Spatial-spectral enhanced mamba with diffusion prior for spectral super-resolution

Hyperspectral remote sensing is pivotal for precise Earth observation, yet its widespread application is often constrained by the inherent limitations of imaging hardware. Spectral super-resolution (SSR) has emerged as a crucial technique to mitigate these issues by reconstructing hyperspectral images from readily available RGB images. Currently, most SSR methods are primarily based on CNN or Transformer architectures, yet these approaches suffer from limited receptive fields and high computational complexity. While the emerging Mamba model offers a new paradigm for SSR due to its long-sequence modeling capability, it exhibits noticeable deficiencies in capturing local spatial-spectral features. To address this, we propose a Spatial-spectral enhanced Mamba with Diffusion Prior (SMDP). This model proposes a channel-aware mamba attention symbiosis module by combining Mamba and attention mechanisms, which serves as the backbone for learning global spatial-spectral context. Then, we incorporate two complementary modules to enhance local information capture: a spatial multi-scale partition convolution aggregation module, which captures fine-grained local textures through a grouping strategy and multi-scale convolutional kernels, and a spectral latent diffusion enhancement module, which incorporates a latent diffusion model to learn a low-rank spectral dictionary, thereby enforcing structural spectral consistency and enhancing reconstruction fidelity. Experimental results on four benchmark datasets demonstrate that SMDP achieves competitive performance across multiple metrics.