Developing a Hail Probability Product for the Probabilistic Hazards Information Framework

A major effort underway at CIMMS and NSSL is the development of probabilistic guidance for hazards associated with thunderstorms: lightning, hail, damaging winds, and tornadoes. This effort is generally known as Probabilistic Hazards Information (PHI) and falls under the Forecasting A Continuum of Environmental Threats (FACETs) paradigm. The end-goal for PHI is to use a large historical database of Multi-Radar, Multi-Sensor (MRMS) data called the Multi-Year Reanalysis of Remotely Sensed Storms (MYRORSS) to develop well calibrated probabilities for thunderstorm hazards. This presentation will focus on the hail probability development.

The Severe Hazards Analysis and Verification Experiment (SHAVE) collected high-resolution reports of hail, including reports of ‘no hail’ and non-severe hail sizes. This comprehensiveness of the reports are ideal for developing hail-related probabilities. A database of volume-by-volume analyses of single-radar, WSR-88D data is currently being constructed. Using the manual identification of storm location and motion, difficulties associated with automated identification and tracking can be avoided and allow for easier development of the probabilistic guidance. SHAVE reports will be paired to storm objects by using the storm’s current location and motion to construct a cone that extends down the motion vector for 30-minutes. The sides of the cone are constructed first from a 5-km radius half-circle and a ±22.5° deflection from the storm motion direction. All SHAVE reports within the cone are considered and the time it would take the storm to reach the SHAVE report location is calculated. MRMS data are paired to the storm using a 5-km radius search at the storm’s location and different summary statistics of MRMS and near-storm environment variables are calculated. The final database is a collection of storm identifications with location, motion, and radar-derived attributes, paired to SHAVE reports with “time-to-storm” information attached. Thus, information such as “maximum hail size within the next 15 minutes” can be easily derived.

This presentation will focus on the construction of the database, which lead time binning strategy: general (e.g., hail size X within 15 minutes) or specific (e.g., hail size X in 15 minutes), results in better probabilities, and efforts to group storms producing different hail sizes together using unsupervised learning. Discussions on how to develop the spatial probabilities using fully convolutional neural networks, applications of the findings to MYRORSS, and using MRMS-based verification to expand the database for big data and deep learning applications will be included.