SOFT COMPUTATION OF SATELLITE DATA FOR LAND COVER INFORMATION

Abstract

Land cover information is very important for many technical and societal issues. The synoptic advantages of satellite data make it suitable as a source of land cover information. Conventional methods for extraction of land cover information from satellite data (based on assumption that each picture element (pixel) of satellite image contains a particular type of land cover class) is counterproductive in many cases as the prevalent field condition does not match with the spatial and geometric extension and orientation of the picture elements in an image. Many pixels of satellite images contain mixed land cover types. Over the years, many researchers evolved different methodology to analyze satellite images to extract mixed land cover classes from pixels. Methods based on Artificial Neural Network (ANN) are getting popularized towards extraction of mixed land cover information from pixel data. In this study, three ANN based algorithms — Feed Forward Back Propagation Neural Network (FFBPN), Self Organizing Map (SOM) and Learning Vector Quantization (LVQ) have been used to extract land cover information from satellite image. These methods extract land cover information from pure as well as from mixed pixels. In order to test the quality of land cover information extracted by these methods, LANDSAT TM data of Haridwar, Uttarakhand has been classified by the three ANN methods as well as by conventional maximum likelihood classifier (MXL). Six land cover classes have been considered in this quality assessment. The overall accuracies have been found to be 92,90%, 98.09%, 96.82% and 88.01 % respectively from FFBPN, SOM, LVQ and MXL. Thus, ANN methods provide better classification accuracy than conventional (MXL) classifier. However, the quality of land cover information extracted from ANN methods depend on topology of the ANN.