Mapping Natural Disasters Using Social Media Posts with an Encoder-Decoder Model

Abstract. Real-time mapping of social media posts from users in the affected areas during a natural disaster can generate actionable intelligence for disaster response teams. Given the complexities of natural language used in such posts, modern approaches rely on Artificial Intelligence (AI) to improve accuracy. Extracting actionable geospatial intelligence from unstructured text requires a robust geoparsing pipeline comprising toponym detection and toponym resolution. While general-purpose Large Language Models (LLMs) can be utilized for toponym detection, their operational utility in high-volume, real-time workflows is constrained by high computational costs and a heavy reliance on intensive prompt engineering. To address these limitations, this study presents a highly efficient alternative utilizing custom encoder-decoder models fine-tuned specifically for toponym detection. Leveraging a dataset of 7,400 curated tweets from the 2024 hurricane Helene and another dataset of 50,000 tweets from the 2017 hurricane Harvey, we fine-tuned Google Research’s Flan-T5-base architecture twice. Both finetuned versions of the model demonstrated relatively good robustness, converging to F1 scores of 0.87 and 0.83 for hurricanes Helene and Harvey respectively. For the subsequent toponym resolution stage, we implemented a hybrid pipeline that categorizes extractions into four GIS-aligned classes. Regional-scale entities are matched against authoritative local GIS feature layers using a fuzzy string matching approach to map polygons, while localized features are resolved to point coordinates via GoogleMaps Geocoding API. The results of mapped tweets are examined as dynamic heatmaps that prove to be valuable in generating geospatial intelligence for disaster management.