Japan Geoscience Union Meeting 2021

Presentation information

[J] Poster

A (Atmospheric and Hydrospheric Sciences ) » A-CG Complex & General

[A-CG43] Earth & Environmental Sciences and Artificial Intelligence/Machine Learning

Thu. Jun 3, 2021 5:15 PM - 6:30 PM Ch.03

convener:Tomohiko Tomita(Faculty of Advanced Science and Technology, Kumamoto University), Shigeki Hosoda(Japan Marine-Earth Science and Technology), Ken-ichi Fukui(Osaka University), Satoshi Ono(Kagoshima Univeristy)

5:15 PM - 6:30 PM

[ACG43-P03] Hindcast of the July 2018 torrential rainfall event in Kagoshima City using LSTM

*Genki Shirasawa1, Ryusei Hamasaki1, Shinichiro Kako2, Hirohiko Nakamura3 (1.Kagoshima University, Faculty of Engineering, 2.Graduate School of Science and Engineering, Kagoshima University, 3.Kagoshima University, Faculty of Fisheries)


Keywords:LSTM, Torrential rain, Radar AMeDAS, AMeDAS, Hindcast/Forecast of rainfall event

In this study, we constructed a long short-term memory (LSTM) model to hindcast the torrential rainfall event that occurred in Kagoshima City in July 2018. In this model, automated meteorological data acquisition system (AMeDAS) and radar–AMeDAS rainfall data from Kyushu were used as training data. Our LSTM model could hindcast the time variation of the rainfall event; however, it tended to underestimate the rainfall event observed by AMeDAS, and in some cases the start times of the precipitation events were late. Underestimation and phase lags were probably because the LSTM model was optimized for light rainfall events that occur frequently. To further improve the prediction accuracy of our model, it was necessary to improve the reproducibility of the magnitude of the precipitation amplitude. Therefore, we selected further training data such as physical quantities over the East China Sea that contribute to the forecast/hindcast of torrential rainfall events. The use of radar–AMeDAS rainfall data as training data over the ocean in the southern part of Kagoshima significantly improved the accuracy of our LSTM model compared to that without this additional training data.