Regional Lightning Occurrence Probability Forecasting and Risk Identification Based on Resampling Ensemble Machine Learning

Abstract

Accurate regional lightning-occurrence prediction is important for operational weather-risk management, but its development is challenged by the severe class imbalance of grid-hour lightning samples. This study proposes a repeated random undersampling (RUS) stacking ensemble that combines heterogeneous machine-learning models and produces probabilistic lightning-occurrence forecasts using atmospheric variables from the European Centre for Medium-Range Weather Forecasts (ECMWF) fifth-generation reanalysis (ERA5), together with spatial and temporal predictors. The effects of the undersampling ratio, ensemble size, and positive-class weighting were systematically evaluated, and a configuration with a RUS ratio of 1:15, 20 ensemble members, and a positive-class weight of 2 was selected. Using an operational threshold of 0.591 selected exclusively on the 2024 validation set, the final model achieved an area under the receiver operating characteristic curve (ROC-AUC) of 0.945, an area under the precision–recall curve (PR-AUC) of 0.235, a probability of detection (POD) of 0.521, and an F1-score of 0.302 on the independent 2025 test set. Physical-variable-group ablation and aggregated TreeSHAP (tree-based Shapley additive explanations) analyses were further conducted to interpret the predictions. Removing the spatiotemporal predictors produced the largest reduction in performance, followed by removing cloud and microphysical variables. Both the Light Gradient Boosting Machine (LightGBM) and categorical boosting (CatBoost) models consistently identified longitude, latitude, total-column cloud ice water, and convective available potential energy as the leading predictors, while higher cloud-ice-water content and stronger convective-instability indices generally shifted model outputs towards lightning occurrence. Direct transfer of the ERA5-trained ensemble to corresponding ECMWF forecast fields without retraining retained useful predictive skill and substantially outperformed the operational ECMWF lightning product over the collocated evaluation samples, although performance degradation indicated a cross-dataset distribution shift. These results demonstrate the value of combining imbalance-aware ensemble learning with physically interpretable predictors for regional lightning-risk forecasting, while the strong influence of geographic variables indicates that external validation and local recalibration or retraining are required before application to other regions.