Thin Cloud Removal Method Based on Low-Frequency Residual Diffusion and High-Frequency Modulation Refinement by Laplacian Pyramid Decoupling

Optical remote sensing images are often affected by thin cloud cover, which degrades their observation quality. Although recent deep learning-based methods for thin cloud removal have achieved promising results, their performance still requires further improvement. From the frequency-domain perspective, thin-cloud interference is mainly reflected in the low-frequency components, while texture structures are primarily concentrated in the high-frequency components. To guide the model toward thin-cloud removal in the low-frequency layers and detail reconstruction in the high-frequency layers, we propose a divide-and-conquer, two-stage de-clouding method. This approach is developed under the framework of the Laplacian pyramid, which decouples low- and high-frequency information and allows the incorporation of distinct physical priors. In the first stage, a cloud-residual denoising diffusion model is constructed to eliminate thin clouds within the low-frequency domain. By drawing an analogy between the additive model of cloud distortion and the forward diffusion process, the addition of cloud noise is explicitly represented. The reverse process then performs denoising from a non-pure-noise initialization based on the cloud residual, thereby achieving efficient thin-cloud removal at low resolution. This design enhances the interpretability of diffusion models for thin-cloud removal while simultaneously reducing computational cost. In the second phase, statistical analysis revealed that the high-frequency information from the Laplacian pyramid decomposition is highly correlated with brightness. Guided by the brightness information, we progressively modulate and refine the high-frequency features, ultimately generating a cloud-free image through the inverse Laplacian transform. The method presented in this paper is characterized by its intuitiveness and strong interpretability. Experiments conducted on multiple publicly available datasets demonstrate the effectiveness of the proposed approach. The source code is available at: https://github.com/ZzzTD/LP_cloud_removal.