Enhancing the Interpretation of AI Models for Natural Hazard Susceptibility: The case of daily wildfire occurrence prediction

Beginning primarily in the 2010s, the use of machine and deep learning algorithms has seen widespread application across various domains of natural hazard susceptibility, including wildfires, floods, and landslides. Especially in the domain of wildfire science, fire occurrence prediction has been one of the most extensively studied areas. A wide range of algorithms and model architectures have been explored to address this problem as a binary classification task, with primary emphasis on evaluating performance using standard classification metrics. However, when developing applications to predict the likelihood of natural hazard occurrences, it is essential to strengthen trust in the model's output by establishing a well-defined and comprehensive susceptibility index and providing interpretable insights into the model's behavior on real-world data. In this study, we propose model-agnostic approaches aimed at (a) generating a binning of the model's prediction distribution to create a probability-based fire occurrence index satisfying essential requirements for hazard susceptibility maps and (b) applying representative sampling to produce local explanations across the dataset, addressing the computational limitations of explainable AI methods when large volumes of local interpretations are required. The objective is to contribute to the development of reliable applications that generate hazard susceptibility information, facilitating the adoption of ML-based solutions by hazard management professionals.