From logistic regression to deep learning: machine learning modeling of lightning in ERA5 reanalysis data

Abstract. Most lightning parameterization schemes rely on local approaches where the predictors are the atmospheric variables in the same grid cell as the output. To validate the hypothesis that large-scale thunderstorm clusters – such as mesoscale convective systems – are driven by broad spatial predictor patterns, we model lightning occurrence using architectures capable of processing surrounding grid-cell data rather than relying solely on local point-based inputs. This study develops a deep convolutional neural network (U-Net) to model lightning occurrence across Europe using ERA5 reanalysis data. The model is trained on 13 years of data and evaluated with a leave-one-year-out cross-validation strategy. We compare the performance of the U-Net to several local machine learning models of increasing complexity, including logistic regression, generalized additive model, extreme gradient boosting, and multi-layer perceptron. We find that the U-Net outperforms all single grid cell models in overall performance and on the most extreme events. Through a feature importance study, we find that the most important predictors depend on the model type. Finally we show with a spatial sensitivity study that the U-Net captures mesoscale patterns driving lightning occurrence.