and M. Yokozawa, 2007. Detecting temporal changes in the extent of annual floodi ng within the Cambodia and the Vietnamese Mekong Delta from MODIS time-series imagery, Remote Sensing of Environment 109(3):295-313. Sansena, T., and K. Bhaktikul, 2007. The app lication of hydraulic model with geographic information system to create flood risk mapping in Mae Klong river. http://kromchol.rid.go.th/ffd/p apers/Paper%20-Session%202/P206%20PaperThudchai_edit.pdf (last da te accessed: 27 February 2009) Santos, C., and J. P. Messina, 2008. Multi-sensor data fusion for modeling African palm in the Ecuadorian Amazon, Photogrammetric Engineering & Remote Sensing 74(6):711-723. Saraf, A. K., 1999. IRS-1C-LISS-III and PAN data fusion: An approach to improve remote sensing based mapping techniques, International Journal of Remote Sensing 20(10):19291934. Saroglu, E., F. Bektas, N. Musa oglu, and C. Goksel, 2004. Fusion of multisensor remote sensing data: assessing the quality of resulting images, Proceeding of the Geo-Imagery Bridging Continents Istanbul, Turkey. Schetselaar, E. M., 2001. On preserving spectral ba lance in image fusion and its advantages for geological image interpretation, Photogrammetric Engineering & Remote Sensing 67(8):925-934. Schultz, G. A., 1988. Remote sensing in hydrology, Journal of Hydrology 100(1-3):239-265. Schumann, G., P. Matgen, L. Hoffmann, R. Hostache, F. Pappenberger, and L. Pfister, 2007. Deriving distributed roughness values from satellite radar data for flood inundation modelling, Journal of Hydrology 344(1-2):96-111. Sharma, K. D., M. Menenti, J. Huygen, and P. C. Fernandez, 1996. Distributed numerical rainfall-runoff modelling in an arid region using thematic mapper data and a geographical information system, Hydrological Processes 10(9):1229-1242. Sharma, P. K., R. Chopra, V. K. Verma, and A. Thomas, 1996. Technical note flood management using remote sensing tec hnology: The Punjab (India) experience, International Journal of Remote Sensing 17(17):3511-3521. Siegel, H., and M. Gerth, 2000. Satellite-base d studies of the 1997 Oder flood event in the Southern Baltic Sea, Remote Sensing of Environment 73(2):207-217.
109 Singhroy, V., 1995. SAR integrated techniques for geohazard assessment, Advances in Space Research 15(11):67-78. Singhroy, V., K. E. Mattar, and A. L. Gray, 1998. Landslide characterisation in Canada using interferometric SAR and combined SAR and TM images, Advances in Space Research 21(3):465-476. Song, Y. S., H. G. Sohn, and C. H. Park, 2007. Effi cient water area classifi cation using Radarsat1 SAR Imagery in a high reli ef mountainous environment, Photogrammetric Engineering & Remote Sensing 73(3):285-296. Souza Filho, P. W. M., W. R. Paradella, and O. F. M. d. Silveira, 2005. Synthetic aperture radar for recognition of coastal features in the wet tropics: applications in the Brazilian Amazon coast, Bol. Mus. Para. Emlio Goeldi, sr. Cincias Naturais 1(1):201-207. Spruce, J. P., and R. McKellip, 2006. Assessing Hurricane Katrina damage to the Mississippi Gulf Coast using Ikonos imagery. http://calval.cr.usgs.gov/JACIE_files/JACIE06/Files/212Spruc.pdf (last date accessed: 5 May 2009) Srivastava, Y. K., B. Doley, D. K. Pal, R. K. Das, S. Sudhakar, S. Adiga, K. V. Venkatachary, and S. K. Srivastava, 2000. High resolution remote sensing data & GIS techniques in updation of infrastructure details for flood damage assessment a case study, Proceeding of the 21st Asian Conference on Remote Sensing Taipei, Taiwan. Staples, G. C., J. Hurely, and R. S. Burnsa nd, 1998. Coastal zone mon itoring with Radarsat-1, Proceeding of the 19th Asian Conference on Remote Sensing Manila, Philippines. Taragsa, N., C. Mongkolsawat, C. Pawattana, and R. Suwanweerakamtorn, 2004. An application of GIS and hydraulic model to flood plain inundation, Journal of Remote Sensing and GIS Association of Thailand, 5(3):63-75. Townsend, P. A., 2001. Mapping seasonal flooding in forested wetlands using multi-temporal Radarsat SAR, Photogrammetric Engineering & Remote Sensing 67(7):857-864. Townsend, P. A., and J. R. Foster, 2002. A synt hetic aperture radar-b ased model to assess historical changes in lowland floodplain hydroperiod, Water Resour. Res., 38. Townsend, P. A., and S. J. Walsh, 1998. Modeling floodplain inundation using an integrated GIS with radar and optical remote sensing, Geomorphology 21(3-4):295-312. Toyra, J., and A. Pietroniro, 2005. Towards opera tional monitoring of a northern wetland using geomatics-based techniques, Remote Sensing of Environment 97(2):174-191. Toyra, J., A. Pietroniro, and L. W. Martz, 2001. Multisensor hydrol ogic assessment of a freshwater wetland, Remote Sensing of Environment 75(2):162-173.
110 U.S. Army Corps of Engineers, 2005. HEC-GeoRAS GIS tools for support of HEC-RAS using ArcGIS U.S. Army Corps of Engineers, Davis, CA. U.S. Army Corps of Engin eers, 2007. HEC history. URL: http://www.hec.usace.army.mil/whoweare/histor y.html (last date accessed: 5 October 2007) U.S. Geological Survey, 2008. Surface Water and Water Quality Models Information Clearinghouse: Table of Model Char acteristics. http://smig.usgs.gov/cgibin/SMIC/gen_table?dimensions=all&domains=all&constituents=a ll&models=all&list_ge n=SMIChtml (last date a ccessed: 26 February 2009) Van de Griend, A. A., and E. T. Engman, 1985. Partial area hydrology and remote sensing, Journal of Hydrology 81(3-4):211-251. Van der Sanden, J. J., K. S. Schmidt, W. Bakke r, and W. Bijker, 2001. New evironmental remote sensing, Environmental Modelling with GIS and Remote Sensing (A. Skidmore, editor), Taylor & francis. Wang, Y., 2004. Using Landsat 7 TM data acquired days after a flood ev ent to delineate the maximum flood extent on a coastal floodplain, International Journal of Remote Sensing 25(5):959-974. Wang, Y., J. D. Colby, and K. A. Mulcahy, 2002. An efficient method for mapping flood extent in a coastal floodplain using Landsat TM and DEM data, International Jour nal of Remote Sensing 23(18):3681-3696. Weesakul, S., 1995. Real time flood forecasti ng modeling in the south of Thailand. http://www.fao.org/agris/search/d isplay.do?f=./2002/v2802/TH2000003287.xml;TH20000 03287 (last date accesse d: 27 February 2009) Wignyosukarto, B. S., 2006. The hydraulic performa nce of tidal irrigation system in the reclamation of acid sulphate soil. http://b ws.staff.ugm.ac.id/wp-content/bs-wignyosukartoicec2006_b28_.pdf (last date accessed: 7 May 2009) Yang, C., C. Zhou, and Q. Wan, 1999. Deciding the flood extent with Radarsat SAR data and image fusion, Proceeding of the 20th Asian Conference on Remote Sensing Hong Kong, China. Yang, J., R. D. Townsend, and B. Daneshfa r, 2006. Applying the HEC-RAS model and GIS techniques in river netw ork floodplain delineation, Canadian Journal of Civil Engineering, 33(1):19(10). Yongwei, S., and X. Zong-Guo, 1996. A comprehensiv e evaluation of filters for radar speckle suppression, Proceeding of the Geoscience and Remote Sensing Symposium, 1996. IGARSS '96. 'Remote Sensing for a Sustainable Future.', International Lincoln, NE, USA.
111 Zhang, X. M., C. J. S. Cassells, and J. L. va n Genderen, 1999. Multi-sensor data fusion for the detection of underground coal fires, Geologie en Mijnbouw 77(2):117-128. Zhou, G., J. Luo, C. Yang, B. Li, and S. Wang, 2000. Flood monitoring using multi-temporal AVHRR and Radarsat imagery, Photogrammetric Engineering & Remote Sensing 66(5):633-638. Zhu, L., and R. Tateishi, 2006. Fusion of multisenso r multitemporal satellit e data for land cover mapping, International Journal of Remote Sensing 27(5):903-918.
112 BIOGRAPHICAL SKETCH Siripon Kamontum was born and raised in Na khon Ratchasima, a northeastern province in Thailand. She graduated high school from Sura Nari Wittaya and later gained bachelors degree in Applied Mathematics from Prince of Songkl a University, Pattani Campus. In 2000, she finished masters degree in Remote Sensing and GIS from Khon Kaen University and worked as a lecturer at Khon Kaen University, Nongkai Campus for a year. Later, she worked as a Geo-informatics Scientist at GISTDA, Thaila nd. The Royal Thai Government acting through GISTDA sponsored her to pursue a doctoral degree in a field relating to GIS. She received the doctoral degree in Forest Resources and Conser vation concentrated in Geomatics from the University of Florida in the summer of 2009. Upon graduation, she retu rned to Thailand and works for GISTDA and the gator nation is everywhere.