Abstract
Optical remote sensing images are frequently degraded by haze and clouds, which cover approximately 50% of the Earth's surface at any given time. This degradation significantly hinders data interpretation for critical applications like agriculture and surveillance. We present DehazeMamba, a novel framework that leverages Synthetic Aperture Radar (SAR) data to guide the dehazing process. By utilizing the efficient Mamba architecture and a unique "haze decoupling" strategy, DehazeMamba restores clear details in hazy regions while preserving the quality of haze-free areas, establishing a new state-of-the-art in satellite image restoration.
Introduction
Traditional single-image dehazing methods often struggle with complex atmospheric conditions, frequently producing artifacts or failing to recover underlying textures. While Synthetic Aperture Radar (SAR) can penetrate clouds and provide structural information, its integration into optical dehazing is challenging due to the significant domain shift and the potential for SAR noise to corrupt clear optical regions. DehazeMamba addresses these issues by selectively utilizing SAR guidance only where haze is present, ensuring high-fidelity restoration across the entire image.
Methodology
DehazeMamba is built upon the State Space Model (SSM) architecture, specifically the "Mamba" variant. This choice provides a global receptive field necessary for understanding broad haze patterns while maintaining linear computational scaling, which is crucial for high-resolution satellite imagery.
The architecture introduces two primary components:
- Haze Perception and Decoupling Module (HPDM): This module identifies hazy regions by analyzing feature differences between optical and SAR data. It generates a weight map that ensures SAR information is injected only where the haze is thickest, protecting clear areas from unnecessary noise.
- Progressive Fusion Module (PFM): To bridge the gap between modalities, the PFM employs a two-stage process. A "coarse" fusion assesses feature quality, which then guides a "fine-grained" fusion to mitigate the impact of SAR noise and ensure structural alignment.
- DehazeMamba (DM) Block: The core building block combines Vision State Space (VSS) for global context with Multi-Layer Perceptrons (MLP) for local detail enhancement.
Datasets
The study introduces MRSHaze, a significant new benchmark for SAR-guided dehazing. It consists of 8,000 pairs of temporally synchronized and geo-registered SAR (Sentinel-1) and optical (Sentinel-2) images at 512x512 resolution. Compared to existing datasets like SEN12MS-CR, MRSHaze offers superior alignment and higher resolution, providing a more robust platform for evaluating restoration models.
Key Results and Discussion
DehazeMamba demonstrates superior performance both quantitatively and qualitatively. On the MRSHaze dataset, it achieved a PSNR of 33.39 dB, outperforming the previous state-of-the-art (FS-Net) by 0.73 dB.
Visual comparisons highlight the model's ability to recover sharp textures and natural colors, even in scenarios where other models lose target outlines entirely. Furthermore, the effectiveness of DehazeMamba was validated through downstream tasks; dehazed outputs led to significantly improved accuracy in object detection (YOLOv8) and semantic segmentation (ALE-SEG) compared to other dehazing methods.
Conclusion
DehazeMamba represents a significant advancement in remote sensing image restoration by effectively fusing heterogeneous SAR and optical data. By treating fusion as a dynamic, quality-aware process and leveraging the efficiency of the Mamba architecture, the model achieves state-of-the-art results. Future work will focus on further reducing computational overhead and exploring the integration of additional data modalities like thermal or infrared imagery.
Summary updated