Using machine learning to advance next-day probabilistic convective hazard prediction with convection-allowing models

Convection-allowing models (CAMs) provide forecasters with potentially valuable depictions of convective mode and intensity. However, present-day CAM diagnostics, including widely-used updraft helicity (UH) fields, do not sufficiently discriminate between convective morphologies (e.g., supercells vs. mesoscale convective systems). Thus, using CAMs to identify the range of possible modes within a given forecast period is typically accomplished by subjectively interrogating CAM output, such as simulated reflectivity. Although objective automated approaches to identify distinct convective modes have been successfully applied to observed storms, they have not been applied to simulated storms in CAM output. Yet, as the size and update-frequency of CAM ensembles increases, the need for objective techniques to quickly synthesize forecasts of mode from CAMs will become paramount.

Here, we report on initial results to develop an objective convective mode identification system for use with CAM ensembles. The mode identification system will be tested with multiple state-of-the-art machine learning (ML) approaches of various complexity, including random forests and convolutional neural networks, that are well-suited to image classification. The ML systems will also be compared to manually tuned size and orientation thresholds applied to the reflectivity fields to reveal the added value of more complex algorithms.

The classifications from the identification system will be used to improve existing severe hazard guidance (e.g., using surrogate diagnostics) and generate novel probabilistic guidance related to convective mode, such as timing of mode transitions and most likely convective mode products. The utility of the mode identification system and derived guidance will be evaluated in the NOAA Hazardous Weather Testbed (HWT) between 2020–2022; examples of planned comparisons using forecasts from the 2019 HWT will be presented.