Machine learning for one day ahead prediction of high fire weather index conditions from area weighted ERA5 data in mersin

Abstract Severe fire weather places sustained pressure on preparedness in Mediterranean forests. We developed a retrospective benchmark for classifying next-day Fire Weather Index (FWI) conditions in Mersin, Türkiye, from hourly ERA5 reanalysis covering the 2003–2025 fire seasons. Canadian FWI components were calculated independently at 41 ERA5 cells intersecting the province and then aggregated with polygon-intersection area weights. The outcome was High+ conditions on the following day, defined as an area-weighted mean FWI of at least 21.3. Models were developed on 2003–2023 using expanding whole-season validation and evaluated on a prespecified 2024–2025 temporal holdout within the corrected workflow. A 16-predictor regularized logistic model achieved development ROC AUC 0.967, average precision 0.975 and Brier score 0.064, with no clear probability-accuracy advantage from tree ensembles. Corresponding holdout values were 0.946, 0.971 and 0.086, with a calibration slope of 0.962. At the primary 0.50 threshold, recall was 0.929 and precision 0.911. A development-derived F2 threshold of 0.21 increased recall to 0.965 but generated more false alerts. The analysis quantifies predictive information beyond persistence and provides a reproducible regional reference for future forecast-based evaluation. Because the reconstruction uses fixed 1 June initialization, it is not an operational EFFIS forecast.