ML-Based Fire Cause Prediction: Integrating Population Mobility and Meteorological Data

Wildfire management in Central Europe is often hampered by incomplete data; in the Czech Republic, nearly 49% of fire incidents are recorded with "Unknown" causes. This study presents a novel machine learning framework to address this gap by integrating Global Forecast System (GFS) meteorological data, geographic variables, and dynamic population mobility derived from aggregated SIM-card movement. Utilizing a Gradient Boosting model, we achieved an accuracy of up to 53% in predicting ignition causes, audited by a Natural Language Processing (NLP) semantic proxy of emergency call logs, indicating alignment with model predictions in 38–50% of "Unknown" cases, suggesting a recoverable lower bound of latent information in otherwise unstructured data. Variation in predictive accuracy may reflect differences in the underlying structure and heterogeneity of fire occurrence processes, although it is also influenced by data quality, class imbalance, and model limitations. Through hierarchical clustering, we demonstrate that fires in Seniors-Dominated, stable rural communities are the most predictable, while dynamic urban environments introduce stochastic social noise that reduces model performance. These results offer a dual-purpose tool for fire departments: a method for retrospective database population and a framework for localized, demographic-specific risk management.