Development of a new satellite rainfall product HiDRED (Himawari Data Rainfall Estimation using Deep learning) and a fundamental study on its applicability to hydrological models

In many regions, including Southeast Asia, meteorological observation networks remain underdeveloped. While existing satellite rainfall products demonstrate a certain level of accuracy at the macroscale, their accuracy at the watershed scale remains insufficient. This study aims to propose an algorithm that applies deep learning to IR data obtained from Himawari meteorological satellite observations to estimate rainfall with quantitative accuracy at the watershed scale, contributing to predictions of water-related disasters.

The objective of this research is to optimize a deep learning model using meteorological observation data available in abundance in Japan and subsequently apply it to Southeast Asia. The input data consists of IR images from multiple wavelength bands provided by the geostationary meteorological satellites Himawari-8 and 9, as well as elevation data.

The estimated rainfall in the Japanese region, where parameter optimization did not conduct, was evaluated across various watershed scales. As a result, the model outperformed GSMaP in watersheds with areas ranging from approximately 100 km² to 3000 km². Additionally, in tributary watersheds with areas under 100 km², the model was able to qualitatively replicate observed rainfall.