TBNet: Two-Branch Cloud Detection Network for Remote Sensing Imagery

In recent years, convolutional neural network-based cloud detection methods have gradually become mainstream and achieved good performance. However, the detection effect of these methods on cloud edge regions and thin cloud regions is generally poor, mainly because these confusing regions account for a small proportion but have problems of greater intra-class inconsistency and stronger inter-class similarity. To this end, we propose a novel two-branch cloud detection network named TBNet, which focuses more on cloud edge regions by adding a boundary prediction branch to the semantic segmentation branch and jointly optimizing parameters. Specifically, the boundary prediction branch is designed to be shallow and pooling-free to preserve details as much as possible while controlling the computational cost of the model. Further-more, Fusion Module and Semantic Enhanced Module are proposed and used for the interaction between the semantic segmentation branch and the boundary prediction branch. FM can provide the necessary high-level features for the boundary prediction branch, and SEM can provide multi-scale contextual information to enable detection of clouds with different characteristics. In the experiments, our method obtains 97.23% accuracy, 93.20% F1-score, 91.46% Kappa, and 91.92% mIoU on the GF-1 wide field-of-view (WFV) satellite imagery dataset.