Haze removal in satellite images
Cheng Yao Lee · DR-NTU (Nanyang Technological University) · 2026
Haze significantly degrades the quality of Remote Sensing Images (RSIs), severely hindering critical applications like disaster management and agricultural monitoring by obscuring structural details, reducing contrast and colour fidelity. Deep learning models, particularly attention-based Convolutional Neural Networks (CNNs), have demonstrated strong and reliable performance in haze removal, but often at the cost of high computational complexity, or the ineffectiveness in capturing complex and non-uniform haze. To bridge this gap, this paper presents C-SAMBA-Net (Content-guided Spatial Attention maMBA Network), a robust and lightweight dehazing architecture built upon the Dual Attention Network (DA-Net). C-SAMBA-Net features targeted architectural innovations designed to maximise restoration quality without inflating parameter count. It is a hybrid CNN-Mamba model created by implemented two major modifications: (1) Upgrading the network’s bottleneck (Layer 3 with a Detail-Enhanced Attention Block (DEABlock) from the DEA-Net for superior feature weighting and (2) Substituting the Parallel Attention in the Parallel Attention Module with a Mamba-driven Attention State-Space Module (ASSM) to efficiently capture long-range dependencies, preserving spatial structures that standard parallel attention oversimplifies. Extensive evaluations were performed on standard RSI dehazing benchmarks (RSID, RICE, SateHaze1k, UA V). Our optimised C-SAMBA-Net achieved a Peak Signal-to-Noise Ratio (PSNR) of 40.07, Structural Similarity Index (SSIM) of 0.991 and CIE Delta E 2000 (CIEDE) score of 1.191 on the RICE1 dataset. It outperformed the baseline DA-Net, averaging an improvement of 5.41%, 2.12% and 10.32% for PSNR, SSIM and CIEDE respectively across all datasets. C-SAMBA-Net demonstrates superior colour fidelity while delivering strong dehazing performance, while keeping a low complexity with a total parameter count of 1.13M and FLOPS of 11.86G. This establishes the C-SAMBA-Net as a highly efficient and high-quality solution for real-time or resource-constrained RSI processing pipelines.