DisasterTD: Disaster Toponym Disambiguation Using Multimodal LLMs and Cross-View Geolocalization

DisasterTD addresses ambiguous geographic references in social-media imagery during disasters. The framework uses multimodal LLMs to extract toponyms from noisy text and generate candidate locations, then verifies and refines them through cross-view matching among social-media imagery, remote-sensing imagery, and optionally street-view imagery. On a Hurricane Harvey benchmark augmented with remote-sensing and street-view data, it reports accuracies of 47.01% within 50 m, 52.09% within 100 m, and 71.62% within 1 km, with mean and median errors of 11.33 km and 0.68 km.