Automated methods for the detection and creation of frontal systems in atmospheric data
Date issued
Authors
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
License
Abstract
Synoptic scale weather fronts are ubiquitous elements of extra-tropical weather. Their connection to severe weather conditions such as extreme precipitation, cyclones or thunderstorms highlights
the importance of being able to reliably detect these structures to conduct analyses surrounding their effect within the atmosphere. Nonetheless, it is still common to draw fronts via manual analyses by weather services, inevitably introducing biases or being restricted to the corresponding weather service’s analysis region. Common numerical, objective methods on the other hand have been developed for a lower resolution than the currently available datasets. While adjustments to make these algorithms applicable to high resolution datasets exist, fine details may be lost in the process. Furthermore, these traditional methods all only provide frontal information at a single pressure or model level, reducing them to a line in a 2D slice, neglecting their overall three-dimensional structure. Implications surrounding the three-dimensional shape are therefore still scarce, as objective methods for the creation of their three-dimensional structure are very limited. Due to their connection to severe weather conditions, which can lead to serious damages, it is important to provide methods that are able to both localize these fronts on current datasets and being able to recreate their three-dimensional structure. In this work, we present a convolutional neural network, which is capable of locating and classifying atmospheric fronts on the current ERA5 dataset. Our method outperforms a common numerical approach when detecting fronts within both the North American continent and Western Europe. These methods are evaluated with respect to manual front analyses of two weather services, however, our model may also be applied globally, where we could show a high correlation of our detected fronts and extreme precipitation. Additionally, we propose a novel algorithm for the creation of three-dimensional fronts outside of case studies, being - to the best of our knowledge - the first method to allow for statistical evaluation of frontal shape in three dimensions. We further provide an improved implementation of this algorithm which utilizes modern graphics processing units (GPU) to drastically reduce compute time. We show that our implementation can create three-dimensional fronts for our dataset region within the Northern Atlantic in less than 500 seconds, which is vastly faster than our version, that only utilizes central processing units (CPU). The latter took several hours on multiple nodes. Further, we show that both of our implementations exhibit structural features commonly associated with fronts, such as maximum baroclinity and steep temperature gradients. Similar to our deep learning model, we are able to provide information for all commonly used types of fronts: warm, cold, occluded and stationary fronts.