Japan Geoscience Union Meeting 2021

Presentation information

[E] Oral

M (Multidisciplinary and Interdisciplinary) » M-IS Intersection

[M-IS07] Effects of lightning, severe weather and tropical storms

Sat. Jun 5, 2021 10:45 AM - 12:15 PM Ch.03 (Zoom Room 03)

convener:Mitsuteru Sato(Department of Cosmoscience, Hokkaido University), Hisayuki Kubota(Hokkaido University), C. Glenn Vincent Lopez(---), Purwadi Purwadi(Department of Cosmosciences, Hokkaido University, Sapporo 0600810, Japan), Chairperson:Mitsuteru Sato(Department of Cosmoscience, Hokkaido University), Hisayuki Kubota(Hokkaido University)

11:15 AM - 11:30 AM

[MIS07-03] Machine Learning Prediction of Precipitation in Metro Manila, Philippines

★Invited Papers

*Akira Noda1, Yukihiro Takahashi2, Hisayuki Kubota2, Ken-ichi Fukui3, Mitsuteru Sato2 (1.Department of Earth and Planetary Sciences, School of Science, Hokkaido University., 2.Department of Cosmosciences, Graduate School of Science, Hokkaido University, 3.The Institute of Scientific and Industrial Research, Osaka University)

Keywords:torrential rainfalls, machine learning, mesoscale meteorology

It is difficult to accurately predict the occurrence and rain volume of torrential rains such as guerrilla rain, rain band with typhoon and linear precipitation zones even in Japan, where meteorological observations from the ground and space and weather forecasts using numerical models are well established. One of the reasons for this is that the spatial narrowness of the rainfall area and the rapid development of these extreme weather events exceed the temporal and spatial resolution of conventional weather observation networks. Furthermore, in Southeast Asia, where the meteorological observation infrastructure is relatively weak, many disasters such as heavy rains and floods caused by monsoons and typhoons occur every year. There is an urgent need to establish the cost-effective weather forecasting technology. In recent years, the development of machine learning technology has been remarkable owing to the increased processing speed of computers and big data. In addition, in the field of meteorology, the researches on the forecasting method development using machine learning have been actively conducted. Our research group has developed an automatic weather and lightning observation system called as P-POTEKA since 2017, and has been deploying it in Metro Manila, Philippines, which is frequently affected by heavy rains and associated flooding. To date, 35 P-POTEKA units have been installed in Metro Manila and continue the acquisition of meteorological data (precipitation, temperature, pressure, humidity, wind speed, wind direction and solar irradiance every minutes). While AMeDAS (The Automated Meteorological Data Acquisition System) in Japan is deployed with the average interval of 17km, P-POTEKA in Metro Manila is deployed with the average interval of 2-3km, making the observation network with the world’s highest spatial and temporal resolution suitable for capturing extreme weather events. Using the P-POTEKA rainfall data, RGB image data corresponding to the spatial distribution of rainfall in Metro Manila was created by interpolating the data using the ordinary kriging method. By training these time-series rainfall image data on a machine learning model called ConvLSTM (Convolutional Long-Short Term Memory), we predicted the distribution and rainfall amount from the present to 1-hour later with the 10-minutes intervals using the observation data of the past 1-hour. The root mean squared error (RMSE) of the RGB values of pixels corresponding to the predicted and actual observed precipitation distributions were calculated for untrained rainfalls (40 patterns) to evaluate the performance of our machine learning method. As a result, it was found that the prediction using ConvLSTM is relatively accurate up to 30 minutes from the present, but the prediction accuracy of the spatio-temporal change of prediction becomes significantly worse after 40-60 minutes from the present. In this presentation, the details of the machine learning method model using ConvLSTM and the prediction results of precipitation and precipitation distribution predictions will be explained more in detail.