Abstract: The remote sensing-based information recognition of Mars dunes has important significance for the exploration of the interactions between the Martian atmosphere and the dune surface. Aiming at the low accuracy of the automatic information extraction of Mars dunes using the traditional machine learning method, this paper designs a method combining texture feature extraction and deep learning to automatically identify the information of Mars dunes. In detail, this method conducts information extraction based on the texture feature extraction of Mars remote sensing images combined with a deep learning model, thus realizing the end-to-end semantic segmentation of the remote sensing images. According to experiment results, the U-Net method can fully utilize the rich texture information in the remote sensing images and the extraction accuracy of dunes of this method was 96.7%, which was 3.2 percentage points higher than that of the traditional random forest method. Furthermore, compared to the traditional random forest method, the U-net method extracted more accurate and clearer contours of Mars dunes and yielded better extraction effects of highly fragmented dunes. Therefore, the U-net method can be used for accurate and automatic information extraction of Mars dunes.