DFANet: Prototype-driven dynamic feature alignment for natural-to-remote-sensing cross-domain few-shot semantic segmentation

Cross-domain few-shot semantic segmentation (CD-FSS) aims to segment novel classes under limited supervision and domain shifts. In this work, we focus on a more challenging yet practically important subtask, namely natural-to-remote-sensing CD-FSS, where knowledge learned from natural-scene data is transferred to diverse remote sensing target domains. Compared with generic unseen-domain transfer, this setting is particularly challenging due to the complex background structures, large scale variations, and substantial imaging discrepancies in remote sensing imagery, which make accurate segmentation in such scenarios difficult. To address these issues, we propose DFANet, a unified framework tailored for natural-to-remote-sensing CD-FSS. We first design a multi-dimensional prototype reconstruction (MPR) module, which refines class prototypes through query-guided attention to enhance prototype discriminability under complex scene structures and ambiguous foreground-background boundaries. We then introduce a dynamic feature alignment (DFA) module to align support and query features from both local spatial structures and global channel statistics, thereby alleviating cross-scene feature mismatches caused by large scale variations and appearance differences. Furthermore, we develop a test-time anchored support finetuning (TASF) strategy, which performs lightweight target-domain adaptation using a few annotated support samples as anchors to further reduce the residual domain discrepancy. On three representative remote sensing datasets, including Deepglobe, Vaihingen, and Potsdam, DFANet achieves average MIoU scores of 40.24%, 47.07%, and 48.95% under 1-shot, 5-shot and 10-shot settings, respectively, outperforming existing CD-FSS methods for natural-to-remote-sensing evaluation. Code is publicly available at https://github.com/zhangxiaogang-111/DFANet .