Multi-scale strip pooling feature aggregation network for cloud and cloud shadow segmentation

Cloud and cloud shadow detection is a crucial issue in remote sensing image processing. The backgrounds of clouds and cloud shadows are mostly complex in actual remote sensing images. Traditional methods are easily affected by ground object interference, noise interference and other factors, and problems such as missing detection and false detection are prone to occur in the process of cloud detection. In addition, due to insufficient edge information extraction capabilities, traditional methods have very rough segmentation results for cloud and cloud shadow boundaries. In order to improve the accuracy of cloud and cloud shadow detection, a Multi-scale Strip Pooling Feature Aggregation Network is proposed. This method uses the residual network as the backbone to extract different levels of semantic information. And, in order to improve the multi-scale information extraction ability of the network, an Improved Pyramid Pooling module is introduced to mine deep multi-scale semantic information. Then, the Mutual Fusion module is used to guide the fusion of different levels of information. Finally, in view of the problem of rough segmentation boundaries in traditional methods, the Strip Boundary Refinement module is used to repair the boundary information of clouds and cloud shadows. The experimental results conducted on the datasets collected by Landsat-8, Sentinel-2 and a public dataset HRC_WHU show that this method is superior to the existing methods.