Classification of Polarimetric SAR Images

Abstract

Classification of terrain into different ground covers is one of the predominant applications of polarimetric SAR images. TheWishart classifier works well for major classification tasks. However, it is effective only for homogeneous regions. This research aims to improve the classification accuracy when the terrain under observation is heterogeneous. For this purpose, theWishart mixture model (WMM) is employed for modeling the heterogeneity of terrain. The model’s parameters are first estimated using the expectation-maximization (EM) algorithm along with different initialization approaches. Then, the advanced k-maximum likelihood estimator (k-MLE) is employed along with initialization using k-MLE++ and compared with the EM algorithm. The degrees of freedom is one of the crucial parameters of Wishart distribution. Therefore, its impact is analyzed over classification accuracy. The pixel-based classifiers’ performance is greatly affected by inherent speckle. For that, a conditional random field (CRF) based model is proposed for polarimetric SAR data to incorporate spatial-contextual information. It is combined with Wishart and WMM classifiers, namely Wishart-CRF and WMM-CRF, and compared with the traditional Markov random field (MRF) based model. The experiments are performed using six different full polarimetric SAR data sets. The results show that theWMMwith k-MLE parameter estimator and k-MLE++ as initialization methods perform better than the EM algorithm. Furthermore, combining the CRF model exhibits better classification results by significantly reducing the speckle and preserving the details of edges and micro-regions.