Cross-Domain Urban Land Use Classification Via Scene-Wise Unsupervised Multi-Source Domain Adaptation With Transformer

Current land use classification models from Very High Resolution (VHR) remote sensing images often suffer from high sample dependency and poor transferability. To address these challenges, we propose an unsupervised multi-source domain adaptation framework for cross-domain land use classification that eliminates the need for repeatedly using source domain data. Our method uses the Swin Transformer as the backbone of the source domain model to extract features from multiple source domain samples. The model trained on source domain samples, and unlabeled target domain samples are then used for target domain model training. To minimize the feature discrepancies between the source and target domains, we use a weighted information maximization loss and self-supervised pseudo-labels to alleviate cross-domain classification noise. We conducted experiments on four public scene datasets and four new land use scene datasets created from different VHR images in four Chinese cities. Results show that our method outperformed three existing single-source cross-domain methods (i.e., DANN, DeepCORAL, and DSAN) and four multi-source cross-domain methods (i.e., M3SDA, PTMDA, MFSAN, and SHOT), achieving the highest average classification accuracy and strong stability. We conclude that our method has high potential for practical applications in cross-domain land use classification using VHR images.