Spatial-Spectral Attention Mamba U-Net for multisource remote sensing image fusion

Multisource remote sensing image classification is crucial for land-cover mapping and environmental monitoring. Existing methods often struggle to effectively model both long-range spatial dependencies and cross-spectral interactions while maintaining computational efficiency. Although state space models (SSMs) offer a promising alternative with linear complexity in sequence modeling, their plain structure is insufficient for capturing rich spatial-spectral contextual relationships in multisource data. To address these challenges, we propose Spatial-Spectral Attention Mamba U-Net, a hybrid architecture that strategically integrates SSMs with spatial-spectral attention within a U-Net framework. Our approach introduces two components: (1) the Spatial-Spectral Attention Mamba module, a core component that leverages the linear computational complexity of Mamba SSMs to enable efficient long-sequence modeling. It synergistically combines a transformer-based encoder for capturing global spatial contexts and a lightweight channel attention mechanism for highlighting discriminative spectral features; and (2) a lightweight Mamba U-Net backbone that replaces conventional convolutions with efficient sequential modeling, using Pixel Unshuffle for multiscale feature representation and skip connections for detailed spatial recovery. The proposed architecture was evaluated on the Trento, Muufl, and Augsburg datasets, achieving accuracies of 99.67%, 96.22%, and 97.52%, respectively, surpassing other baseline models.