A neural segmentation of multispectral satellite images

A neural model for segmentation of multispectral satellite images is presented, implying the following processing stages: (a) feature extraction using the 2-dimensional discrete cosine transform (2-d DCT) applied on the image segment centered in the current pixel, for each frame of the spectral sequence; (b) a neural self-organizing map having as input the concatenation of the feature vectors assigned to the projections of the current pixel in all the image bands (computed in stage (a)). The software implementation of the model for multispectral satellite images SPOT leads to interesting experimental results.