A polarimetric SAR data classification method using neural networks

Recently, neural network approaches have been adopted into polarimetric SAR data classification methods. A feature vector considering scattering effects, powers, and relative phases between polarimetries is presently being devised to discriminate more detailed categories. The authors propose a neural network classifier using polarization signatures and the above-mentioned feature vector. The polarization signature is composed of like- and cross-PSDs (Polarization Signature Diagram) which fully represent polarimetric features from scatters. The proposed method employs a maximum of /spl sigma//sup 0/ in the like-PSD and a minimum of /spl sigma//sup 0/ in the cross-PSD as input data for the neural network. The LVQ neural network is adopted as a classifier. Multi-frequency polarimetric SAR data observed by the quad-polarization mode of SIR-C were employed for the experiments. The proposed and conventional approaches are compared with average accuracies computed by classifying test data. As a result, the authors show that the proposed method is more useful and effective in producing classification accuracies.