Clouds Determine Diffusion: A Cloud-Deterministic Diffusion Model for Progressive Cloud Removal in Optical Remote Sensing Images

Abstract

Cloud removal is crucial for preprocessing optical remote sensing images affected by natural cloud contamination, which is highly diverse and variable. Clouds of varying thickness often occur simultaneously, where the boundary between these layers is ambiguous and hard to quantitatively define. Consequently, advanced deep learning (DL) techniques, especially diffusion models, have been employed to address complex cloud contamination. They primarily focus on conditional generation and standard Gaussian noise diffusion. Yet this paradigm fails to fully capture the inherent diversity, complexity, and variability of natural cloud contamination, nor does it align with the reconstruction process from severe to mild states, leading to limited generalization in real-world scenes. To overcome these limitations, we propose a cloud-deterministic diffusion and progressive cloud removal (CDPCR) model. It introduces a novel diffusion paradigm where cloud contamination serves as the sole forward degradation source. This allows the model to precisely simulate the morphology and dynamic evolution of real-world cloud contamination, thereby enhancing network exposure to diverse cloud-affected patterns and continuously varying intensities in the training stage. During inference, an enhanced sampler iteratively applies the restoration network and the cloud-deterministic diffusion operator, progressively reversing the cloud degradation while effectively preserving residual details. Leveraging dynamic cloud-driven transitions and iterative updates, CDPCR robustly tackles heterogeneous real-world cloud contamination and achieves more realistic and efficient sequence-to-sequence (S2S) reconstruction for optical time series. Crucially, it operates without requiring external cloud masks while alleviating the strong dependence on clean temporal data. Extensive experiments across different datasets validate the superiority and robustness of the proposed method.