A Random Forest-Based Early Warning System for Wildfire Risk in the Southern Aegean and Western Mediterranean Regions of Türkiye

The fires in İzmir that occured in June 2025 and resulted in widespread evacuations and extensive ecological damage, showed that forest fires are becoming more frequent in Türkiye's Aegean and Mediterranean coastal regions, seriously damaging local ecosystems, socioeconomic activity and vital infrastructure. Therefore accurately predicting regional fire hazard using meteorological data has become critically important as climate change intensifies these vulnerabilities. This study uses two methods: long-term climate risk projections using the Fire Weather Index (FWI) and the development of a high-precision, artificial intelligence-based Decision Support System (DSS). Future wildfire dangers (2026–2096) under RCP 2.6, 4.5, and 8.5 scenarios were projected using high-resolution EURO-CORDEX regional climate models and compared to a historical reference period (1971–2005). Quantitative analysis of severe risk thresholds (FWI > 45) reveals a profound geographical expansion of fire hazards. While the historical average stood at 50.48 extreme-risk days, projections indicate an increase to 55.22 (+9.4%) under RCP 2.6 and 61.71 (+22.2%) under the pessimistic RCP 8.5 scenario. Crucially, coastal hotspots are expected to endure up to 234.92 extreme-risk days annually under RCP 8.5. In this projection, the traditional summer fire season shifts into a nearly constant hazard lasting approximately 65% of the year, demanding immediate proactive adaptation strategies. The next stage of this research presents a proactive early warning strategy using the Random Forest machine learning method to handle these extended fire seasons. The system synchronizes dynamic ERA5 meteorological variables with 30m high-resolution National Aeronautics and Space Administration Shuttle Radar Topography Mission (NASA SRTM) topography data (elevation, slope, and aspect). This integration is achieved through bilinear interpolation into a unified 1-km spatial grid. Trained on 199,606 balanced instances from National Aeronautics and Space Administration Fire Information for Resource Management System (NASA FIRMS) (2010–2020), the model’s dependability was rigorously assessed through an out-of-time validation using 34,265 data points from the extreme fire year of 2021. The DSS achieved an outstanding overall accuracy of 95.81%, successfully detecting 15,709 genuine fire events while maintaining a remarkably low false positive count of only 49, effectively eliminating false alarms that strain public resources. Feature relevance ratings identified surface temperature (0.30 weight) and elevation (0.25) as the primary drivers of fire vulnerability. Ultimately, this research serves as a foundational framework for evidence-based policy-making and strategic land-use planning. By identifying non-linear risk patterns and producing real-time, high-resolution vulnerability maps, the DSS enables forestry and emergency directorates to shift from reactive firefighting to proactive governance. This model provides the scientific justification for modernizing national fire-fighting protocols and optimizing the strategic allocation of aircraft and ground resources, ensuring that mitigation policies are dynamically aligned with the emerging reality of a continuous fire season.