$S^2$Prompt: Sparse Semantic-to-Structural Prompting for Semi-Supervised Remote Sensing Segmentation

CLIP provides transferable class semantics for low-label remote-sensing segmentation, whereas DINOv3 preserves local image structure. However, some existing dual-student training methods couple these complementary priors mainly through confidence-filtered pseudo-labels, leaving soft class uncertainty and its structure-aware propagation underused. We introduce $S^2$Prompt, a sparse semantic-to-structural prompting framework that uses the full CLIP class distribution as a semantic seed and a local top-$k$ graph derived from detached DINOv3 features as structural transport. $S^2$Prompt propagates each class independently, retains the original seed through residual mixing, and injects the raw and refined priors into the DINO decoder through a zero-initialized, gradient-isolated prompt. Both branches remain deployable, with a fixed validation-guided rule selecting the stronger candidate. Across 12 low-label protocols on WHDLD, LoveDA, and Potsdam, the selected predictor achieves the highest validation mIoU among all compared methods. Under fully matched settings, it outperforms the strong baseline by up to 2.9 points, while prompting improves the DINO branch by 14.4–36.4 mIoU. Controlled experiments on three designated protocols show that the raw CLIP prior drives most of this recovery. Feature-derived DINO affinity provides a further average gain of 2.2 mIoU over a parameter-matched random local graph.