A Unified Sentinel-2 Imagery Thick Cloud Removal and Rescaling Framework From a Continuous Perspective

Wei-Hao Wu, Ting-Zhu Huang, Xi-Le Zhao, Xingwen Quan, Yu-Bang Zheng, Deyu Meng

Abstract: Cloud contamination presents a significant challenge in remote sensing of land surface, severely compromising data quality and hindering applications such as spectral unmixing, data fusion, and target detection. However, leveraging the full spectral bands of multitemporal Sentinel-2 images with different spatial resolutions (SRs) for thick cloud removal remains a longstanding challenge. To address this challenge, we propose a unified thick cloud removal and rescaling (UTCR2) framework from a continuous perspective, which can recover cloud-free multitemporal Sentinel-2 images with different SRs in a continuous manner. Specifically, we decompose the Sentinel-2 images with different SRs into the intrinsic abundance maps (i.e., the proportion of different materials at each pixel) with different SRs and discrete spectral signatures. To capture the intrinsic abundance maps with different SRs, we cleverly leverage the implicit neural representation (INR), which can naturally continuously represent data and fuse the spatial information across the full spectrum of multitemporal Sentinel-2 images with different SRs. To capture the discrete spectral signatures at different time nodes, we utilize fully connected networks (FCNs). Extensive simulated and real experiments on Sentinel-2 datasets demonstrate that the proposed UTCR2 significantly outperforms existing state-of-the-art methods and benefits downstream applications.

External IDs:dblp:journals/tgrs/WuHZQZM25