Using Machine Learning for Nowcasting Severe Weather Hazards

Despite improvements in computational power, nowcasting remains a challenging problem, particularly as it pertains to short-term forecasts of severe weather hazards. This work is part of a larger project that transitions research to forecasting applications and focuses on the development of statistical and machine-learning based nowcasting techniques across CONUS. The results of the project will provide experimental guidance about storm longevity and classification to forecasters generating probabilistic forecasts for severe storms in the Hazardous Weather Testbed (HWT). This will be accomplished by incorporating the results into the prototype Probabilistic Hazard Information (PHI) tool for subsequent testing and evaluation in the HWT.

For our current work, we analyze and track one year of storm objects identified from the Multi-Radar Multi-Sensor (MRMS) system at five-minute intervals across CONUS. The data utilized by this project are a subset of the Multi-Year Reanalysis of Remotely Sensed Storms (MYRORSS) project.

Machine learning is used to predict storm longevity as a function of the current storm observations and the near storm environment. We compare the results of several machine learning methods. Additionally, we plan to explore how methods of segmenting storm objects into identifiable groups, such as storm classifications, can affect the performance of these machine learning methods.