Unsupervised Classification of PolSAR Imagery via Kernel Sparse Subspace Clustering

Unsupervised classification is very important for the fully polarimetric synthetic aperture radar (PolSAR) image interpretation. The PolSAR covariance matrices, as one of the most widely used representations for PolSAR data, are Hermitian positive definite (HPD) and form a Riemannian manifold when endowed with an appropriate metric. Considering their geometric properties, we propose a new clustering algorithm by embedding the HPD matrices into Hilbert space and introduce sparse subspace clustering in the newly formed highly dimensional space to recover the latent cluster structure. Moreover, an improved scalable scheme is presented to classify large-scale PolSAR images, which involves dictionary learning and spatially reinforced joint coding for robustness against the speckle noise. Experimental results on real fully PolSAR data sets demonstrate the effectiveness of the proposed method.