Learning Coastlines as Geometric Curves with Vision-Language Models

Coastline detection in remotely sensed imagery is commonly formulated as pixel-wise segmentation, even though coastlines are ultimately used as geometric vector boundaries in coastal monitoring and analysis. We revisit this representation choice and formulate coastline extraction as a direct geometric boundary localization problem. We introduce CoastlineVLM-7B, a vision-language model that predicts ordered coastline polylines while also supporting coastline-presence detection and geomorphic proxy classification. Experiments show that U-Net is stronger for strict local boundary proximity, while CoastlineVLM-7B achieves better worst-case and global structural alignment, with lower Hausdorff distance and Earth Mover’s Distance. The model also shows promising zero-shot cross-region generalization to an independent Australian coastline dataset without additional fine-tuning. These results support direct ordered-polyline grounding as a viable alternative representation for coastline localization and a step toward model outputs that align more directly with vector-based coastal monitoring workflows.