Toward Cross-Regional Land Cover Mapping Using Regional Consistency and Certainty-Based Active Domain Adaptation With Limited Annotations

Semantic segmentation can automatically classify in remote sensing(RS) images by pixel level into different types of land cover. However, in cross-domain semantic segmentation, how to more accurately identify the boundaries between types of surface features remains a critical challenge to be addressed. In this paper, we introduce a region-aware active selection strategy incorporating region consistency and certainty for RS segmentation. The method, Cross-Regional Active Land Cover Mapping, enhances feature discriminability by maintaining local prediction consistency between pixels and their nearest neighbors in the source image. More accurate identification of boundaries between two types of ground objects positively contributes to improving segmentation accuracy. We conducted four sets of transfer annotation tasks on two representative RS datasets and validated our proposed method through these tasks, and our proposed method has achieved a significant improvement of 1.03% to 38.23% compared to existing methods in the transfer tasks mentioned above and has narrowed the performance gap with fully supervised learning to a minimum of just 0.63%. Our method achieves outstanding performance with minimal annotation effort, offering practical and valuable information for advancing the field of RS.