Remote Sensing Image Dehazing Using Content-Driven State Space Modeling With Scale-Aware Aggregation

Haze leads to a substantial decrease in the quality of remote sensing images, thereby hindering high-quality earth observation and other vision-related applications. Recently, state space models have demonstrated great potential in image restoration tasks, offering advantages in global information modeling while maintaining low computational complexity. However, we observe that existing Mamba-based image dehazing methods rely on a fixed-scale spatial-aware scanning mechanism, which presents limitations when applied to complex remote sensing scenarios characterized by objects and features at varying spatial scales. In this paper, we propose an effective scale-aggregation content-driven state space model for remote sensing image dehazing, called SC-DehazeMamba. Specifically, the proposed framework integrates both coarse-to-fine and multi-patch strategies into a unified parallel multi-branch architecture, where each branch leverages multi-scale learning to enhance robustness against diverse degradation appearances. In addition, we design a content-driven dynamic adaptive scanning mechanism that can adaptively determine the scanning order and regions based on the input remote sensing image. Through a series of extensive experiments on different benchmark datasets, our proposed framework has shown its effectiveness and favorable performance compared to state-of-the-art solutions.