Source-Free Domain Adaptive Semantic Segmentation of Remote Sensing Images With Diffusion-Guided Label Enrichment

Source-free domain adaptation for remote sensing semantic segmentation is highly challenging due to the inaccessibility of source data and the absence of target annotations during adaptation. In SFDA for semantic segmentation, pseudo-label-based self-training has been widely adopted, but most existing methods refine the full set of pseudo-labels, which often contains substantial noise and thus severely limits their effectiveness. To address this issue, we propose a novel pseudo-label optimization framework, termed Diffusion-Guided Label Enrichment (DGLE), which starts from a small set of reliable seed pseudo-labels and transforms them into complete and high-quality pseudo-labels. Specifically, DGLE first performs confidence filtering on the predictions from the original and super-resolved target images. The retained predictions are then cross-validated and fused to produce the seed pseudo-labels. Building on these sparse seed pseudo-labels, DGLE trains an image-conditioned diffusion model to progressively generate complete and high-quality pseudo-labels with improved regional consistency and boundary integrity. In this way, DGLE avoids directly optimizing the entire noisy pseudo-label set and provides more reliable supervision for target-domain adaptation. Experiments show that DGLE achieves state-of-the-art performance on three SFDA semantic segmentation benchmarks: Vaihingen →Potsdam and Rural →Urban for remote sensing, and GTA5 →Cityscapes for urban street scenes. It surpasses the previous best methods by 2.45%, 2.18%, and 2.5% in mIoU, respectively.