DecloudFormer: Quest the key to consistent thin cloud removal of wide-swath multi-spectral images

Mingkai Li, Qizhi Xu, Kaiqi Li, Wei Li

Abstract: Highlights•DecloudFormer solves cross-patch color differences in wide-swath cloud removal.•Group Layer Norm preserves spatial and channel info for feature extraction.•Checkerboard Mask boosts robustness to cloud variation by reshaping attention maps.•Checkerboard Attention removes color differences via local-global feature fusion.•DecloudFormer outperforms SOTA methods in cloud removal, with no color differences.