Author(s): Narongrit Luangdilok; Watin Thanathanpon; Ticha Lolupiman; Kachapond Chettanawanit; Piyamarn Sisomphon; Sathit Chantip; Apimook Mooktaree; Theerapol Charoensuk

Linked Author(s): Narongrit Luangdilok, Watin Thanathanphon

Keywords: Rainfall; IMERG; GPM; Bias-correction; Thailand

Abstract: A bias correction is commonly used to reduce differences between satellite-based and ground-based precipitation data. This study aimed to develop an operational bias correction system in Thailand with automatic rain gauges for daily satellite-based precipitation data from the Integrated Multi-satellitE Retrievals for GPM (IMERG) products. To find a suitable approach for the near real-time system, bias correction methods were compared. As representatives of modern machine-learning techniques, the Random Forest (RF) algorithm and the Convolutional Neural Network Autoencoder (AE) algorithm were chosen. The comparison also included the Linear Scaling (LS) and Quantile Mapping (QM) methods to see how much the machine learning approach can improve the performance of traditional statistical techniques. Due to the availability of satellite data, the analysis in this study only used rain gauge and IMERG precipitation data from 2012 to 2020. The results show that the RF approach is the best choice for the operational purpose, outperforming other techniques in terms of performance and ease of implementation. In more detail, Quantile Mapping and LS perform similarly in overall results (10.9-12.2 mm/day of RMSE) but significantly worse than RF and AE methods (8.1-8.2 mm/day of RMSE). Furthermore, all bias-corrected rainfall datasets were fed into the Nedbor Afstrmnings Model (NAM), a hydrological model, to simulate rainfall runoff during heavy rain events. As a result, the possibility of implementing these bias-corrected satellite-based rainfall estimates into an operational system that predicts rainfall runoff is presented.

DOI: https://doi.org/10.3850/978-90-833476-1-5_iahr40wc-p0794-cd

Year: 2023