High Impact Weather in the Mid‐Latitudes: A Neural Network Approach to Identifying North Atlantic Dry Intrusion Outflows

Dry intrusions are coherent airstreams that originate from the upper troposphere, or lower stratosphere, and descend towards the surface where they can influence the dynamics of mid‐latitude weather. Notably, the occurrence of these airflows with atmospheric fronts and extratropical cyclones can exacerbate their impacts, leading to increased precipitation and stronger surface winds. Therefore, it is of interest to understand how dry intrusions may respond to our changing climate. Traditional identification methods with Lagrangian trajectory analysis, however, cannot always be applied to climate projection data due to the computational cost of this approach and a lack of necessary available data from climate models. This research explores the alternative application of image segmentation concepts to build a machine learning classification model for dry intrusion outflows in the North Atlantic. A U‐Net convolutional neural network (CNN) is trained to predict the presence of dry intrusion objects from ERA5 atmospheric data, including temperature and relative humidity. With a catalogue of labelled dry intrusion objects calculated from trajectory analysis as predictands, the ability of this CNN to identify individual dry intrusion footprints, capture their size and shape, and recreate long‐term climatologies is evaluated with Matthew's correlation coefficient and intersection over union. Compared with a multiple logistic regression model, the CNN outperforms across all metrics and compares more favourably with the target data. However, the CNN struggles to predict small dry intrusion signatures and limitations are encountered outside of the spatial training domain in a high‐impact case study. Despite this, these results provide proof‐of‐concept for an alternative way to identify dry intrusion outflows that uses less data, is fast and easy to implement, and could be utilised to study the possible futures of dry intrusions and extreme mid‐latitude weather.