Tackling the Wildfire Prediction Challenge: An Explainable Artificial Intelligence (XAI) Model Combining Extreme Gradient Boosting (XGBoost) with SHapley Additive exPlanations (SHAP) for Enhanced Interpretability and Accuracy

Abstract

The intensification of global climate change, combined with increasing human activities, has significantly increased wildfire frequency and severity, posing a major global environmental challenge. As an illustration, Guizhou Province in China encountered a total of 221 wildfires over a span of 12 days. Despite significant advancements in wildfire prediction models, challenges related to data imbalance and model interpretability persist, undermining their overall reliability. In response to these challenges, this study proposes an explainable wildfire risk prediction model (EWXS) leveraging Extreme Gradient Boosting (XGBoost), with a focus on Guizhou Province. The methodology involved converting raster and vector data into structured tabular formats, merging, normalizing, and encoding them using the Weight of Evidence (WOE) technique to enhance feature representation. Subsequently, the cleaned data were balanced to establish a robust foundation for the EWXS model. The performance of the EWXS model was evaluated in comparison to established models, such as CatBoost, using a range of performance metrics. The results indicated that the EWXS model achieved an accuracy of 99.22%, precision of 98.48%, recall of 96.82%, an F1 score of 97.64%, and an AUC of 0.983, thereby demonstrating its strong performance. Moreover, the SHAP framework was employed to enhance model interpretability, unveiling key factors influencing wildfire risk, including proximity to villages, meteorological conditions, air humidity, and variations in vegetation temperature. This analysis provides valuable support for decision-making bodies by offering clear, explanatory insights into the factors contributing to wildfire risk.

Topics & Concepts