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

Wide-swath images contain clouds of various shapes and thicknesses. Existing methods have different thin cloud removal strengths in different patches of the wide-swath image. This leads to severe cross-patch color inconsistency in the thin cloud removal results of wide-swath images. To solve this problem, a DecloudFormer with cross-patch thin cloud removal consistency was proposed. First, a Group Layer Normalization (GLNorm) was proposed to preserve both the spatial and channel distribution of thin cloud. Second, a CheckerBoard Mask (CB Mask) was proposed to make the network focus on different cloud-covered areas of the image and extract local cloud features. Finally, a two-branch DecloudFormer Block containing the CheckerBoard Attention (CBA) was proposed to fuse the global cloud features and local cloud features to reduce the cross-patch color difference. DecloudFormer and compared methods were tested for simulated thin cloud removal performance on images from QuickBird, GaoFen-2, and WorldView-2 satellites, and for real thin cloud removal performance on images from Landsat-8 satellite. The experiment results demonstrated that DecloudFormer outperformed the existing State-Of-The-Art (SOTA) methods. Furthermore, DecloudFormer makes it possible to process thin cloud covered wide-swath image using a small video memory GPU. The source code are available at the link.