ELECTRONIC OFFPRINT

Use of this pdf is subject to the terms described below

This paper was originally published by IWA Publishing. It is an Open Access work,
and the terms of its use and distribution are defined by the Creative Commons
licence selected by the author.

Full details can be found here: http://iwaponline.com/content/rights-permissions

Please direct any queries regarding use or permissions to [email protected]

 340 © 2021 The Authors Journal of Hydroinformatics | 23.2 | 2021

Hydrological model optimization using multi-gauge

calibration (MGC) in a mountainous region
Sead Ahmed Swalih and Ercan Kahya

ABSTRACT

It is a challenge for hydrological models to capture complex processes in a basin with limited data Sead Ahmed Swalih (corresponding author)
Ercan Kahya
when estimating model parameters. This study aims to contribute in this field by assessing the Civil Engineering Department,
Istanbul Technical University,
impact of incorporating spatial dimension on the improvement of model calibration. Hence, the main Istanbul,
Turkey
objective of this study was to evaluate the impact of multi-gauge calibration in hydrological model
E-mail: [email protected]
calibration for Ikizdere basin, Black Sea Region in Turkey. In addition, we have incorporated the
climate change impact assessment for the study area. Four scenarios were tested for performance
assessment of calibration: (1) using downstream flow data (DC), (2) using upstream data (UC), (3)
using upstream and downstream data (Multi-Gauge Calibration – MGC), and (4) using upstream and
then downstream data (UCDC). The results have shown that using individual gauges for calibration (1
and 2) improve the local predictive capacity of the model. MGC calibration significantly improved the
model performance for the whole basin unlike 1 and 2. However, the local gauge calibrations
statistical performance, compared to MGC outputs, was better for local areas. The UCDC yields the
best model performance and much improved predictive capacity. Regarding the climate change, we
did not observe an agreement amongst the future climate projections for the basin towards the end
of the century.
Key words | climate change, hydrological modelling, multi-gauge calibration, SWAT

HIGHLIGHTS

• Four calibration techniques were investigated: (1) downstream calibration (DC), (2) upstream
calibration (UC), (3) calibration with both upstream and downstream data (MGC), and (4)
calibration with first from upstream data, then downstream data sequentially (UCDC).

• UC and DC improve the predictive capacity of the model only for the region where calibration
data is used. They gave a better statistical performance for the particular region compared with
MGC and UCDC.

• MGC calibration significantly improved the model performance for the whole basin unlike the
local gauge calibrations.

• UCDC technique gave best model performance, where the model performance improved for the
whole study area far better than the other calibration techniques.

• No agreement observed amongst future climate projections for the study area.

This is an Open Access article distributed under the terms of the Creative
Commons Attribution Licence (CC BY-NC-ND 4.0), which permits copying
and redistribution for non-commercial purposes with no derivatives,
provided the original work is properly cited (http://creativecommons.org/
licenses/by-nc-nd/4.0/).

doi: 10.2166/hydro.2020.034
 341 S. A. Swalih & E. Kahya | Evaluating hydrological model calibration techniques in a mountainous region Journal of Hydroinformatics | 23.2 | 2021

INTRODUCTION

Hydrologic models have been widely used in hydrological addition, the impact of the model calibration techniques
research and water resource management studies. Hydrolo- on the hydrological modeling of the basin and climate
gical models are increasingly used to simulate changes in change impact assessment will be studied.
the basin management, not only for investigating the
impacts of external influences, but also for discovering
the impacts of future data series (observing climate LITERATURE REVIEW
change, drought and flood studies) (Zhang et al. ).
The physical hydrological model, namely Soil and Water A number of attempts have been conducted by hydrologists
Assessment Tool (SWAT), is a river basin scale model to incorporate model parameters spatially to improve model
requiring specific information about weather, soil proper- accuracy. In 2005, a multi-variable and multi-site approach
ties, topography, vegetation, and land management to the calibration and validation of the SWAT model for
practices occurring in a basin (Arnold et al. ; Neitsch the Motueka catchment, New Zealand, was conducted
et al. ; Winchell et al. ). It has gained international and found that the method improved the model calibration
acceptance as a robust interdisciplinary basin modeling and validation (Nash–Sutcliffe coefficient of 0.78 and 0.72
tool. However, calibration has been one of the major chal- respectively) (Cao et al. ). In another study, spatially dis-
lenges in developing hydrological models in such a way tributed calibration at sub-basin level and temporal
that simulations will be in good agreement with observed validation at the stream gauges outlet points was conducted
flows. The SWAT model calibration has been improving to incorporate the spatial and temporal hydrological pat-
due to continuing advancements in the calibration tech- terns for two river basins in the Ohio, USA, region and
niques (Van Griensven et al. ; Arnold et al. a, found out that spatially distributed calibration and vali-
b; Ercan et al. ). dation of the basins improve the model predictive capacity
In past decades, many calibration methods have been (Santhi et al. ). Similar studies conducted by Lu et al.
developed for SWAT, including manual calibration and (), Bai et al. () and Zhang et al. () in the Yin-
automated procedures using the shuffled complex evol- gluoxia, Miyun and Baihe basins, China, comparing single-
ution (SCE) method and others. Arnold et al. (a, and multi-site calibration and validation of the SWAT
b) developed the SWAT-CUP program that facilitates model, showed that the parameters determined from multi-
the calibration and validation of SWAT model using the site calibration and validation were far better than those
semi-automated approach (SUFI2) (Abbaspour ). The from the single site calibrations. Haas et al. () studied
sensitivity analysis of SWAT-CUP helps check model par- a joint multi-metric calibration of flow simulated by SWAT
ameters that have a high impact on the simulations and in Treene basin, Germany. The results showed that the
make the necessary changes accordingly and model par- approach used has improved the calibrated predictions,
ameters must only be calibrated in a reasonable range where adequate model runs with good performance for
that could give meaning on the ground (Van Griensven different hydrological conditions for the flow were detected.
et al. ). For calibration studies, it is important to Contrary to the above, Shresthaa et al. () indicated that
spatially account for hydrological processes in order to multi-site calibration did not improve simulations of flow
improve the efficiency of model run time and incorporate and sediments compared to single-site calibration.
the impact of uncertainty in the model (Klemes ). This Changes have occurred in several aspects of the atmos-
task is one of the major challenges that enhances SWAT phere and surface that alter the global energy budget of the
model calibration. Therefore, the aim of this study is to Earth causing the climate to change (Solomon et al. ).
investigate the impact of incorporating hydrological pro- The Intergovernmental Panel on Climate Change (IPCC),
cesses spatially in order to improve model calibration. In set up by the World Meteorological Organization (WMO)
 342 S. A. Swalih & E. Kahya | Evaluating hydrological model calibration techniques in a mountainous region Journal of Hydroinformatics | 23.2 | 2021

to provide information on climate change, has confirmed (UCDC). We analyzed the impact of these techniques on
with high confidence that the gradual increase in the aver- the hydrologic regime of the basin. Hydrological model
age temperature of the Earth’s surface is due to the simulations and statistical analysis, as well as climate
growing concentration of greenhouse gases (GHGs) in the change impact assessments, have been used in the study.
atmosphere, attributed to burning of fossil fuels and changes
in land use and land cover (IPCC ). Today, Regional Cir-
culation Models (RCMs) are applied to determine the likely STUDY AREA AND DATA
effects of these changes in the GHGs. It is to be mentioned
here that a GCM/RCM is a Global/Regional Circulation The Rize province is located in the northeastern Anatolian
Model that could depict different pictures for the future mainland of Turkey, so-called the eastern Black Sea
weather over a region, and consequently affect the hydrolo- region, with an area of 3,920 km2. The average annual pre-
gic regime of basins (Elshamy et al. ). Increases have cipitation in the region is nearly 2,250 mm and total
occurred in the number of heavy precipitation events. The annual runoff is about 2,745 million m3 based on the refer-
availability of observational data restricts the types of ence period 1960–1996 (Sen & Kahya ). The study
extremes that can be analyzed. Confidence in these esti- area is a mountainous region with elevations reaching
mates is higher for some climate variables (e.g. higher than 3,000 m (Figure 1) and a very wet climate
temperature) than for others (e.g. precipitation). (Kahya et al. a). In addition, Rize province falls into
Some studies have been conducted to investigate the the homogeneous streamflow region (Kahya et al. b).
effects of multiple stream gauging stations’ data on basin-
wide calibration and validation in Turkey. The impact of Digital elevation model (DEM), soil and land use maps
size of sub-basins on the hydrologic parameters and their
spatial variability in the estimation of the hydrologic par- The Digital Elevation Model (DEM), the soil map, and the
ameters in the western Black Sea region were studied by land use map of the basin were obtained from Istanbul Tech-
Kocyigit et al. () and Akay et al. (). The studies indi- nical University (ITÜ) – Hydraulics and Water Resource
cated that multi-site calibration and validation resulted in Engineering Laboratory. The digital elevation model
satisfactory outcomes for the direct flow hydrograph but (DEM) with 28-meter spatial resolution was used which is
not the peak flow prediction. These studies have only quite sufficient for hydrological modelling studies. A total
focused on the application of hydrological models, and of 22 sub-basins were identified after delimiting the basin
very limited attempts were made in the model parametriza- boundary in ArcSWAT (Figure 1).
tion and calibration technique studies for basins in Turkey.
Therefore, there is an obvious need to assess the impact of
spacial dimension and calibration and validation techniques DATA AND METHODOLOGY
in hydrological modeling.
Consequently, the main objective of this study is to Weather and hydrological data
evaluate the impact of multi-site calibration on the perform-
ance of hydrological model calibration and validation. To The observed precipitation data was accessed from the Turk-
achieve this, we used four different calibration techniques: ish Meteorological Directorate (MGM). We used the
(1) calibration using gauge data only from the downstream maximum and minimum temperature data from the Climate
area (DC); (2) calibration using gauge data only from the Forecast System Reanalysis (CFSR) which was proven to
upstream area (UC); (3) calibration using gauge data from have good accuracy with the observed temperature (Duana
both the upstream and downstream areas of the basin et al. ). The precipitation data in the vicinity of the Ikiz-
(Multi Gauge Calibration – MGC); and (4) calibration dere basin was used for the SWAT simulations. In addition,
using gauge data first from the upstream area only, then the daily flow data of three flow gauging stations found in
using the data from the downstream area sequentially Ikizdere basin were received from the Turkish General
 343 S. A. Swalih & E. Kahya | Evaluating hydrological model calibration techniques in a mountainous region Journal of Hydroinformatics | 23.2 | 2021

Figure 1 | The Rize Province, delineation of the study area, and sub-basins of Ikizdere basin.

Directorate of State Hydraulic Works (DSI) (for the years from the official IPCC data repository (www.ipcc-data.org/
1971–1996) to carry out the model calibration and vali- sim/gcm_clim/SRES_AR4/index.html)
dation. The flow gauging stations of Camlıkderesi or
Ikizdere (station no. E22A015) and Tözköyderesi (station Model calibration and sensitivity analysis
no. E22A033) were used to assess the impacts of spatial
streamflow data on the overall model optimization perform- SWAT is a river basin scale model developed to predict the
ance. The third gauge of Cimil-Koknar (station no. impact of basin management on water, sediment, nutrients,
D22A096) in Ikizdere basin was insufficiently small to be and agricultural and chemical yields (Arnold et al. ). In
used for the study. After failing to simulate the Kaptanpaşa this study, the ArcSWAT program was used in setting up the
basin of Rize province within acceptable accuracy, due to model for the study area. The observed precipitation data
a shortage of observed climate and streamflow data, the Ikiz- from the Turkish Meteorological agency stations was
dere basin was selected for our study as sufficient data is employed in the SWAT model calibration. The model uses
available for hydrological modelling, which is also rec- only one climate station in each sub-basin when a single
ommended by Kahya et al. (). Moreover, the mean station is available within the sub-basin; otherwise, it
monthly values of temperature, wind speed, solar radiation demands the mean data of nearby stations. Once the deli-
and humidity were generated from the Climate Forecast neation of the basin is completed, Hydrological Response
System Reanalysis (CFSR) data, which can be accessed Units (HRU) were categorized with respect to the soil,
from the website of US National Centers for Environmental land use and slope maps overlay. Categorizing all the similar
Prediction (NCEP-CFSR). The climate change data for the parcels of land with similar hydrology, SWAT employs a
three climate models and two scenarios were downloaded lumped calculation for parameters instead of a distributed
 344 S. A. Swalih & E. Kahya | Evaluating hydrological model calibration techniques in a mountainous region Journal of Hydroinformatics | 23.2 | 2021

calculation for the sake of saving time and resources. Once the relative sensitivities based on the linear approximations,
the HRU and sub-basin analysis is completed, the ArcSWAT providing only partial information about the sensitivity of
interface then inputs climate data (such as precipitation and objective function to the model parameters. In general, the
temperature) to finalize the SWAT model setup. The two larger the absolute value of t-stat and the smaller the
streamflow series in Camlıkdere station located at the down- p-value, the more sensitive is the parameter. The most sensi-
stream and Tözköydere station located at the upstream of tive model parameters are essential in model calibration as
the study area were used for the model calibration and these parameters will affect the model more than less sensi-
validation. tive parameters. The four model calibration methods were
The SWAT model setup based on the study area was simu- tested for performance. In each method, a set of five cali-
lated using precipitation data for the period 1971–1996. The bration steps were conducted until the simulated
observed precipitation was assumed the only dominant cli- streamflow at a desirable performance is attained as com-
mate variable that has direct and significant impacts on pared to the observed streamflow. It is important to note
streamflow downstream. The simulations were run on a that each calibration step consists of 500 iteration runs. The
monthly time step and a warm-up period of three years for final step was validation to evaluate the performance of the
the model to adjust itself according to the data and basin con- calibrated model with a separate precipitation forcing.
ditions. The output of the model in the warm-up period was not
used for further analysis. Following successful setting pro-
cedures for SWAT, it was calibrated and validated using the Performance analysis
remaining historical streamflow data. More specifically, the
first three years (1971–1973) were used in the model warm- To evaluate the performance of the model calibration tech-
up, and the next 12 years (1974–1990) in the model calibration, niques, the following five statistical methods were used:
and finally the remaining six years (1991–1996) for the model (i) Nash–Sutcliff Efficiency (NS), which measures the close-
validation. The model was calibrated using the following four ness of the simulated value to the observed value (Equation
methods: the first one using Ikizdere streamflow data (Down- (1)); (ii) Coefficient of Determination (R 2) (Equation (2));
stream Calibration – DC), the second using Tözköydere (iii) Root Mean Square Error (RMSR) (Equation (3));
streamflow data (Upstream Calibration – UC), the third using (iv) Percentage Bias (PBIAS) (Equation (4)), and finally
both Ikizdere and Tözköydere streamflow data (Multi-Gauge (v) p-factor (the percentage of observed time series covered
Calibration – MGC), and finally the fourth calibration using by the 95% prediction uncertainty, or in short 95PPU),
Tözköydere streamflow before the calibration using Ikizdere which is a measure of uncertainty in the model (input
streamflow data (Upstream Calibration before Downstream data, parameter, model, etc.). The NS, a major objective
Calibration – UCDC). From the review of relevant literatures, function criterion, was used to assess the degree of fitness
the technique employed in this research is quite different from exhibited by the simulation with that observed. These
what has been done in the past (at least for the Black Sea methods are discussed by Moriasi et al. () in detail. In
region). Thus, it will have a significant contribution in high- addition, Molina-Navarro et al. () advised the use of
lighting the significance of incorporating spatial dimension NS as an objective function when addressing a multi-site
in hydrological model parametrization studies. and multi-variable calibration studies:
In the phase of model calibration and sensitivity analysis, 2 3
Pn
n 2
we adopted the SWAT-CUP auto-calibration tool (Abbaspour 6 (Yi  Yi ) 7
obs
6 7
). The Latin Hypercube One-factor-At-a-Time design NS ¼ 1  6 ni¼1 7 (1)
4 P obs mean 2 5
(Yi  Yi )
(LH-OAT) method proposed by Morris () was i¼1
implemented to carry out the sensitivity analysis method 2 32
P P P
which uses a stratified sampling approach that better covers 6 n (Gi Si )  ð Gi Þð Si Þ 7
R2 ¼ 6  P    ffi7
4rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 5 (2)
the sampling hypercube with fewer samples (White & P P 2  P
n G2i  ð Gi Þ2 n Si  ð Si Þ2
Chaubey ; Van Griensven et al. ). The results are
 345 S. A. Swalih & E. Kahya | Evaluating hydrological model calibration techniques in a mountainous region Journal of Hydroinformatics | 23.2 | 2021

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi each month of the year in the SWAT input/output files

P n
(Yiobs  Yin )
2
(TxtInOut folder). We copied the ‘swat.exe’ executable file
RMSE i¼1
to run the SWAT model from the TxtInOut folder after
RSR ¼ ¼ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi (3)
STDEVobs P obs
n
(Yi  Y mean )
2 making the necessary change in all the sub-basin (.sub) files.
i¼1

2 3
P
n
6 (Yobs  Yni )7
6
i
7 RESULTS AND DISCUSSION
PBIAS ¼ 1006i¼1 n 7 (4)
4 P 5
(Yobs
i )
i¼1 In this study, we used the SWAT model using the SWAT-
CUP model optimization tool to assess our four model
where Yiobs is the ith observation; Yin is the ith simulated value; optimization techniques. In addition, we assessed the
Yimean is the mean of observed rainfall data; Gi is ground pre- impact of climate change for the study area.
cipitation measurements; Si is satellite precipitation estimates,
and n is the total number of data pairs or observations.
Model calibration and sensitivity analysis

Climate change assessment The SWAT model of the study area was set-up and calibrated
using the observed flow data. The output of the model cali-
In this study, three regionally downscaled GCM climate bration, fitted parameter values and ranks of the sensitivity
models (MIMR, INCM3 and HADCM3) and two scenarios analysis are illustrated in Supplementary Material, Annex 1.
(A2 and A1B) were used. It is evident that the three climate Some of the parameters are significantly affected by the stream-
models are not consistent in their estimation of future pre- flow stations used in the model calibration. CH_N2 and CN
cipitation change. Except for the A1B scenario of INCM3, parameters have shown to be sensitive, which indicated the
all the other climate change scenarios predicted the same direct impact of river channel characteristics on streamflow.
or decreased precipitation across the seasons. The summary In addition, SFTMP and SNO50COV were found to be
of IPCC climate change data used in our study is presented sensitive which indicates the effect temperature on the flow
in Figure 2. The values indicate the projected precipitation regime of the basin. The majority of model parameters are
in mm at the end of the century (2080–2099) with respect less sensitive when it comes to the location of gauge data
to the baseline period (1961–1990). We implemented the cli- used for calibration.
mate change analysis by directly substituting the percentage Irrespective of the four model calibration techniques
change in precipitation forecast by the climate model for used, the best parameter values fitted for each calibration

Figure 2 | The three climate change models and two scenarios of IPCC used in the study.
 346 S. A. Swalih & E. Kahya | Evaluating hydrological model calibration techniques in a mountainous region Journal of Hydroinformatics | 23.2 | 2021

technique are quite similar along the calibration techniques

6.1 (6.5)

6.3 (7.2)
used. The only exceptions are: CN2, CH_K2, LAT_TTIME,

Valid.

0.84
0.84
0.40
6.20
0.42
ALPHA_BF, GWQMN, PLAPS, SLSOIL and SOL_K.

Tözköydere

6.4 (6.4)

6.7 (6.7)
Most of these parameters are characterized by their effect

 0.10
Calib.

0.70
0.72
0.55

0.66
on the river channel and soil characteristics.

11.4 (13.8) 5.9 (6.5) 5.6 (6.5) 11.4 (13.1) 10.6 (13.8) 6.2 (6.4) 5.7 (6.5) 11.7 (13.1) 11.3 (13.8)

11.6 (12.9) 6.0 (6.9) 6.5 (6.9) 10.6 (11.4) 10.9 (12.9) 5.9 (6.7) 6.0 (7.2) 10.3 (11.4) 10.9 (12.9)
Model performance

18.40
Valid.

0.75
0.79
0.50

0.48
Five statistical parameters: NS, R2, RSR, PBIAS and

Ikizdere
p-factor, were calculated to evaluate the performance of

UCDCd

Calib.

10.40
0.74
0.76
0.51

0.53
the model and the calibration techniques (Table 1). Accord-
ing to Moriasi et al. (), model simulation can be judged

Valid.

11.70
as satisfactory if NS > 0.50 and RSR < 0.70, and if PBIAS <

0.64
0.65
0.60

0.55
Tözköydere
±25% for streamflow analysis. The SWAT model was cali-

Calib.
brated using four methods (DC, UC, MGC, and UCDC).

0.75
0.75
0.50
3.90
0.81
The calibration with Tözköydere flow data (UC) gave a
good NS parameter (0.8 and 0.7 for calibration and vali-

Valid.

11.70
0.60
0.64
0.60

0.50
dation respectively). That is due to low flow intensity
upstream of the basin caused by a smaller contribution
Ikizdere

area for the river outlet. This makes calibration easier for
Calib.

13.20
MGCc

0.71
0.74
0.54

0.82
the objective functions to achieve a large number in a
short period of time. The calibration with only Ikizdere
Valid.

13.90
0.74
0.76
0.51

0.58

The calibration was conducted using Tözköyere flow (upstream) before using the Ikizdere flows (downstream).
flow (DC) gave the NS to be 0.7 both for calibration and vali-
Tözköydere

dation. The DC calibration technique resulted in a smaller

Calib.

NS value due to a larger contribution area for flow causing

0.81
0.82
0.43
8.40
0.76

high flow at the outlet. Maximizing the objective function

The statistical evaluation of model parametrization using flow data for Ikizdere Basin

(NS) for a larger basin is more difficult than a basin of

Calib. Valid.

0.66
0.70
0.58
17.1
0.60

small area. The third technique (MGC) used both Ikizdere

Ikizdere

and Tözköydere flow for the model calibration. The objec-

UCb

tive function for both gauging stations were lower than

–
–
–
–
–
6.19 (6.50) –

5.93 (7.17) –

The calibration was conducted using both Ikizdere and Tözköyere flows.
that achieved in the previous steps (0.7 and 0.6 for Ikizdere
calibration and validation respectively, 0.8 and 0.6 for Töz-
Calib. Valid.
Tözköydere

0.69
0.69
0.55

0.57

The parameters were calculated for the flows of Tözköydere.

4.7

köydere calibration and validation respectively). The

The parameters were calculated for the flows of Ikizdere.

fourth method calibrated the model using Tözköydere flow

–
–
–
–
–
10.7 (13.8) –

11.4 (12.9) –

before calibrating with Ikizdere flow values. The final

model parameters best fits calibrated using the Tözköydere
Valid.

22.70
0.69
0.75
0.56

0.72

flows were not changed when calibrating using Ikizdere

flows. That way, the interdependence of the upstream and
11.8 (13.1)

11.1 (11.4)
Ikizdere

downstream flows was maintained. Although the objective

Calib.

10.10
0.72
0.75
0.52

0.86
DCa

function was lower for the third technique (MGC), the

mean, standard deviation and PBIAS values were found to
(StDev_Obs)
(Mean_Obs)

perform better than the first two calibration techniques.

StDev_Sim
Mean_Sim
parameter
|

Statistic

p-factor
Table 1

PBIAS

The best objective function performance was observed for

RSR
NS
2

the fourth technique (UCDC), which gave far better NS

R

d
a

c
 347 S. A. Swalih & E. Kahya | Evaluating hydrological model calibration techniques in a mountainous region Journal of Hydroinformatics | 23.2 | 2021

values both for the upstream and downstream regions of the the algorithm adjusts all model parameters providing the
basin. The other observation was calibrating the model best optimization value (NS in our case) (Figure 5).
using the upstream gauge data (UC) which gave more accu- In addition, the temporal analysis proved the superiority
rate model performance when compared to the calibration of the fourth technique (UCDC) as compared to the first
using the downstream flow data (DC). The calibration of three techniques, since the hydrological model was opti-
the model with the flow data of Ikizdere resulted in mized using two flow gauging stations one after the other
improved NS values at the Tözköydere location too. How- (Figure 6). This allows the model to be optimized sequen-
ever, the reverse did not show a similar outcome (Table 1). tially, starting from the upstream region down to the
downstream region. The other techniques optimize the
model with local gauge data contrary to the holistic
Temporal performance approach implemented in UCDC. This results in loss of infor-
mation about the characteristics of the basin which could
In the temporal analysis, the monthly average simulated otherwise be useful to improve the model performance. In
flow values are plotted against time. The calibration of the the UCDS technique, we made sure not to change
model using the upstream flow (UC) (NS of 0.81 for cali- the parameters that were set using the upstream flow when
bration) compared to the calibration using the downstream calibrating using the downstream flow in order to maintain
flow (DC) (NS of 0.72 for calibration) demonstrated good the inter-dependence of upstream–downstream flow.
accuracy (Figures 3 and 4). The main reason is that the
upstream flow intensity is lower, making it less difficult for Climate change assessment
the algorithm to capture the distribution including the
peak flow values. When the flow value increases down- The monthly average flow values for the simulated flow,
stream, it becomes difficult to estimate the flows, as well as six climate change scenarios compared with
especially the peak flows resulting in lower calibration per- the observed flow at Ikizdere sub-basin outlet, is plotted
formance. When both the upstream and downstream flow in Figure 7. There is no consistent trend in the forecasted
data are used for model calibration simultaneously (MGC), flow for the various climate model precipitations. For the

Figure 3 | Observed flow comparison with simulated flow at Ikizdere station, for calibration and validation periods (a) and (b) respectively; and that of Tözköydere station validation (c) and
(d) (for DC – only Ikizdere flow used for model optimization).
 348 S. A. Swalih & E. Kahya | Evaluating hydrological model calibration techniques in a mountainous region Journal of Hydroinformatics | 23.2 | 2021

Figure 4 | Observed flow comparison with simulated flow at Ikizdere station, for calibration and validation periods (a) and (b) respectively; and that of Tozkoydere station validation (c) and
(d) respectively (for UC – only Tözköydere flow used for model optimization).

Figure 5 | Observed flow comparison with simulated flow after calibration and validation (a) and (b) respectively for Ikizdere station; (c) and (d) for Tözköydere station (for MGC – both
Ikizdere and Tözköydere flow used for model optimization).

winter, all the climate models forecast a decrease in flow CONCLUSIONS

for the river. The other seasons forecasted a slight
increase in flow. All in all, we did not observe an agree- It is quite challenging for a hydrological model to capture
ment amongst the climate projections forecast for the the complex processes in a basin, especially for mountai-
Rize basin. nous regions with limited input data. Estimating the
 349 S. A. Swalih & E. Kahya | Evaluating hydrological model calibration techniques in a mountainous region Journal of Hydroinformatics | 23.2 | 2021

Figure 6 | Observed flow comparison with simulated flow after calibration and validation (a) and (b) respectively for Ikizdere station; (c) and (d) for Tözköydere station (for UCDC – Töz-
köydere flow used before using Ikizdere for model optimization).

Figure 7 | Seasonal average flow of the observed, baseline simulation and simulations with climate change forcing at Ikizdere outlet.

parameters that represent the various hydrological pro- The results of our analysis have shown that using individ-
cesses is one of the main issues hydrologists need to ual gauges for calibration improve the predictive ability of the
solve since it is impossible to measure all of the hydrologi- model for the particular region. The DC calibration technique
cal parameters. In order to assess the impact of calibration resulted in a lower performance compared to the UC, due to
techniques on the model performance, four methods were a larger drainage area for the outlet causing high flow. In cali-
tested, calibration using flow data: (1) only from upstream bration, maximizing the objective function (NS) for a large
area (US), (2) only from downstream area (DC), (3) basin area is more difficult than for a basin of small area.
using both upstream and downstream areas of the basin– Model calibration using the MGC technique significantly
multi-gauge calibration (MGC), and (4) first the upstream improved model performance as a whole, unlike the single
then downstream flow data (UCDC). The study was gauge calibrations. This is due to the fact that the model
aimed at assessing the performance of multi-gauge cali- obtains much wider information on the basin characteristics
bration over single gauge calibration techniques. to adjust its parameters when more than one gauging station
 350 S. A. Swalih & E. Kahya | Evaluating hydrological model calibration techniques in a mountainous region Journal of Hydroinformatics | 23.2 | 2021

is used simultaneously for calibration. The UCDC calibration REFERENCES

technique gave the best model performance since the model
performance was improved step by step starting from the Abbaspour, K. C.  SWAT-CUP: SWAT Calibration and
upstream area going to the downstream area of the basin. Uncertainty Programs – A User Manual. Eawag – Swiss
This technique not only improved the model for each Federal Institute of Aquatic Science and Technology,
Switzerland.
region, but also the interdependence of the two regions was
Akay, H., Koçyiğit, M. B. & Yanmaz, A. M.  Effect of using
maintained since the model parameters calibrated by the multiple stream gauging stations on calibration of hydrologic
upstream flow were not changed when calibrating the parameters and estimation of hydrograph of ungauged
model with the downstream flow. Our study findings are in neighboring basin. Arabian Journal of Geosciences 11,
282–293.
agreement with what has been reported by Cao et al.
Arnold, J. G., Srinivasan, R., Muttiah, R. S. & Williams, J. R. 
(), Santhi et al. (), Haas et al. (), Lu et al. Large area hydrologic modeling and assessment part I:
(), Zhang et al. (), Bai et al. () and Akay et al. model development. Journal of the American Water
Resources Association 34 (1), 73–89.
(). They all reported improvements in model calibration
Arnold, J. G., Moriasi, D. N., Gassman, P. W., Abbaspour, K. C.,
performance with multi-site (multi-gauge) calibration com- White, M. J., Srinivasan, R., Santhi, C., Harmel, R. D., van
pared with single-gauge calibrations for their studies. Griensven, A., Van Liew, M. W., Kannan, N. & Jha, M. K.
Evaluating the results of previous studies and our study, we a SWAT: model use, calibration, and validation.
American Society of Agricultural and Biological Engineers
can confidently say multi-gauge calibration gives much
(ASABE) 55 (4), 1491–1508.
better model calibration performance than single-gauge cali- Arnold, J. G., Kiniry, J. R., Srinivasan, R., Williams, J. R., Haney, E.
bration techniques for the Ikizdere basin, Black Sea region. B. & Neitsch, S. L. b Soil & Water Assessment Tool –
However, we still recommend more studies to be carried Input/Output Documentation, Version 2012. Texas Water
Institute, Texas, USA.
out on other basins to reach a comprehensive conclusion
Bai, J., Shen, Z. & Yan, T.  Comparison of single- and multi-
about multi-gauge calibration techniques for the Black Sea site calibration and validation: a case study of SWAT in the
region. Regarding the climate change assessment, we have Miyun Reservoir watershed, China. Journal of Frontiers of
Earth Science 11 (3), 592–600.
not observed a consistent trend in the forecasted flow by
Cao, W., Bowden, W. B., Tim, D. T. & Fenemor, A.  Multi-
the various climate model predictions. The IPCC climate pro- variable and multi-site calibration and validation of SWAT in
jections forecast did not show a visible shift in the hydrologic a large mountainous catchment with high spatial variability.
regime for the Rize basin. Journal of Hydrological Processes 20, 1057–1073.
Duana, Z., Tuoa, Y., Liu, J., Gaod, H., Song, X., Zhang, Z., Yangb,
L. & Mekonnen, D. F.  Hydrological evaluation of open-
access precipitation and air temperature datasets using
ACKNOWLEDGEMENTS SWAT in a poorly gauged basin in Ethiopia. Journal of
Hydrology 569, 612–626.
Elshamy, M. E., Seierstad, I. A. & Sorteberg, A.  Impacts
We would like to acknowledge TUBİTAK (the Turkish of climate change on Blue Nile flows using bias-corrected
Scientific and Technological Research Council) for GCM scenarios. Hydrology and Earth System Sciences 13,
providing the funds to carry out this study. We are also 551–565.
Ercan, M. B., Goodall, J. L., Castronova, A. M., Humphrey, M. &
indebted to the Turkish State Meteorological Service Beekwilder, N.  Calibration of SWAT models using the
(MGM) and General Directorate of Hydraulic Works cloud. Journal of Environmental Modelling & Software 62,
(DSI) for the observed weather and flow data access. 188–196.
Haas, M. B., Guse, B., Pfannerstill, M. & Fohrer, N.  A joined
multi-metric calibration of river discharge and nitrate loads
with different performance measures. Journal of Hydrology
DATA AVAILABILITY STATEMENT 536, 534–545.
IPCC  Technical Summary: Climate Change Impacts,
Adaptation and Vulnerability: A Report of Working Group II.
Data cannot be made publicly available; readers should con- Intergovernmental Panel on Climate Change (IPCC),
tact the corresponding author for details. Geneva, Switzerland.
 351 S. A. Swalih & E. Kahya | Evaluating hydrological model calibration techniques in a mountainous region Journal of Hydroinformatics | 23.2 | 2021

Kahya, E., Kalaycı, S. & Piechota, T. C. a Streamflow Sen, O. & Kahya, E.  Determination of flood risk: a case study
regionalization: case study of Turkey. Journal of Hydrologic in the rainiest city of Turkey. Journal of Environmental
Engineering 13 (4), 205–214. Modelling & Software 93, 296–309. doi:10.1016/j.envsoft.
Kahya, E., Demirel, M. C. & Bég, A. O. b Hydrologic 2017.03.030.
homogeneous regions using monthly streamflow in Turkey. Shresthaa, M. K., Recknagela, F., Frizenschaf, J. & Meyer, M. 
Earth Sciences Research Journal 12 (2), 181–193. Assessing SWAT models based on single and multi-site
Kahya, E., Özger, M., Ş eker, D. Z., Karaca, M., Can, I., Kömüşcü, calibration for the simulation of flow and nutrient loads in
A. Ü., Bozkurt, D., Ş en, O., Mehr, A. D., Erdem, H. & the semi-arid Onkaparinga catchment in South Australia.
Bagheri, F.  Determination of Flood Risk on Basin Journal of Agricultural Water Management 175, 61–71.
Boundaries of the Rize Province: Current and Future Solomon, S., Qin, D., Manning, M., Alley, R. B., Berntsen, T.,
Situation by Climate and Hydrological Models (Turkish). Bindoff, N. L., Chen, Z., Chidthaisong, A., Gregory, J. M.,
Scientific and Technological Research Council of Turkey Hegerl, G. C., Heimann, M., Hewitson, B., Hoskins, B. J.,
(TÜBİTAK), Istanbul. Joos, F., Jouzel, J., Kattsov, V., Lohmann, U., Matsuno, T.,
Klemes, V.  Dilettantism in hydrology: transition or destiny? Molina, M., Nicholls, N., Overpeck, J., Raga, G.,
Journal of Water Resources Research 22 (9), 177S–188S. Ramaswamy, V., Ren, J., Rusticucci, M., Somerville, R.,
Kocyigit, M. B., Akayl, H. & Yanmaz, A. M.  Effect of Stocker, T. F., Whetton, P., Wood, R. A. & Wratt, D. 
watershed partitioning on hydrologic parameters and Technical summary. In: Climate Change 2007: The Physical
estimation of hydrograph of an ungauged basin: a case study Science Basis. Contribution of Working Group I to the Fourth
in Gokirmak and Kocanaz, Turkey. Arabian Journal of Assessment Report of the Intergovernmental Panel on
Geosciences 10, 331–344. Climate Change (S. Solomon, D. Qin, M. Manning, Z. Chen,
Lu, Z., Zou, S., Xiao, H., Zheng, C., Yin, Z. & Wang, W.  M. Marquis, K. B. Averyt, M. Tignor & H. L. Miller, eds).
Comprehensive hydrologic calibration of SWAT and water Cambridge University Press, Cambridge, UK and New York,
balance analysis in mountainous watersheds in northwest NY, USA.
China. Journal of Physics and Chemistry of the Earth 79–82, Van Griensven, A., Meixner, T., Grunwald, S., Bishop, T., Diluzio,
76–85. M. & Srinivasan, R.  A global sensitivity analysis tool for
Moriasi, D. N., Arnold, J. G., Van Liew, M. W., Bingner, R. L., the parameters of multi-variable catchment models. Journal
Harmel, R. D. & Veith, T. L.  Model evaluation of Hydrology 324, 10–23.
guidelines for systematic quantification of accuracy in basin White, K. L. & Chaubey, I.  Sensitivity analysis, calibration,
simulations. American Society of Agricultural and Biological and validations for a multisite and multivariable SWAT
Engineers (ASABE) 50 (3), 885–900. model. Journal of the American Water Resources Association
Molina-Navarro, E., Andersen, H. E., Nielsen, A., Thodsen, H. & (JAWRA) 41 (5), 1077–1089.
Trolle, D.  The impact of the objective function in multi- Winchell, M., Srinivasan, R., Di Luzio, M. & Arnold, J. 
site and multi-variable calibration of the SWAT model. ArcSWAT Interface for SWAT2012 – User’s Guide. Blackland
Journal of Environmental Modelling & Software 93, 255–267. Research and Extension Centre & Grassland, Soil and Water
Morris, M. D.  Factorial sampling plans for preliminary Research Laboratory, USDA Agricultural Research Service,
computational experiments. Journal of Technometrics 33 (2), Texas, USA.
161–174. Zhang, L., Jin, X., He, C., Zhang, B., Zhang, X., Li, J., Zhao, C.,
Neitsch, S. A., Arnold, A. D., Kiniry, A. S. & Williams, D. A.  Tian, J. & DeMarchi, C.  Comparison of SWAT and
Soil and Water Assessment Tool – Theoretical DLBRM for hydrological modeling of a mountainous
Documentation Version 2009. Texas Water Resources watershed in arid northwest China. Journal of Hydrologic
Institute, Technical Report No. 406. Texas. Engineering (ASCE) 21 (5), 04016007 (1–11).
Santhi, C., Kannan, N., Arnold, J. G. & Di Luzio, M.  Spatial Zhang, Y., Shao, Q., Zhang, S., Zhai, X. & She, D.  Comparison
calibration and temporal validation of flow for regional scale of single- and multi-site calibration and validation: a case study
hydrologic modeling. Journal of the American Water of SWAT in the Miyun Reservoir watershed, China. Journal of
Resources Association (JAWRA) 44 (4), 829–846. Frontiers of Earth Science 301, 54–61.

First received 6 February 2020; accepted in revised form 19 October 2020. Available online 15 December 2020