1. Introduction
Precipitation has been recognized as the most critical meteorological parameter in relation to developing hydrological models, because its spatiotemporal variability has a significant impact on hydrological behavior and water distribution [
1,
2,
3]. Previous research studies have illustrated that having less precipitation information uncertainty has a sizable effect on stabilizing model parameterization and calibration [
4,
5,
6]. However, there are severe limitations to describing rainfall inputs’ true spatiotemporal variability of a river basin accurately, such as the rainfall pattern influenced by the complex topography and impacted by a hierarchy of regionally dominated atmospheric cycles [
7,
8].
Precipitation observed from a rain gauge, in general, is considered to be actual rainfall [
9,
10]. In most cases, point rainfall measurements are spatially interpolated to illustrate the rainfall field at a basin scale, and hence they are used as inputs in spatial-distributed hydrological models [
11,
12]. Field rainfall obtained from such interpolation, however, can represent the true distribution of precipitation well only if the rain gauges are deployed with reasonable density and uniform distribution [
13]. Unfortunately, in most areas, especially in remote and developing areas, rain gauges are distributed irregularly and sparsely [
14,
15,
16,
17]. Consequently, the true rainfall field is poorly represented through interpolation, challenging the application of hydrological models. The accidental missing of the ground observations also exacerbate this challenge [
18,
19].
Recently, the feasibility of satellite-based data as alternatives for describing the temporal and spatial variability of the true rainfall field has been frequently tested. For example, Hur et al. [
20] compared two high-resolution satellite rainfall datasets (TRMM 3B42 v7.0 and GSMaP v5.222) with rain gauge observations in Singapore. It was found that TRMM 3B42 v7.0 and GSMaP v5.222 both tended to overestimate the light rain and frequency but underestimate high-intensity precipitation when extreme precipitation was analyzed. Jiang et al. [
21] researched a middle-latitude basin in South China, pointing out that rainfall was overall largely underestimated when using TMPA 3B42RT, Precipitation Estimation from Remote Sensing Information using Artificial Neural Network (PERSIAN), and the NOAA/Climate Precipitation Center Morphing Technique (CMORPH). Duncan et al. [
22] assessed the accuracy of satellite-derived precipitation estimation (TRMM) over Nepal and found that though the precipitation of TRMM was significantly correlated with ground-based observations in all seasons, satellite precipitation estimates consistently overestimated the amount of precipitation and inaccurately detected extreme precipitation events.
The distributed hydrological model is beneficial for understanding the hydrological process [
23,
24,
25]. The most habitually utilized distributed hydrological models have been appeared to effectively consolidate information from rain gauges, whereas satellite-based precipitation has been persistently moved forward and integrated into distinctive modules that assess its execution in simulating watershed streamflow [
26,
27]. The Soil and Water Assessment Tool (SWAT) is the most widely used distributed hydrological model among all the various hydrological models [
28,
29,
30,
31]. Huang et al.’s [
32] study in the German state of Baden-Württemberg used three precipitation datasets with different time scales (daily, sub-daily, and diurnal) as inputs to drive a SWAT model to simulate the runoff, and found that there is a positive correlation between model performance and higher precipitation resolution. Yeganantham et al. [
33] found that Climate Hazards Group InfraRed Rainfall with Station (CHIRPS) performed better than Climate Forecast System Reanalysis (CFSR) in simulating streamflow when using the SWAT model in ten watersheds located in the USA, Brazil, Spain, Ethiopia, and India. Hamoud et al.’s research [
34] showed that the applicability of CHIRPS and TRMM 3B42 in runoff simulations were better than that of CFSR, Artificial Neural Networks–Climate Data Record (PERSIANN-CDR), and European Atmospheric Reanalysis (ERA-5) in the Highland Region of Yemen. Moreover, the performances of the satellite-based data are various in different areas. For example, Mararakanye et al.’s research in the lower Vaal River Catchment area (South Africa) [
35] found that the CFSR performed well in simulating runoff by using a hydrological model, while according to Dao et al.’s study in the Cau River Basin (North Vietnam) [
36], the performance of CFSR in runoff simulation was unsatisfactory. Gao et al. [
37] proved that the performance of PERSIANN-CDR as an input to drive the SWAT model to simulate runoff was not suitable for the Xiang River Basin (China); however, its performance when simulating runoff was good in the Lancang River Basin (China). Like the studies above, the results simulated using the data-based SWAT model are heterogeneous and the performance of satellite-based datasets to simulate runoff should be evaluated for the specific basin.
Originating from the Q-DM, the Danjiang River Basin (DRB) is the main water source of the central route projects of the South-to-North Water Diversion Project [
38]. This project is one of the most important hydraulic engineering projects in China and aims to improve the water shortage problem in northern China and improve the ecological environment along the related region. The quantity and quality of the water delivered are influenced by the erosion of the DRB [
39,
40]. Therefore, the DRB is considered to be a sensitive area in terms of water quality and safety with regard to the watershed. However, the uneven distribution of the meteorological stations in Q-DM makes it difficult to understand the real hydrological process. A previous study [
41] used CFSR-driven SWAT models to simulate the runoff in the Bahe River Basin (Q-DM area) and found that the runoff simulated by uncorrected CFSR data were only satisfactory in this basin, while corrected data performed better. This indicates that it is necessary to verify the applicability of meteorological data in the DRB (Q-DM area).
Here, this study explores the results of the CMADS, TRMM, and rain gauge data when simulating rainfall estimation and surface runoff at monthly and daily scales in the DRB. The study aims to verify the applicability of the CMADS data and TRMM data in the DRB, and it can, therefore, serve as a reference for choosing the precipitation datasets in watersheds similar to the DRB where the ground-based rain gauge data are unavailable. With the objectives above, this study involves (1) a comparison of rainfall estimations from CMADS, TRMM 3B42 data, and rain gauge observations (Gauge) at monthly, daily, and spatial scales, (2) setting up a SWAT model with CMADS, TRMM 3B42 data, and rain gauge observations to simulate monthly and daily runoff, (3) calibrating and validating the simulated streamflow at three hydrological stations using the SWAT Calibration Uncertainties Program (SWAT-CUP) which uses the Sequential Uncertainty Fitting ver.2 (SUFI-2) algorithm, and (4) evaluating the multi-statistical performance of the simulation against the observed streamflow data. The main goal of this study is to evaluate the use of satellite-based and reanalysis precipitation products as model operation driving data, and assess whether they can drive the model in a watershed similar to the DRB where the gauge observations are limited.
2. Materials and Methods
2.1. Study Area
The largest tributary of the Hanjiang River, the Danjiang River, is a mountain river that covers a drainage area of 8887 km
2. The total length of its main stream is 280 km. Originating from the South Qinling Mountains and flowing into the Hanjiang River [
42], the Danjiang River flows through the Shaanxi, Henan, and Hubei Provinces. It stretches between 33°04′10″ N and 34°11′09″ N and across 109°30′08″ E and 111°15′51″ E. The Danjiang River Basin (DRB) features a high-rising west and a low-lying east, with a relative elevation difference of 1915 m. The continental monsoon climate contributes to the distinct seasons of the DRB. According to the records from 1950 to 2015, the long-term annual precipitation of the DRB is 732.29 mm and the spatial distribution difference shows an increasing trend from the west to the east. Rainfall is concentrated in the period from May to October, accounting for about 80% of the annual precipitation. Moreover, the annual average temperature ranges from 7.8 °C to 13.9 °C and the annual runoff is 14.36 × 108 m
3.
Forestland occupies the largest area in the DRB, followed by the cropland. The yellow-brown soil and sandy loam are the dominant soil types in the DRB [
38]. There are 3 hydrological stations (Majie Station upstream, Danfeng Station midstream, and Jingziguan Station downstream) and 58 ground-based rain gauges in the study area. The digital elevation model (DEM), stream network, weather stations, and hydrological stations are shown in
Figure 1.
2.2. Hydrological Model and Data Sources
In this study, the SWAT model was used for hydrological modeling, which was developed by USDA-ARS. Because the SWAT model is designed for long-term simulations on a daily scale, it is suitable for evaluating the performance of three precipitation products. To ensure the accuracy of relative changes induced by different precipitation inputs, all input parameters, such as temperature, wind, solar radiation, and humidity, were kept the same, except precipitation. Additionally, the temperature, wind, solar radiation, and humidity inputs were simulated by the internal weather generator of SWAT.
Moreover, the target watershed is required by the SWAT model to be divided into sub-watersheds. Each sub-watershed may include one or more Hydrologic Response Units (HRUs). On the basis of the 30 m DEM and by choosing the Jingziguan Station as the outlet, the controlled watershed was delineated. The threshold to discretize the sub-watershed was based on the 2% area. Other input parameters, such as soil type and land use, were downloaded from websites (
Table 1). The data of measured runoff were obtained from the Department of Hydrology of the Ministry of Water Resources of China. Additionally, the SPWA (Soil–Plant–Air–Water) software was used to analyze the soil–water characteristics of each soil type.
Daily rainfall data were collected from weather gauge stations and the two satellite-based and reanalysis precipitation products used were CMADS and TRMM 3B42 version 7.
Daily precipitation data obtained from the fifty-eight rain gauges in the DRB were available from the website of the Department of Hydrology of the Ministry of Water Resources of China. The rain gauge data covered from 1964 to 2015.
The dataset CMADS introduces the technology of The Space and Time Mesoscale Analysis System (STMAS) assimilation algorithm. Multiple technologies and scientific methods were used to develop CMADS [
43,
44]. The dataset, containing information relating to precipitation, temperature, and other variables, can be used to run hydrological models such as SWAT. The precipitation data of CMADS are merged with the hourly precipitation data collected by the China National Meteorological Information Center using the CPC MORPHing technique (CMORPH). CMADS stations provide information throughout the day from 2008 to 2016 in the areas between 0–65° N and 60–160° E. There are a total of 19 CMADS stations in the study area.
In late 1997, the TRMM satellite was launched by the National Aeronautics and Space Administration (NASA) and the Japanese Aerospace Exploration Agency (JAXA) to monitor precipitation [
45]. TRMM 3B42 is one of the RMM Multi-satellite Precipitation Analysis (TMPA) products [
46]. It provides daily precipitation data from 50° S to 50° N at a resolution of 0.25° spatially and temporally from 1998 to 2015 [
47,
48]. There is a total of 19 TRMM 3B42 pixels in the study area. Further information about TRMM and CMADS can be found in
Table 2.
The SWAT model uses data from the station nearest to the centroid of each sub-basin to categorize precipitation data into sub-basins [
47].
2.3. Model Calibration and Evaluation
When all parameters were entered into the SWAT model, the SWAT model ran with three precipitation products (rain gauge data, CMADS dataset, and TRMM dataset) separately at the monthly and daily scale. The watershed was divided into a total of 237 sub-catchments by the SWAT model, and these sub-watersheds were further divided into 980 HRUs on the basis of the land use, soil type, and slope classes. The simulated period was selected to be the period from 2008 to 2015 to ensure its consistency, because the available gauge data, CMADS data, and TRMM data were, respectively, collected from 1964 to 2015, 2008 to 2018, and 1998 to 2015. Here, 2008 was taken as the warm-up period.
The SUFI-2 algorithm in SWAT-CUP was used in the calibration procedure. On the basis of Duan et al.’s research [
48] and the official guide, 17 parameters were selected. Considering the influence of elevation on precipitation, the precipitation lapse rate (PLAPS) was introduced [
49]. Moreover, the simulated results of the Majie Station, Danfeng Station, and Jingziguan Station were calibrated together. The model was calibrated by first using the initial value ranges of each parameter and then using the suggested ranges of the previous simulation. The simulations were calibrated five times with 500 iterations each.
In this study, the coefficient of determination, Nash–Sutcliffe efficiency (NSE), and percent bias (PBIAS) were used to evaluate the accuracy of runoff modeling results. The formulas are as follows [
35]:
where
Qi is the observed value,
Si is the simulated value, and
and
are the mean values of the observed and simulated values. The statistical threshold values that were used to evaluate the performance of the model are shown in
Table 3.
4. Discussion
Precipitation inputs play an important role in runoff simulation, and the errors can influence the accuracy of the hydrographical outputs [
52]. Generally, precipitation inputs are evaluated on the basis of their predictable performances with hydrological parameters at the watershed scale, which avoids the scale difference found when using ground-based observations for validation [
53]. This study evaluated the performance of RO and satellite-based precipitation datasets (CMADS and TRMM) in driving the SWAT model to simulate streamflow in the DRB on both monthly and daily scales. All modeling scenarios were calibrated and validated against runoff data measured at Majie Station, Danfeng Station, and Jingziguan Station. The SWAT-Cup’s SUFI-2 algorithm was used for calibration and validation. Indices, including NSE, R
2, and PBIAS, were selected to evaluate the efficiency of simulation runoff outcomes.
We found that compared with rainfall gauge observations, TRMM tended to underestimate the precipitation on both monthly and daily scales, while CMADS tended to overestimate the rainfall on the monthly scale but understate the rainfall on the daily scale. These findings are similar with previous studies. For example, Jiang et al. [
21] found that TRMM underestimated precipitation on a daily scale in the Mishui Basin, Jiang et al. [
47] found an average bias of −20.5% for CMADS over Xixian Basin, and Guo et al. [
38] calculated an average bias of −28.7% for CMADS over Jinhua River Basin. The reason why CMADS underestimated the precipitation was the underestimation of the background field CMORPH data [
54]. Additionally, the ability of TRMM and CMADS to identify the torrential rain events was worse than that of Gauge. In summary, the performance of CMADS in precipitation simulation was better than that of TRMM, maybe because the correct process of TRMM was simpler than that of CMADS [
55]. The spatial distribution of the precipitation varied from dataset to dataset, namely, the rainfall of CMADS increased from the center to the surroundings and its rainfall in the central north was the highest. TRMM increased from upstream to downstream and the highest rainfall occurred in the east, but there is no obvious spatial distribution pattern with the rainfall of Gauge. This result can be explained by the different distributions of the rain gauge. In addition, all meteorological data were categorized into each sub-basin by the “nearest-distance” principle in the SWAT model [
49], which contributed to the difference in the precipitation data from CMADS, TRMM, and Gauge as well. Moreover, the similarity between the rainfall of CMADS and Gauge was higher than that of the TRMM on both the monthly and daily scales, which is consistent with Wang’s [
56] research in the Ganjiang River Basin, where the area and elevation are similar to the DRB. Wang et al. [
56] found that CMADS performed better than TRMM in precipitation estimation because of the different development processes of these two products. Only 500 stations were used to correct the TRMM data, while 2421 stations were used to correct the CMADS data [
57,
58]. Song et al. [
59] conducted research on the Qujiang River Basin, (38,900 km
2) finding that the spatial distribution of CMADS and TRMM was different from that of Gauge, which is consistent with our study.
Pre-calibration results showed that CMADS and TRMM were reliable enough to estimate runoffs on the monthly scale at Majie Station and Jingziguan Station, while they performed unsatisfactorily in simulating streamflow at Danfeng Station. The performances of Gauge in estimating runoff on the monthly scale in the middle stream and downstream were both unreliable, and only its performances in runoff simulation at the Majie Station was satisfactory. The performances of that on the daily scale, however, were all unsatisfactory. Moreover, all three datasets seriously overestimated the runoff in the middle stream and downstream on the monthly scale. Moreover, underestimation is probably better attributed to poor representation of the spatial variability of precipitation patterns in the middle and downstream, thereby causing the low ratio of streamflow to precipitation. According to Vu et al.’s [
60] research, the underestimation can be attributed to the spatiotemporal uncertainty of the precipitation inputs. Previous studies indicated that the spatiotemporal uncertainty of the catchment rainfall was one of the main sources of uncertainty in runoff simulation using rainfall–runoff models [
61,
62,
63]. Additionally, satellite-based precipitation estimations have their own uncertainties [
13]. This means that the satellite-based rainfall estimation affects the runoff simulation significantly [
64]. Moreover, Rivera’s research [
65] found that previous conditions were more important before extreme floods, while previous conditions had little effect on conditions after extreme floods. It explains the result that the average observed streamflow was lower than the average simulated runoff and the average measured runoff was higher than the average observed runoff throughout the year in the upstream, except for September, because the underestimation in September in the Majie Station is the continuous precipitation before an extreme flood occurred in September.
When it comes to the post-calibration results, all three products tended to overestimate the runoff in the upper and middle reaches and underestimate the downstream at the monthly scale, while all three products overestimate the streamflow upstream and underestimate the runoff in the lower and middle streams at the daily scale. The overestimation at the monthly scale and the underestimation at the daily scale may be due to the overall inaccurate estimation of precipitation with the CMADS and TRMM data. In most cases, RO performed better than satellite precipitation in runoff simulation, even in sparsely gauged areas, when SWAT was used for modeling both monthly and daily scales, such as Vu et al.’s [
60] research. Namely, they tested the accuracy of four satellite precipitation products, including TRMM 3B42 V7, PERSIANN, PERSIANN-Climate Data Record (PERSIANN-CDR), and CMADS, by using these four products to drive the SWAT model and comparing the runoff simulation results with the runoff simulated by gauged rainfall data in the Han River Basin in South Korea. Their results illustrated that the application of TRMM and CMADS in runoff simulation was worse than that of the gauges. However, our results vary from theirs. It is found in this paper that CMADS-SWAT was superior to the other two precipitation products in both monthly and daily runoff simulations, but Gauge-SWAT performed the worst in both monthly and daily streamflow simulations. This finding is not uncommon; for example, Song et al.’s [
59] research on the Qujiang River Basin (38,900 km
2) proved that the CMADS-SWAT performed best in the whole basin, followed by TRMM-SWAT and Gauge-SWAT, which performed the worst. That was mainly because of the non-uniform distribution of the gauges. According to Wang et al.’s [
66] research, when the number of the stations is similar, the more uniform distribution of rainfall stations, the greater the NSE. In this study, the distribution of the Gauge was the most nonuniform, causing the performance of Gauge-SWAT to be the least satisfactory among the three products. Moreover, the gauge data only represent the observed rainfall at a specific station, whereas the CMADS and TRMM data represent precipitation averages over a large area [
67]. For the variation of the topography that causes the precipitation variations over a short distance, the heterogeneity of the landscape of the weather forecasts by CMADS and TRMM is better. In addition, CMADS is a combination of the gauge and satellite data; therefore, its accuracy is higher than that of Gauge and TRMM.
At present, few studies have compared the performance of CMADS and TRMM data by using the SWAT model, because CMADS only covers East Asia and is a newly released dataset. Additionally, most studies are focused on evaluating the performance of CMADS, CHIPRS, and CPC data or only studying the applicability of CFSR data by using the SWAT model [
35,
68]. Using different models or inputting different parameters will cause different results [
69]. Therefore, the satellite or satellite-based products (including CMADS and TRMM) can be applied to the SWAT model or other models in the future to ensure that their replications of runoff are accurate and their predictions of rainfall are credible. Nonetheless, the datasets evaluated in this study can serve as viable alternatives in watersheds similar to the DRB where the observed precipitation data are unavailable.