Enhancing Mamba with discriminative graph convolution for hyperspectral image classification
Abstract
Accurate hyperspectral image (HSI) classification requires long-range dependency modeling without losing irregular spatial context. Existing Mamba-based approaches may linearize two-dimensional neighborhoods, whereas graph-based approaches can depend strongly on a single graph definition; simply combining the two architectures does not resolve these limitations. We propose DGCM, a single-modality framework that assigns complementary roles to state-space sequence modeling, trainably fused superpixel topology, and representation-level regularization. Unlike recent graph-Mamba hybrids organized around graph sequences, parallel branches, or multimodal fusion, DGCM preserves an explicit spatial graph while modeling long-range dependencies through a separate linear-complexity pathway. Experiments on four public within-scene HSI benchmarks show consistent improvements over representative CNN-, Transformer-, Mamba-, and graph-based methods, together with stable behavior across the evaluated training sizes and model settings. These results support the effectiveness of separating topological and sequential modeling under the investigated within-scene benchmark protocol; cross-scene and operational generalization remain to be evaluated.