Segmentation and symbolic description for a classification of agricultural areas with multispectral scanner data

A procedure that overcomes some of the disadvantages of conventional pixel-based classification is proposed. Several methods for data preprocessing are tested, and the edge-preserving smoothing filter is found to give the best results. The segmentation module uses an easy and efficient region-based algorithm based on a comparison of the intensity value of a single pixel in the 4- or 8-adjacency neighborhood. This algorithm discriminates 97% of the regions in the test area and shows in some selected regions a mean area deviation of 7.8%. Within the symbolic description, 127 attributes of morphologic and internal descriptors are tested on four main criteria. It is shown that the selection of the descriptors has to be done in consideration of the data as well as the application