Learning a multi-scale vision Mamba for weather-degraded remote sensing image restoration

Adverse weather conditions consistently compromise the quality of remote sensing images and hinder downstream vision-based tasks. Recent progress in remote sensing image restoration has been driven by Convolutional Neural Networks and Transformers. Nonetheless, these approaches face challenges such as constrained receptive fields or high computational costs with quadratic complexity, resulting in a trade-off between performance and efficiency. In this paper, we propose an effective multi-scale vision Mamba for remote sensing image restoration by modeling long-range pixel dependencies with linear complexity. Specifically, we develop a bidirectional Mamba network architecture that effectively explores intra-scale and inter-scale information interactions. In addition, we design an efficient multi-scale 2D scanning mechanism to better facilitate image restoration across different scales. Extensive experiments show that the proposed method performs favorably against state-of-the-art models.