*4.2. Simulation Results Using Different Meteorological Inputs*

Tables 2 and 3 summarize the model performance for the eighteen scenarios at daily and monthly timescale, respectively. It can be found that using the same precipitation data but different meteorological data, the streamflow simulation results are similar. Using different precipitation data and the same meteorological data, the results tend to be quite different. This finding shows the dominant role of precipitation data in streamflow simulation. The scenarios using gauged precipitation data had the best performance for both daily and monthly streamflow simulation. During the calibration and validation periods, almost all the simulated results using gauged precipitation data reached the level of very good performance (NSE > 0.75) at daily and monthly timescales (Tables 2 and 3). It is worth noting that among the three scenarios using gauged precipitation data, the scenario with gauged precipitation data together with gauged meteorological data yielded the best performance with NSE higher than 0.87 for the calibration period and with NSE of 0.82 for the validation period.


**Table 2.** Evaluation statistics for the performance of eighteen scenarios in daily streamflow simulation.

The models using CMADS and CMORPH precipitation data as inputs performed satisfactorily in simulating daily streamflow with NSE larger than 0.55 for both the calibration and validation periods. The models using CMADS precipitation data had larger bias values than those using CMORPH precipitation data, so CMORPH is preferred for daily streamflow simulation. For monthly streamflow simulation, the models using TRMM and CHIRPS precipitation data as inputs performed satisfactorily with all except one NSE value larger than 0.5. These results suggested that different gridded precipitation datasets should be used to achieve optimal results for daily and monthly streamflow simulations. Models using CFSR precipitation data as inputs performed the worst among all considered scenarios. All NSE values were less than 0.50 with even negative values for monthly simulation during the validation period (Table 3), suggesting the unsatisfactory model performance.


**Table 3.** Evaluation statistics for the performance of eighteen scenarios in monthly streamflow simulation.

Although the models using different meteorological datasets had comparable performance, Gauge performed the best and CFSR usually performed better than CMADS, especially on the monthly scale. This may be related to the large bias of CMADS in wind speed (Figure 4). These results suggest that the CFSR meteorological data should be used in this region when gauge measurements are not available.

Figures 7 and 8 compare the measured and simulated streamflow obtained from the CFSR meteorological (excluding precipitation) data and different precipitation data at the daily and monthly timescales. These diagrams give information about the temporal consistency between the measured streamflow and simulated streamflow. In general, most of the models can well reproduce the seasonal changes in streamflow. The consistencies of simulated streamflow and measured streamflow using all the open-access gridded precipitation datasets were worse than those using gauged precipitation data. Underestimation persists in the simulation of the peak flow, especially for the extreme precipitation events such as those in July and September of 2010. This is due to the underestimation of the precipitation amount of these extreme events. CMADS had the best performance among open-access gridded precipitation datasets in capturing the precipitation amounts and streamflow during extreme events, indicating CMADS should be used for the simulation of extreme events if gauge measurements are not available.

It should be noted that the PBIAS values in this study were large with values beyond the generally acceptable thresholds (i.e., −15% < PBIAS < +15%) in most cases (Tables 2 and 3). There may be two reasons for the large biases: firstly, the Nash–Sutcliffe efficiency coefficient (NSE) was adopted as the target of parameter calibration, and PBIAS was not set as an optimization goal. Secondly, paddy fields and ditches are widely distributed in the study area, and the related hydrological processes are not well represented in the SWAT model.

and streamflow during extreme events, indicating CMADS should be used for the simu-

lation of extreme events if gauge measurements are not available.

**Figure 7.** Comparison of daily measured (black line) and simulated (red line) streamflow from models using the CFSR meteorological (excluding precipitation) datasets and six different precipitation datasets for the period 2009–2014, including (**a**) Gauge, (**b**) CFSR, (**c**) CMADS, (**d**) TRMM, (**e**) **Figure 7.** Comparison of daily measured (black line) and simulated (red line) streamflow from models using the CFSR meteorological (excluding precipitation) datasets and six different precipitation datasets for the period 2009–2014, including (**a**) Gauge, (**b**) CFSR, (**c**) CMADS, (**d**) TRMM, (**e**) CMORPH, and (**f**) CHIRPS.

CMORPH, and (**f**) CHIRPS.

**Figure 8.** Comparison of monthly measured (black line) and simulated (red line) streamflow from models using the CFSR meteorological (excluding precipitation) datasets and six different precipitation datasets for the period 2009–2014, including (**a**) Gauge, (**b**) CFSR, (**c**) CMADS, (**d**) TRMM, (**e**) **Figure 8.** Comparison of monthly measured (black line) and simulated (red line) streamflow from models using the CFSR meteorological (excluding precipitation) datasets and six different precipitation datasets for the period 2009–2014, including (**a**) Gauge, (**b**) CFSR, (**c**) CMADS, (**d**) TRMM, (**e**) CMORPH, and (**f**) CHIRPS.

#### CMORPH, and (**f**) CHIRPS. *4.3. Comparison of Calibrated Parameters and Water Balance Components*

It should be noted that the PBIAS values in this study were large with values beyond the generally acceptable thresholds (i.e., −15% < PBIAS < +15%) in most cases (Tables 2 and 3). There may be two reasons for the large biases: firstly, the Nash–Sutcliffe efficiency Tables 4 and 5 show the optimal parameters for the eighteen scenarios calibrated by SWAT-CUP. The scenarios using the same precipitation data (e.g., Gauge, CFSR, CMADS, and CMORPH) and different meteorological (excluding precipitation) data often had almost the same optimal parameters, which showed the dominant role of precipitation to drive

the model. The scenarios using the CFSR meteorological (excluding precipitation) data and two different precipitation datasets (i.e., TRMM and CHIRPS) also had the same parameters, which suggested that the influence of meteorological data also should not be overlooked. The parameters related to channel routing (e.g., ALPHA\_BNK, CH\_K2, and CH\_N2) were relatively close to each other among different scenarios, while there were large differences for parameters related to soil properties (e.g., SOL\_BD, SOL\_K, and SOL\_AWC) and groundwater (e.g., GW\_DELAY). These results suggested that the parameters related to subsurface water movement have large uncertainty because they are difficult to measure in a spatially explicit way.

**Table 4.** Optimal parameters calibrated for the scenarios using the Gauge, CFSR, and CMADS precipitation data. The scenarios using the same precipitation data and different meteorological (excluding precipitation) data often had similar optimal parameters (with the different ones displayed in bold). "Gauge\_P" represents using gauged precipitation data, "Gauge\_M" represents using gauged meteorological data, and so on.


**Table 5.** Optimal parameters calibrated for the scenarios using the TRMM, CMORPH, and CHIRPS precipitation data. The scenarios using the same precipitation data and different meteorological (excluding precipitation) data often had similar optimal parameters (with the different ones displayed in bold). "Gauge\_P" represents using gauged precipitation data, "Gauge\_M" represents using gauged meteorological data, and so on.


Figure 9 shows the comparison of different water balance components of the eighteen modelling scenarios. There are eight subgraphs, each of which represents the values of a water balance component of eighteen modelling scenarios. As we can see, the scenarios using the same precipitation data and different meteorological (excluding precipitation) data usually yielded similar water balance components. The scenarios using the same meteorological data but different precipitation data had an obvious difference in WYLD (water yield, which is the net amount of water that leaves the subbasin and contributes to streamflow in the reach), SUR\_Q (surface runoff contribution to streamflow), GW\_Q (groundwater contribution to streamflow), LAT\_Q (lateral flow contribution to streamflow), and PERC (water percolating past the root zone) components, slight differences in the

precipitation and AET (actual evapotranspiration) components, and almost no difference in the PET components. The scenarios using the TRMM and CHIRPS precipitation data had similar values to those using gauged precipitation data in WYLD, GW\_Q, and PERC, while the scenarios using the CMADS precipitation data had large differences from those using gauged precipitation data in most balance components. The scenarios using the TRMM and CHIRPS precipitation data also had similar spatial distributions in precipitation to those using gauged precipitation data (Figure 10). The scenarios using all the gridded precipitation data can well simulate the pattern that the southwest part of the watershed had the largest WYLD values (Figure S1) because the steep terrain in that region is conducive to runoff generation. *Water* **2022**, *14*, x FOR PEER REVIEW 17 of 23

**Figure 9.** Annual mean water balance components for simulations using eighteen scenarios during

color used the Gauge, CFSR, and CMADS meteorological data from left to right, respectively. AET actual evapotranspiration; WYLD—water yield, the net amount of water that leaves the subbasin and contributes to streamflow in the reach; SUR\_Q—surface runoff contribution to streamflow; GW\_Q—groundwater contribution to streamflow; LAT\_Q—lateral flow contribution to stream-

flow; PERC—water percolating past the root zone.

TRMM, CMORPH, CHIRPS), and the three bars displayed in the same color used the Gauge, CFSR, and CMADS meteorological data from left to right, respectively. AET—actual evapotranspiration; WYLD—water yield, the net amount of water that leaves the subbasin and contributes to streamflow in the reach; SUR\_Q—surface runoff contribution to streamflow; GW\_Q—groundwater contribution to streamflow; LAT\_Q—lateral flow contribution to streamflow; PERC—water percolating past the root zone. *Water* **2022**, *14*, x FOR PEER REVIEW 18 of 23

**Figure 10.** Spatial distributions of annual mean precipitation using different precipitation data during 2009–2014, including (**a**) Gauge, (**b**) CFSR, (**c**) CMADS, (**d**) TRMM, (**e**) CMORPH, and (**f**) CHIRPS. **Figure 10.** Spatial distributions of annual mean precipitation using different precipitation data during 2009–2014, including (**a**) Gauge, (**b**) CFSR, (**c**) CMADS, (**d**) TRMM, (**e**) CMORPH, and (**f**) CHIRPS.

It should be noted that, although the scenarios using the CFSR precipitation data yielded similar WYLD to the scenarios using gauged precipitation data, there are great differences among the water balance components (Figure 9). This shows the limitation of model calibration only based on measured streamflow at the outlet, which is very common in state-of-the-art hydrological modelling, as streamflow is usually the only measured water balance component available in many watersheds. Therefore, the simulation results of individual water balance components calibrated with only outlet streamflow may contain large uncertainty [1,53]. This also highlights that there is a clear need for measurements of other water balance components with the community's effort for innovation in all realms from ground-based to remote sensing [5]. Once such data are more available, multivariable and multisite calibration is a promising way to reduce the afore-It should be noted that, although the scenarios using the CFSR precipitation data yielded similar WYLD to the scenarios using gauged precipitation data, there are great differences among the water balance components (Figure 9). This shows the limitation of model calibration only based on measured streamflow at the outlet, which is very common in state-of-the-art hydrological modelling, as streamflow is usually the only measured water balance component available in many watersheds. Therefore, the simulation results of individual water balance components calibrated with only outlet streamflow may contain large uncertainty [1,53]. This also highlights that there is a clear need for measurements of other water balance components with the community's effort for innovation in all realms from ground-based to remote sensing [5]. Once such data are more available, multivariable and multisite calibration is a promising way to reduce the aforementioned uncertainty.

#### mentioned uncertainty. **5. Discussion**

#### **5. Discussion**  *5.1. Comparison with Existing Studies*

*5.1. Comparison with Existing Studies*  The performance of gridded meteorological and precipitation datasets often varied with location. It was reported that CFSR typically performs well in the United States and South America, but performs poorly for Asia and Africa compared with other products, such as CMADS [54]. For example, Fuka et al. (2014) reported that the stream discharge The performance of gridded meteorological and precipitation datasets often varied with location. It was reported that CFSR typically performs well in the United States and South America, but performs poorly for Asia and Africa compared with other products, such as CMADS [54]. For example, Fuka et al. (2014) reported that the stream discharge simulations forced by CFSR precipitation and temperature data performed as good as or

simulations forced by CFSR precipitation and temperature data performed as good as or better than those forced by data from traditional weather gauging stations in four small

better than those forced by data from traditional weather gauging stations in four small watersheds in the USA [19]. Liu et al. (2018) and Zhang et al. (2020) found CMADS had better performance than CFSR for streamflow simulation in the Yellow River Source Basin, China, and the Muda River Basin, Malaysia [17,55]. While this study found that CFSR meteorological data usually performed better than CMADS, especially at the monthly scale caused by the large bias of CMADS in wind speed, but the CFSR precipitation data performed the worst among all the precipitation datasets evaluated in this study. Gao et al. (2018) also found the performance of CFSR precipitation data was not satisfactory in the Xiang River Basin of China [56], and similar results were also reported in two watersheds in the USA [24,26] and the upstream watersheds of Three Gorges Reservoir in China [30,32]. The model driven by gauged precipitation data together with gauged meteorological data yielded the best performance, which further confirms previous findings that ground measurements are the most reliable input data compared to open-access gridded products [1,16,21].

An important finding of this study is that the CMORPH precipitation data had overall the best performance for daily streamflow simulation among gridded precipitation datasets; TRMM had overall the best performance for monthly streamflow simulation; and CMADS performed the best in capturing the precipitation amounts and streamflow during extreme events. Therefore, different gridded precipitation datasets should be used for different aims. The good performance of CMORPH on the daily scale and that of TRMM on the monthly scale were also reported in the Meichuanjiang watershed of the Ganjiang Basin, China [2], in the Adige Basin, Italy [7], and in Australia [57]. It was also reported that the performances of CMORPH and TRMM were similar in some regions (e.g., Central Thailand) [58]. The capability of CMADS precipitation data in capturing extreme events can be attributed to the incorporation of in situ precipitation measurements from dense weather stations [36].

#### *5.2. Uncertainty of the Evaluation*

It has been widely acknowledged that considerable uncertainties exist in the results of hydrological modeling, which come from model structure, parameters, and inputs [59–61]. To quantify these uncertainties, various approaches, such as GLUE (generalized likelihood uncertainty estimation [62]) and MCMC (Markov chain Monte Carlo [63]), have been proposed. Using these methods, a confidence band instead of a single curve can be obtained to represent the uncertainty of a hydrological variable, such as streamflow. Nevertheless, searching for a single optimal parameter set through model calibration is still the commonly used approach for the inter-comparison of gridded meteorological and precipitation datasets [1,16,28]. In this study, a single-model structure was used following previous researchers, as this approach is relatively simple and at the same time can reasonably reflect the accuracy of different meteorological/precipitation products. In order to obtain more comprehensive and reliable evaluation results, uncertainty analysis should be conducted in the future. In addition, the usage of the differential split-sample test (the DSS-test, e.g., using periods with apparent different climate conditions, e.g., dry/wet or cold/warm, where calibration is performed on one period and validation is performed on another period) will also improve the reliability of evaluation [64].

#### **6. Conclusions**

This study evaluated the performance of two mainstream open-access gridded meteorological datasets (i.e., CFSR and CMADS) and three satellite-based precipitation datasets in driving the SWAT model in streamflow simulation at daily and monthly timescales in the Fengle watershed, Anhui Province, China. Eighteen modelling scenarios, which are generated through the combination of six precipitation datasets (i.e., Gauge, CFSR, CMADS, TRMM, CMORPH, and CHIRPS) and three meteorological (excluding precipitation) datasets (i.e., Gauge, CFSR, and CMADS), were used to study the impact of different

input data on simulation results. After comprehensive comparison of these scenarios, the following conclusions can be drawn:


Currently, the development of data fusion and machine learning techniques provides unprecedented opportunities to design novel methods to reduce the uncertainties of the precipitation dataset. While the focus of this study was to evaluate the performance of existing datasets on hydrological modeling, in the future, novel methods should be developed to construct more accurate precipitation datasets. In addition, it is important to obtain the right simulation results for the right reasons, especially for policy makers. Although the simulation results in this study were generally good based on indicators such as NSE and R<sup>2</sup> , the SWAT model did not adequately represent the impacts of widely distributed paddy fields and ditches on hydrological processes in this study area. In the future, the model should be improved to better characterize these agriculture-related processes and make the simulation results more reliable for decision making.

**Supplementary Materials:** The following supporting information can be downloaded at: https:// www.mdpi.com/article/10.3390/w14091406/s1, Figure S1: Annual mean WYLD (water yield) for simulations using CFSR meteorological data and different precipitation data during 2009–2014, including (a) Gauge, (b) CFSR, (c) CMADS, (d) TRMM, (e) CMORPH, and (f) CHIRPS.

**Author Contributions:** Conceptualization, J.L. and Y.L.; methodology, J.L., L.Y. and Y.L.; software, J.L. and L.Y.; validation, L.Y.; formal analysis, J.L. and L.Y.; investigation, Y.Z. and L.Y.; resources, L.Y.; data curation, J.L. and L.Y.; writing—original draft preparation, J.L. and L.Y.; writing—review and editing, J.L., L.Y. and Y.L.; visualization, Y.Z. and L.Y.; supervision, Y.L.; project administration, Y.L.; funding acquisition, Y.L. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was funded by the National Natural Science Foundation of China (Project No.: 51879130 and 42077449), Discipline Innovation and Talent Introduction Base Project of Colleges and Universities in Henan Province of China (Project No.: CXJD2019001).

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Informed consent was obtained from all subjects involved in the study.

**Data Availability Statement:** The data used to support the findings of this study are included within the article.

**Acknowledgments:** We are thankful to the anonymous reviewers for their valuable comments.

**Conflicts of Interest:** The authors declare no conflict of interest.
