Unsupervised classification of PolSAR data using large scale spectral clustering

In this paper, a spectral clustering based unsupervised classification scheme is proposed for processing large scale polarimetric synthetic aperture radar (PolSAR) data. Due to its high computational complexity, spectral clustering can hardly handle large PolSAR image. To overcome this bottleneck, a representative points based scheme is introduced. Instead of building pairwise affinity graph on the whole data set, we first build a bipartite graph between data points and a small set of selected representative points. Then an approximate large graph is constructed based on this bipartite graph. After that, spectral analysis on the approximate graph is solved efficiently by singular value decomposition (SVD). To integral context information, Markov random fields (MRF) model based smoothing is also performed to get the final clusters. We test the proposed approach on DLR ESAR data set. Experimental results demonstrate its effectiveness and efficiency.