Unsupervised Classification of PolSAR Images by Combining Feature Selection and Large Scale Spectral Clustering

Unsupervised classification is an important tool for PolSAR image interpretation,which is easily affected by the high dimensional features.Thus, we propose an unsupervised classification algorithm for PolSAR images by combining feature selection and large scale spectral clustering.First, the widely used features are extracted from PolSAR images including the features based on the simple linear transformation of raw data and polarimetric target decomposition. Secondly, in order to remove redundant information and facilitate later analysis, a cluster forest based feature selection algorithm is used to perform dimension reduction. Finally, the representative points are generated by an oversegmentation method, and the bipartite graph is constructed between raw data points and these representative points.The unsupervised classification result is obtained by using a large scale spectral clustering method which employs a representative point based scheme.The experiments show that the proposed method can select an effective feature set and obtain a satisfactory result for unsupervised classification of PolSAR images.