A new classifier for polarimetric SAR images

Abstract

This paper proposes a hybrid classifier for polarimetric SAR images. The feature sets consist of span image, the H/A/α decomposition, and the gray-level co-occurrence matrix (GLCM) based texture features. Then, the features are reduced by principle component analysis (PCA). A 3-layer neural network (NN) is constructed, trained by resilient back-propagation (RPROP) method to fasten the training and early stop (ES) method to prevent the overfitting. The results of San Francisco and Flevoland sites compared to Wishart Maximum Likelihood and wavelet-based method demonstrate the validness of our method in terms of confusion matrix and overall accuracy. In addition, NNs with and without PCA are compared. Results show the NN with PCA is more accurate and faster.