Classification of Earth Observation Data and Flood Encroachment Analysis *R. Hrishikesh Mahadev and +Hugues Sassier *Research Associate, The Energy and Resources Institute, India Habitat Centre Lodhi Road, New Delhi - 11 00 03, India +System Engineer, "Image Quality & Processing", Observation & Sciences System Architecture, Alcatel Space 26 avenue JF.Champollion, BP1187 - 31037, Toulouse cedex France ABSTRACT Flood management deals with understanding the movement of floods and necessary steps to minimize adverse impacts. The study aims to find out best possible solution with good accuracy to prepare land use map with minimum available data and time for extreme events and to calculate depth of water in flooded areas. The study area is Herault watershed, which is 2500 km² located in South of France. Two SPOT (August and November) and one Landsat (August) images have been used as multi-temporal and single date classification. The study is divided into two parts: first part deals with finding out best land use classification algorithm out of three algorithms, parallelepiped, Minimum distance to mean (MDM) and Maximum likelihood Classification (MLC) for two set of images in ERDAS and ENVI, which comes to be classification of 12 images. Accuracy assessment results indicated that classification of multi-temporal SPOT image (79.20 % for MLC) is better than single date Landsat image (69.53 % for MLC). The second part deals with finding out depth of water in inundated areas due to floods in different land use. Flow data is fed in HEC RAS (Hydrologic Engineering Centre River Analysis System) and Steady flow analysis is performed which uses river flow data in HEC RAS and geometrical data in ARC VIEW for generating water level. The water surface profile is created in ARC VIEW for a return flood frequency data of 50 years, which is then compared with flood risk prevention map created by DIREN (Direction Regionale De L'environnement). Affected area was 37.94 Sq Km from calculated data where as DIREN map shows an inundated area of 23.58 Sq Km. The difference in flooded areas of river is due to fact that many tributaries were not considered, as data was not available. The resultant flood inundated map was then used to estimate land use and administrative blocks affected due to flood, which is helpful in allocation of resources during floods. DIREN map shows aerial distribution of flooded areas but the calculated map gives depth of water at each pixel apart from aerial distribution of flooded areas. Hence, such studies are very important in planning, prioritisation and taking suitable measures to minimise the impacts in watersheds prone to floods across world. INTRODUCTION Several natural disasters occur throughout the world round the year and cause huge loss to the lives, property and natural resources. Assessment of damage caused by these disasters needs to be carried out accurately in minimum possible time frame. Space technology is one of the best-suited means for the assessment of damage brought about by natural disasters. Remote sensing satellites provide synoptic view, repetitive coverage and high-resolution images. This advanced high-resolution sensor technology has provided immense scope to the earth resource scientists world-wide for mapping and analysis of earth surface feature details using Remote Sensing and Geographic Information System (RS and GIS). In processing of multi-temporal remote sensing images, accurate and convenient classification is among one of difficult tasks in practical applications (Baber, 1985). Maximum likelihood classification has traditionally been used as a baseline for the classification of remotely sensed data. For instance, Apan (1997) uses maximum likelihood classification to assess the utility of Landsat data for mapping forest rehabilitation and Basham May et al. (1997) use maximum likelihood classification to compare the effectiveness of Landsat and SPOT data for vegetation classification. Dwivedi et al (2004), carried out study to evaluate the potential of the Gaussian maximum likelihood classifier, Mahalanobis minimum-distance classifier, minimum-distance classifier, and artificial neural network (ANN) classifier in deriving information on land-use/land-cover over part of Ethiopia using various band combinations of Landsat TM data. Yang et al (2001) made an attempt to develop an integrated methodology for flood prediction using geographic information systems (GIS) and hydrodynamic modelling and to obtain flood information for flood emergency planning. Brivio et al (2002) carried out a unique study, which describes a synergetic use of satellite radar images, and ancillary information to detect flooded areas at their peak and evaluates its potential with mapping. Blasco et al 1992, studied extent of flood coverage in Bangladesh using SPOT data. Townsend and Marsh (1998) carried out study using Synthetic aperture radar images from multi-temporal L-band JERS-1 and C-band ERS-1 satellites, a Landsat Thematic Mapper (TM) time-series, and GIS coverages were used in an integrative approach to model the potential of flood inundation within the lower Roanoke River floodplain, North Carolina. Statistical results indicate that the GIS-derived models successfully identified flooded areas as mapped by the radar change-detections. Further, statistical tests assessed the ability of individual radar and optical (Landsat TM) images to discriminate flooding as predicted by the GIS models. Both JERS-1 and ERS-1 images identified areas of inundation at different flood levels. The present study aims at immediate availability of land use / land cover maps during disaster events where such maps are not commonly available in many places across the world. Such maps are required to estimate total disaster in the region within limited time and report to concern authorities for prevention and relief measures. Normally, such studies are not accompanied with ground truths. The present study aims to find out best possible solution with good accuracy to prepare land use map with minimum available data under extreme events. The present study is a preliminary study; hence no real time data were used in analysis. The area under interest is Herault Watershed, South-East of France. The following were the objectives of the project.
STUDY AREA The Herault River Basin is a Mediterranean watershed of 2500 km² located in the Languedoc Region, south-east of France. The Herault river stretch from the Cévennes Mountains in north to the Mediterranean is 150 km long and has 9 main tributaries. Most part of the mountains is covered by forest and in plains agriculture dominates. In this region, vineyards cover an area of 128,000 hectares (60% of agricultural land) and involve more than 93% of farms. Mountain areas in this region have higher rainfall than the plains of Herault watershed and temperatures depend on altitude. Rainfall is medium at around 700 mm (28 in) per year on the plain, more in the mountains. In September 2002, 680 mm (27 in) of rainfall was recorded in 24 hours near Anduze causing severe flooding further down the river valleys. This rainfall is equal to the annual rainfall in London. Floods 2002 was one of worst hit floods in the region. It was reported that almost 40% of the vines in Gard and 10% in Vaucluse, a total of 40,000 hectares were damaged in the floods, according to Guy Carpenter (2003). DATA AND METHODOLOGY The remote sensing data used for classification is SPOT and Landsat data. Table 1 gives details of satellite data used in the present study. 
Classification was carried out using 2 SPOT data and 1 Landsat data separately, and then using flood frequency data, encroachment analysis was carried out. To meet the objectives of the study, following software were used:
Floods encroachment analyses were carried out using HEC RAS and ARC VIEW extension HEC Geo RAS. HEC-RAS stands for Hydrologic Engineering Centre River Analysis System, developed by US army Corps of Engineer. This software has several aspects of hydrologic engineering including rainfall runoff, river hydraulics, flood damage analysis and Real Time River forecasting for reservoir operations. CLASSIFICATION To perform a supervised classification it was necessary to identify training areas representing the spectral characteristics of the defined categories. The division of land cover classes was based on the BRGM (Bureau de recherches géologiques et minières) land cover nomenclature, but with less number of classes. As a clear identification of the categories by means of the SPOT and Landsat images, most of the training data assessment took place by taking samples from already classified data, considering the following principles.
In this way 243 test sites were assessed and assigned to 8 broad land cover categories. There were around 40 classes were generated, which represented 8 major classes. Then these classes were merged in the output file after classification. The remaining classes were identified by means of the Landsat and SPOT images. The class assignments were gathered from BRGM classified map and additional features like development type; crop species were not been able to register, as ground truth was not possible. In order to improve classification results a number of additional training areas were assigned in an iterative process, respecting the results of several classification test runs from Landsat image and multi-temporal SPOT images. The master set of signature files were finalized after running several test-classified images. This set of signature were used for Landsat in ERDAS and then exported in vector format to ENVI to classify same image in ENVI. Similarly, the results were produced for SPOT in ERDAS and ENVI. Regarding the result of the test runs and seperabilty matrix. This matrix represents the statistical distance between two signatures. This matrix produces results calculated for any combination of band that is used in the classification. It seemed advisable to unite or delete some of the classes, when the transformed divergence where less than 1700 (ERDAS 2002). Categories like e.g. winter crops (unplanted fields) and urban areas could not be divided properly due to their similar spectral characteristics. ACCURACY ASSESSMENT Accuracy assessment was carried out for all 12 classified maps: 3 algorithms, 2 software and 2 sets of images, derived from ERDAS and ENVI. The maps derived from ENVI was exported and then imported in ERDAS to have unbiased assessment of all maps. Stratified random sampling was carried out for a single classified image and the same points were used in all other classified maps. The advantage of this method is same points will be tested across software and images. But the drawback is that stratification of random sampling gets lost once satellite image or software changes. Applying the reference image classification, 250 random points, 10 per class, were generated. As stratified random sampling was carried out for MLC, the points of each class reduced to 3, which finally served as input for the accuracy assessment. FLOODS The encroachment study was carried out in ARC VIEW HEC Geo RAS extension and HEC RAS. HEC-Geo RAS facilitates the generation of GIS themes from exported HEC-RAS simulation results. DEM is one of the major inputs for floods analysis. HEC Geo RAS takes the elevation data in terms of TIN (Triangulated Irregular Network). Thus, DEM was converted into TIN in ARC View. The goal of performing a floodplain encroachment analysis is to determine the limits of encroachment that will cause a specified change in water surface. The main inputs for flood encroachment analysis is stream line centre and cross sectional data. The other inputs used in the study are Channel banks and flow path.
Analysis in HEC RAS and HEC Geo RAS After creating these themes, RAS GIS import file is generated. This file will contain all the themes and will be imported to HEC RAS. The pre-processing menu also builds topology and assigns attributes (stream length, cross section, distance between main channel and banks, etc.). Prior to performing steady flow hydraulic computations in HEC RAS, the geometric data is imported and completed and flow data was entered. Floodplain delineation and water depth themes are created from exported cross-sectional water surface elevation using HEC Geo-RAS. This file is again exported to ARC VIEW. Here, the first step is to create water surface TIN for each water surface profile. The water surface profile is created for a flood frequency of 50 years. The water surface TIN created is based on water the surface elevation at each cross section specified in the RAS GIS Export file. The floodplain may then be delineated for each water surface profile for which water surface profiles TIN that was created in previous step. The flood plain delineation procedure converts the water surface TIN and Terrain TIN to lattice with the same cell size and origin. A depth grid is then created wherever the water surface grid is higher than the terrain grid. The depth grid is converted into floodplain polygon. The resultant flood inundated map was then used to estimate land use and administrative blocks affected due to flood. Comparison between calculated and original flood inundated areas was also done in the study. RESULTS AND DISCUSSION Classification resulting from ERDAS and ENVI has almost similar results. But results are different for different algorithm used. The decision rules for classification used was Maximum Likelihood, Minimum distance to Mean and Parallelepiped. MLC gives a better accuracy than the others as it uses probability theory. It assumes that the input data (i.e. training data) are normally distributed and independent. These criteria are met; the MLC is well suited for accurate classifications. From the Table 2 and Table 3, it is clear that multi-temporal image gives a better accuracy than single date image. 
However, as a single measure of accuracy, the overall accuracy (or percentage classified correctly) gives no insight into how well the classifier is performing for each of the different classes (Fitzgerald and Lees, 1994). In particular, a classifier might perform well for a class, which accounts for a large proportion of the test data and this will bias the overall accuracy, despite low class accuracy for other classes. For example, overall accuracy of MDM classifier for SPOT in ERDAS is 73.89, but some of the classes have been poorly classified. There not much variation in kappa coefficients of ERDAS and ENVI for same algorithm, but there is significant variation across algorithm. The parallelepiped classifier shows the lowest kappa coefficient of the three algorithms. The reason is there are some outliers, which lies away from mean .of pixels being classified in a class. FLOODS From the Chart 1 and Chart 2, there exists significant difference in calculated and predicted values of floods. In few Blocks like Montagnac, Lezignan-La-Cebe, calculated flood data matches with Original data. But in some cases as in Saint-Andre-De-Sangonis, Gignac Calculated flood values were considerably higher than original values, which can mislead in allocation of resources during flood event. In the present study, flow data was available only for 6 stations. Due to which the streamlines were drawn only for few tributaries. The results of analysis can be improved by using more flow data and streamline data. The error due meandering of river can be minimised by taking more cross-sectional data across river. In case of Land use, calculated values were greater than original values. In both cases, total affected area was 37.94 Sq Km (including river: 39.63 Sq Km) from Calculated data where as original data shows and inundated area of 23.58 Sq Km (including river: 28.23 Sq Km). The difference in area of river is due to fact that many tributaries were not considered, as data was not available. The present study aims at finding best classification algorithm during flood events. The aim of study was to classify land use rapidly with better accuracy and to develop inundated maps so as to report to authorities for proper allocation of resources.
REFERENCES
 Chart 1: Percent of Administrative Blocks Affected due to Floods: Comparison between Original and Calculated Values  Chart 2: Land Use Affected due to Floods: Comparison between Original and Calculated Values  Figure 3: Flood inundated areas in Herault Watershed  Figure 4: Comparison between reference Flood map and Calculated flood map.
Page 1 of 1
|