Two-level Alignment-based Unsupervised Domain Adaptation for Semantic Segmentation of Remote Sensing Images

Semantic segmentation is an essential analysis task for understanding remote sensing images. Recently, many supervised semantic segmentation models have achieved high performance. However, this performance tends to decline when there is a distribution shift between the source and target domains, such as a change in the geographical area or sensor mode. One solution to overcome this issue is to use unsupervised domain adaptation, which transfers the grasp of a model trained on a source domain with accessible labels to the target data domain without label access. This paper proposes a new unsupervised domain adaptation method for remote sensing images. The proposed approach leverages a combination of Fourier transform-based image-to-image translation to diminish the shift in the input-level space and the fine-grained domain discriminator to address the shift in the class-based feature-level space. The experimental results demonstrate that our proposed method effectively improves the performance of cross-domain remote sensing semantic segmentation tasks.