Post-Hoc Explanation of AI Predictions in Wildfire Risk Mapping Through an Interactive Web-Based GeoXAI System

This case study presents an innovative approach for explaining wildfire susceptibility through a web-based Geospatial eXplainable Artificial Intelligence (GeoXAI) system. By addressing limitations in traditional GeoXAI tools, such as the lack of geographical context for model predictions and local explanation, this system integrates state-of-the-art XAI methods with open-source geospatial technologies. Applied to the wildfire-prone regions of Berlin and Brandenburg, Germany, the system combines environmental, topographic, and meteorological features derived from high-resolution geospatial data for training a Random Forest (RF) model. The web-based GeoXAI system enables interactive exploration of the model output and its features, allowing users to visualize wildfire susceptibility, examine feature contributions, and correlate predictions with spatial patterns through post-hoc interpretability. By employing post-hoc explanation methods like SHAP, the system offers clear insights into model predictions by analyzing feature contributions after training, which helps users better understand AI-driven outcomes. Designed with a user-centered approach, the platform promotes trust and usability through transparent predictions, interactive geovisualizations, and local explanations, allowing users to navigate spatial data intuitively by exploring overviews, focusing on specific regions, and accessing detailed insights on demand. This work highlights the potential of combining GeoXAI with machine learning to improve decision-making in wildfire prevention and management.