Tensorial Independent Component Analysis-Based Feature Extraction for Polarimetric SAR Data Classification
Mingliang Tao,Feng Zhou,Yan Liu,Zijing Zhang
DOI: https://doi.org/10.1109/TGRS.2014.2360943
2015-01-01
Abstract:For polarimetric synthetic aperture radar (PolSAR) data, various polarimetric signatures can be obtained by target decomposition techniques, which are of great help for characterizing the land cover. It is straightforward to combine these polarimetric features together and formulate them as a third-order polarimetric feature tensor. However, how to make full use of the abundant information provided by these polarimetric features remains a challenge. A feasible solution is applying feature extraction (FE) techniques on the high-dimensional polarimetric manifold to obtain a lower dimensional intrinsic feature set. Common FE methods, such as principal component analysis (PCA), independent component analysis (ICA), etc., use matrix linear algebra and require rearranging the original tensor into a matrix. This leads to the loss of the spatial information of the PolSAR data. In this paper, to jointly take advantage of the spatial and feature information, a novel FE scheme incorporating ICA with the tensor decomposition techniques is proposed. After applying the proposed FE method on the third-order polarimetric feature tensor, each PolSAR image pixel is represented by a low-dimensional intrinsic feature vector. Furthermore, these feature vectors are fed to the k-nearest neighbor (KNN) classifier and support-vector-machine classifier for supervised classification. Simulated data, together with two measured data sets, i.e., Flevoland of Airborne Synthetic Aperture Radar (AIRSAR) and Québec City of Uninhabited Aerial Vehicle Synthetic Aperture Radar (UAVSAR), are utilized to evaluate the performance of the proposed method. For comparison purpose, several classical and advanced FE methods, such as PCA, ICA, Laplacian eigenmaps, and LRTAdr - (K1,K2,p), are also applied. The experimental results demonstrate the superiority of the proposed FE method in three folds: 1) The extracted features by the proposed method are more discriminative- characterized by the high separability in the scatterplots; 2) the classification accuracy is improved as much as approximately 7% compared with the complex Wishart classifier; and 3) the proposed method is computational efficient and has fast convergence.