Comparison of land cover classification using high-resolution TerraSAR-X and optical imagery

Abstract:
TerraSAR-X is capable of acquiring imagery of one meter resolution. In data of such kind man-made objects become visible and typical land cover classification classes appear in high detail. Our aim is to find out if TerraSAR-X imagery may complement optical images for automatic land cover classification. Thus, we classify imagery of both data types into classes settlement, agriculture, streets, and forested areas and compare classification performances. We use Markov Random Fields (MRF) as learning based probabilistic framework to classify optical and SAR data. In case of the TerraSAR-X amplitude data we model the likelihood function with Fisher distributions, whereas texture measures are evaluated using Gibbs probability distributions for the optical images. First results show that high-resolution TerraSAR-X imagery may complement land cover classification.