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

• Published in: Earth's future 2025-01, Vol.13 (1), p.n/a
• Main Authors: Pourmohamad, Yavar, Abatzoglou, John T., Fleishman, Erica, Short, Karen C., Shuman, Jacquelyn, AghaKouchak, Amir, Williamson, Matthew, Seydi, Seyd Teymoor, Sadegh, Mojtaba
• Format: Article
• Language: English
• Subjects: Arson; Biological effects; Climate change; Deep learning; Fire prevention; Forest & brush fires; Ignition; Jurisdiction; Machine learning; Prevention; risk mitigation; Social discrimination learning; Topography; Vegetation; wildfire; wildfire attributes; wildfire prevention; Wildfires
• ISSN: 2328-4277; 2328-4277
• DOI: 10.1029/2024EF005187

Effective wildfire prevention includes actions to deliberately target different wildfire causes. However, the cause of an increasing number of wildfires is unknown, hindering targeted prevention efforts. We developed a machine learning model of wildfire ignition cause across the western United States on the basis of physical, biological, social, and management attributes associated with wildfires. Trained on wildfires from 1992 to 2020 with 12 known causes, the overall accuracy of our model exceeded 70% when applied to out‐of‐sample test data. Our model more accurately separated wildfires ignited by natural versus human causes (93% accuracy), and discriminated among the 11 classes of human‐ignited wildfires with 55% accuracy. Our model attributed the greatest percentage of 150,247 wildfires from 1992 to 2020 for which the ignition source was unknown to equipment and vehicle use (21%), lightning (20%), and arson and incendiarism (18%). Plain Language Summary Prevention of human‐caused wildfires, which account for >60% of ignitions across western United States (WUS), is an effective way of reducing wildfire risk. However, the reported cause of an increasing number of wildfires (>50% in recent years) is unknown, challenging the development of targeted prevention strategies. Here, we leverage the spatial and temporal characteristics of wildfires to infer the cause of 150,247 wildfires from 1992 to 2020 in WUS with an unknown cause. Each ignition cause was associated with distinct attributes. For example, lightning‐ignited fires generally start at higher elevations and lower vegetation greenness, whereas wildfires started by power infrastructure mainly ignite during dry, hot, windy weather. Model accuracy in assigning an ignition cause to each wildfire among the 12 available classes was >70% when applied to test data. Across WUS, global human modification index, elevation, discovery day of year, fire year, and temperature on the day of ignition made the greatest contributions to determining the causes of wildfires. The contribution of weather was lower because wildfires start during a subset of weather conditions that are generally common between wildfires. The overall accuracies of models developed for individual states ranged from 60% (California) to 81% (Nevada), and the most influential attributes varied among states. Key Points The spatial and temporal distribution of wildfires is captured by associated biological, physical, social, and management attribut