Amanda Burke - March 6

Name:     Amanda Burke Title:    Improving Machine Learning-Based Probabilistic Hail Forecasts through Monthly Weighting Location: NWC 5600 Date:     2020/03/6 Time:     3:00 PM Series:   Convective Meteorology (Mesoscale Dynamics) Abstract: Forecasting severe hail using machine learning (ML) models during the spring and summer season over the contiguous United States (CONUS) has shown improved skill

Start

March 6, 2020 - 3:00 pm

End

March 6, 2020 - 4:00 pm

Address

120 David L Boren Blvd, Norman, OK 73072   View map
Name:     Amanda Burke
Title:    Improving Machine Learning-Based Probabilistic Hail Forecasts through Monthly Weighting
Location: NWC 5600
Date:     2020/03/6
Time:     3:00 PM
Series:   Convective Meteorology (Mesoscale Dynamics)
Abstract: Forecasting severe hail using machine learning (ML) models during the spring and summer season over the contiguous United States (CONUS) has shown improved skill compared to traditional techniques. However, annual hail fall can vary spatially and temporally based on different localized environments. Leveraging the computational efficiency of ML models on specific temporal scales could further improve hail forecasting skill for various local environments, and better allow the public to take action and reduce risk.
The ML-based method uses random forests to predict the probability of severe hail, based on inputs from the High-Resolution Ensemble Forecast system version 2 (HREFv2) and observations from maximum expected size of hail (MESH) data.  Rather than limiting the training dataset, the random forests are trained on all identified CONUS storms and weighted around a specific time frame. The environments occurring during these time periods are then considered more important when forecasting, while still maintaining information about general hail formation.  The weighted and non-weighted severe hail forecasts are compared during the desired time frame, to determine if temporal weighting results in superior forecasting performance. The weighted ML forecasts initially outperform the non-weighted forecasts during June, July, and August.  However, the non-weighted and weighted models produce objectively similar forecasts in May.