May 21 - July 31
Do Forecasters Add Value to Machine Learning Algorithms of Cloud-to-Ground Lightning?
Cara Gregg
What is already known:
- From 2008-2017, there was an average of 28 deaths and 156 injuries caused by cloud-to-ground lightning annually, with 40 deaths in 2016.
- The National Weather Service is currently not required to provide lightning information to the public.
- A probabilistic algorithm was developed by researchers at the National Severe Storms Laboratory to create an automated system that predicts cloud-to-ground lightning.
What this study adds:
- Verification of the automated system’s and forecasters’ abilities to forecast cloud-to-ground lightning.
- Forecasters increase the probability of detection over the automated system through the modification of objects generated by the automated system and by creating their own lightning objects.
- Based on where the forecasters are adding the most value, such as identifying smaller storms and modifying the storm size, severity, and probabilities, the automated system can be improved by addressing these factors.
Abstract:
Cloud-to-ground lightning is an extremely dangerous weather phenomenon resulting in 28 deaths annually over the last decade; currently, there are no requirements for National Weather Service to communicate light- ning dangers or hazards to public. A probabilistic algorithm was developed at the National Severe Storms Laboratory using machine learning to create an automated system that generates objects around areas where it predicts cloud-to-ground (CG) lightning will occur. In spring 2017, nine forecasters from the National Weather Service tested a Probabilistic Hazard Information prototype in the Hazardous Weather Testbed in which they used the guidance of the automated system, modified these objects from the system, and created their own objects to ideally create better forecasts of CG lightning. These forecaster and automated objects were verified and aspects of their performance, such as the probability of detection, were compared to see if the forecasters added value to the automated system. Forecasters added value to the system by adding discussion to the objects and through modifying the size, severity, duration, and probability of the lightning storms. However, forecasters found the task particularly tedious to complete. The areas where the forecasters are adding the most value could be used to improve the automated system’s performance at predicting CG lightning, further reducing forecaster workload.