Exploring Severe Weather Nowcasting Predictions Using Atmospheric Observations and Deep Learning

Machine learning applications now abound in the atmospheric sciences on all predictive scales. In the realm of nowcasting, tree-based and computer vision methods perform well for a variety of problems. For example, the experimental ProbSevere version 3 uses a tree-based method to make severe weather forecasts in a storm-object-tracking framework. National Weather Service forecasters at the Hazardous Weather Testbed have strongly indicated a preference for the tree-based ProbSevere v3 versus the operational ProbSevere v2. This research will further explore probabilistic predictions for severe weather hazards (hail, wind, tornado) using computer-vision deep-learning methods (e.g., convolutional neural networks) with radar, satellite, and lightning observations, and short-range NWP output. Several benefits of this approach include: 1) providing more explicit predictions of “when and where” a hazard may occur, 2) more efficient learning of salient spatial features in the data, and 3) avoiding issues inherent in object identification and tracking (e.g., splitting/merging objects). While there are still challenges, preliminary results show promise. We demonstrate that models can learn features that are difficult to extract in an object-based framework, and that models can learn motion from a single timestamp of data.