A1178
Title: Model-agnostic feature selection for high-dimensional bushfire Data
Authors: Houying Zhu - Macquarie University (Australia) [presenting]
Abstract: Bushfires are a frequent and destructive natural hazard in Australia, posing serious threats to communities, ecosystems, and the economy. Modeling bushfire risk and associated losses are particularly challenging due to the integration of diverse and high-dimensional data sources, including satellite imagery, weather sensor data, digital terrain models, historical fire records, loss statistics, and infrastructure information. Although rich bushfire-related data are available, building reliable and interpretable models that effectively synthesize these heterogeneous inputs remains a significant methodological hurdle. Ensemble learning methods such as random forests can improve predictive performance but often function as black boxes, limiting transparency around the contribution of individual variables. This lack of interpretability is especially problematic in disaster risk contexts, where transparent and explainable models are essential for informed decision-making. The aim is to present a model-agnostic feature selection framework based on the Shapley value to identify and explain key drivers of bushfire risk. This approach enables a clearer interpretation of complex models by quantifying the marginal contribution of each feature to the prediction. By enhancing both accuracy and transparency, the method supports more informed and accountable decision-making in bushfire risk management.
