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

<jats:p>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.</jats:p>