Lightning-ignited wildfire prediction in the boreal forest of northeast China

Abstract

Lightning-ignited fires are the leading fire type in boreal forests, where early warning systems are essential for effective fire suppression and loss reduction. However, the prediction of lightning ignitions and the identification of contributing factors have not been thoroughly investigated in the boreal forest of northeast China, a region that experienced the most frequent lightning fires and the largest burned areas in the country. This study develops a prediction model using the eXtreme Gradient Boosting (XGBoost) algorithm. The model integrates the cases of igniting and non-igniting lightning, along with datasets of weather, soil, topography, vegetation, and lightning in 2019–2023. An optimized repeated random undersampling method was implemented to address the imbalanced population of the three cases. The most accurate classifier (MAC) was obtained from training 1000 XGBoost classifiers, which achieves a prediction accuracy of 88.7 %. The MAC performance remains robust when tested on individual lightning fire days and within the entire study period, indicating its reliablity for lightning ignition nowcasting. Using the Shapley Additive exPlanations (SHAP) framework, we quantified the relative contributions of wildfire variables and their marginal effects on the lightning ignition. Results indicate that low surface soil moisture (an indicator of fuel dryness) and low lightning density (associated with little precipitation) are the dominant factors for lightning ignition. Overall, the MAC significantly outperforms traditional fire danger rating indices, suggesting that weather conditions alone are inadequate for lightning ignition prediction, and the effects of surface soil moisture and lightning activity should be considered.