Probabilistic Wildfire Susceptibility from Remote Sensing Using Random Forests and SHAP

AI can identify wildfire risk factors, but real-world performance lags behind lab results: The model achieved near-perfect accuracy (AUC = 0.996-0.997) when tested on familiar data, but performance dropped significantly when applied to new geographic areas (AUC = 0.54-0.62), highlighting the importance of testing AI systems in realistic conditions before deployment in critical decisions.

Explainability tools reveal which environmental factors matter most: By using SHAP analysis, researchers identified specific drivers of wildfire risk—such as soil carbon and tree cover in forests, and temperature and elevation in grasslands—making the AI's reasoning transparent rather than a "black box," which is essential for stakeholders to trust and act on predictions.

Different landscapes require different risk models: The study found that grasslands and forests have distinct wildfire susceptibility patterns driven by different environmental factors, suggesting that one-size-fits-all AI approaches may be inadequate and that responsible AI deployment requires ecosystem-specific customization.

Geographic variation in risk requires localized attention: The analysis identified specific high-risk districts (Central Valley and Northern Buttes for grasslands; Northern Buttes and North Coast Redwoods for forests), demonstrating how AI can help target prevention and preparedness resources to vulnerable regions most effectively.