PFDiff: Physics- and Frequency-Guided Residual Diffusion for Remote Sensing Image Dehazing

Abstract

Haze, aerosols, and thin cloud-like atmospheric interference reduce contrast and obscure structural cues in optical remote sensing images, degrading both visual quality and downstream interpretation. Direct-regression networks may not sufficiently exploit physically interpretable haze and frequency cues, whereas full-image diffusion requires iterative reconstruction of the complete clean image. This article presents PFDiff, a physics- and frequency-guided residual diffusion framework for remote sensing image dehazing. A haze-aware restoration U-Net first predicts a coarse restoration using a dark-channel map, a transmission-like proxy, and a Laplacian high-frequency cue. Conditional diffusion then models the lower energy residual between the clean target and the coarse output, conditioned on an 11-channel tensor comprising the hazy input, coarse restoration, and the three priors. A learned gate applies spatially varying residual fusion. We further construct HazeRS45 using real-haze-derived masks, multiple transmission patterns, and atmospheric-light variations. Experiments on paired remote sensing dehazing datasets, a cross-degradation benchmark, unpaired real-world images, and downstream tasks demonstrate strong fidelity and structural preservation. Across RSID and HazeRS45, PFDiff achieves the highest average peak signal-to-noise ratio and structural similarity of 26.17 dB and 0.94, respectively, together with the best average deep image structure and texture similarity among the compared methods. It also shows favorable cross-degradation and no-reference real-world performance. Controlled ablations support the residual target and identify ten-step sampling as a balanced quality–efficiency setting. The source code and dataset are available online.