Superpixel-Based Unsupervised Classification of Polsar Images with Adaptive Number of Terrain Classes

This paper proposes a superpixel-based unsupervised classification framework for polarimetric synthetic aperture radar (PolSAR) images. First, the PolSAR image is over-segmented into a variety of superpixels, based on which mean Freeman decomposition and HSI color feature vectors are extracted and stacked directly into a high-dimensional feature vector. Second, based on a distance matrix constructing from the high-dimensional feature vector, visual assessment of tendency using diagonal tracking (VATdt) is adopted to adaptively estimate the number of terrain classes and automatically capture the cluster structure. Third, spectral clustering and a reduction technique as well as complex Wishart classifier are performed to obtain the final classification results. Experiments conducted on one simulated and one real-world PolSAR images demonstrate the superiority and effectiveness of the proposed method.