Data-Driven Polinsar Unsupervised Classification Based on Adaptive Model-Based Decomposition and Shannon Entropy Characterization

We introduce a data-driven unsupervised classification algorithm that uses polarimetric and interferometric synthetic aperture radar (PolInSAR) data. The proposed algorithm uses a classification method that preserves scattering characteristics. Our contribution is twofold. First, the method applies adaptive model-based decomposition (AMD) to represent the scattering mechanism, which overcomes the flaws introduced by Freeman decomposition. Second, a new class initialization scheme using a histogram clustering algorithm based on a Dirichlet process mixture model is applied to automatically determine the number of clusters and effectively initialize the classes. Therefore, our algorithm is data-driven. In the first step, the Shannon entropy characteristics of the PolInSAR data are extracted and used to calculate the local histogram features. After applying AMD, pixels are divided into three canonical scattering categories according to their dominant scattering mechanism. The histogram clustering algorithm is applied to each scattering category to obtain the number of classes and initialize them. The iterative Wishart classifier is applied to refine the classification results. Our method not only can obtain promising unsupervised classification results but also can automatically assign the number of classes. Experimental results for E-SAR L-band PolInSAR images from the German Aerospace Center demonstrate the effectiveness of the proposed algorithm.