Evaluating applicability of multi-source precipitation datasets for runoff simulation of small watersheds: a case study in the United States

Small watersheds are ideal objects for studying the evolution of hydrological and water resources at small scales. Whether precipitation products can meet the runoff simulation of small watersheds is the main purpose of this study. With NOAA-CPC-US precipitation as a reference, in nine small watersheds of the United States, accuracy of the precipitation products such as PERSIANN, PERSIANN-CDR, TRMM-3B42V7, GPM-IMERG, StageIV, and ERA5 is analyzed. By driving the CREST hydrological model using these datasets, the runoff simulation effects were evaluated. Result shows the precipitation products match the NOAA-CPC-US from high to low in the order of StageIV, PERSIANN-CDR, GPM-IMERG, PERSIANN, ERA5, and finally TRMM-3B42 V7. These datasets have relatively low accuracy in the northern high latitude area and the western mountains, while accuracy is better in the central, southern, and eastern parts of the United States. In the runoff simulation effectiveness evaluation, the daily runoff of watersheds is simulated during the same verification period. The results show that: NOAA-CPC-US and StageIV have good simulation effect in small watersheds. In the northern and western United States, the PERSIANN, PERSIANN-CDR, GPM-IMERG, and ERA5 for runoff simulation should be used with caution. TRMM-3B42V7 is not suitable for runoff simulation in small watershed.


Introduction
Precipitation, one of the main links in hydrological cycle, is also the main data input in hydrological model research.Precipitation plays an important role in the research, monitoring, and prediction of climate, meteorology, hydrological forecasting, and hydrological disasters (Islam, 2013).Affected by climatic factors and underlying surface factors, precipitation has uneven formation and distribution in space and time, which in turn affects water cycle processes of river basins such as land runoff, evapotranspiration, soil moisture, and groundwater.Therefore, accurate precipitation estimation and large-scale rapid acquisition of continuous precipitation observation data with high spatial and temporal resolution can not only help us better understand the water cycle, but also means important significance for the simulation and prediction of hydrological processes in the basin.At present, there are four methods for obtaining precipitation data: observation from ground precipitation stations, precipitation detection by ground radar, satellite remote sensing for precipitation and numerical simulation of precipitation based on climate system models (Tapiador et al., 2012).These four methods have respective advantages and disadvantages.Where observation by ground precipitation stations is direct and highly accurate.However, due to density of the stations and their spatial distribution, it is difficult to provide complex and variable spatial distribution of precipitation.Precipitation detection by ground radar can directly detect spatial distribution of precipitation, which can track the precipitation center and spatial changes in real time, but it is susceptible to terrain occlusion and radar ray uplift.Although satellite remote sensing for precipitation has a lower accuracy than ground observation of precipitation, it can quickly and conveniently acquire continuous precipitation data with certain space-time accuracy on a large scale.It can make up for the deficiency of ground observation of precipitation to a certain extent, which provides an effective source of precipitation data especially for areas with no data or few ground observation data.The climate system model provides an important way to simulate and predict climate change.However, due to the impact of global climate complexity, there are still uncertainties in the simulation results (Stocker et al., 2013).Current climate models can simulate the large-scale change features of global precipitation, but there are still many deficiencies in the simulation of regional-scale precipitation (Ying & Chong-Hai, 2012;Zhen et al., 2015).The precipitation simulation data of some climate models cannot be directly used for hydrological utility assessment (Lü et al., 2016).
Recently, with the rapid development of satellite remote sensing technology, radar rain measurement technology, climate numerical models, as well as the application of machine learning algorithms in precipitation retrieval, a large number of precipitation products with different spatial and temporal resolutions have appeared (Hou et al., 2014).At the same time, many scholars have carried out a lot of research around evaluation of precipitation accuracy, analysis of precipitation changes at different time and space scales using precipitation products, and simulation of hydrological utility (Balsamo et al., 2018;Kumar et al., 2017;Maldonado, 2011;Nelson et al., 2016;Zhao et al., 2017).The global or large and medium regionalscale runoff simulation and real-time prediction system developed based on precipitation product are positively evaluated in the prediction and reproduction of multiple hydrological disasters (Omranian et al., 2018;Paska et al., 2017;Tian & Zou, 2018;Yang et al., 2017).However, there is still significant spatial variability in current precipitation products, and significant uncertainties still exist in hydrological applications at different temporal and spatial scales (Hong et al., 2006;Jiang et al., 2014Jiang et al., , 2012;;Krajewski et al., 2010;Mei et al., 2014;Sarachi et al., 2015).Large floods may also occur in small watersheds (Michaud et al., 2001); therefore, the ability in hydrological simulation and forecasting on small-watershed scale is still worth in-depth evaluation.
Small watersheds are ideal objects for studying the evolution rule of hydrological and water resource systems at small and micro scales.As important geographical units for soil erosion control and ecosystem restoration research, they are minimum units for calculating the production of water and sediment by rivers, which provide optimal geographical scale for research and management of hydrology and soil erosion.The importance of small watersheds can be illustrated by the relationship between cells and organisms, small tributaries, and large rivers.The use of different precipitation products to accurately simulate and predict runoff in small watersheds plays a very important supporting role for the above research work.This study selected small watersheds distributed in nine different geographic regions in the United States.Based on the precipitation reanalysis data of Climate Prediction Center, National Oceanic and Atmospheric Administration, United States (NOAA-CPC-US) ground precipitation stations, the accuracy of satellite precipitation products such as the precipitation estimation from remotely sensed information using artificial neural networks (PERSIANN), the precipitation estimation from remotely sensed information using artificial neural network climate data record (PERSIANN-CDR), the precipitation version 7 derived from tropical rainfall measuring mission 3B42 research version (TRMM-3B42V7), the precipitation of integrated multi-satellite retrievals for the global precipitation measurement (GPM-IMERG), radar precipitation Stage IV data of the National Centers for Environmental Prediction, United States (StageIV) and the precipitation product of the fifth-generation reanalysis dataset released by the European Center for Medium-Term Weather Forecasting (ERA5) is evaluated.Combining the coupled routing and excess storage (CREST) distributed hydrological model, the daily runoff process of the nine small watersheds is simulated and reproduced to evaluate the hydrological simulation effectiveness of multi-source precipitation products at the small-watershed scale, thereby providing references for research and applications of the above-mentioned multi-source precipitation product in the hydrological industry.

Study area
There is no uniform definition for the concept of small watershed at home or abroad.Different countries and organizations have different watershed standards for different research purposes and perspectives.The small watersheds discussed herein are based on the U.S. Stream Classification System (USSCE) of the US Department of Energy's Oak Ridge National Laboratory, USA.This system divides nearly 2.6 million rivers in the United States into 8 levels.From large to small, they are Great River, Large River, Mainstem, Medium River, Small River, Large Creek, Creek, and Headwater (McManamay & DeRolph, 2019).Indeed, there are many small watersheds in a climate zone.We used a random sampling method to select small watersheds in the middle of the climate zone without any restrictions.In this study, nine closed watersheds are selected from different geographic regions of the continental United States (Figure 1).The rivers in these small watersheds belong to headwater and creek level in the river classification system.See table (Table 1) for information like geographical zone, watershed area, and annual average flow of these nine small watersheds.

Data used
This study selects seven types of precipitation data: NOAA-CPC-US, PERSIANN, PERSIANN-CDR, TRMM-3B42V7, GPM-IMERG, StageIV, and ERA5.The data period is from 1 January 2002 to 30 June 2018.In the study, the time resolution was set to the daily scale, and the spatial resolution was 0.0083°× 0.0083° (about 1 km 2 ).Due to the different temporal and spatial resolutions of the abovementioned 7 types of precipitation data, during the data preprocessing process, data with temporal resolution smaller than the daily scale is cumulatively converted to the daily scale, and the spatial resolution is unified to 0.0083°× 0.0083° via bilinear interpolation.

TRMM-3B42V7 satellite precipitation
Due to the exhaustion of fuel in April 2015, TRMM satellite officially ended its observation mission and reentered the Earth's atmosphere in June of the

R E T R A C T E D
same year, but falling into the South Indian Ocean.
Due to the success of TMPA precipitation product, TRMM team did not immediately stop the TMPA precipitation data product due to termination of satellite observation.Instead, various methods are adopted to extend TMPA precipitation products, which are planned to be updated until mid-2019 (Huffman, 2016).The selection of TRMM-3B42v7 precipitation product in this paper is also intended to assess the ability of this dataset to simulate hydrological utility under current conditions.TRMM-3B42 V7 data has a spatial resolution of 0.25°, a temporal resolution of 3 h, and a precipitation unit of mm/h.This data is downloaded from NASA's TRMM satellite data website (https://pmm.nasa.gov/trmm/).

GPM-IMERG satellite precipitation
GPM is a follow-up global satellite precipitation observation program for TRMM led by the National Aeronautics and Space Administration (NASA).
With the core observation platform launched on 28 February 2014, GPM can provide rain and snow observation data within 3 hours worldwide.
Compared with TRMM, GPM covers a large area and extends to the poles of the earth.With high spatial-temporal resolution and superior sensor performance, it can accurately detect trace precipitation.
For other characteristics of this precipitation product, please refer to the technical documentation provided by its portal (https://pmm.gsfc.nasa.gov/GPM),and the Level 1-3 products of GPM data can also be downloaded from this website.This paper selects integrated multi-satellite retrievals for GPM (IMERG, version: 05B_Final) for global 30-minute rain and snow data product.The precipitation data was produced by the latest GPROF2017 algorithm, which calibrated and integrated all microwave, infrared, ground observation, and other possible sensor data of the GPM constellation (Huffman et al., 2015).It has a spatial resolution of 0.1°, a time resolution of 30 minutes and a precipitation unit of mm/h.

Methodology
The methodology consists of three major steps.First, the NOAA-CPC-US precipitation products were taken as a reference to conduct precipitation accuracy assessment for other precipitation products (StageIV, PERSIANN, PERSIANN-CDR, TRMM-3B42V7, GPM-IMERG, ERA5) of the same period.Second, the precipitation accuracy of different rain intensity levels was evaluated to indicate which precipitation products have good or poor estimation accuracy.Finally, the performance of hydrologic simulations in small watersheds was evaluated with multi-resources precipitation datasets by a distributed hydrologic model, Coupled Routing and Excess Storage (CREST).

CREST distributed hydrological model
The Coupled Routing and Excess Storage model (CREST) was jointly developed by the University of Oklahoma's Hydrometeorology and Remote Sensing Laboratory (HyDROS) and the NASA SERVIR project team.As a distributed hydrological model, it divides the whole basin into several grids, calculates runoff by grid using variable permeability curve, and calculates surface and groundwater confluence by grid using multi-linear reservoirs, and finally simulates and reproduces runoff process by coupling runoff elements and confluence structure (Wang et al., 2011).The model demonstrates excellent hydrological simulation and prediction capabilities at global scale, large regional scale, and small and medium scales (Khan et al., 2010;Meng et al., 2014;Shen et al., 2017;Xue et al., 2013).When the CREST model is running, the data that needs to be input include data watershed information (such as watershed boundary, DEM, flow direction, etc. ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi P n i¼1 ðObs i À ObsÞðSim i À SimÞ ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi P n i¼1 ðObs i À ObsÞ Wang et al. ( 2011)

Wang et al. (2011)
In the above formula, Sim represents multi-source precipitation estimation data or runoff data simulated by CREST model, Obs represents reference precipitation data or runoff station observation data; n is the simulation value or the total observed value, and i is the i-th simulation value or observed value.When the four evaluation indicators meet Bias = 0%, RMSE = 0, CC = 1, NSCE = 1, it represents the optimal evaluation results.The NSCE value can be divided into 4 intervals to represent the utility of 4 levels of hydrological simulation: poor (NSCE≤0.5),applicable 376 K. FENG ET AL.

Precipitation accuracy assessment
Accuracy of precipitation estimates by precipitation products has a significant impact on hydrological simulation effectiveness.Therefore, the precipitation accuracy of each precipitation product should be evaluated first.NOAA-CPC-US precipitation dataset collects precipitation observation information from tens of thousands of ground observation stations.Then, data is optimized for interpolation based on consideration to terrain effect, followed by quality control.
Therefore, this study takes the NOAA-CPC-US precipitation products from 1 June 2014 to 30 June 2018 as a reference to conduct precipitation accuracy assessment for other precipitation products of the same period: StageIV, PERSIANN, PERSIANN-CDR, TRMM-3B42V7, GPM-IMERG, ERA5.

Precipitation accuracy
For each small watershed, the correlation coefficient between the precipitation products to be evaluated and NOAA-CPC-US precipitation was calculated at a grid scale of 1 km 2 (Figure 2, Figure 3).The vertical axis of Figures 2 and 3 is the correlation coefficient, and the horizontal axis is the precipitation product.
The box in the figure represents the maximum, minimum and average value of the correlation coefficient.The RMSE and MAE of each precipitation product in each study area relative to NOAA-CPC-US precipitation (Table 2) also reflect the difference in precipitation accuracy.For mean value of root mean square error and absolute error of each precipitation product in the nine small watersheds, Stage IV precipitation has a RMSE of 5.78 mm and a MAE of 2.15 mm; PERSIANN-CDR precipitation has a RMSE of 6.52 mm; and a MAE of 6.39 mm; PERSIANN precipitation has a RMSE of 12.14 mm and a MAE of 6.90 mm; GPM-IMERG satellite precipitation has a RMSE of 23.27 mm and a MAE of 6.44 mm; ERA5 precipitation has a RMSE of 53.02 mm and a MAE of 6.84 mm; TRMM-3B42 V7 satellite precipitation has a RMSE of 65.92 mm and a MAE of 7.06 mm.
To sum up, the statistical index analysis shows that there is a difference in the correlation between the six precipitation products and NOAA-CPC-US precipitation.Such difference has regional characteristics and shows spatial variability of precipitation products.Precipitation products have slightly lower estimate accuracy in regions such as high latitude area (northeast) and western mountains of the United States.The precipitation products match the NOAA-CPC-US precipitation from high to low in the order of Stage IV radar precipitation, followed by PERSIANN-CDR and GPM-IMERG, followed by PERSIANN and ERA5, and finally TRMM-3B42V7.From the statistical results in Figure 4, it can be seen that relative to NOAA-CPC-US precipitation, occurrence frequency of different precipitation intensity differs for StageIV, PERSIANN, PERSIANN-CDR, TRMM-3B42V7, GPM-IMERG, and ERA5 precipitation products in nine small watersheds of different regions.There are overestimates and underestimates to varying degrees, but the errors do not exceed ±10%.In general, most precipitation products are underestimated for no rain (precipitation<1 mm) and drizzle (1 mm≤precipitation<2 mm) levels, and are overestimated for heavy rain (10 mm≤precipitation<50 mm) and rainstorm (precipitation≥50 mm).StageIV has higher estimation accuracy than PERSIANN, PERSIANN-CDR, TRMM-3B42V7, GPM-IMERG, ERA5 for each precipitation intensity level.

Hydrological simulation effectiveness assessment
After assessing the precipitation accuracy, this paper uses the above seven precipitation products to drive the distributed hydrological model CREST, respectively, so that runoff processes in nine small watersheds in different regions of the United States can be simulated, and hydrological simulation effectiveness of each precipitation product can be assessed.In runoff simulation scenarios, 1 January 2002 to 31 December 2003 is set as the model warm-up period, with 15 January 2004 30 Juneto 2014 as the model rate period, 1 July 2014 30 Juneto 2018 as the model verification period.In the same calibration period, the CREST parameters were respectively calibrated using seven types of precipitation products, and the runoff was simulated during the same verification period based on the parameter sets calibrated by the respective precipitation products.Statistical indicators of NSCE, Bias, SD, and CC were used to evaluate the hydrological simulation effectiveness.This scenario setting helps independent assessment of the simulation results of the basin hydrological process for each precipitation data.
Figures 5, 6 and 7) illustrate the daily runoff calibration and simulation results of the precipitation products in 9 small watersheds in different regions of the United States, which are relative to the standard deviation SD of the measured runoff, the root mean square error RMSE, and the correlation coefficient CC, Nash efficiency coefficient NSCE, and relative deviation Bias.
From Figure 5, Figure 6, Figure 7, and Table 3, the hydrological simulation results of the multi-source precipitation products are analyzed for the small watersheds of each region, and the precipitation product has better runoff simulation performance in calibration period than in verification period, and each statistical

Conclusion
It is a main trend to simulate and predict hydrological processes at different scales by obtaining precipitation estimation data via remote sensing technology and driving distributed hydrological models in the study of hydrological and water resources in river basins.
The typical advantages of the remote sensing precipitation data lie in fast acquisition speed, large spatial range, continuous time, and certain accuracy guarantee.These data are also important for the prediction of natural disasters such as floods, hurricanes, and landslides; it is also an effective source of precipitation data for areas with no data or few ground observation data.
Compared with large-scale watersheds, small watersheds also play a very important role in the earth's hydrosphere, biosphere, and lithosphere.Small watersheds have complete water cycle process, which are ideal objects for studying small and micro-scale hydrology, water resources, and water environment.It is also an important means to find out the evolution rules of small and micro hydrological water resource systems and find sustainable water resource utilization methods.In recent years, scholars have carried out many researches on hydrological and water resources on the global

Figure 1 .
Figure 1.Small watersheds selected from different geographic regions of the United States.
ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi ffiffi P

Figure 2 .
Figure 2. Correlation coefficients between multi-source precipitation products and NOAA-CPC-US.

Figure 3 .
Figure 3. Correlation coefficients between multi-source precipitation datasets and NOAA-CPC-US in different regions.

Figure 4 .
Figure 4. Frequency error of multi-source precipitation products relative to NOAA-CPC-US precipitation at various precipitation intensity levels.
verification period experiences attenuation.These precipitation products have different runoff simulation performance in small watersheds in different regions of the United States.NOAA-CPC-US and StageIV precipitation maintain good runoff simulation characteristics, which reproduces the daily runoff process during the flood discharge of the basin.PERSIANN, PERSIANN-CDR, GPM-IMERG, and ERA5 precipitation are less capable of simulating runoff than NOAA-CPC-US and StageIV precipitation.The simulation of peak discharge has obvious fluctuations and lacks accuracy.The simulation performance is better in central, southern United States, but poor in northern high latitude area, west, and southwest of United States.TRMM-3B42V7 is not ideal for runoff simulation in multiple small watersheds due to sensor accuracy and resolution.

Figure 5 .
Figure 5. Statistical values in runoff simulation of multi-source precipitation products in small watersheds in different regions of the United States.

Figure 6 .
Figure 6.Nash efficiency coefficients for multi-source precipitation products in runoff simulation.

Figure 7 .
Figure 7. Relative deviations of multi-source precipitation datasets in runoff simulation.

Table 1 .
Basic information of the small watersheds.

runoff, topography and potential evapotranspiration data
(Lin, 2017)ity control, and finally the regional data were synthesized into quantitative radar precipitation estimation data for the continental United States(Lin, 2017).The data has a spatial resolution of 4 km, a temporal resolution of 1 h, 6 h, and 24 h, and a precipitation unit of mm.In this study, 6 h estimation data with StageIV time resolution is selected(www.emc.ncep.noaa.gov/mmb/ylin/pcpanl/stage4/).It has created a milestone for humans to use satellite technology to predict precipitation, but it has failed.Its successor is the more excellent GPM satellite precipitation observation program.ERA5 represents the precipitation data output by the climate model.The precipitation products of climate models are also important data sources for runoff simulation and forecasting.The watershed boundary, digital terrain DEM, and runoff data of the nine small watersheds herein adopts the Watershed Boundary Dataset V2.2.1 (US Geological Survey and US Department of Agriculture, Natural Resources Conservation Service, 2013) provided by the US Geological Survey (USGS), the US 10 m precision digital elevation model (DEM), radars across the United States and measurements from approximately 5,500 ground-based rain gauges, which is produced by the 12 river forecasting centers distributed throughout the continental United States by region.Data collection started at 33rd minute per hour, and cumulative calculations were performed with time steps of 1 h and 6 h, followed by iterative review and Please refer to the CREST model website of the HyDROS Laboratory for detailed information about model data preprocessing, model parameter value ranges, and so on (http://hydro.ou.edu/research/crest/).
), precipitation, potential evapotranspiration, watershed exit flow, initial conditions, and model parameters.Model output results include soil water content, soil moisture, and surface runoff.
US data, with an average value of 0.73.The correlation coefficient changes from 0.65 to 0.81 for small watersheds in different regions; the correlation coefficient is lower in the West and Southwest regions.Next comes PERSIANN-CDR and GPM-IMERG products, whose correlation coefficients with NOAA-CPC-US data are 0.51-0.69,0.53-0.73,respectively.
Among the six precipitation products, StageIV has the highest correlation coefficient with NOAA-CPC-

Table 2 .
Errors of multi-source precipitation products relative to NOAA-CPC-US precipitation in each study area (unit: mm).

Table 3 .
Nash efficiency coefficient of the runoff simulation utility of multi-source precipitation products in each small watershed.