DDSNet: a dual-domain supervised network for remote sensing image dehazing

Haze shrouds remote sensing images with a thick veil, severely affecting the extraction of valuable information and posing many obstacles to subsequent high-level vision tasks. However, current methods frequently concentrate solely on spatial information while neglecting frequency domain information. To tackle the above problem, we propose a novel model in this study that combines information from the spatial and frequency domains. Unlike most existing methods, We also investigate the relationship between phase and amplitude spectrum components in the frequency domain and haze degradation and use this connection to design a network structure. We have meticulously designed a central Spatial-frequency block containing a Global frequency supervised block (GFS), a Local spatial supervised block (LSS), and a Spatial frequency fusion block (SFF) and utilized parameter-free normalization representation to improve the model's capacity to manage instances with varying attributes. To evaluate the effectiveness of our proposed method, we conduct extensive experiments on two remote sensing image dehazing datasets: SateHaze1k and RICE-1. The results indicate that our network performs exceptionally well, surpassing previous techniques in both quantitative assessments and visual quality. Our DDSNet demonstrates remarkable effectiveness through quantitative analysis, achieving the highest performance across three subsets of the SateHaze1k dataset, with measured values of 24.0053 dB PSNR and 0.9661 SSIM, 26.6054 dB PSNR and 0.9696 SSIM, and 21.3015 dB PSNR and 0.9208 SSIM.