Abstract: Satellite images stereo matching is one of the important steps for large-scale earth surface reconstruction, there are relatively few existing studies, and there are problems such as poor matching effect and poor model generalization ability. A deep learning-based satellite image stereo matching algorithm is proposed to perform stereo matching, including dataset construction, building stereo matching network, multi-level transfer learning and post-processing. Dataset construction contains disparity offset and and cropping. The cropped patches are then fed into attention volume network, which includes deep feature extraction, attention volume construction, disparity estimation. The network is trained by multi-level transfer learning, adapts to different data sources, and predicts the disparity maps. The false matches are filtered out by post-processing. The experiments were carried out with Jilin1-GF02 and Jilin1-GF04 images. The accuracy of the disparity maps obtained from the experimental results is better than one pixel. It shows that the algorithm proposed in this paper can obtain accurate and efficient results, which determines the generation of subsequent high-quality digital surface model.

Key words: satellite stereo image, disparity estimation, convolutional neural network, attention cost volume, transfer learning