Abstract
Wildfire has increasingly been recognised as a coupled environmental and social hazard rather than a purely ecological disturbance. In this study, a multiclass classification task with a four-class target structure was formulated, in order to predict broad wildfire cause categories in the United States using historical wildfire occurrence records. After several supervised learning models were evaluated, it was found that tree-based models substantially outperformed linear baselines in the reported experiments, with random forest achieving the strongest mean cross-validated accuracy. At the same time, the results were found to raise a more difficult methodological question: whether cause was being predicted prospectively, or whether it was being inferred partly from post-ignition attributes already embedded in the administrative record. The principal contribution of the study lies not only in obtaining moderate predictive performance, but in exposing the distinction between operationally useful inference and genuinely prospective wildfire-risk prediction.