Predictive Modeling for Fire Severity to Optimize Emergency Resource Allocation

This study develops and validates a predictive framework for urban fire severity assessment to support data‐driven resource allocation in firefighting operations. Traditional decision‐making approaches often rely on heuristics and experience, limiting their adaptability to the complex, nonlinear factors influencing modern fire dynamics. To address this gap, we propose an interpretable predictive model that integrates Extreme Gradient Boosting (XGBoost) with Geographic Information Systems (GIS) to analyze the fire incidents recorded in a certain city (2010–2020). Unlike prior GIS–ML studies that emphasize static mapping or single‐phase validation, the proposed approach operationalizes predictive outputs within a realistic resource allocation simulation, linking prediction accuracy with measurable performance gains. This methodological synthesis establishes a new standard for connecting predictive analytics, spatial modeling, and operational decision support in urban fire management. The model's underlying mechanism combines gradient‐based optimization and regularization to capture nonlinear relationships among temporal, structural, and spatial predictors while preventing overfitting. Comparative experiments against Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN) models were conducted under identical conditions using five‐fold cross‐validation. XGBoost achieved the highest overall performance (accuracy = 85.1%, recall = 84.3%, AUC = 80.5%), outperforming alternative models in both predictive precision and operational reliability. Simulation results further indicate potential reductions of 24.5% in property damage, 17.8% in firefighter injuries, and 12.5% in response times when the model guides resource deployment. The study contributes methodologically by introducing a tri‐level validation framework (temporal, geographic, and operational) and theoretically by linking evidence‐based decision‐making with spatial predictive analytics. These findings establish a robust, interpretable, and scalable foundation for integrating machine learning into real‐time firefighting management and broader urban safety planning.