Cloud removal with compact diffusion models: A residual block-based approach

Abstract

Satellites are powerful tools for remote sensing, as they enable the imaging of large areas with high quality. However, satellites can be prone to artifacts such as clouds, which can negatively influence the analysis of these images. Thus, researchers have widely investigated cloud removal techniques to mitigate these artifacts, leveraging the rise of generative artificial intelligence methods. These techniques, although powerful, require a high computational cost, which limits their use in real-time applications, embedded devices, and environmental monitoring systems, where computational resources are often limited. Therefore, this work presents an approach based on compact latent diffusion, where the denoising model uses attention channels and residual block operations. In addition, we evaluated different training loss functions, which help the model perform cloud removal across various land cover types. Considering a resource-constrained approach, we investigated different experimental configurations using Pareto Front to optimize the most promising experiments. Our results demonstrate a balance between reconstruction quality and computational cost compared to baseline. Our approaches have between 48% and 82% fewer parameters while presenting competitive results for similarity, noise, and perceptual metrics.