Blending multi-satellite, atmospheric reanalysis and gauge precipitation products to facilitate hydrological modelling

• Published in: Journal of hydrology (Amsterdam) Vol. 593; p. 125878
• Main Authors: Yin, Jiabo, Guo, Shenglian, Gu, Lei, Zeng, Ziyue, Liu, Dedi, Chen, Jie, Shen, Youjiang, Xu, Chong-Yu
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.02.2021
• Subjects: Atmospheric reanalysis; Bias correction; China; Data fusion; Hydrological modelling; Satellite precipitation; Hydrological modelling; Bias correction; Satellite precipitation; Data fusion; Atmospheric reanalysis; China
• ISSN: 0022-1694; 1879-2707
• DOI: 10.1016/j.jhydrol.2020.125878

•A dynamic three-stage blending scheme is developed to simultaneously correct precipitation occurrence and intensity.•A new daily 0.25° precipitation dataset over mainland China is produced by merging four multi-satellite/reanalysis products.•Different data merging methods are systematically evaluated in 238 catchments over China.•Our developed approach can facilitate hydrological modelling by substantially improving the KGE of simulated streamflow. Satellite-retrieved and atmospheric reanalysis precipitation can bridge the spatiotemporal gaps of in-situ gauging networks, but estimation biases can limit their reliable applications in hydrological monitoring and modelling. To correct precipitation occurrence and intensity simultaneously, this study develops a three-stage blending approach to integrate three multi-satellite precipitation datasets (IMERG Final, TMPA 3B42V7 and PERSIANN-CDR), the ERA5 atmospheric reanalysis product and a gauge dataset within a dynamic framework. Firstly, the systematic biases of the four members were individually corrected by combining the local intensity scaling and ratio bias correction methods. Then, the “state weights” used for determining wet/dry events were optimized by evaluating a score function of the four bias-corrected members. Thirdly, the “intensity weights” were optimized using the cuckoo search (CS) algorithm and the Bayesian Model Averaging (BMA) method, respectively. The three-stage blending approach produced dynamic weights varying both spatially and temporally, and the performance was thoroughly evaluated over mainland China. Results show that the three-stage dynamic scheme performs better than individual datasets and two-stage blending methods in terms of all eight statistical metrics, and the CS algorithm outperforms the BMA method in the third stage. By randomly sampling validation sites using K-fold experiments, the developed algorithm also demonstrates a superior performance in ungauged regions. After interpolating and normalizing blending parameters of all gauges to entire domain using ordinary kriging, a new blended precipitation dataset with a daily 0.25° scale was produced. Four hydrological models are forced by blended and primary precipitations in 238 catchments over China, further confirming that the developed approach can facilitate hydrological modelling demonstrated by improving the Kling-Gupta efficiency of simulated streamflow by 12–35%.