A Novel Feature Extraction and Land Cover Classification Method Based on Tensor Algebra for Multi-band SAR Imagery

Multi-band synthetic aperture radar (SAR) can provide more reliable results than the single-band case through feature level fusion for land cover classification. The existing feature level fusion approaches of multi-band SAR simply stack the features of each pixel in different bands, which are limited in capturing inherent correlation among different bands, leading to suboptimal classification performance. In order to classify land cover more accurately, a novel classification method based on tensor algebra for multi-band SAR is proposed in this paper. First, each pixel is represented as a fourth-order tensor constructed by combining multi-features of different bands in the neighborhood window. Then, a novel tensor discriminant locality preserving projection (TDLPP) is proposed for feature extraction from tensors, integrating local manifold structure and global label information. Finally, the lower-dimensional features extracted by the TDLPP are classified using the classic classifier, such as K-Nearest Neighborhood (KNN) classifier. The experimental results on the real multi-band SAR dataset confirm the effectiveness of the proposed method.