Using Deep Learning to Improve Prediction and Understanding of High-Impact Weather

May 20, 2020

Using Deep Learning to Improve Prediction and Understanding of High-Impact Weather

Ryan Lagerquist, who has studied machine-learning applications in atmospheric science for 8 years, will demonstrate that deep learning can be used to advance both the prediction and understanding of high-impact weather.

Specifically, he will describe the application of convolutional neural networks (CNN), a type of deep-learning method, to high-impact weather. CNNs are specially designed to learn directly from spatial grids, which improves both skill and interpretability.

Lagerquist develops and tests CNNs for two tasks. The first is tornado prediction, where two CNNs predict next-hour tornado occurrence for a given storm, using datasets similar to those used by forecasters in real-time operations. The tornado models achieve an area under the receiver-operating-characteristic curve (AUC) of 0.94 and critical success index (CSI) of ~0.3. This is competitive with a machine-learning model currently used in operations, which suggests that the CNNs would also be suitable for operations. Specialized machine-learning-interpretation methods highlight the importance of a deep reflectivity core and strong mesocyclone, as well as low-level instability and wind shear in the surrounding environment. Also, interpretation methods suggest that a rear-flank downdraft with too much precipitation and negative buoyancy can lead to tornadogenesis failure, which corroborates some previous literature.

The second task is front detection, where a CNN draws warm and cold fronts in reanalysis data. Lagerquist uses the CNN-detected fronts to create a 40-year climatology over North America. On a large scale, fronts are most common in the mid-latitude cyclone track, which migrates poleward from winter to summer, equatorward during El Niño, and poleward during La Niña. Also, the cyclone track appears to be migrating poleward as a consequence of global warming. These results are broadly consistent with the few pre-existing climatologies, although there are some discrepancies that should be investigated in the future.

Webinar access:

Meeting number: 907 998 014

Password: STARSeminar

Join by phone: 415-527-5035