Early wildfire escalation prediction with temporal validation and statistical benchmarking in a mediterranean region

ObjectiveTo evaluate an incident-level framework for predicting whether a newly reported wildfire would reach at least 10 ha, emphasizing temporal validation, statistical benchmarking, and transparent optimizer assessment using only information available at or near incident registration.MethodsThis retrospective modelling study integrated 1,901 operational wildfire records documented by the Mersin Regional Directorate of Forestry during 2004-2025 with event-day meteorology and Canadian Fire Weather Index variables. Burned area was known for 1,898 incidents, including 68 major fires. Random Forest, Gradient Boosting, and XGBoost were evaluated under default settings, Random Search, Optuna/TPE, and repeated Enhanced Ruppel's Fox Optimization (eRFO), together with regularized logistic regression using the identical early-stage predictors. Three expanding-window folds within 2004-2020 guided development, while 471 incidents from 2021-2025, including 17 major fires, formed a locked recent-year evaluation.ResultseRFO produced the highest development PR-AUC for Random Forest (0.142) and Gradient Boosting (0.167), whereas Optuna/TPE achieved the highest XGBoost score (0.170). In the locked recent-year evaluation, the logistic benchmark achieved the highest discrimination (ROC-AUC 0.870; PR-AUC 0.217), compared with ROC-AUC 0.813 and PR-AUC 0.210 for XGBoost-eRFO and ROC-AUC 0.774 and PR-AUC 0.145 for Gradient Boosting-eRFO. Reviewing the highest-ranked 20% of incidents captured 9 of 17 major fires with Gradient Boosting-eRFO and 10 of 17 with XGBoost-eRFO. Out-of-sample SHAP analysis identified wind speed, relative humidity, district, temperature, and drought-related Fire Weather Index components as leading model associations.ConclusionsIncreased learner and optimizer complexity did not guarantee superior temporal generalization. The principal contribution is a leakage-controlled evaluation framework combining early-available predictors, temporally ordered development, simple-model benchmarking, calibrated probabilities, and interpretable risk ranking for human-in-the-loop prioritization of newly reported wildfires.