Inference of Wildfire Causes from Their Physical, Biological, Social and Management Attributes

Wildfire disasters have increased markedly in the western US (WUS) in the past several decades. Prevention of human-started wildfires, which account for >60% of all ignitions in WUS, is one of the most effective ways of reducing wildfire risk. Wildfire prevention strategies, however, are specific to ignition type, and an increasing number of wildfires (>50%) have been recorded with a missing/unknown ignition cause in recent years. Here, we leverage the spatial and temporal characteristics of wildfire incidents, captured through a cohort of biological, physical, social and management attributes, to infer the possible ignition source for 150,247 wildfires from 1992-2020 in WUS with a missing/unknown ignition cause. Each ignition cause is associated with specific attributes, for example lightning-started fires generally start at higher elevations and lower vegetation greenness, whereas power-started wildfires mainly occur during dry-hot-windy weather. Trained on wildfires with known causes, our model achieves >70% accuracy in determining specific wildfire causes among a cohort of 12 for test data. If only natural versus human-started wildfires are of concern, our model exceeds 93% overall accuracy. Across WUS, global human modification index, elevation, discovery day of year, fire year, and temperature at the day of ignition were the most important features that helped determine the causes of wildfires. Lower weights assigned to weather is because all wildfires start during “dry-hot enough” weather. Models developed for individual states in WUS show overall accuracies ranging from 60% (California) to 81% (Nevada), with markedly different feature importances.