Unsupervised Multi-Source-Free Domain Adaptation for High-Resolution Remote Sensing Scene Classification

The existing studies on cross-domain classification of remote sensing scenes face challenges of low classification accuracy and strong dependence on source domain samples, particularly in complex urban scenes. In this paper, we proposed an unsupervised multi-source domain adaptation method for cross-domain classification of high-resolution remote sensing image scenes. Our method starts from extracting source domain featuresusing Swin Transformer as the backbone. We then used a cross-domain classification loss to train the target domain model. The loss consists of weighted information maximization and pseudo-labeling loss. The weighted information maximization loss is used to obtain the source domain feature distributions from the source domain model and combine multiple source domain models, while the pseudo-labeling loss is used to further align the feature distributions of the target and source domains. Our method can fully utilize the existing source domain samples to achieve effective classification of the target domain, while avoiding over-reliance on source domain data. We conducted multiple cross-domain classification experiments on four publicly available remote sensing scene classification datasets (AID, NWPU- RESISC45, U C Merced, and WHU-RS19) as well as the Fuzhou GF-2 dataset. The experimental results show that our proposed method achieves satisfactory cross-domain classification results for remote sensing scenes and remains applicable in complex urban scenes.