DGAM: Dual-modal Group Attention Mixer for Hyperspectral–Multispectral Collaborative Classification

Fusing hyperspectral images (HSI) and multispectral images (MSI) enhances the performance of land cover classification by combining rich spectral and spatial information. However, existing methods often struggle with insufficient collaborative feature selection and imbalanced utilization of complementary features. To address these challenges, this paper proposes a Dual-modal Group Attention Mixer (DGAM) framework specifically. DGAM introduces a spectral-spatial group attention (SSGA) module that applies differentiated attention mechanisms: spectral grouping for HSI to enhance channel-specific sensitivity and spatial grouping for MSI to improve the ability of multi-scale context modeling, effectively resolving resolution mismatches. A bidirectional interactive modality attention (BIMA) module enables symmetric cross-modal interaction, enhancing common features while preserving modality-specific characteristics. Furthermore, a multi-stream correlation (MSC) classifier is designed to integrate modality-specific and fusion classifiers, which are optimized via a multi-task loss combining cross-entropy, feature consistency, and contrastive objectives. Experimental results on three datasets demonstrate that the proposed DGAM significantly outperforms other state-of-the-art methods.