Event-level attribution of wildfire ignitions using a machine-learning framework
This study develops a machine-learning classification framework (MLCF) to attribute the cause of wildfire ignitions as either human- or lightning-induced on a global, event-level scale. The researchers compiled over 1.03 million wildfire records from 17 regional datasets, spanning countries such as the United States, Canada, Australia, and others, and matched them with 16 meteorological, geographic, and socioeconomic predictors. The framework uses an ensemble of seven machine-learning models and is validated through in-sample testing, as well as spatial and temporal out-of-sample experiments. The analysis also includes comparison with conventional empirical lightning-attribution methods and interprets feature contributions using Shapley additive explanations (SHAP).
The main findings show that the MLCF achieved an in-sample overall accuracy of 94.2%, generalizes well across different regions and years, and outperforms conventional lightning data-informed attribution methods. Recent lightning strikes, air temperature, and human activity indicators like nighttime lights and distance to roads are among the most important predictors for ignition cause. The authors recommend that future improvements to global wildfire ignition attribution should focus on enhancing the geographic coverage and quality of ignition-labeled fire records, increasing the resolution and range of input features, and using class-ratio training strategies. The framework can operationally support fire-cause inference and prevention, and underpins the creation of global wildfire datasets with event-level ignition information.