TCME: Thin Cloud Removal Network for Optical Remote Sensing Images Based on Multidimensional Features Enhancement

Existing thin cloud removal methods mainly rely on generative paradigms and discriminative paradigms. The generative paradigm has issues with training instability, while the convolutional neural network (CNN) structure in the discriminative paradigm can only focus on local features, neglecting global ones. This article proposes a thin cloud removal network for optical remote sensing images based on multidimensional feature enhancement (TCME), which combines the advantages of convolution and transformer. Specifically, TCME adopts a transformer backbone to capture features with long-range dependencies. In the transformer block, a space-channel attention module based on sparse selection mechanism [top-K sparse mixed-dimension head transposed attention (TKSMA)] and a dual convolutional gated feedforward network (DCGFN) are designed to efficiently enhance global perception and feature aggregation capabilities. In addition, a full-dimensional feature enhancement module with convolutional structure [omnidimensional feature enhancement block (ODFEB)] is designed to enhance local features for detailed recovery. Finally, a color angle loss function is used to avoid color loss issues. Extensive experiments on the public dataset WHUS2-CR and a custom dataset demonstrate that our model, TCME, achieves the best performance compared to other methods. The source code and datasets are available at this https://github.com/nannan2323/TCME-NET.