Enhancing earth target classification in hyperspectral imagery using graph convolutional neural networks and graph-regularized sparse coding

Abstract

As hyperspectral remote sensing technology continues to advance, classification approaches using hyperspectral images (HSIs) have become increasingly important in earth target identification, mineral mapping, and environmental management. The strength of HSIs lies in their capacity to provide a detailed understanding of a target's composition. However, challenges such as high dimensionality, redundancy in HSI datasets, and potential class imbalances complicate their effective utilization. In this study, a novel framework combining graph convolutional neural networks (GCNNs) and graph-regularized sparse coding (GSC), referred to as GCNN-GSC, is proposed to address these challenges in HSI-based earth target classification. HSIs often exhibit irregular spatial structures, making traditional grid-based methods less effective. GCNNs excel in handling irregular grids, making them well-suited for hyperspectral data where spatial pixel arrangements deviate from regular patterns. GSC complements GCNNs by mitigating high dimensionality through compact and informative feature representation. To evaluate the efficacy of the proposed approach, a comparative study was conducted using key performance metrics, including overall accuracy, per-class accuracy, and Cohen's Kappa coefficient. The results demonstrate that GCNN-GSC outperforms state-of-the-art methods, achieving notable improvements across multiple benchmark datasets. Specifically, for the Indian Pines dataset, GCNN-GSC achieved increases of 5.74%, 5.49%, and 7.89% in Cohen's Kappa coefficient, per-class accuracy, and overall accuracy, respectively. Similar enhancements were observed for the Kennedy Space Center, Pavia University, and Houston 2013 datasets, with respective improvements of 6.58%, 6.55%, and 6.15% in Cohen's Kappa coefficient; 6.24%, 6.30%, and 5.57% in per-class accuracy; and 6.24%, 6.54%, and 6.30% in overall accuracy. These results underscore the robustness and effectiveness of GCNN-GSC in hyperspectral image classification tasks.