Effective Cloud Removal for Remote Sensing Images by an Improved Mean-Reverting Denoising Model with Elucidated Design Space

Cloud removal (CR) remains a challenging task in remote sensing image processing. Though diffusion models (DMs) have achieved promising progress in image generation, their applications to CR are suboptimal, as they employ the vanilla DMs that generate cloudless images from pure noise, ignoring the valuable information in cloudy images. To overcome this drawback, we develop a new CR method $\textbf{EMRDM}$ based on mean-reverting diffusion models (MRDMs) to establish a direct diffusion process between cloudy and cloudless images. Compared to current MRDMs, EMRDM offers a well-elucidated design space with a reformulated forward process and a new ordinary differential equation (ODE)-based backward process. We redesign key MRDM modules to boost CR performance, focusing on restructuring the denoiser and redesigning the training process via preconditioning techniques. We also introduce novel deterministic and stochastic samplers. Additionally, to support the multi-temporal CR task, we develop a denoising network for simultaneously denoising sequential images. We evaluate EMRDM on both mono-temporal and multi-temporal CR tasks. Extensive experiments on various datasets show that EMRDM achieves the state-of-the-art (SOTA) performance.