Abstract: Most cloud detection methods aim at specific sensors or rely on multiple bands, which require high parameters. However, domestic high-resolution satellite images usually contain a small number of bands, and most cloud detection methods are not applicable. In this paper, the deep learning method is adopted, the fused GF-1 images are innovatively applied to the “dualattention mechanism” model for cloud detection,and compared with the test results of manual drawing and full convolutional network model. Theoretical analysis and research results show that: first, the cloud detection efficiency results of the “dual attention mechanism” model are compared with the results of manual drawing, the accuracy rate is 0.986 4. Second, by increasing the number of cloud samples and the number of non-cloud samples, the model can effectively solve the problem of misdetection of roads, rivers, residential areas. Third, compared with the full convolutional network model, the dual attention mechanism model has more accurate cloud boundaries and stronger model applicability. Using less band information for cloud detection provides a reference for other domestic high-resolution satellite images cloud detection.

Key words: high-resolution, sample, cloud detection, fully convolutional network, visual attention