Interpretable artificial intelligence models for predicting lightning prone to inducing forest fires

Abstract

Specific types and intensities of lightning are significant causes of forest lightning fires. Analyzing the relationship between these lightning events and the climatic conditions that favor their occurrence is crucial for predicting and preventing forest lightning fires. However, there is a lack of research in this area for the Greater Khingan Range in Northeast China. This study utilized data from three lightning location networks and the ERA5 meteorological dataset to analyze the historical climate and lightning data in the Greater Khingan Mountains region of China from 2021 to 2023, focusing on the impact of various climatic factors on the density of target lightning—lightning that is prone to cause forest lightning fires. Four machine learning models—SVM, RF, XGBoost, and LightGBM—were evaluated, with RF demonstrating the best predictive performance, achieving R² of 0.83, MAE of 1.91, and MSE of 14.90. Additionally, the prediction results of the RF model were evaluated using the Kruskal-Wallis test to determine if the results are statistically significant. Using SHAP values to interpret the model, it was found that the K-index (kx) and Convective Available Potential Energy (CAPE) are the most significant predictors of target lightning density, followed by the leaf area index for high vegetation (lai_kv), surface pressure (sp), cloud base height (cbh), temperature at 2 m (t2m), and coverage of high vegetation (cvh). Approximately 70% of the total average absolute SHAP values are attributed to kx and CAPE, highlighting their crucial role in the prediction process. This study provides insights into the environmental factors influencing lightning frequency and emphasizes the importance of interpretable machine learning models in predicting future lightning occurrences and forest lightning fires. Visualization tools, including SHAP summary plots and force plots, were used to provide a detailed illustration of each feature's contribution to the model predictions.