Quantifying the Influence of Climate on Storm Activity Using Machine Learning

Overview

This study quantifies the relative contributions of seasonal climatology and synoptic conditions to midlatitude storm activity using 84 years of ERA-5 reanalysis data and convolutional neural networks. The research distinguishes between factors controlling mean storm activity and those governing individual storm properties, and examines how long-term climate trends specifically influence storm characteristics.

Methods and approach

Convolutional neural networks were applied to 84 years of ERA-5 reanalysis data to assess the relative importance of seasonal climatology versus synoptic conditions in controlling storm activity. The models evaluated both aggregated mean storm activity and individual storm properties. A separate analysis isolated the impact of long-term climate trends on individual storm characteristics by examining their contribution to storm-intensity variability and heat anomalies associated with storms.

Results

The models successfully predicted over 90% of the variability in mean storm activity, demonstrating that climatic conditions dominate averaged storm activity. In contrast, only one-third of the variability in individual storm properties is attributed to climatic factors, indicating synoptic conditions account for approximately two-thirds of the variability in individual storm characteristics. Long-term climate trends contribute minimally to storm-intensity variability but contribute over three times more substantially to storms' associated heat anomalies, suggesting that variables directly linked to global warming provide a more direct pathway for detecting climate influence on individual storm properties.

Implications

The differential attribution of climatic versus synoptic control across different metrics indicates a layered dominance structure: seasonal climatology governs the background envelope of storm activity, while synoptic conditions determine specific storm realization and intensity. This distinction has consequences for seasonal predictability, where mean activity may be more predictable from climatic indicators than individual storm events.