Deep Learning for Identifying High-Impact Dry Intrusions: European Storm Case Study

Oral Presentation 

Dry intrusions are coherent airstreams that originate in the upper troposphere and advect relatively dry, cool, and sometimes high potential vorticity air towards the boundary layer. Here, they are associated with atmospheric fronts and extratropical cyclones, where their co-occurrence can give rise to extreme precipitation and severe wind gusts. This was observed in October 2014 when a dry intrusion descended slantwise from southern Greenland: moving equator-ward over the United Kingdom, central Europe, and spreading over the Mediterranean and North Africa. Along its path, severe weather was reported from multiple regions including accounts of extreme wind gusts, infrastructural damage, and downstream flash flooding in local areas.Traditionally, these airstreams are identified with Lagrangian trajectory analysis, but it is data intensive and computationally expensive. As an alternative, this research applied computer vision methods to ERA5 data to identify low-level dry intrusion outflows from atmospheric fields, such as temperature and relative humidity, using convolutional neural networks. The success of this deep learning model was then evaluated on the European case study in comparison to target data. It was found that the neural network could accurately capture the size, shape, and trajectory of the dry intrusion outflow, with an approach that was cheaper to run and required significantly less data than trajectory analysis. As a result, this work could allow investigations into the potential future of dry intrusions by examining their representation in climate model data. This research is otherwise inaccessible with current methods due to its demands for high resolution spatial and temporal data. Therefore, the application of machine learning in this field could expand our knowledge on the future of high-impact weather in the mid-latitudes.