Cloud Removal Using Patch-Based Improved Denoising Diffusion Models and High Gray-Value Attention Mechanism

In recent years, diffusion-based methods have outperformed traditional models in many cloud removal tasks due to their strong generative capabilities. However, these methods face the challenges of long inference time and poor recovery effect in cloud regions. To address this issue, this letter proposes a patch-based improved denoising diffusion model with a high gray-value attention for cloud removal in optical remote sensing images. We introduce an overlapping fixed-sized patch method in the improved denoising diffusion model. The patch-based diffusion modeling approach enables size-agnostic image restoration by employing a guided denoising process with smoothed noise estimates across overlapping patches during inference. Additionally, we introduce a high gray-value attention module, specifically designed to focus on thick cloud regions, enhancing attention on areas with relatively high gray values within the image. When compared with other existing cloud removal models on the RICE dataset, our model outperformed them in terms of both peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) index. Qualitative results demonstrate that the proposed method effectively removes clouds from images while preserving texture details. Ablation studies further confirm the effectiveness of the high gray-value attention module. Overall, the proposed model delivers superior cloud removal performance compared to existing state of the arts (SOTA) methods.

I. Introduction

Remote sensing information technology, with its extensive coverage, high data acquisition efficiency, and diverse data types, occupies a crucial role in Earth observation tasks [1]. However, despite its significant advantages, optical remote sensing imagery is often hindered in practical applications by natural factors such as cloud cover and atmospheric conditions. These factors severely impede the extraction of effective information, leading to distorted image data that fail to meet the urgent demand for high-quality remote sensing information in various fields [2]. Therefore, the development of an algorithm capable of efficiently and accurately removing cloud contamination and achieving high-quality reconstruction of remote sensing imagery is critical.