Interpretable machine learning insights into wildfire damage drivers in California, USA

The escalating threat posed by wildfires to civil infrastructure and communities highlights the urgent need for efficient wildfire-induced damage prediction tools that can identify key damage drivers and be integrated into wildfire risk models. This paper addresses these challenges by developing an interpretable machine learning-based framework for probabilistic wildfire damage modeling in wildfire risk models. The framework also provides data-driven insights into wildfire damage drivers. First, a suite of machine learning and statistical models, including deep neural networks, symbolic neural networks, random forests, balanced random forests, XGBoost, support vector machines, k-nearest neighbors, and logistic regression, were trained and validated using the CAL FIRE damage inspection data collected between 2013 and 2024. These models classified wildfire damage into two states: unaffected-to-minor and moderate-to-severe, using twenty-one selected features across three risk dimensions: hazard, vulnerability, and exposure. Second, the model performance was evaluated using the CAL FIRE damage inspection data from the January 2025 Southern California wildfires, with the deep neural network outperforming others. Finally, transfer learning and cross-validation further enhanced the generalizability and robustness of deep neural network on the entire CAL FIRE damage inspection dataset. To improve interpretability, the Shapley additive explanations were employed, offering both local and global insights into key wildfire damage drivers. Overall, this interpretable machine learning-based framework bridges advanced predictive modeling with practical wildfire risk assessment needs. It provides a robust, data-driven tool that can be seamlessly integrated into wildfire risk models, contributing to improved wildfire hazard preparedness and resilience in California.