Unsupervised Classification of PolSAR Images using NMF-FLD Features

The purpose of this research is to provide a new and efficient way to identify targets in polarimetric synthetic aperture radar (PolSAR) images using the inherent properties of these images and to reduce the time and cost of calculation. In this paper, a new method is proposed to feature extraction and classification based on non-negative matrix factorization (NMF), Fisher linear discriminant analysis (FLD), and support vector machines (SVM) for polarimetric SAR images. At the first phase, non-negative features of polarimetric SAR images, including the local spatial structure of targets, are extracted by the NMF algorithm. Next, the FLD method is applied to the extracted features, thus separating the features and classifying them by the SVM method. At the second phase, polarimetric features are extracted by Fisher criteria. At the final phase, each pixel in the classified images with extracted polarimetric features is assigned to a class with the shortest distance, and the final classification is done. The proposed algorithm is applied to C and L-band AIRSAR and RADARSAT-2 polarimetric SAR data. Comparing the experimental results with the Wishart demonstrated the effectiveness and classification accuracy of the proposed method to identify targets in polarimetric SAR images. The methods used in the proposed algorithm are linear and thus they reduce the time and cost of calculation.