Evaluation of Machine Learning Classifiers for Predicting Deep Convection

Abstract

The realistic representation of convection in atmospheric models is paramount for skillful predictions of hazardous weather as well as climate, yet climate models especially suffer from large uncertainties in the parameterization of clouds and convection. In this work, we examine the use of machine learning (ML) to predict the occurrence of deep convection from a state-of-the-art atmospheric reanalysis (ERA5). Logistic regression, random forests, gradient-boosted decision trees, and deep neural networks were trained with lightning data to predict thunderstorm occurrence (TO) in Central and Northern Europe (2012–2017) and in Sri Lanka (2016–2017). Up to 40 input variables were used, representing, for example, instability, humidity, and inhibition. Feature importances derived for the various models emphasize the high importance of conditional instability for deep convection in Europe, while in Sri Lanka, TO is more strongly regulated by humidity. The Precision-Recall curve indicates more than a twofold improvement in skill over convective available potential energy for short-term (0–45 min) predictions of TO in Europe by using neural networks or gradient-boosted decision tree and a larger improvement in the tropical domain. The diurnal cycle of deep convection is closely reproduced, suggesting that ML could be used to trigger convection in climate models. Finally, a strong relationship was found between area-mean monthly TO and ML predictions, with correlation coefficients exceeding 0.94 in all domains. Convective available potential energy has a similar level of correlation with monthly thunderstorm activity only in Northern Europe. The results encourage the use of reanalyses and ML to study climate trends in convective storms.