Development of a Machine Learning-Based Tornado Detection Algorithm for the WSR-88D Network

A New Tornado Detection Algorithm (NTDA) is currently under development and is proposed to replace the existing NSSL Tornado Detection Algorithm (TDA) at WSR-88D sites nationwide. The goal is to improve the ability of NEXRAD radars to automatically identify and track tornadic vortex signatures while reducing the number of false detections. Work toward a prototype NTDA has included development of a random forest machine-learning technique to predict tornado occurrence in real-time. Initial rotation objects are identified using a combination of velocity-based fields, including NSSL’s single-radar azimuthal shear, and are then classified as tornadic or non-tornadic by the random forest. The random forest was developed using a multi-year dataset of both tornadic and non-tornadic storms and analyzes a number of radar-derived object attributes, including divergent shear, shear diameter, gate-to-gate velocity shear, and others. This presentation summarizes the scientific approach to tornado detection used in the NTDA, an initial summary of its performance, and preliminary plans for its operational evaluation and implementation. The NTDA is scheduled for testing within the 2019 Spring Experiment of the Hazardous Weather Testbed.