Hierarchical classification of SAR data using a Markov random field model

A general framework is presented for classifying coastal environments using synthetic aperture radar (SAR) data. This framework addresses two main issues associated with the accurate classification of SAR data: 1) the variability in radar backscatter of a given pixel caused by the presence of speckle in the imagery and 2) the characteristic decrease in intensity as a function of incidence angle. To combat the effect of speckle on a given pixel's backscatter, a Markov Random Field (MRF) model is used to incorporate contextual information from the imagery by considering neighbor pixel statistics in the classification process. To address the class-specific decrease in backscatter as a function of angle, a hierarchical two-level classifier is considered to compensate for the highly variable water class and the less influenced land classes. Preliminary results are shown from the hierarchical MRF-based classifier and are compared to single level MRF and radial basis function (RBF) classifiers. For the test site presented, classification accuracy only improves slightly in using the hierarchical architecture, but does show the potential for application to coastal areas with larger percentages of upland and urban land cover types.