Spatial-Topological-Semantic alignment for cross domain scene classification of remote sensing images with few source labels

Domain adaptation is crucial for information integration of remote sensing systems, such as satellite constellations and space stations, to intelligently achieving full domain awareness. The conventional methods focus on aligning spatial features without fully considering the topological structure and semantic information in the scene, resulting in loss of useful information and suboptimal classification results. This situation becomes more severe and further complicated to deal with under the condition of few labels available in the source domain. To address the above problems, a spatial-topological-semantic alignment method called STSA is proposed to implement unsupervised domain adaptation (UDA) with few source labels, fully exploring multiple types of information and their complementarity in remote sensing images (RSIs). The proposed method is applied to complete the classification task on a multi-modal cross-domain datasets with synthetic aperture radar (SAR), thermal infrared (TI), near infrared (NI), and short wavelength infrared (SW) images derived from Chinese Tiangong-2 manned spacecraft, as well as a Single modal cross-domain datasets with optical images. Compared with the state of the art UDA methods, even with only one labeled RSI in the source domain, the proposed methods still perform better and achieve satisfying accuracy. It properly explores valuable knowledge from unlabeled RSIs and improves the robustness and flexibility of the model, which is more suitable for UDA with few source labels in RSIs scene classification.