Cloud detection algorithm based on point by point refinement

Abstract In order to limit the interference of cloud noise on ground scene information, detection has been a hot issue in research remote sensing image processing. Cloud labels clouds images at pixel level. The majority early systems rely manually created feature and threshold segmentation with limited generalizability. Remote based deep learning improved accuracy speed thanks quick development convolutional neural networks, but it is still unable satisfy practical application requirements when dealing sceneries variable block size sparse distribution. To this end, study proposes algorithm point-by-point refinement idea coarse fine. Specifically, firstly, residual module introduced U-Net network extract more features; secondly, designed filter out areas where are easily detected wrongly for optimization re-prediction, then produce finer-grained accurate results. quantitative qualitative experiments validate effectiveness proposed method.