[A-CG38] Satellite Earth Environment Observation

[ACG38-08] Development of a precipitation estimation product with machine learning specializing in the linear rainband in summer Japan

*Hitoshi Hirose¹, Takuji Kubota¹, Koichi Toyoshima², Atsushi Higuchi³ (1.Japan Aerospace Exploration Agency, 2.The University of Tokyo, 3.CEReS, Chiba University)

Keywords:Precipitation, Geostationary meteorological satellite, machine learning

Geostationary meteorological satellites (GEOs) are essential for frequent global precipitation estimation by satellites. Most of the traditional precipitation estimation methods using GEOs have used only single infrared (IR) band with a wavelength of 10.4 μm. However, it has been reported that precipitation estimation relying on cloud top temperature information obtained from a single IR band underestimates heavy rainfall from relatively shallow clouds. With the launch of the latest GEOs, more and more IR-band are now available over a wider area. Therefore, we developed Himawari-8 precipitation estimation algorithm (HPA; Hirose et al. 2019) by applying the Random Forest machine learning method to nine IR bands of Himawari-8 and precipitation observation of GPM KuPR. The accuracy of the HPA was verified using radar-AMeDAS as truth, and it was shown that the three water vapor bands (6.2, 6.9, 7.3 μm) of Himawari-8 were particularly effective in estimating linear precipitation bands. In addition, we are developing an HPA specialized for estimating linear rainbands by collecting training data only from the near Japan region in summer. The case study for the heavy rainfall in 2020-2021 showed that the HPA was able to qualitatively estimate the linear rainbands in the Kyushu region well. Since the HPA tended to underestimate the maximum precipitation intensity, we corrected the estimated precipitation intensity using the histogram matching method and verified how effective the statistical correction method is in underestimating the extreme precipitation.