In order to improve the accuracy of cloud detection in multispectral remote sensing images, this paper first proposes a deep learning framework to solve the problem of cloud detection accuracy in remote sensing images. This framework benefits from the Full Convolutional Neural Network (FCN), Cloud regions in Landsat 8 images are pixel-level labeled. Secondly, in view of the difficulty in distinguishing clouds from snow and ice regions during cloud detection, a method for removing snow and ice regions based on a gradient recognition method is proposed. Experiments show that a hybrid based on the above two methods (based on gradient recognition and deep learning) can improve the performance in the cloud recognition process, and the method can be automatically generated without manual correction. The Jeckard index and recall rate increased by an average of 4.36% and 3.62%, respectively.