Polarimetric SAR Classification: Fast Learning with k-Maximum Likelihood Estimator

Abstract

Classification of polarimetric synthetic aperture radar (SAR) images into different ground covers becomes challenging when the terrain under consideration is heterogeneous in nature. A finite mixture model-based Wishart mixture model (WMM) classifier can efficiently incorporate heterogeneity of terrain using multiple Wishart components. The expectation-maximization (EM) is the most widely used algorithm for learning parameters of such statistical models. However, the convergence of the EM algorithm is very slow. Moreover, the information contained in the polarimetric SAR images is in complex numbers and the size of the data is often very large. Therefore, they incur a large amount of computational overhead. The training of the classifier becomes very slow due to these reasons. In this paper, a k-maximum likelihood estimator (k-MLE) algorithm is employed for learning of the parameters in WMM classifier. As k-MLE is an iterative process, different algorithm initialization approaches such as random, global K-means, and k-MLE++ are analyzed. The algorithm is compared with the traditional EM algorithm in terms of classification accuracy and computational time. The experiments are performed on six different full polarimetric SAR datasets. The results show that the classification accuracy using k-MLE is comparable to the traditional EM algorithm while being more computationally efficient. The initialization using k-MLE++ results in better modeling compared to random and global K-means.