Evaluation of open access precipitation and temperature products using SWAT in Shiyang river basin, Northwest China

Precipitation and temperature are significant inputs for hydrological models. Currently, many satellite and reanalysis precipitation and air temperature datasets exist at different spatiotemporal resolutions at a global and quasi-global scale. This study evaluated the performances of three open-access precipitation datasets (gauge-adjusted research-grade Global Satellite Mapping of Precipitation (GSMaP_Gauge), Climate Hazards Group Infrared Precipitation with Station data (CHIRPS), Climate Forecast System Reanalysis(CFSR)) and CFSR air temperature dataset in driving the Soil and Water Assessment Tool (SWAT) model required for the monthly simulation of streamflow in the upper Shiyang River Basin of northwest China. After a thorough comparison of six model scenarios with different combinations of precipitation and air temperature inputs, the following conclusions were drawn: (1) Although the precipitation products had similar spatial patterns, however, CFSR differs significantly by showing an overestimation; (2) CFSR air temperature yielded almost identical performance in the streamflow simulation than the measured air temperature from gauge stations; (3) among the three open-access precipitation datasets, CHIRPS produced the best performance. These results suggested that the CHIRPS precipitation and CFSR air temperature datasets which are available at high spatial resolution (0.05), could be a promising alternative open-access data source for streamflow simulation in the case of limited access to desirable gauge data in the data-scarce area.

Therefore, it is important to explore the overall performance when CFSR air temperature data is combined with other better precipitation data to drive the SWAT model, particularly in regions where weather stations are sparse.
Besides GSMaP and CFSR datasets, there are many other satellite or reanalysis precipitation datasets in different resolutions at global or qusia-global scales [27]. It is worth noting that among the recently released datasets, the Climate Hazards Group Infrared Precipitation with Station data (CHIRPS) precipitation dataset  stands out by providing daily precipitation at the finest spatial resolution of 0.05° (one grid representing around 25km 2 ) from 1981 to date [8,13].
CHIRPS merges global climatology, satellite estimates, and in-situ observations to produce precipitation data, which has been evaluated and compared with other precipitation products in many regions such as Tibetan Plateau [2,28], mainland China [29], Ethiopia [8], Northeast Brazil [30] among others. CHIRPS was found to be more accurate and better than other commonly used satellite or reanalysis precipitation products. Given its high resolution, long duration, and good performance, CHIRPS can be a good alternative data source in many aspects.
In this study, our objective is to evaluate the performance of using different combinations of precipitation datasets including the relatively new CHIRPS with CFSR air temperature datasets in driving the SWAT model for streamflow simulation in the Shiyang river basin. This will enable us to select suitable weather data from satellite and reanalysis sources.

Study Area
Shiyang river begins from the northern slope of the Qilian Mountains and ends at the Minqin oasis (Fig. 1). The Shiyang river basin which covers an area of about 41,400 km 2 , is one of the three inland river basins in the Hexi corridor [31,32]. Precipitation in this region has high spatio-temporal heterogeneity where in the southern mountain regions, the annual precipitation reaches approximately 600-300mm, while in the lower reaches, it's lower than 150mm, with nearly 81-89% occurring from May to October [33]. The upper reaches with an elevation of 2000-5000m are main source region for runoff, which has an area of about 1.1×10 4 km 2 , consisting of eight main sub-basins located upstream, including the Xida River basin, Dongda River basin, Xiying River basin, Jintai River basin, Zamu River basin, Huangyang River basin, Gulang River basin and Dajing River basin [34,35].

In-situ measurements from gauge stations
The gauged meteorological data (daily precipitation, daily maximum temperature, minimum temperature) of five meteorological stations (Wuqiaoling, Menyuan, Wuwei, Yongchang, and Minqin) for 1984-2012 were provided by the National Meteorological Information Center of China Meteorological Administration (http://data.cma.cn/). Observed monthly discharge data for the study periods were obtained from the Hydrological Service of Gansu province. Also, we collected a spatially interpolated datasets of annual precipitation from the Resource and Environment Data Cloud Platform (http://www.resdc.cn/data.aspx?DATAID=229) as reference data.
GSMaP_Gauge product is corrected based on GSMaP-MVK (a pure satellite-based GSMaP product without correction by gauge data) and adjusted by Climate Prediction Center (CPC) global daily gauge data analysis [38,39]. GSMaP _Gauge with 0.1° spatial resolution and 1-h temporal resolution covers latitude 60°N to 60°S and longitude 180°W to 180°E [6,36]. More information about the GSMaP products is available in the G-Portal data service system (http://www.gportal.jaxa.jp).
The 1-hour GSMaP data were integrated into daily precipitation data.
The CFSR products are available at a spatial resolution of 0.31° with 64 levels extending from the surface to 0.26 hPa [8]. The CFSR data spanning from 1979 to 2014 are available on the website (https:// globalweather.tamu.edu/) and popularizes its application in SWAT modelling community because it provides readily weather data (precipitation, air temperature, relative humidity, wind speed, and solar radiation) required by SWAT in the ready-to-use format.  [29,41].

SWAT setup
The SWAT was developed by the USDA to simulate the quantity and quality of surface and groundwater and predict the impacts of land use, climate change, and land management practices on water resources, sediments, and agricultural production [32,[42][43][44]. SWAT model divides the basin into several sub-basins and then constructs hydrological response units (HRU) based on land use and soil types to model hydrological processes at each HRU. The official theoretical documentation [45] and a review paper [46] gives more details regarding the SWAT model.  [47].
By varying different precipitation and temperature data as inputs, a total of 6 models were set up using a combination of two temperature data (gauge and CFSR) and three precipitation data (GSMaP_Gauge, CHIRPS, and CFSR) to study the effects of different weather products on streamflow simulation.

Model calibration and evaluation
The SUFI-2 algorithm in the SWAT-CUP software package is employed for the SWAT model calibration, uncertainty, and sensitivity analysis. The SUFI2 algorithm covers all uncertainties (parameter, conceptual model and input) of the parameters [48]. It calculates 95% PPU at the 2.5% and 97.5% levels of the cumulative probability distribution of an output variable obtained through Latin hypercube sampling disallowing 5% of the very bad simulations [48,49]. P-factor and R-factor are used to evaluate the strength of model calibration and uncertainty. The P-factor varies from 0 to 1, where 1 indicates 100 bracketings of the measured data within model prediction uncertainty [50].
The R-factor represents the ratio of the average width of the 95% PPU band and the standard deviation of the observed variable. There are ten objective functions such as R 2 , Nash-Sutcliff (NS) to evaluate the simulated results. In this paper, we used R 2 , NS, and (Percent Bias) PBIAS to evaluate the accuracy of the result. The formulas are as follows: where Q is a variable (discharge in here), m and s stand for the measured and simulated variable, i represents the ith measured or simulated data, and the bar stands for average. The optimum values of R 2 , NS, and PBIAS are 1, 1 and 0 respectively. For an even comparison, models are evaluated after an equal number of iterations.

Comparison of precipitation and temperature products
The comparison of different satellite-based precipitation products can help to identify the consistency and discrepancy that exist between them. Figure 2 shows the spatial patterns of annual precipitation estimates from gauged and 3 satellite products. Generally, these precipitation products have similar patterns, showing a decreasing trend from southwest to northeast. However, the precipitation from the 3 satellite products displayed large differences over the study area. The mean annual precipitation for the GSMaP_Gauge, CHIRPS, and CFSR are 255.2, 238.4 and 351.4 mm/yr, respectively. The CFSR precipitation generally has the largest value followed by CHIRPS and GSMaP_Gauge, and the maximum difference in annual precipitation between the two products was larger than 100 mm/yr.  (rain=0 mm/day), which are 49%, 52%, 68% and 20% for gauge, GSMaP_Gauge, CHIRPS, and CFSR, respectively. Overall, GSMaP_Gauge and CHIRPS had similar distribution for all precipitation intensity. CFSR product was significantly different from other products except for high wet days (>15mm). CFSR data set had the highest probability (11%) of precipitation beyond 5mm/day, while others had less than 5%.  months that occurred June-September. However, CFSR generally had more precipitation than the other three datasets for the wet months throughout the entire period. The variation and magnitude of monthly precipitation for GSMaP_Gauge and CHIRPS precipitation data were much similar, which were also similar to gauged precipitation, but their peaks were usually higher than gauged precipitation.

Streamflow simulation without calibration
We first evaluated the performance of six model scenarios without calibration in three stations, and for conciseness, the streamflow graphs without calibration are not shown here. Table 1 shows the evaluation statistics for the performance of six models in three stations. Moriasi et al. [51] indicated that model simulation can be judged as satisfactory if NSE > 0.5 and PBIAS < ±25% for streamflow. According to these criteria, using CHIRPS as inputs generated the best performance. In values. This is due to the high overestimation of CFSR precipitation (Fig. 2, 4), which indicated that CFSR precipitation is unsuitable for this area.

Streamflow simulation after calibration
Figs. 7-9 presents a comparison of monthly measured and simulated streamflow from six model scenarios in three stations respectively. For brevity, we did not separate the calibration and validation periods but showed the evaluation statistics for the two periods in Table 1. In Fig. 7-9, the first row showed the comparison of measured and simulated streamflow using gauged temperature as input with different precipitation datasets, and the second row showed the same as the first row except that CFSF air temperature was used as input instead of gauge data.
Generally, most models were able to capture seasonal patterns very well. Hydrograph showed reasonable agreement between the observed and simulated streamflow using CHIRPS and GSMaP_Gauge precipitation as inputs except for overestimation and underestimation in a very few events. CHIRPS precipitation performed best in streamflow simulation with NS >0.60 in all the three stations (Table 1). In Jiutiaoling station and Zamu stations, NS >0.80 and PBIAS <5% using CHIRPS and gauge temperature as inputs (Table 1). In Jiutiaoling station, GSMaP _Gauge precipitation even performed better than CHIRPS with NS >0.84 and PBIAS < 5% at the calibration period. However, CHIRPS had overestimation at peak flow period in some years, for example in the year 2010 and 2011 in Xida station. The cause of this phenomenon is the overestimation of daily rainfall from CHIRPS in these months (Fig. 4). Using CFSR precipitation yielded the worst performance in all the three stations.
The overestimation of precipitation by CFSR seems to be more severe during the validation period (2008)(2009)(2010)(2011)(2012), and thus the calibrated parameters could not compensate for such overestimation.
In summary, using several precipitation datasets as inputs to SWAT had a larger impact on streamflow simulation than using different air temperature datasets in the study area. Using CFSR temperature as input can deliver equal performance to gauge temperature. This is a good sign for researchers and managers that they have an alternative when gauge temperature products are scarce.
Regarding the selection of precipitation dataset, we do recommend CHIRPS as input for streamflow simulation in this area. And the combination of CHIRPS precipitation and CFSR air temperature is

Comparison of calibrated parameters
We compared the differences in parameters from different models taking Jiutiaoling station as a case study. Table 2 shows the optimal parameters after calibration in Jiutiaoling station. Models using gauge temperature and CFSR temperature had similar values when using the same precipitation data to drive the SWAT model. Different models driven by different precipitation products generate diverse optimal parameters, but still yielding similarly good performance, which also reflects the effect of parameters. For example, the optimal parameter CN2 is positive 0.2 with GSMaP_Gauge driving SWAT model and negative with CHIRPS and CFSR. The decrease in CN2 could generally result in less runoff by SWAT. For the parameter SOL_AWC that is responsible for the available water capacity of the soil layer, GSMaP precipitation led to an increase, and CHIRPS and CFSR precipitation datasets resulted in a decrease. The decrease in SOL_AWC could also result in less runoff. For the groundwater delay time (GW_DELAY) parameter, GSMaP precipitation had higher values, which resulted in a slow recharge of the shallow aquifer and discharge to the stream. During calibration, however, different parameter values compensated the differences in precipitation inputs to increase the performance of streamflow simulation at the basin outlet, which may lead to different hydrological components. This is an intrinsic deficiency using only streamflow to calibrate hydrological models. Unfortunately, this is a common practice in hydrological modelling due to the scarcity of other observed hydrological components (soil moisture, evaporation). Therefore, the multivariable (such as satellite-based evapotranspiration or soil moisture data) or multi-site calibration should be made to overcome the uncertainty of the parameters. Several studies have attempted to examine the feasibility of improving hydrological modelling using satellite soil moisture and evaporation [52][53][54][55]. This practice (model calibration with satellite-based evapotranspiration or soil moisture data) is also within our plans for further study in the Shiyang river basin.

Discussion
Precipitation and temperature are the two most significant inputs for streamflow simulation, and inaccuracy from their measurements could lead to large uncertainties in streamflow simulations and predictions [8,56,57]. Therefore, the hydrological evaluations of the climatic inputs are based on the hypothesis that errors in the climatic inputs can propagate into the hydrographic outputs [58].
Evaluating climatic inputs through their predictability on streamflow is usually carried out at the watershed scale, which avoids the scale discrepancy compared to using gauge observations for validation [2]. This study compared the model performances for streamflow simulations driven by different precipitation and temperature products derived from satellite data in the Shiyang river basin. Several studies have evaluated the performance of different precipitation datasets in driving hydrological SWAT model in streamflow simulation in the same basin. Guo and Su [34] evaluated the performances of five different satellite and reanalysis precipitation datasets in streamflow simulation of Jiutiaoling station and indicated that CHIRPS and CMDF performed better than other datasets (i.e., TRMM, CFSR, and PGF). They found that CFSR overestimated precipitation and produced unsatisfactory streamflow simulation, which was consistent with our study. Similar overestimation in CFSR precipitation datasets has also been reported in many other regions like Ethiopia [6,59], Singapore [60], Xiang river basin and Qu river basin in China [61] and Mekong river basin [62], with limited research reporting much better performances of CFSR precipitation datasets in streamflow simulations [63]. In contrast, Musie found that the CFSR product underestimated precipitation amounts in the Meki watershed of Ethiopia [59]. That results indicated that CFSR precipitation data have large uncertainty and should be used with caution despite its global availability and free access. Haihe River Basin, China [67] were also in line with this result. In the same vein, these studies also agree with our study that the CHIRPS product had a good performance. GSMaP _Gauge and CHIRPS precipitation datasets had similar performance despite their differences in space and time.
A viable explanation is that the differences in the precipitation inputs were buffered inside the hydrological model via parameter calibration.
It should be noted that the CHIRPS product only gives daily precipitation despite their finer resolution (0.05°), which hinders its application on streamflow simulation at sub-daily scales. In contrast, CFSR and GSMaP provided hourly rainfall products that are assumed to have beneficial potentials for applications on streamflow simulation at an hourly scale. Compared to CFSR, GSMaP is an assuring product to simulate sub-daily streamflow due to its finer resolution and higher precision. Further research is, therefore, necessary to evaluate the performances of gridded precipitation products at sub-daily scales in future studies.

Conclusions
Due to the inadequacy of measured weather datasets in mountainous regions, this research compared the performances of three remote sensing precipitation products (GSMaP_Gauge, CHIRPS, CFSR) and one temperature product (CFSR) in driving SWAT model for simulating streamflow in Shiyang river basin. After a thorough comparison in three hydrologic stations located in upstream, the following conclusions were drawn: 1. The annual precipitation from the three satellite products displays a large disparity over the study area. The CFSR precipitation generally has the largest values, followed by gauged precipitation, CHIRPS and GSMaP_Gauge products. CFSR data set agreed better with gauge measurements for the daily minimum air temperature than for maximum air temperature.
2. The CFSR temperature produced almost the same simulated results with the observed temperature in streamflow simulation for the three stations. This proved the accuracy of using CFSR air temperature for hydrological modelling in this region. This is useful for researchers and the local community as the availability and quality of observed temperatures are often worse than that of precipitation, especially on a daily scale.
3. Among the three precipitation products, CHIRPS yielded the best performance and it was the only one that can achieve satisfactory simulation without calibration. The CFSR overestimated precipitation values and resulted in inconsistent and unsatisfactory performance for monthly streamflow simulation.

4.
The CHIRPS precipitation product with a high spatial resolution (0.05°) with the CFSR air temperature can be a promising alternative open-access data source for streamflow simulation with SWAT in data-scarce areas.
This study did not analyze the performance of using complete CFSR weather data (wind speed, solar radiation, and relative humidity) for driving the SWAT on account of non-availability of