LLM-enhanced disaster geolocalization using implicit geoinformation from multimodal data: A case study of Hurricane Harvey

Timely and accurate geolocalization of natural disasters is crucial for effective emergency response, which is foundational for risk mitigation and resilience development. Although social media texts have been widely used to recognize and resolve disaster geolocations, the implicit geoinformation in social media images remains largely underexplored. In this paper, we propose a novel large language model (LLM)-enhanced disaster geolocalization method that considers both explicit and implicit geoinformation from multimodal data. Based on the recognition of geolocations in disaster-related images and texts, geolocalization results were obtained by combining LLMs with map services. The selection of a geolocalization strategy depends on the available geoinformation modality and the presence of spatial relationships. A multimodal dataset of 1,000 images and 1,000 texts from the Hurricane Harvey Twitter dataset was constructed to evaluate geolocalization accuracy through error distance. The results demonstrated that the proposed method achieves significant improvements over baseline geocoding and toponym retrieval methods, with overall accuracies of 81.45%, 78.40%, 74.60%, 65.20%, and 44.95% within 161, 100, 50, 10, and 1 km, respectively. These findings confirm the potential of LLMs in enhancing geolocalization by considering implicit geoinformation from multimodal data for future disaster response and broader GeoAI applications.