Proposal of Hourly Rainfall Estimation Method by CNN Using Meteorological Satellite Himawari and Its Evaluation of Areal Rainfall in a Watershed Scale

 Many regions, including developing countries, have limited meteorological observation networks and still lack quantitative rainfall data with basin-scale accuracy that can contribute to water-related disaster prediction.

 This study aims to develop satellite precipitation products with quantitative accuracy in basin-averaged precipitation for water-related disaster forecasting. In recent years, deep learning has been utilized in many fields as an experiential statistical model, and CNN is a useful model for estimating precipitation from meteorological satellites. The purpose of this study is developing a satellite precipitation estimation method that can be used for predicting water-related disasters by using CNN and the brightness temperature of clouds and water vapor from the Himawari meteorological satellite.

 The data used were precisely geometrically corrected data from the Himawari meteorological satellite and elevation data from MERIT DEM. The training period was four months during the summer of 2015 through 2021 (July through October), and the validation period was the summer of 2022. The training domain was the northeastern part of Japan, and the validation watersheds were the Arakawa River in the Kanto region (within the training domain) and the Chikugo River in the Kyushu region (outside the training domain). As a result, this study was able to reproduce the basin-averaged precipitation quantitatively with high accuracy within the training domain. Outside of the training domain, precipitation of rainfall events could be reproduced qualitatively and generally, and some rainfall cases were more accurate than GSMaP's accuracy, however there were cases where no rainfall events were misclassified as rainfall events, therefore we still have room of improvement.