DehazeMamba: SAR-guided Optical Remote Sensing Image Dehazing with Adaptive State Space Model

TL;DR

DehazeMamba tackles optical remote sensing image dehazing under non-uniform haze by leveraging haze-free SAR references with a selective fusion strategy. It introduces the Haze Perception and Decoupling Module (HPDM) and the Progressive Fusion Module (PFM) within a dual-branch Mamba backbone to selectively fuse optical and SAR features in haze-affected regions. The work also provides MRSHaze, a large-scale, high-resolution SAR–optical dataset with precise alignment, enabling robust evaluation. Experiments show a PSNR improvement of $0.73$ dB over state-of-the-art methods and notable gains in downstream segmentation tasks, highlighting the practical impact for remote sensing analysis.

Abstract

Optical remote sensing image dehazing presents significant challenges due to its extensive spatial scale and highly non-uniform haze distribution, which traditional single-image dehazing methods struggle to address effectively. While Synthetic Aperture Radar (SAR) imagery offers inherently haze-free reference information for large-scale scenes, existing SAR-guided dehazing approaches face two critical limitations: the integration of SAR information often diminishes the quality of haze-free regions, and the instability of feature quality further exacerbates cross-modal domain shift. To overcome these challenges, we introduce DehazeMamba, a novel SAR-guided dehazing network built on a progressive haze decoupling fusion strategy. Our approach incorporates two key innovations: a Haze Perception and Decoupling Module (HPDM) that dynamically identifies haze-affected regions through optical-SAR difference analysis, and a Progressive Fusion Module (PFM) that mitigates domain shift through a two-stage fusion process based on feature quality assessment. To facilitate research in this domain, we present MRSHaze, a large-scale benchmark dataset comprising 8,000 pairs of temporally synchronized, precisely geo-registered SAR-optical images with high resolution and diverse haze conditions. Extensive experiments demonstrate that DehazeMamba significantly outperforms state-of-the-art methods, achieving a 0.73 dB improvement in PSNR and substantial enhancements in downstream tasks such as semantic segmentation. The dataset is available at https://github.com/mmic-lcl/Datasets-and-benchmark-code.

Figures (13)

• Figure 1: Comparative analysis of dehazing performance. DehazeMamba demonstrates superior texture recovery and detail preservation through SAR-guided dehazing, producing results that closely match the ground truth (GT) compared to state-of-the-art methods.
• Figure 2: Architecture overview of DehazeMamba. The network comprises dual-branch encoders that process hazy optical and SAR images separately to extract multi-level features. These features undergo fusion through the HPDM and PFM modules in the Fusion Layer. The decoder then reconstructs the fused features into a haze-free output image, with residual connections preserving spatial details across different scales.
• Figure 3: Architectural details of the DehazeMamba (DM) Block. The module combines VSS for global context modeling with MLP for local feature enhancement, incorporating residual connections with learnable parameters to facilitate efficient information flow.
• Figure 4: Synthetic haze generation process. We combine clear optical remote sensing images from Sentinel-2 L2A Product with realistic haze patterns derived from cloud masks extracted from Sentinel-3 SLSTR Product and transformed via a pre-trained CycleGAN model.
• Figure 5: Statistical analysis of haze region proportions in the MRSHaze dataset. The horizontal axis represents percentage intervals of haze coverage per image, while the vertical axis shows the proportion of dataset images falling within each interval. The dataset exhibits a balanced distribution across different haze coverage ranges.