An explainable ANN–MCDM framework for wildfire susceptibility prioritization and preliminary response-center candidate generation in Muğla, Türkiye

Abstract This study presents an explainable hybrid artificial neural network (ANN)–multi-criteria decision-making (MCDM) framework for wildfire susceptibility prioritization and preliminary response-center candidate generation in Muğla, Türkiye. A matched grid–date case–control dataset comprising 65,020 observations was constructed from FIRMS/VIIRS active-fire detections, common-protocol ERA5 fire-weather variables, and structural environmental predictors. Predictive performance was evaluated using leakage-safe nested spatial validation, repeated 5-, 10-, and 20-km block settings, calibration analysis, and spatial-supergroup bootstrap confidence intervals. The ANN achieved a mean ROC-AUC of 0.6601 and a mean PR-AUC of 0.6191, remaining competitive with random forest, penalized logistic regression, and LightGBM, but showing no decisive superiority over random forest. Cross-fitted grouped permutation importance and SHAP analyses for both ANN and random forest identified daily minimum relative humidity, daily maximum wind speed, and daily maximum temperature as the dominant predictive-relevance core, jointly accounting for 67.76% of the four-source mean relevance mass. MCDM robustness was assessed across 144 prespecified weight, utility, normalization, TOPSIS, and VIKOR scenarios. The representative consensus TOPSIS and VIKOR rankings showed strong global agreement but only moderate overlap in the extreme upper tail, indicating non-redundant and scenario-dependent reprioritization. Temporal-to-spatial aggregation identified 117 robust 1-km priority blocks within 28 connected hotspot components. Priority-weighted clustering then generated ten hotspot-based screening points for subsequent response-center planning. These outputs support transparent spatial screening, but do not constitute calibrated wildfire probabilities or final facility-location decisions.