Investigating the Locational Accuracy of Crowdsourced Data in the Context of NSW Wildfires Using NLP and Geocoding Techniques

Abstract

Wildfires are major natural disasters, and timely monitoring is essential for effective response. Recent progress in Crowdsourced Data (CSD) and Remote Sensing (RS) has improved wildfire monitoring, however their combined use for spatial validation is still limited. This study examined the locational accuracy of Twitter-based wildfire reports during the 2019-2020 New South Wales (NSW) - Australia wildfire season by comparing them with satellite-derived hotspots. Wildfirerelated tweets were processed using Natural Language Processing (NLP), Named Entity Recognition (NER), and geocoding through the Nominatim API, while hotspot data was obtained from the VIIRS sensor via Digital Earth Australia. Advanced data cleaning and keyword filtering methods were applied to extract relevant geolocated tweets from a large raw dataset, and geocoded locations were rigorously clipped to the NSW boundary to improve spatial relevance. A 5 km grid overlay and statistical tests, including chi-square, binomial tests and Poisson-Based tests, were used to measure the spatial relationship between the two datasets. The results showed a strong positive association, indicating that Twitter reports often align with satellite-validated hotspots, particularly in densely reported areas. Although CSD can be uneven in coverage and limited by geocoding accuracy, a combined approach using NLP, spatial filtering, and statistical validation improves its reliability. The study highlighted the value of integrating social media data with RS, proposed a reproducible framework for spatial accuracy assessment, and provided practical guidance relevant to real-time wildfire monitoring and disaster management systems.