CaDS Seminar 2022 - Nov. 8
Dr. Randy Chase (School of Computer Science and School of Meteorology University of Oklahoma, USA)
Machine learning estimation of storm updrafts
Abstract:
Diagnosing a storm’s severe potential can be challenging. Many observational methods to diagnose a storm's severe potential are ultimately related to the storm's updraft and are effectively proxies for assessing the storm’s updraft intensity and width. For example: overshooting tops and above anvil cirrus plumes have been linked to severe storm hazards; and Zdr columns have been correlated to tornado formation and intensity. The updraft of a storm could be explicitly retrieved using multi-Doppler analyses, but issues with radar locations (i.e., baselines) and the amount of quality control required usually prevent high quality and timely three dimensional wind retrievals. Thus, retrievals of updrafts are usually unavailable to forecasters to diagnose a storm’s severe potential.
Here a machine learning model, which could be run in an operational environment, is trained to estimate the updraft of a storm from radar data. More specifically, we train a deep learning model, called a UNET, to approximate the maximum vertical velocity from three dimensional gridded radar reflectivity. The deep learning model is trained using output from a convective allowing model (CAM) ensemble, namely the NSSL Warn-on-Forecast-System (WoFs). This investigation will determine if a machine learning model can accurately reproduce the CAM simulated updrafts.