Event‐Level Attribution of Wildfire Ignition Cause: A Machine‐Learning Approach

Wildfire ignition causes, namely human or lightning, largely determine fire behavior and impacts, but are sparsely reported worldwide. Event‐level fire causes have been inferred regionally using high‐accuracy lightning observations, but global generalization remains limited. Here we develop a machine‐learning classification framework to attribute individual wildfire ignitions using 16 predictors covering meteorological, terrestrial, and socioeconomic states. A soft‐voting ensemble built from seven base learners is trained and tested using 1.03 million fire records compiled from 17 regional data sets. The MLCF achieves an in‐sample accuracy of 94.2% while maintaining reasonable generalization skills across both spatial and temporal out‐of‐sample evaluations. Feature importance analysis identifies recent lightning strikes, air temperature, and human footprint as leading and interacting predictors of fire cause. Overall, this framework provides a basis for global event‐level wildfire ignition attribution that could inform process‐level understanding of the drivers of wildfire ignition.