A region-based approach is tried out on a subimage of a Thematic Mapper frame in order to investigate its ability to eliminate some major drawbacks of standard multispectral classification. To overcome the difficulties arising from the rather low signal to noise ratio and the poor dynamic range of video-values within each spectral band a new multispectral edge-preserving smoothing algorithm has been developed. A gradient-based segmentation partitions the smoothed image into homogeneous regions by grouping neighboring pixels with similar spectral properties, the basic intrinsic characteristics of which are scene. A gradient-based algorithm with split-and-collected in a property table. This structural description of the image serves as a starting point for a subsequent knowledge-based interpretation in terms of land-use categories.