DisasterTD: Multimodal LLM framework geolocates disaster images with 71% accuracy within 1km

Social media imagery provides timely ground-level views during disasters, but vague geographic references (e.g., 'downtown' or 'near the river') make automated geolocalization unreliable. A new framework called DisasterTD solves this by combining two complementary techniques: multimodal large language models (MLLMs) for semantic toponym extraction and candidate generation, followed by cross-view geolocalization that matches the image against satellite (RSI) and optional street-view (SVI) imagery to verify and refine the location. The MLLM extracts place names from noisy text captions and proposes a set of candidate coordinates, which are then scored via visual similarity with overhead and ground-level reference images.

The framework was evaluated on a curated subset of the Hurricane Harvey dataset, augmented with RSI and SVI, and split into four categories based on toponym clarity (clear, ambiguous, etc.). DisasterTD achieved geolocalization accuracies of 71.62% within 1000 m, 62.36% within 500 m, and 47.01% within 50 m, with mean and median errors of 11.33 km and 0.68 km respectively. These results handily beat both MLLM-only and cross-view-only baselines, with the biggest gains seen on ambiguous toponyms where semantic reasoning combined with visual verification drastically reduced candidate dispersion. The work demonstrates a practical path to fine-grained disaster geolocalization that could directly aid emergency response teams.