Patch Tensor-Based Geometric Structure Representation for Hyperspectral Imagery Classification

Abstract

Hyperspectral remote sensing images (HSIs) have nanometer-level spectral resolution, and the rich spatial and spectral information they contain makes them possible to perform detailed land-cover analysis. In this letter, a patch tensor-based geometric structure representation (PTGSR) method was proposed for HSI classification. At first, based on the high-order structure of tensor samples, a novel representation learning (RL) model based on tensor neighborhood structure is constructed to capture the relationships between different tensor samples. Then, by utilizing the spatial coordinates of central pixels in the tensor samples, the distribution probabilities between different samples are computed to improve the effectiveness of the representation coefficients. Finally, the classification errors of each class across different tensor orders are integrated to enhance the overall classification performance. Experimental results on three HSI datasets demonstrate that the PTGSR algorithm outperforms several classification methods with overall accuracy (OA) improvements of 1.78%–31.43%, 1.55%–8.54%, and 2.14%–15.72% for Indian Pines, Fanglu, and Houston2013, respectively.