SEGMENT-BASED LAND-USE CLASSIFICATION FROM SPOT SATELLITE DATA

Spectrally heterogeneous land-use categories cannot be adequately clas sified from high resolution satellite data, using conventional multisp ectral classification techniques. In this study, built-up land is extr acted based on the spectral and the spatial properties of the segments in a spectrally classified satellite image. Object structures and cla ssification rules have been implemented in an expert system shell - Ne xpert Object. The approach has been tested using SPOT multispectral da ta covering an area south of Stockholm, Sweden, The classification acc uracy for built-up land improved significantly after the application o f a few relatively simple rules, based on segment size and on relation s between segments. The object oriented implementation in Nexpert Obje ct is flexible but slow, and performance problems were encountered due to the large number of objects being processed.