A Semi-Supervised Dimension Reduction Method for Polarimetric SAR Image Classification

Aiming at the problem of feature redundancy in polarimetric synthetic aperture radar (SAR) application, a semi-supervised dimension reduction algorithm: semi-supervised local discriminant analysis (SliDA) is proposed by combining the thoughts of linear discriminant analysis (FDA) and locally linear embedding (LLE). Firstly, the regularization term is established based on local preserving property of LLE to avoid overfitting problem during learning. Then, discriminant analysis with regularization is performed on labeled data set in order to improve the generalization ability and preserve the local geometric structure in original space for the whole data. Dimension reduction experiments arc performed on all polarimetric SAR data from Flevoland regions obtained by RADARSAT-2 and AIRSAR satellites. The results show that the low dimensional features extracted by SLDA has the characteristics of "intra compactness and inter separation". Further classification experiment results show that SliDA can make the classification accuracy reach about 90% only with 1 parts per thousand-2 parts per thousand labeled samples, and the classification performance of SliDA is superior to other comparison algorithms.