Abstract: Cloud detection is an effective measure to improve the utilization rate and application range of remote sensing images. However, most of the existing cloud detection algorithms face with two problems: the difficulty in distinguishing complex underlying surfaces like ice and snow from cloud and the requirement of a large number of manually labeled cloud samples for model training. To improve the performance of cloud recognition, we propose a cloud detection algorithm based on Attention U-Net for Landsat images. Firstly, convolution operation is conducted to extract the shallow features of cloud in coding modules. Then, deconvolution, jump connection and attention mechanism are integrated to mine deeper cloud features in decoding modules. Finally, a small number of public Landsat image cloud samples are used for training to achieve end-to-end pixel-level cloud recognition. Compared with traditional machine learning algorithms, experimental results indicate that the proposed algorithm has higher overall accuracy, lower false detection rate and missed detection rate of thin cloud and shadow.

Key words: Landsat images, cloud detection, U-Net network, attention mechanism, endto-end