Methodology
This study consists of two parts: in the first (precipitation product
evaluation) we aimed to evaluate the quality of TRMM, IMERG, CMADS, and
CFSR precipitation products at grid and watershed-scales based on GO; in
the second (streamflow simulation evaluation), 12 precipitation
scenarios were created to drive the hydrological model (Table
2 ). Scenarios S1 to S7 were used to study the runoff simulation effect
of each precipitation dataset; Scenarios S8 and S9 were SWAT models
driven by corrected precipitation data to study the influence of
precipitation data correction on runoff simulation (Section
4.2.2 describes the reasons for correcting only CMADS and CFSR
precipitation data). Scenarios S10, S11, and S12 cover CMADS
precipitation data combined with GO1, corrected CFSR precipitation data
combined with GO1, and IMERG precipitation data combined with GO2,
respectively: these were designed to study the effects of precipitation
data combination on runoff simulation (Section 4.2.3 describes
the reasons for choosing these three combinations). The analysis process
used herein is shown in Fig. 2 .
3.1 Precipitation data
evaluation
To quantitatively evaluate the accuracy of the TRMM, IMERG, CFSR, and
CMADS precipitation products in the YRHR, the precipitation derived from
the four precipitation products is directly compared with GO. Six
statistical metrics, including the root mean square error
(RMSE), percent bias (PBIAS),
correlation coefficient (CC), probability of detection (POD), false
alarm ratio (FAR), and critical success index (CSI), were utilized to
evaluate the agreement between the GO and the four precipitation
products. The calculation
equations, units, ranges, and optimal values of the evaluation
indicators are listed in Table 3 .
3.2 SWAT model and model
setting
The SWAT is a semi-distributed, physics-based eco-hydrological model,
which runs in daily, monthly, or annual time steps (Arnold et
al ., 1998), and has been widely used in hydrological processes (Grussonet al ., 2015), soil erosion (Song et al ., 2011), and
nutrient transportation (Wang et al ., 2018). Previous studies
have proven that dividing the YRSR into 25 (Liu et al ., 2018), 29
(Hao et al ., 2013), and 97 (Mengyaun et al ., 2019)
sub-basins would yield reliable simulation results. Therefore, the YRSR
was divided into 26 sub-basins to reduce unnecessary calculation. SWAT
was originally developed to evaluate water resources in large
agricultural basins, and was not designed to model heterogeneous
mountain basins typical of the western United States (Fontaine et
al ., 2002). Ten elevation zones (each covering an change in elevation
of 500 m) were established in the present work, to divide each sub-basin
to reduce the influence of topography on precipitation. According to
previous research (Fontaine et al ., 2002; Zhenchun et al .,
2013), the snowfall temperature (SFTMP), snow melt base temperature
(SMTMP), maximum melt rate for snow during year (SMFMX), minimum melt
rate for snow during the year (SMFMN), snow pack temperature lag factor
(TIMP), and minimum snow water content that corresponds to 100% snow
cover (SNOCOVMX) in the snowmelt module have been adjusted to reduce the
influence of snowmelt on the model (Table 4 ).
3.3 Parameter calibration and model
evaluation
Calibration and uncertainty analyses of the simulation results from the
model were performed using Sequential Uncertainty Fitting Version 2
(SUFI2 ) in the SWAT calibration and uncertainty program (SWAT-CUP)
(Abbaspour et al ., 2015). According to previous studies on
hydrological modeling in alpine basins (Bhatta et al ., 2019;
Mengyaun et al , 2019; Shuai et al ., 2019; Zhenchunet al ., 2013), 30 sensitive parameters were initially selected.
Sixteen parameters with the highest sensitivity were then selected using
the Latin hypercube and one-factor-at-a-time sampling (LH-OAT) method
for calibration (Table 5 ). Due to limitations of space, we do
not present any analysis of the calibration parameters. According to
Abbaspour (2015), the model was calibrated using three iterations with
400 simulations (necessitating a total of 1200 simulations during
calibration) using the Nash-Sutcliffe Efficiency (NSE) (Nash and
Sutcliffe, 1970) and coefficient of determination
(R 2) as the objective function. The range of
each parameter was modified after each iteration, according to both new
parameters suggested by SWAT-CUP and their reasonable physical ranges.
The criteria proposed by Moriasi et al (2015) was adopted to
classify model performance into the respective categories, “very good”
(NSE > 0.80; PBIAS < ±5%), “good” (0.70
< NSE ≤ 0.80; ±5% ≤ PBIAS < ±10%),
“satisfactory” (0.50 < NSE ≤ 0.70; ±10% ≤ PBIAS <
±15%), and “unsatisfactory” (NSE ≤ 0.50; PBIAS ≥ ±15%).
3.4 Precipitation data
pre-processing
Before modeling, we preprocessed the precipitation data:
(1) The numbers of grids or stations with precipitation products of
TRMM, IMERG, CMADS, and CFSR located in the YRSR are 200, 1027, 198, and
122, respectively. Considering that SWAT only uses data from the one
weather station closest to the centroid of the sub-basin (Masih et
al ., 2011; Villarán, 2014). It is impractical to divide the watershed
into 1027 sub-watersheds and correspond thereto on a one-by-one basis.
Therefore, virtual weather stations
were constructed for each sub-basin (Ruan et al ., 2017; Tuoet al ., 2016). The specific methods are as follows:
- Based on the ArcGIS platform, satellite raster or reanalysis station
precipitation data falling in each sub-basin were extracted;
- The arithmetic average method was used to calculate the areal rainfall
of each sub-basin, giving precipitation data pertaining to each
virtual precipitation station;
- The centroid of each sub-basin is the location of the virtual
precipitation station [Fig. 1 (c) ].
(2) Considering that the starting period of SWAT-CUP calibration must be
a whole year, the periods of coincidence of CMADS (1 January 2008 to 31
December 2016) and CFSR (1 January 1974 to 31 December 2014) data are
only six years (1 January 2008 to 31 December 2013), deducting the
warm-up period of the SWAT model (12 years), the final simulation
time will be shorter (45 years), which does not reflect the quality
of the data. Therefore, we added meteorological data from 1 January 2008
to 31 December 2010 and 1 January 2006 to 31 December 2007 for the
warm-up of the SWAT model, so that the data time-span used for the
simulation becomes six years (1 January 2008 to 31 December 2013).