Unsupervised domain adaptation alignment method for cross-domain semantic segmentation of remote sensing images

Deep learning models rely on a large number of homogenous labeled samples, i.e., limiting the training and testing data to obey the same data distribution. However, when facing large-scale and diverse remote sensing data, it is difficult to guarantee the requirement of homogeneous distribution among data, and the segmentation accuracy of deep learning models decreases significantly. To address this problem, this paper proposes an unsupervised domain adaptation (UDA) method for semantic segmentation of remote sensing images. When the distributions of training data (source domain) and testing data (target domain) are different, the proposed method improves the accuracy of semantic segmentation in the target domain by training deep learning models using only source-domain labels. The method introduces optimal transport theory and global alignment in image, feature and output spaces to mitigate the domain shift between the source and target domains. The experiments employ the Potsdam and Vaihingen datasets provided by the International Society for Photogrammetry and Remote Sensing (ISPRS) to validate the performance. The results show that the method in this paper achieves higher accuracy compared with existing methods. Based on the ablation study, the effectiveness of the optimal transport theory is demonstrated in the UDA framework for semantic segmentation driven by deep learning. © 2023 SinoMaps Press. All rights reserved.