Machine Learning to predict Downbursts in Japan Based on Total Lightning and Ground Precipitation Data

In this study, using total lightning data from the Japan Total Lightning Network (JTLN) and precipitation data from X-band MP radar, machine learning with a Random Forest model was used to successfully classify the occurrence of downbursts in Japan. TL data associated with the event is collected from JTLN which consists of 11 Earth Networks Total Lightning Sensorsover Japan (data set in 2017, currently 16 stations nationwide) deployed by the UECand jointly operated with Earth Networks. These sensors can detect lightning pulses with a spatial resolution of 500 m. TL parameters such as types of lightning (IC and CG), time of occurrence (UT), location (latitude-longitude), and lightning polarity were collected for each lightning discharge. Ground precipitation data (temporal resolution of 1min, spatial resolution of 250m) from the Ministry of Land, Infrastructure, Transport, and Tourism’s eXtended RAdar Information Network (XRAIN) composed of 26 C-band radars and 39 X-band multiparameter (X-MP) radars are used. Fourteen thunderstorms causing gusty winds and 33 of those without causing gusty winds that occurred in Japan from 2014 to 2017 were analyzed. AITCC (Algorithm for the Identification and Tracking of Convective Cells) was applied to track both precipitation cell and associated lightning discharges. By using Random Forest model, the importance of variables was derived, and it was shown that lightning jump is one of the most important variables for predicting downbursts. This implies that the updrafts inside the clouds are closely related to the occurrence of a significant increase in lightning, followed by a downburst. The prediction accuracy was highest for models that included both lightning and precipitation, followed by lightning-only and precipitation-only models, confirming the importance of data fusion for improving prediction accuracy.