Applicability Of Multi-Source Satellite Precipitation Products In Streamflow Modeling In The Yangtze River Basin

Precipitation is one of the most important climatological(meteorological)variable.Although precipitation is closely associated with drought and flood disasters,water resources management and flood forecasting,which fundmentally determine national economy and the people's livelihood,it's hard to measure and estimate precipitation due to its active change in space and time.Lacking spatio-temporal distribution of precipitation with high precision restricts the relavant hydrological application.With the development of remote sensing and meteorological satellite techeques,the satellite precipitation estimates(SPE)become potential alternatives of ground observation of rain gauges.As compared with rain-gauge observation,larger errors are inherent in current SPE.Therefore,before input into hydrological systems,the modeling precision can be improved via systematical bias correction of SPE as pre-processing.On the other hand,since biased SPE inevitably induce modeling unertainty when driving hydrological models,analyzing the multi-source uncertainty components related to SPE is beneficial to diagnosing the primary contributer of modeling uncertainty and enhancing the reliability of modeling.However,current investigation of SPE impact on hydrological modeling is mainly concerntrated on preliminary assessment of SPE and verification of hydrological applicability of SPE.Thus,the quantification and reduction of errors and uncertainty related to input and output in rainfall-runoff modeling were insufficiently considered in previous study.Hundreds of millions of people live in the Yangtze River basin where water disaster frequently occurs.It's critical for social-economy development to accurately capture the spatio-temporal precipitation and apply to hydrological modeling and forecasting.To explore the potential applicability of SPE in hydrological modeling,this study evaluated the precision and hydrologcial response of global SPE products in the Yangtze Rive basin and its subcatchments,respectively;then corrected the systematical bias of SPE as pre-processing;and verified the correction effect using statistical evaluation metrics and streamflow response;and finally analyzed the multi-source uncertainties of hydrological modeling driven by SPE as post-processing.The main conclusions are summarized as follows:(1)Taking the ground rain gauge observations from the China Meteorology Administration as benchmarking,the accuracy of error characterization of four TRMMera SPE products,including TRMM3B42 RT,TRMM3B42v7,GsMap and PERSIANN,and four GPM-era SPE products,including IMERG-E,IMERG-F,GsMap and PERSIANN,were separately quantified over the Yangtze River basin.Four continuous metrics,including the Root Mean Square Error(RMSE),Correlated Coefficient(CC),Relative Bias(pBIAS)and Kling-Gupta Efficient coeficient(KGE),and three categorical metrics,including Probabality of precipitation Detection(POD),False Alarm Rate(FAR)and Critical Success Index(CSI),were selected.The results of various metrics between SPE and corresponding gauge observations showed that 1)the deviation of SPE from ground reference,characterized by overestimation of TRMM3B42 RT and PERSIANN in TRMM-era products and by underestimation of IMERG-E and overestimation of PERSIANN in GPM-era products,was largest over the western area of the basin;2)SPE of both TRMM era and GPM era performed overall better in wet period than in annual period,and better in annual period than in dry period;3)GsMap performed best among both four TRMM-era products and four GPM-era products,while PERSIANN performed lowest due to its high FAR in both TRMM and GPM eras.(2)The response of simulated steamflow derived from lumped hydrological model GR6 J and distributed CREST to the SPE input was verified in two typical subcatchments of the Yangtze River basin.The results showed that 1)the precision of simulated streamflow driven by GsMap in Ganjiang River basin was highest among all the SPE inputs(NSE in calibration period: 0.77 and 0.83 for GR6 J and CREST,respectively;NSE in validation period: 0.83 and 0.89);2)the simulated streamflow driven by IMERG-F in the Wujiang River basin generally performed best among all the SPE inputs(NSE in calibration period: 0.56 and 0.58 for GR6 J and CREST,respectively;NSE in validation period: 0.54 and 0.70).(3)An improved bias correction approach,i.e.,the Censored Shifted Mixture Distribution Mapping(CSMD),was established to statistically reduce the systematical bias of SPE.CSMD was applied to correct the GPM early near-real-time product IMERG-E over the Yangtze River basin.Contrasted to the traditional Bernoulli and gamma mixture distribution mapping(BerGam),the results showed that both the improved bias correction CSMD and traditional BerGam could significantly reduce the systemarical bias of raw IMERG-E and the correction of CSMD to SPE,especially to highly extreme precipitation,performed better than that of BerGam in various metrics.Furthermore,the hydrological response of corrected SPE was analyzed using the lumped GR6 J and distributed CREST rainfall-runoff models.It was found that two correction appraoches,especially for CSMD,exerted positive imapct on streamflow simulation via GR6 J and CREST models through reducing bias of IMERG-E,compared to the performance of streamflow modeling driven by both the raw IMERGE and post-processed IMERG-F products.(4)To quantify the SPE input uncertainty and its interaction effects with other uncertainty sources in hydrological modeling,a framework based on variance decomposition was constructed to partition the multi-source uncertainties in modeling.The implementation of the framework can be separated into three steps,including 1)determining the number and interaction of multi-source uncertainties,2)configurating the hydrological model to define single factors in variance decomposition,i.e.,input uncertainty,parameter uncertainty and model structure uncertainty,and 3)using GR and CREST to simulate and obtain ensemble streamflow outputs,and calculating the magnitude and relative contribution of the multi-source uncertainty components according to 1)and 2).The precipitation inputs used to drive hydrological model are near-real-time TRMM3B42 RT,post-real-time TRMM3B42v7 and the interpolated ground precipitation.The results showed that the input uncertainty of CREST modeling driven by three precipitation sources was lower than that of GR modeling driven by corresponding input.Among six scenarios combining three precipitation inputs and two hydrological models,input uncertainty in TRMM3B42v7-driven CREST was lowest,while that in ground precipitation-driven GR was highest.These results indicated that the distributed CREST model was capable of making better use of the spatial distribution advantage of SPE especially for the post-real-time TRMM3B42v7 product.This study provides a whole-process framework from pre-processing to postprocessing and multi-case evaluation for the hydrological application of quasi-global SPE in specific regions.