People as Sensors: An End-to-End AI Framework for Real-Time Wildfire Detection, Classification, and Geospatial Intelligence from Social Media

Source abstract

Wildfires are increasing worldwide, and the first sign of a fire is often a citizen’s social media post, not a satellite. This paper conceptualizes “Citizens as sensors” as a complete, operational system: an end-to-end, Greek-language NLP pipeline that transforms raw, unlabeled X posts into a live, geolocated wildfire map, with no human intervention at inference. The pipeline identifies active fires, categorizes them as wildland, urban, or mixed, resolves the referenced toponym, and groups fire reports into incident zones on an interactive, time-sliced satellite map. On the first-ever corpus of 3,277 manually double-annotated Greek posts, validated by a formal inter-annotator audit (Cohen’s κ = 0.978), fine-tuned transformers achieved F2 = 0.911 for fire detection and macro-F1 = 0.874 for fire-type classification, decisively outperforming two LLMs and classical baselines on the same test sets, while being more accurate, over twice as fast, and free of per-call cost. Spatiotemporal toponym disambiguation raised geocoding accuracy from 54.68% to 80.58% (p < 0.001), and a later repeatability test revealed measurable gazetteer drift. Applied to raw posts from 12 August 2024, the pipeline automatically identified and mapped a major Attica wildfire in near real time, resolving hundreds of citizen reports into a single incident cluster.