Estimating Mediterranean Cyclone Activity via Explainable Machine Learning

Intense cyclones in the Mediterranean drive most of the region's rainfall and wind-wave extremes, exerting a significant socio-economic impact. Currently there are no established analytical tools for estimating Mediterranean cyclone activity from climatological fields. To bridge this gap, we develop a convolutional neural network (CNN) that, given large-scale climatological fields as input, learns to estimate monthly Mediterranean cyclone activity, which we quantify using the Accumulated Cyclone Energy—a metric integrating the intensity, frequency, and duration of all cyclonic systems in a region. The purpose of this model is to provide a computationally cheap estimate of the expected severity of a cyclonic season given forecasts or projections of key climatological fields, akin to the Genesis Potential Indices used for tropical cyclones. Model performance is optimized through nested cross-validation combined with Bayesian Optimization for both hyperparameter tuning and input variable selection. The resulting CNN accurately captures both peak cyclone activity (mean squared error = 0.01) and the overall temporal evolution (r = 0.83 monthly; r = 0.63 yearly). Input variable selection highlights absolute vorticity as the most informative input variable, followed by relative humidity and mean sea-level pressure. Furthermore, we visualize the CNN's internal activations and find that, when analyzing the input variables, the model focuses on regions historically associated with cyclogenesis, indicating that the model has captured at least some aspects of the underlying physical processes. This study represents a significant step toward leveraging explainable artificial intelligence for studying Mediterranean cyclones, offering new insights into the drivers of extreme weather in the region.