In most remote sensing image cloud detection methods，the training data needs to mark each pixel of the image，which is very expensive. In order to reduce the cost of manual labor marking data in deep learning remote sensing image cloud detection，the image block label replaces the pixel label for in-depth learning training. First，the remote sensing images of various underlying surfaces are cut into image blocks and labeled，and the labeled image blocks are used as data sets. Then，the block data set is trained to improve the VGG deep learning network，and the trained network is used for cloud detection of large remote sensing images. Finally，multiple medium resolution satellite images are selected for cloud detection comparison experiments using improved VGG and VGG networks. The results show that the improved VGG remote sensing image cloud detection method can effectively detect fragmented and thick clouds，with an average accuracy of over 90% for the entire cloud area. The use of tagged image blocks not only reduces manual labor，but also effectively detects clouds in remote sensing images，which can provide a reference for the research of remote sensing images with weak supervision and deep learning.