SAR Image Classification Through Information-Theoretic Textural Features, MRF Segmentation, and Object-Oriented Learning Vector Quantization

Segmentation of optical images may be obtained through algorithms based on image prior models that exploit the spatial dependencies of land covers. In synthetic aperture radar (SAR) images, speckle conceals such spatial dependencies and segmentation algorithms suitable for optical images may become ineffective. Textural features may be used to emphasize spatial dependencies in the data and hence to improve segmentation. Once segmentation has been accomplished, a number of shapes is available. In this paper, the problem is tackled through the joint use of information-theoretic (IT) SAR features, of a segmentation algorithm based on tree structured Markov random fields (TS-MRFs), and of object-oriented classification achieved through learning vector quantization (LVQ). The proposed system works with one or more coregistered images, not necessarily all SAR, and one or more spatial maps of pixel features derived from each input image. A unique partition into connected regions, or segments, is achieved from the plurality of input channels, either images or feature maps. From each segment, representing a shape, geometric, radiometric, and textural parameters are extracted and fed to an LVQ classifier, trained through a partial reference ground truth (GT) of the scene. Classification results on a textured SAR image of a city and its surroundings validate the proposed object-oriented approach. Good performances can be achieved with small sizes of training sets, but they can be improved by using a decision fusion through majority voting (MV) of the outcomes of several experiments.