The role of frequency and polarization in terrain classification using SAR data

摘要

The expected accuracies of land-cover classification are evaluated for existing and potential orbital SAR systems. Land-cover classifications are compared for ERS-1, JERS-1, SIR-C and X-SAR. In addition, SIR-C/X-SAR data from a largely forested test site in northern Michigan are used to simulate the expected performance of potential orbital SAR systems such as Envisat, PALSAR and LightSAR. The classification approach uses orthorectified and filtered SIR-C/X-SAR data overlain with known polygons subdivided into spatially independent training and testing populations. For each potential sensor configuration, the relevant feature vectors are subsampled for a portion of the image and used to generate unsupervised clusters. These clusters are then assigned to the known classes of the training population using maximum likelihood criteria with equal probabilities. Contingency tables are produced for the testing population using minimum distance criteria. The classification results show that longer wavelengths (such as L-band) are of greatest value for discriminating general land-cover classes on the basis of biomass and roughness since there is a greater dynamic range relative to these attributes. Shorter wavelengths (C-band or X-band) are more sensitive to smaller scattering elements such as foliage and small stems and are therefore of importance in discriminations related to crown-layer architecture (i.e., leaf size and shape). The best results are achieved when classification is based upon multiple frequency data.