Algorithms for feature extraction from synthetic aperture radar data

Earth's surface consists of land features such as vegetation, soil, water, etc. Modeling of the earth's surface requires identification and understanding of the dynamics of land features. Analysis of land feature dynamics would reveal the changes that occur due to human induced activities or natural phenomenon. This plays a major role in providing up-to-date information of the natural resources. Data acquired remotely through space-borne sensors at regular intervals in visible and microwave bands aid in spatial mapping of the land features. Data acquired in visible and IR (Infrared) bands have been used for land use and land cover analysis. However, these data fails when there are cloud cover due to non-selective scattering. In this context, RADAR remote sensing would be useful as it provide information during all seasons due to long penetration properties. In present study, RADARSAT-2 single polarized HH (i.e., Horizontal to Horizontal with C-band) has been used to derive land features with spatial extent. Radar data interpretation and analysis is considered challenging and have both advantages and disadvantages in land use feature extraction. This study assess the performance of classification algorithms (Gaussian Maximum likelihood classifier (GMLC), Neural network classifier, Decision tree classifier (DTC), Contextual classification using sequential maximum a posteriori (SMAP) estimation for feature extraction using multi-temporal single polarized RADARSAT data, texture extracted data and fused data (optical sensor -LANDSAT ETM+ with SAR data). Accuracy assessments suggest that fused data perform better with all algorithms.