Spectral Gradient Field-Guided Bayesian Adaptive Graph Convolution for Hyperspectral Image Classification

Abstract

Graph-based deep learning methods play a significant role in hyperspectral image classification (HSIC). Nonetheless, the paucity of prior information to construct graph model structures continues to severely challenge the classification of images characterized by significant spatial heterogeneity. To achieve effective adaptive construction of graph model structures, this article proposes a spectral-driven self-adapting graph construction framework. A spectral gradient field (SGF) is introduced to quantify the local rate and direction of spectral variation, enabling dynamic adjustment of node connections according to spatial–spectral characteristics. The method adaptively expands or contracts neighborhood scopes based on local spectral gradients, effectively addressing over- and under-segmentation problems in complex regions. Furthermore, a beta distribution is employed to model the probabilistic connectivity between nodes, enhancing the robustness of graph topology. For large hyperspectral scenes, constructing and storing a full graph leads to prohibitive memory consumption and computational cost. To address this issue, we extend the proposed method to a block-wise and pruned Bayesian adaptive graph convolutional network (BP-BAGCN), in which images are divided into nonoverlapping or lightly overlapping spatial blocks; then the local maps are constructed, refined, and inferred independently. Experiments are conducted on three hyperspectral datasets: a benchmark dataset and two newly constructed datasets derived from GF-5 satellite observations over complex coastal wetlands, providing challenging real-world scenarios with rich spectral variability and ecological diversity. Experimental results demonstrate that the proposed method consistently outperforms state-of-the-art approaches under limited labeled samples. Across five independent random seeds, the proposed method achieves average overall accuracies of 98.79% $\pm ~0.70$ %, 99.35% $\pm ~0.13$ %, and 92.27% $\pm ~0.65$ % on the three datasets, respectively, demonstrating both superior classification performance and stable generalization ability.