Cloud and cloud shadow are the primary factors that affect the application of remote sensing images, and they have always been problems encountered in remote sensing image processing. This article puts forward a new cloud removal strategy, whose data is from the Landsat multi-source remote sensing images and based on an improved BP neural network. Compared with the previous cloud removal methods, the selection value of BP neural network training is changed to reduce human participation. The previous gray-scale value group marked by classification (Vegetation, water body, bare land, residential land, and fields, etc.) is changed to the gray-scale value group of the two images’ common areas without cloud. The experimental results show that the de-cloud image got by our method has higher SSIM and Cosine similarity with the reference image.