Hydrological Sciences Journal

ISSN: (Print) (Online) Journal homepage: www.tandfonline.com/journals/thsj20

Optimization of the SWAT+ model to adequately

predict different segments of a managed
streamflow hydrograph

Tibebe B. Tigabu, Ate Visser, Tariq Kadir, Shalamu Abudu, Philip Cameron-
Smith & Helen E. Dahlke

To cite this article: Tibebe B. Tigabu, Ate Visser, Tariq Kadir, Shalamu Abudu, Philip Cameron-
Smith & Helen E. Dahlke (2024) Optimization of the SWAT+ model to adequately predict
different segments of a managed streamflow hydrograph, Hydrological Sciences Journal, 69:9,
1198-1217, DOI: 10.1080/02626667.2024.2364714

To link to this article: https://doi.org/10.1080/02626667.2024.2364714

© 2024 The Author(s). Published by Informa

UK Limited, trading as Taylor & Francis
Group.

Published online: 24 Jun 2024.

Submit your article to this journal

Article views: 1026

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=thsj20
HYDROLOGICAL SCIENCES JOURNAL
2024, VOL. 69, NO. 9, 1198–1217
https://doi.org/10.1080/02626667.2024.2364714

Optimization of the SWAT+ model to adequately predict different segments of a

managed streamflow hydrograph
a
Tibebe B. Tigabu , Ate Visserb, Tariq Kadirc, Shalamu Abuduc, Philip Cameron-Smithb and Helen E. Dahlkea
a
Department of Land, Air, and Water Resources, University of California Davis, Davis, California, USA; bLawrence Livermore National Laboratory,
Livermore, California, USA; cBay-Delta Office, Modeling Support Branch, California Department of Water Resources, Sacramento, California, USA

ABSTRACT ARTICLE HISTORY

Complete representation of rainfall–runoff responses in complex, large watersheds using a single- Received 8 September 2023
objective parameterization approach in watershed models is often unachievable. In this study we present Accepted 16 May 2024
a calibration approach for the SWAT+ model that independently fits model parameters for different flow EDITOR
segments of the hydrograph. The approach is demonstrated for the Feather River, California, USA, using A. Fiori
daily streamflow from the Lake Oroville Reservoir outlet gage. Results show that when model parameters
were independently fitted for different flow segments the KGE, NSE, PBIAS, and RSR values improved to ASSOCIATE EDITOR
0.96, 0.99, −3.3, and 0.10, respectively, compared to 0.72, 0.66, −9.30, and 0.53, respectively, achieved K. Schroeter
under a multiobjective and full hydrograph (average hydrograph) calibration. The results highlight that KEYWORDS
when considering the average hydrograph and flow duration curves, a more balanced representation of Feather River; managed
both poorly and well-performing segments is achieved, emphasizing the importance of segment-specific streamflow; optimization;
parameterization and multi-objective evaluation for accurately representing different flow conditions. flow segment; multi-
objective functions; SWAT+

1 Introduction
investigate the impacts of climate and land use changes on
Climate change will impact water resources globally, especially hydrological processes (e.g. Kannan et al. 2019, Aliye et al.
in high-elevation regions that serve as the water towers of the 2020, Tigabu et al. 2020, Liu et al. 2021, Mahmoodi et al. 2021,
world (Immerzeel et al. 2010, 2020). Predicting the impact of Wagner et al. 2022, Bailey et al. 2023). Due to the growing
climate change on the timing and magnitude of streamflow in interest of users and the availability of a strong user group for
a warmer, low- to no-snow future (Barnett et al. 2005, Siirila- the SWAT model, SWAT model developers have been advan­
Woodburn et al. 2021) is critical to adapt water resources cing the model structure over time. Most recently, Bieger et al.
infrastructure (Hedden-Nicely 2022) and secure safe and reli­ (2017) completely restructured and advanced the SWAT
able sources of water for human consumption, agriculture, model into the SWAT+ version, which is the most prominent
ecosystem health, and industry. Advancements in large-scale revision in its history (Wagner et al. 2022). SWAT+ incorpo­
integrated hydrological models are required to quantify the rates new features, including landscape unit and stream con­
current and future water supply and water quality conditions nectivity, to better represent various scales of watersheds and a
(Bailey et al. 2023). Prediction of current and future stream­ consolidated file structure (White et al. 2022). SWAT+ is also
flow is one of the most important tasks in water resources capable of simulating managed flows, impacted by dams and
management. Hydrologists have been using data-driven and reservoir operations, via a set of reservoir operation rules (Wu
physically-based hydrological models to simulate streamflow et al. 2020). Despite the extensive progress in advancing the
and other hydrological processes in catchments. There are SWAT model structure and a growing number of studies on
several such models that are being used globally. Among calibration and uncertainty analysis, a holistic calibration
others, the Soil and Water Assessment Tool (SWAT) devel­ approach is still missing due to the heterogeneity and complex­
oped by Arnold et al. (1998), MIKE SHE, which is the ity of watershed hydrological processes encountered through­
European Hydrological System Model (Refsgaard and Storm out the world. Moreover, calibration of hydrological models,
1995), and the Agricultural Policy/Environmental Extender including SWAT, is always challenging due to uncertainties
abbreviated as APEX (Gassman 2009) are a few examples of that arise from model inputs, model structure, parameters, and
watershed-scale models used worldwide (Golmohammadi et outputs (Wu et al. 2021). The challenge is far higher in regions
al. 2014). with extensive water management infrastructure for flood
The SWAT model has been extensively used to predict the protection and water delivery, such as dams, surface water
terrestrial hydrological cycle of watersheds, to evaluate best reservoirs, and hydropower operations, and water diversion
management practices, to simulate environmental flow, and to as can be found in the State of California, USA.

CONTACT Tibebe B. Tigabu [email protected] Department of Land, Air, and Water Resources, University of California Davis, 1 Shields Ave, Davis, CA
95616, USA
© 2024 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group.
This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivatives License (http://creativecommons.org/licenses/by-nc-nd/4.0/),
which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited, and is not altered, transformed, or built upon in any way. The
terms on which this article has been published allow the posting of the Accepted Manuscript in a repository by the author(s) or with their consent.
 HYDROLOGICAL SCIENCES JOURNAL 1199

California is known for having the most complex water Notably, Pfannerstill et al. (2014, 2017), Shafii and Tolson
delivery systems in the world (Avanzi 2018). According to (2015), Donnelly et al. (2016), Alemayehu et al. (2022), and
Avanzi (2018), the water delivery systems in California are Dal Molin et al. (2023) all have considered FDCs as calibration
implemented through a network of reservoirs, aqueducts, objectives for model performance evaluation.
and groundwater pumps that deliver water from the head­ Donnelly et al. (2016) incorporated flow signatures in the
waters in the northern and eastern portions of the state to evaluation of the performance of a multi-basin model across
population centers and agricultural land in the western and various sites within the domain, utilizing several model
southern parts of the state. These numerous water infrastruc­ performance metrics to better understand dominant catch­
tures, including a collection of canals, pipelines, reservoirs, and ment processes. Their study indicated that simulated flows
hydroelectric power facilities, deliver clean water to 38 million based on the semi-distributed and process-based HYPE
Californians, 3.2 million hectares of farmland, and businesses model (Lindström et al. 2010) successfully captured observa­
throughout the state (DWR 2023). Moreover, the high seasonal tions, with dominant temporal variability also represented
and inter-annual hydrological availability and extreme weather when considering all flow signatures simultaneously. Dal
events in the state are other factors that make the water Molin et al. (2023) investigated the efficacy of a streamflow
delivery system the most intricate and complex (Hanak and signature-based model calibration in predicting streamflow
Lund 2012). for six catchments in the Thur basin, in northeastern
The Feather River watershed is one of the most important Switzerland. Their findings demonstrated that signature-
watersheds in California; it provides a third of all water dis­ based calibration of precipitation–streamflow models ade­
tributed by the Metropolitan Water District of Southern quately predicts streamflow for ungaged catchments.
California through Oroville Reservoir and canals (Avanzi Similarly, Alemayehu et al. (2022) found that utilizing
2018). The complex hydrology, intensive water use system, FDCs from historical records and recent remote sensing-
Mediterranean climate, and large elevation range (including based evapotranspiration data in model calibration enhances
the rain–snow transition) of the Feather River make water the efficiency of hydrological models in simulating catch­
management very challenging. To accommodate these chal­ ment hydrological processes. This underscores the superior
lenges, the California Department of Water Resources uses efficiency of FDC-based calibration approaches for the
various models to track streamflow and water deliveries to SWAT model when integrating streamflow signatures and
the State Water project, the nation’s largest state-owned remote sensing information. Consequently, the authors sug­
water and power generator and user-financed water system gest employing historical FDC records and recent remote
(DWR 2023). Application of a physically-based hydrological sensing-based evapotranspiration data in model calibration
model like SWAT+ is needed to sustain water supply services to optimize the efficiency of hydrological models in simulat­
under such multifold problems. ing catchment hydrological processes. Overall, hydrological
For the purpose of water management, it is also essential to model calibration frameworks that prioritize FDCs and
calibrate and validate the model to accurately represent the remote sensing information as calibration objectives are
spatial and temporal heterogeneities of hydrological processes essential for overcoming limitations associated with conven­
(Mengistu et al. 2019). To achieve an adequate calibration and tional calibration approaches.
to reduce uncertainty in model simulations (Kannan et al. Despite numerous calibration and validation studies using
2019, Guse et al. 2020), both calibration and validation periods the SWAT model, achieving satisfactory calibration for large,
should include wet, average, and dry years (Arnold et al. 2012). complex watersheds remains challenging, as a single para­
Various calibration approaches have been applied to predict meter set often fails to capture the diverse signatures of the
hydrological processes in watersheds. The traditional approach streamflow hydrograph. As highlighted by Westerberg et al.
involves calibrating the model based on streamflow data at a (2011), conventional calibration performance measures suf­
catchment outlet point (Daggupati et al. 2015). However, new fer from four main limitations: uncertainty in observed
approaches have been developed, such as seasonal clustering of streamflow, variable sensitivity of model performance across
daily streamflow for calibration (Lakshmi and Sudheer 2021, different flow segments, the influence of input/output errors,
Tigabu et al. 2023), multi-metric calibration for different parts and the inability to evaluate model performance when
of hydrographs (Pfannerstill et al. 2014), and incorporating observed and simulated flow magnitudes do not overlap in
streamflow signatures, flow duration curves (FDCs), and spa­ time.
tially distributed remote sensing data in the model calibration Therefore, the current study aims to demonstrate a novel
(Westerberg et al. 2011, Alemayehu et al. 2022). More recently, calibration approach that can better represent the distinct flow
the multicriterion sequential calibration and uncertainty ana­ signatures of managed streamflow using the SWAT+ model.
lysis (MS-CUA) method was developed to better optimize Here, we independently parameterized the high flow, middle
SWAT simulations and to provide balanced uncertainty ana­ flow, and low flow segments of a managed streamflow hydro­
lyses compared to other calibration approaches (Wu et al. graph. We believe that testing this calibration exercise is cru­
2021). Multi-variable calibration approaches have also been cial, given the limited calibration exercises tested for the
shown to improve the performance of SWAT, particularly in SWAT+ model. This study does not utilize remote sensing
predicting snow-affected streamflow (Liu et al. 2021, Chen et evapotranspiration and soil moisture data in the calibration
al. 2023). Similarly, the utilization of streamflow signatures and validation process because the scope of the study is to
(FDC) and remote sensing information in hydrological introduce a new method of model calibration to reproduce
model evaluations have emerged as crucial calibration options. observed streamflow.
1200 T. B. TIGABU ET AL.

The study is conducted using the Feather River watershed (1.613 km3). The Feather River watershed is characterized by a
in California, USA, as an example. SWAT+ has a wider flex­ Mediterranean climate with warm, dry summers and cool, wet
ibility to include manmade structures such as reservoirs, weirs winters (Koczot et al. 2012). The mean long-term total annual
and ponds, which can help to simulate regulated flows in a precipitation is 1078 mm, while the mean monthly maximum
watershed (Wu et al. 2020). Multiple parameter sets will be and minimum temperatures are 21 and 4.8°C, respectively.
proposed to improve the efficiency of the SWAT+ model in Most of the watershed area is at elevations where winter
reproducing each flow segment independently, aiming to iden­ temperatures can fluctuate from below to above freezing, and
tify the segment of the FDC that most significantly influences slight temperature changes can affect snow formation and
the overall predictive performance of the model for managed melting. This feature of the watershed leads to complex
streamflow. streamflow variability, making changes in streamflow depen­
This study assesses the feasibility of these methods using long- dent on both temperature and precipitation lapse rates since
term managed streamflow data from the Feather River at the temperature affects snowmelt and precipitation form. The land
Oroville gauging station, juxtaposed with SWAT+ simulated cover of the watershed primarily consists of coniferous trees,
streamflow at the same location. Oroville, the second-largest with some areas of shrubs and grassland mainly located in the
reservoir in California, serves crucial roles in water supply storage, agricultural valleys (Koczot et al. 2004).
hydropower generation, and flood control along the Feather River
(Nelson et al. 2016). Given its well-documented reservoir opera­
2.2 Data
tion rules in the SWAT+ model and extensive historical stream­
flow records at the outlet point, our calibration approach focuses For this study, daily gridded precipitation and maximum and
on regulated flow at the outlet of the reservoir. Moreover, minimum temperature data were obtained from the California
although calibration and validation were not practical due to Department of Water Resources (DWR 2022) based on the
insufficient records at the outlet points of upstream water bodies, 4 km resolution Parameter-elevation Relationships on
we have integrated these water bodies into our modeling frame­ Independent Slopes Model (PRISM) product (PRISM 2014).
work. Hence, the specific objectives of the study are as follows: Spatial data, including land use/cover of the National Land
Cover Database (NLCD 2001 Land Cover), and the Soil Survey
(1) To identify the most influential parameters for simulat­ Geographic Database (SSURGO), and DEMs, were obtained
ing high flow, middle flow, and low flow segments of from publicly available resources listed in Table 1.
the managed streamflow hydrograph using the SWAT+ Long-term streamflow records from 1953 to the present at
model. the Oroville Dam gauging station’s outlet point were obtained
(2) To propose multiple parameter sets that can improve from the United States Geological Survey (USGS) National
the efficiency of the SWAT+ model in reproducing Water Information System (NWIS, https://waterdata.usgs.
each flow segment independently. gov/nwis/) to calibrate and validate the SWAT+ model. The
(3) To evaluate the practicability of the proposed calibra­ spatial (Fig. 2), climate, and hydrological data were used to
tion and parameterization methods using long-term establish and calibrate the SWAT+ model. All data used to
managed streamflow data from the Feather River at establish and calibrate the SWAT+ model and their basic
the Oroville gauging station and SWAT+ simulated characteristics are presented in Table 1.
streamflow at the same location.
2.3 Hydrological model set-up
2 Materials and methods This study was conducted with the SWAT+ model (Bieger et
al. 2017), which is a completely revised version of SWAT
2.1 Study area
(Arnold et al. 1998, Arnold and Fohrer 2005). SWAT+ is
The Feather River is a large tributary to the Sacramento River capable of simulating spatially distributed water balance and
(71 432 km2), the largest river in northern California. It serves as nutrient cycles based on hydrological response units (HRUs).
the primary source of surface water for the state, flowing into The SWAT+ model for the Upper Feather watershed was
Oroville Reservoir (Koczot et al. 2004, Huang et al. 2012). The constructed based on the pre-existing files of the SWAT2009
Upper Feather River Watershed, above Oroville Dam, is situated obtained from the California State Department of Water
in the Sierra Nevada Mountains of California. Using a digital Resources (DWR 2022). We used the long-term release of
elevation model (DEM) with a resolution of 30 m by 30 m, we QGIS (version 3.22.10) to delineate watersheds and the burn
delineated the watershed to Oroville Dam and found that it function was applied to enforce the existing stream networks
covers an area of 9427 km2, with altitudes ranging from 256 to in the delineation of the DEM-based stream network.
2826 m above sea level. Moreover, we used the DEM Inversion function to classify
There are three reservoirs and five powerhouses upstream of the watershed into either landscape or flood plain zones and
the North Fork Feather River, as depicted in Fig. 1. The most the Add Lake function to append the existing reservoir system.
prominent natural reservoir, Lake Almanor, is a spring-fed lake SWAT+ incorporates a set of rules governing reservoir
that has been expanded by the construction of Canyon Dam operations for surface water reservoirs in the United States.
(Avanzi 2018). The other two reservoirs are Butt Valley, located These rules are utilized without modification to simulate
on Butt Creek, with a capacity of 49 891 acre-feet (0.062 km3); release scenarios. The default release rules, outlined in a
and Lake Almanor, with a capacity of 1.308 million acre-feet decision table specific to the Oroville Reservoir (Table 2),
 HYDROLOGICAL SCIENCES JOURNAL 1201

Figure 1. Map showing the location of the Upper Feather River Watershed within the Sierra Nevada Mountains, California, USA, along with reservoirs, river networks,
and the boundary of the modeled watershed.

Table 1. Spatial and hydrometeorological data sources used this study, and their availability.
Data type Variable Temporal/spatial resolution Availability Data source
Spatial data Land use and land cover 2001/ 30 m × 30 m Accessed July 2022 NLCD 2001 (Homer et al. 2004)
https://www.mrlc.gov/data/nlcd-2001-
land-cover-conus
Soil 1:50 000 Accessed July 2022 https://gdg.sc.egov.usda.gov
Digital elevation model (DEM) 30 m × 30 m Accessed July 2022 http://earthexplorer.usgs.gov/
Climate data Precipitation Daily, 4 km × 4 km 1915–2022 PRISM Obtained from California Department
and temperature (max/min) of Water Resources (DWR 2022)
Hydrology Streamflow At the outlet point of Oroville Lake 1953–2022 https://waterdata.usgs.gov/nwis/ (USGS 2022)

encompass five conditions, seven alternatives, and five action volume for drawdown days (dyrt) for multiple-use flood
options. To facilitate the model’s utilization of the decision (multiple_use_fl). Similarly, the determination of action
table, key parameters such as the conditional variable, con­ types for the remaining alternatives follows suit, relying on
dition limits (limit variable), limit operator, and limit con­ conditional statements derived from the interplay of the
stant must be clearly defined (Arnold et al. 2018). In Table 2, conditional variable, limiting variable, limit operation, and
the release rate is determined as a function of both reservoir limit constants as outlined in Table 2. Here, it is worth
volume and storage volume (e-pv). The conditional variable, noting that calibration of parameters connected to reservoir
representing reservoir volume, and the limit variable, denot­ storage and release are not within the scope of this paper. As
ing storage volume in hectare-meters (ha-m), jointly dictate we used the default parameterization for the reservoir, there
the selection of alternatives and corresponding action might be uncertainty associated with releases from reservoir
entries. For instance, the implementation of the first alter­ (see the Appendix, Fig. A1).
native is contingent upon the satisfaction of the following The watershed delineation step resulted in 59 sub-basins
conditional statement: if reservoir volume (conditional vari­ and 583 channels. Following the watershed delineation step,
able) > e-pv (storage volume in ha-m) = −6.761(limit con­ the HRUs were created by combining the land use, soil, and
stant) and reservoir volume < e-pv = 0.356 and month < slope classes. To capture topographic effects on watershed
6.892, then the action entry involves releasing the base processes, we classified the DEM into five slope classes as 0–
1202 T. B. TIGABU ET AL.

Figure 2. Spatial SWAT+ model input data including (a) a digital elevation model (DEM) showing topography, (b) land use/land cover, and (c) the soil hydrological
groups within the Upper Feather River watershed, in the Sierra Nevada Mountains, California, USA.

Table 2. Default decision rule for Oroville Reservoir release as available in SWAT+ model (Bieger et al. 2017) on multipurpose use.
Name Conds Alts Acts
Oroville 5 7 5
var obj obj_num lim_var lim_op lim_const alt1 alt2 alt3 alt4 alt5 alt6 alt7
vol res 0 e-pv = −6.761 > > > – – – –
vol res 0 e-pv = 0.356 < < < > > > –
vol res 0 e-pv = 1.02 – – – < < < >
month null 0 null – 6.892 < – > < – > –
month null 0 null – 8.043 – > < – > < –
act_typ obj obj_num name option const const2 fp outcome
release res 0 multiple_use_fl dyrt 233.16170 0.13998 con1 y y n n n n n
release res 0 multiple_use_nf dyrt 402.9186 0.31844 con1 n n y n n n n
release res 0 sfl_cont+mu_fl dyrt 53.09800 0.31844 con2 n n n y y n n
release res 0 sfl_cont+mu_nf dyrt 153.4471 2.34415 con2 n n n n n y n
release res 0 efc_cont dyrt 0.69298 5.11667 con3 n n n n n n y
conds: conditions; alts: alternatives; acts: actions; var: variable;lim_var: limit variable; lim_op: limit operator; lim_const: limit constant; act_typ: action type; const:
constant; fp: file pointer; e-pv: storage volume in ha-m; dyrt: day rate; multiple_use_fl: multiple use flood; multiple_use_nf: multiple use non-flood; sfl_cont+mu_fl:
seasonal flood control multiple use flood; sfl_cont+mu_nf: seasonal flood control + multiple usenon-flood; efc_cont: emergency flood control.

2% (flat to very gently sloping), 2–5% (gently sloping), 5–8% 2.4 Calibration and parameterization
(sloping), 8–15% (strongly sloping), and >15% (moderately
2.4.1 SWAT+ parameters, calibration data, and objective
steep to very steep) based on the Food and Agricultural
functions
Organization of the United Nations (FAO) slope guidelines
The SWAT+ model of the Upper Feather River watershed is
(Jahn et al. 2006). Next, the model input files were compiled
using the SWAT+ editor version 2.0.4. This resulted in 59 sub- described by 19 parameters in total. For all parameters, an
basins, 583 channels, 1131 routing units, 95 241 HRUs, and initial value based on prior studies is available in SWAT+. To
117 aquifer objects for the first groundwater layer. To mini­ limit the number of parameters under study here, we focus on
mize the computing time, only the dominant HRUs were the 19 parameters listed in Table 3.
considered for calibration. The Hargreaves equation was The model was calibrated and validated against daily obser­
used for calculating evapotranspiration (Hargreaves and vations of managed streamflow below Oroville Dam. Five years
Samani 1985), the variable storage method was used for chan­ of data (2005–2009) were used as the model warm-up period to
nel routing, and the soil moisture function was used to calcu­ define appropriate initial conditions and to attain equilibrium
late the average daily runoff curve number (CN). conditions for the model. The 2010 to 2020 period was used for
 HYDROLOGICAL SCIENCES JOURNAL 1203

Table 3. List of sensitive model parameters identified during the manual calibration step that were subsequently calibrated during the automatic calibration steps for
the high flow, middle flow, and low flow segments based on multi-criterion evaluation statistics. Parameters highlighted in bold were calibrated separately.
Calibrated parameter values for different segments of streamflow hydrographs and objective functions

Value ranges NSE KGE PBIAS

Parameter Min Max Change method Average High Middle Low Average High Middle Low Average High Middle Low
snomelt_tmp −5 5 Replace 2.90 0.42 0.11 2.32 1.25 3.80 0.11 2.32 0.11 3.16 3.80 0.12
snofall_tmp −5 5 Replace 1.07 2.38 −2.20 −2.93 0.10 1.35 −2.20 −2.93 −2.65 −1.94 1.35 −2.54
snomelt_max 0 10 Replace 3.15 2.19 6.83 4.57 5.99 4.01 6.83 4.57 3.10 3.77 4.01 6.64
snomelt_min 0 10 Replace 1.99 0.85 4.73 3.38 3.98 1.82 4.73 3.38 1.82 3.48 1.82 4.24
snomelt_lag 0 1 Replace 0.47 0.79 0.72 0.42 0.48 0.43 0.72 0.42 0.64 0.62 0.43 0.44
plaps 0 200 Replace 19.65 26.42 22.92 26.07 17.10 24.60 22.92 26.07 31.33 30.56 24.60 29.03
tlaps −10 10 Replace 4.00 −2.42 7.93 −9.69 6.01 −0.30 7.93 −9.69 7.88 −0.23 −0.30 −6.06
cn2 −20 20 Relative change 6.29 13.53 9.76 −4.64 2.58 4.84 9.76 −4.64 14.91 0.58 4.84 1.40
esco 0 1 Replace 0.28 0.15 0.01 0.56 0.42 0.25 0.01 0.56 0.45 0.54 0.25 0.52
epco 0 1 Replace 0.79 0.71 0.36 0.44 0.49 0.72 0.36 0.44 0.68 0.34 0.72 0.43
lat_ttime 0 180 Replace 48.97 44.76 69.37 70.36 40.54 31.17 69.37 70.36 46.14 8.93 31.17 43.39
perco 0 1 Replace 0.18 0.14 0.21 0.20 0.17 0.14 0.21 0.20 0.20 0.16 0.14 0.17
awc 0 1 Relative change 0.08 0.09 0.44 0.20 0.32 0.33 0.44 0.20 0.20 0.46 0.33 0.34
alpha 0 1 Replace 0.30 0.39 0.11 0.08 0.16 0.54 0.11 0.08 0.48 0.65 0.54 0.38
flo_min 0 50 Replace 15.24 14.65 17.58 31.79 19.65 28.06 17.58 31.79 14.53 22.85 28.06 21.67
revap 0 1 Replace 0.07 0.05 0.06 0.04 0.03 0.05 0.06 0.04 0.09 0.03 0.05 0.03
rchg_dp 0 1 Replace 0.29 0.25 0.42 0.60 0.21 0.25 0.42 0.60 0.42 0.74 0.25 0.47
revap_min 0 500 Replace 18.97 14.80 10.15 17.87 7.36 15.37 10.15 17.87 14.07 24.34 15.37 22.87
latq_co 0 1 Replace 0.10 0.16 0.16 0.39 0.06 0.14 0.16 0.39 0.43 0.18 0.14 0.51
Model parameters –cn2: Condition II curve number; lat_ttime: exponential of the lateral flow travel time; esco: soil evaporation compensation factor; epco: plant water
uptake compensation factor; snomelt_tmp: snowmelt temperature; snofall_tmp: snowfall temperature; snomelt_max: maximum snowmelt temperature; snomelt_­
min: minimum snowmelt temperature; snomelt_lag: snowpack temperature lag factor; perco: percolation coefficient; awc: available water capacity of soil layer;
alpha: alpha factor for groundwater recession curve; Condition II curve number; revap: threshold depth of water in shallow aquifer required to allow revap to occur;
rchg_dp: recharge to deep aquifer (the fraction of root zone percolation that reaches the deep aquifer; revap_min: water table depth for revap to occur; flo_min:
water table depth for return flow to occur; plaps: precipitation lapse rate (mm per km of elevation difference); tlaps: temperature lapse rate (°C per km of elevation
difference); latq_co: plant ET curve number coefficient; replace: absolute value; relative change: add/subtract the value to the default one. Change methods: “relative”
indicates adding/subtracting the value to/from the default, and “replace” indicates replacing the default value with the absolute value of the new number.

calibration and the five years for validation using the first
(1995–1999) as the warm-up period.
In accordance with Pfannerstill et al. (2014), the FDC of daily
streamflow data was categorized into very high flow (0–5%), high
flow (5–20%), middle flow (20–80%), low flow (80–95%), and very where ED is the Euclidian distance from the ideal point,
low flow (95–100%) segments based on the exceedance probability of δs μ
average daily streamflows. This segmentation approach, as advocated α ¼ , β ¼ s , and r = correlation coefficient.
σo μo
by Pfannerstill et al. (2014), allows for a process-based calibration, PBIAS (Equation 4) (Gupta et al. 1999) is one of the objec­
capturing dominant watershed processes manifested in various parts tive functions related to error measure of simulations with
of the hydrograph. The threshold for high flows was 2000 m3/s, reference to observation points. It is a popular method fre­
coinciding with the 80th flow percentile, and the threshold of low quently used in determining whether a model simulation is
flows was 37 m3/s, coinciding with the 10th flow percentile. overestimated or underestimated. PBIAS values can vary
Four objective functions were used for the automatic calibra­ between −∞ and +∞, whereas its optimal value is zero.
tion to daily streamflow data: Nash-Sutcliffe efficiency (NSE),
Kling-Gupta efficiency (KGE), percent bias (PBIAS) and standar­
dized root mean square error (RSR). NSE is a single metric that
captures errors of timing and magnitude between the simulated
and observed mean daily streamflow (Equation 1) (Nash and
Sutcliffe 1970). KGE (Equation 2) is a modified version of NSE
(Gupta et al. 2009), which targets maximizing its value to 1 based where “n”, “Q”, “s”, and “m” stand for the number of sampling
on the decomposition of the mean squared error into three points, streamflow, simulated, and observed, respectively.
factors: mean (β), variability (α), and dynamics (correlation coef­ The standardized root mean square error (RSR) is another
ficient r) (Equation 3) (Gupta et al. 2009). error evaluation criterion; it represents the ratio of the root
mean square error (RMSE) to the standard deviation of the
observations (Equation 5). It is useful to understand the varia­
tion between the observed data and simulated data.

where “n” refers to the number of observations, “Q” stands for

streamflow, and subscripts “s” and “o” refer to simulated and � m “s”, and “m”, represent streamflow, observed
where “Q”, Q
observed, respectively. mean, simulated, and observed, respectively.
1204 T. B. TIGABU ET AL.

Both the NSE and KGE are measures of goodness of fit and snowmelt parameters was the annual basin average
need to be maximized. PBIAS and RSR are error coefficients precipitation.
and need to be minimized toward zero. In the third stage, keeping calibrated snow, temperature, and
precipitation parameters constant, we performed 1000 iterations
for the remaining 12 parameters using parameter sets generated
2.4.2 Sensitivity analyses
through LHS method in the R FME package (Soetaert and
To identify the most sensitive model parameters, we adjusted
Petzoldt 2010). We chose not to employ the Sequential
the default value of each parameter (one by one) following
Uncertainty Fitting ver.2 (SUFI-2) calibration algorithm due
appropriate change methods (Abbaspour et al. 2015) and
to its tendency to produce only a single local solution.
compared the effects of these parameter changes on stream­
The automated calibration was performed following a
flow and water yield prediction with the default parameter set.
methodology outlined by Guse et al. (2020). This approach
One parameter at a time was considered and changes were
allowed the exploration of potential model performances
applied within the SWAT+ editor interface. Several parameter
within the specified parameter space. Streamflow simulated
value changes were applied for each parameter to quantify
by SWAT+ with these parameter sets was compared against
sensitivity and determine the maximum and minimum values
observed managed streamflow, leading to the selection of new
for the automatic calibration. Finally, we identified the model
parameter value ranges for the fourth iteration, comprising
parameters that showed the largest effect on streamflow and
500 parameter sets (see the Appendix, Table A2). Finally, in
basin water balance components (Table 3) and further opti­
the fourth iteration, parameters were optimized for high, mid­
mized their values via automatic calibration.
dle, and low flow segments using the last 500 parameter sets
and multiple model evaluation statistics such as NSE, KGE,
2.4.3 Manual and automatic calibration RSR, and PBIAS for managed streamflow prediction (Table 4).
We performed the model calibration using the SWATplusR To identify the optimal set of model parameters suitable for
package (Schürz 2019) and a multiple flow segment and multi- the high, middle, and low flow segments, ensuring acceptable
objective calibration approach using performance metrics and statistical indices for all objective functions, the hydroGOF
signature metrics (Pfannerstill et al. 2014a, 2014b, Haas et al. package in R (Zambrano-Bigiarini 2017) was employed.
2016). The SWATplusR can automatically initiate multiple Throughout this process, the historically observed streamflow
SWAT+ simulations with varying calibration parameter hydrograph and FDCs were used to select suitable set of para­
values. Moreover, it enables us to manage changes in model meters for the different flow segments. These parameter sets
parameters, simulation periods and time steps and to store the are calibrated to best fit the flows of the hydrograph within
simulated output variables. those segments, while other parts of the hydrograph might be
The sensitive parameters shown in Table 3 control different less accurately simulated.
aspects of the hydrological functioning of the Upper Feather Statistical indices for the specified objective functions were
River watershed. The sensitive parameters identified during computed, and iterations with NSE values greater than 0.5
the manual calibration were classified based on their relation from the last 500 iterations were systematically selected.
to the model’s spatial entities, such as HRUs, aquifers, soils, Subsequently, we scrutinized parameter sets that met the cri­
snow areas, and basin. These parameters were calibrated in teria of a KGE value greater than 0.5, PBIAS within the range
four stages. In the first stage, we conducted 500 iterations (first of −25% to 25%, and RSR less than 1.
iteration) using wide ranges of parameter values generated Ultimately, we derived parameter sets that reasonably satis­
through a Latin hypercube sampling (LHS) method in R with fied all objective functions for the average hydrograph and
the FME package (Soetaert and Petzoldt 2010) for 19 para­ across the different segments of the hydrograph, including
meters. The results helped determine new ranges for the sub­ high flow, middle flow, and low flow (see the Appendix,
sequent phase (see the Appendix, Table A1). In the second Table A2). Figure 3 summarizes the sequential modeling
stage, we conducted 1000 iterations (second iteration) focusing approach adopted in this study. Here we want to emphasize
solely on calibrating snowmelt, snowfall, precipitation lapse that the simulated hydrograph is not stitched together from
rate, and temperature lapse rate parameters independently. different flow segments. Each parameter set is used to simulate
This was done for two reasons. First, these parameters influ­ an entire hydrograph but with portions of the hydrograph
ence water contribution to the system, introducing a level of either simulated very well or less accurately. The user can
uncertainty that may arise from the model’s input data and then choose which parameter set to run the model with,
introduce uncertainty from the model (Abbaspour et al. 2017). depending on the desired application (e.g. if model is used to
Therefore, pre-parameter fitting was conducted to address this simulate low flows or high flows or all flows).
concern. Second, the decision to calibrate these parameters
separately was influenced by the recognition that employing
many model parameters during automatic calibration could 3 Results
lead to equifinality or difficulty in determining the optimal
3.1 Model parameterization
model parameterization from the available set (Casado-
Rodríguez and del Jesus 2022). Optimal parameter values Figure 4 illustrates the impact of changing model parameter
were identified by comparing the performance of the snowfall, value on water yield responses as a default model parameter
snowmelt, and lapse rate parameters to the annual basin aver­ value is adjusted to a new value. For instance, reducing the
age precipitation. At this stage, our reference to calibrate the default condition II curve number (CN2) value by 15 and
 HYDROLOGICAL SCIENCES JOURNAL 1205

Table 4. Optimized objective function values as calibrated based on the average hydrograph and the very high, high, middle, low, and very low flow segments of the
hydrographs. Values of interest are highlighted in bold.
Parameters adjusted based on the average hydrograph Parameters adjusted based on flow segments

Model parameterization based on best NSE

Hydrograph segment NSE PBIAS KGE RSR NSE PBIAS KGE RSR
Average 0.64 −9.30 0.62 0.59
Very high flow 0.45 −28.80 0.35 0.83 0.88 0.74 0.70 0.35
High flow 0.72 −10.20 0.79 0.63 0.99 0.20 0.97 0.10
Middle flow 0.98 7.00 0.91 0.14 0.99 −0.50 0.98 0.06
Low flow 0.56 0.60 0.36 0.83 0.99 −0.10 0.98 0.05
Very low flow −2.73 −21.10 0.50 0.76 0.96 −1.50 0.91 0.19
Model parameterization based on best KGE
Average 0.57 −2.5 0.74 0.65
Very high flow 0.77 −7.9 0.54 0.48 0.88 −3.8 0.88 0.35
High flow 0.76 9.0 0.79 0.48 0.97 −3.3 0.96 0.18
Middle flow 0.98 −4.8 0.91 0.14 0.99 −0.5 0.98 0.06
Low flow −3.63 −39.7 0.51 2.15 0.99 −0.1 0.98 0.05
Very low flow −29.22 −61.7 0.23 5.48 0.84 −3.2 0.96 0.34
Parameterization based on best PBIAS
Average 0.57 −0.10 0.65 0.66
Very high flow 0.56 −18.40 0.39 0.66 0.82 0.20 0.60 0.41
High flow 0.97 −2.90 0.95 0.18 0.79 −0.20 0.56 0.46
Middle flow 0.93 13.60 0.86 0.25 0.96 0.20 0.94 0.08
Low flow −2.79 35.80 0.51 1.94 0.95 0.83 0.00 0.22
Very low flow −20.89 52.60 0.38 4.67 0.94 −1.00 0.81 0.25

increasing the available water capacity of soil layer (awc) value Towner et al. (2019), hydrological simulations concluding with
by 50% significantly decreases and increases the monthly basin a KGE value falling within the range of 0.75 to 0.5 (0.75 ≥ KGE ≥
water yield, respectively. Likewise, modifying other default 0.5) are classified as intermediate, while simulations yielding a
parameter values results in either an increase or a decrease in KGE value of 0.75 or higher are deemed good. About 10% of
basin water yield (Fig. 4, also see the Appendix, Table A3). By simulations from a total of 500 model runs resulted in satisfac­
systematically adjusting one parameter at a time and analyzing tory or above satisfactory model performances when we con­
the simulated water balance components, it was possible to sidered the average hydrograph. However, both high and low
refine the maximum and minimum boundaries for each para­ flow segments of the hydrographs were consistently underesti­
meter (refer to Appendix Table A2). Among the 19 sensitivity mated when using a single parameter set to capture the average
model parameters that we identified during the preliminary hydrograph. Middle flows were relatively well captured by the
sensitivity analyses, CN2, latq_co, snomeltmax, plaps, and model and achieved the best NSE (Figs. 6 and 7). Similarly,
temperature lapse rate tlaps were identified to have the most optimization of the KGE to the average hydrograph resulted in
influence on streamflow (Fig. 4). satisfactory or more than satisfactory model performance, with
Although all 19 parameters were identified as sensitive approximately 13% of the parameter sets in the range of ade­
during the preliminary (manual) calibration, some of them quate model performance. Optimizing for PBIAS resulted in
did not show clear patterns (Fig. 5) in the automatic calibra­ nearly 12% of model simulations having a more than satisfac­
tion because the relative sensitivity of each parameter is tory model performance. About 4% of the runs achieved a
affected by the values of other parameters and the number of PBIAS value between −10% and 10%, which can be interpreted
runs (Abbaspour et al. 2017). as having achieved a very good model performance according to
Moriasi et al. (2007). About 6% of the model runs resulted in a
satisfactory model performance in terms of all four metrics
3.2 Model performance
(NSE, KGE, PBIAS, and RSR) for the average hydrograph.
The calibrated SWAT+ model performance was evaluated We further examined the best calibrated models for the
using multi-metric evaluation techniques that included average hydrograph based on NSE and KGE metrics
multi-objective functions and focused model calibration on (Table 4, left side). The parameter set with best overall
different segments of the streamflow hydrograph. Calibrating NSE (0.64) shows the lowest NSE for the lowest flows
the model in simulating the average hydrograph within the (−2.73) and highest flows (0.45), but a very high NSE
calibration period (2010–2020) resulted in acceptable model for the middle flows (0.98). Similarly, the parameter set
performances. Based on Moriasi et al. (2007), we found several with the best overall KGE (0.74) shows the lowest KGE
parameter sets that achieved adequate model performance (0.5 for the lowest flows (0.23), low flows (0.51) and highest
≤ NSE ≤ 0.65) when NSE was considered as the main objective flows (0.54), but a very high KGE for the middle flows
function in simulating the daily streamflow. Building on (0.91). This suggests that, overall, model calibration and
Moriasi et al. (2007), our assessment of model performances model performance were largely influenced by the middle
also incorporated the criteria outlined in Towner et al. (2019), flow segment and that the very low flow and very high
with the KGE serving as the chosen objective function. As per flow segments were largely omitted during the calibration
1206 T. B. TIGABU ET AL.

Figure 3. Flowchart outlining the steps involved in the modeling process, including data preparation, watershed configuration, manual calibration, and multi-segment
multi-objective function parameterization and calibration. The flowchart emphasizes the iterative nature of the calibration process, which enhances model
performance and confidence.

when the model was calibrated using the average hydro­ corresponded to the best NSE value (0.64), indicating that
graph (Table 4). Conversely, the best model performance having lower KGE values does not mean poor model perfor­
metrics under PBIAS optimization (−0.1) were influenced mance. Based on the recommendations of Moriasi et al. (2007)
by the high flow segment (−2.9) whereas the low and and Knoben et al. (2019), the optimization of KGE has been
lowest flow segments had higher absolute PBIAS values shown to yield more acceptable model runs compared to the
(35.8 and 52.6, respectively). optimization of the NSE objective.
According to Knoben et al. (2019), −0.41 is a benchmark About 40% of the parameter sets resulted in good model
value for KGE to decide whether a simulation is acceptable or performances, whereas only 10% of the parameter sets led to
not, whereas its equivalent value for NSE is zero based on satisfactory model performance in the case of the NSE optimi­
Moriasi et al. (2007). Because NSE and KGE include different zations. Because the KGE is the modified version of the NSE
metrics in the objective function, parameter sets that achieve that includes correlation, variability bias, and mean bias factors
high NSE values do not also automatically achieve high KGE (Knoben et al. 2019), the KGE optimization results found in this
values. For example, in our simulation a KGE value of 0.62 study are more robust to different segments of the hydrograph.
 HYDROLOGICAL SCIENCES JOURNAL 1207

Figure 4. Example graph showing the effects of different SWAT+ parameters on the simulated commutative monthly water yield values as compared to the default
simulation. Each graph was generated from a monthly simulation output based on changing one parameter at a time.

Figure 5. Scatter plots showing the association between NSE (as computed using simulated streamflow against observed managed streamflow records) and values of
different parameters.

In contrast, the mean streamflow value and coefficient of varia­ study yielded a deeper understanding of the model perfor­
tion are controlling factors in the NSE optimization that may mance. We found that the use of multiple objective functions
result in disparity between observed and simulated hydrological in evaluating the performance of the SWAT+ model increases
variables in a highly complex watershed with seasonal stream­ the reliability of the calibration and finally resulted in calibrated
flow dynamics. Because a single objective function cannot com­ model that was able to capture the seasonal and inter-annual
prehensively address these issues of model calibrations, using variations in streamflow that are characteristic for the Feather
multi-objective functions and multi-segment calibration in this River watershed.
1208 T. B. TIGABU ET AL.

Figure 6. Hydrographs showing a comparison of daily observed streamflow data versus simulated streamflow under different calibration options that optimized model
parameters based on the average hydrograph, or individual high flow, middle flow, and low flow segments that were independently optimized.

Figure 7. Flow duration curves of daily observed streamflow data versus simulated streamflow under different calibration options such as all flows (average), high flow,
middle flow, and low flow segments considered independently; multi-objective functions are optimized.

Calibration and parameterization of the model for different high flow segment independently. Similarly, all other model
segments of the streamflow hydrograph resulted in significantly performance metrics were improved when applying separate
improved model performances (Table 4, right side). For exam­ parameterization for each of the flow segments.
ple, the NSE value increased from 0.72 to 0.99 for the high flow Figure 8 illustrates the impact of various parameter sets on
segment when we calibrated the model parameter sets for the the predictive efficiency of the SWAT+ model in replicating
 HYDROLOGICAL SCIENCES JOURNAL 1209

Figure 8. Hydrographs showing a comparison of daily observed streamflow data in water year 2014 versus streamflow simulated with parameter sets that are
optimized to best capture either all flows (average), or the high flow, middle flow, and low flow segments. The dashed lines indicate the streamflow thresholds that
were used to optimize the high flow (> 240 m3/s), middle flow (40 to 240 m3/s), and low flow (<40 m3/s) segments.

the high, middle, and low flow segments of the observed functions to the model parameters. The results show that
streamflow. Calibration of model parameters specifically for optimal parameter values resulted depending on which objec­
high and low flow segments led to enhanced curve fitting in the tive function was chosen. For instance, values of parameters
hydrograph for both high flow and low flow segments, com­ which resulted in the best NSE did not necessarily result in a
pared to the calibration based on the average hydrograph. satisfactory KGE, PBIAS, and RSR at the same time. On one
However, the middle flow segment showed no significant hand, the best NSE solution (0.64) was achieved when the
changes, as it was optimally represented during the calibration default CN2 value was increased by an absolute value of 6.29.
based on the average hydrograph (average). On the other hand, the best KGE solution (0.74) was achieved
The calibrated model, which we achieved by joining the by increasing the default CN2 by an absolute value of 2.57.
parameterization of multi-objective functions and flow seg­ Similarly, parameter values which led to optimal solutions for
ments, was more than satisfactory for the validation period PBIAS differed from parameter values that resulted in optimal
(Fig. 9) than the model that was calibrated to the average solutions for RSR. A 20% increase in the default value of
hydrograph as per the calibration guideline proposed by available water content was required to get the best optimal
Moriasi et al. (2007). The performance metrics improved to solution for PBIAS, which is 2 times greater than the increase
0.62, 0.67, −3.6, and 0.61, for NSE, KGE, PBIAS, and RSR, required to achieve an optimum solution for RSR. Although
respectively. However, we found that the overall performance there were differences in parameter selections to obtain opti­
of the model during the calibration period (2010–2020) was mal solutions for NSE, KGE, and PBIAS, the parameter sets
higher than during the validation period (1995–2009), This that led to the best NSE solution also resulted in a good RSR.
lower model performance for the validation period (Table 5) This means direct optimization of parameter sets for NSE also
might be due to changes in precipitation and temperature that resulted in an optimum solution for RSR. Figure 10 demon­
occurred during that period since California entered a much strates the relationship between the NSE, RSR, KGE, and
drier climate regime in early 2000 (Pierce et al. 2016). PBIAS values that were computed under different model
runs and parameter sets. In maximizing the NSE values, the
RSR values are minimized on the 1:1 line and the optimum
3.3 Correlations between parameters and objective
NSE value is achieved at the smallest RSR value. Likewise,
functions
higher KGE values are associated with lower RSR values
Different objective functions were used to evaluate the good­ although there is more scatter at lower RSR and higher KGE
ness of fit of a calibrated model in simulating streamflow. In values, and the smallest RSR value does not result in the
this study, we evaluated the connection of the multi-objective optimal KGE value. Thus, considering one objective function
1210 T. B. TIGABU ET AL.

Figure 9. Hydrographs showing a comparison of daily observed streamflow data versus simulated streamflow for the validation period (2000–2009) which resulted
from combining multi-objective functions (NSE, KGE, and PBIAS) with the calibration of different flow segments based on the average hydrograph or for individual high
flow, middle flow, and low flow segments that were independently optimized.

Table 5. Optimized objective function values as validated based on the average, high, middle, and low flow segments of the hydrographs. Values of interest are
highlighted in bold.
Model parameterization based on best NSE Model parameterization based on best KGE

NSE PBIAS KGE RSR NSE PBIAS KGE RSR

Average 0.62 −3.60 0.67 0.61 0.52 2.50 0.74 0.69
High flow 0.92 −3.20 0.77 0.28 0.25 17.50 0.76 0.86
Middle 0.97 7.50 0.90 0.18 0.95 −0.20 0.78 0.22
Low −2.63 −23.10 −0.01 1.90 −26.03 −67.40 0.22 5.20
Model parameterization based on best PBIAS Parameters optimized to the best of NSE for high flow segment
Average 0.50 −18.60 0.52 0.70 0.54 3.10 0.67 0.68
High 0.19 −17.90 0.71 0.90 0.92 5.20 0.87 0.29
Mid 0.92 −11.70 0.80 0.28 0.95 11.30 0.89 0.22
Low −8.41 −38.70 0.16 3.07 −0.29 −2.80 −0.11 1.14
Parameters optimized to the best of NSE for middle flow Parameters optimized to the best of NSE for low flow
Average 0.51 2.90 0.68 0.70 0.41 −23.30 0.34 0.77
high 0.84 8.00 0.78 0.40 −1.85 −33.20 0.60 1.69
mid 0.93 8.30 0.68 0.26 0.82 −6.90 −0.77 0.42
low −1.96 −21.30 0.38 1.72 0.41 23.25 0.34 0.77

alone may result in statistically acceptable calibration and objective function, the optimal KGE parameter set has an even
validation results, which can affect other objective functions smaller PBIAS. Overall, we found that the parameter set with
or some of the hydrological processes. the highest KGE value also showed good results for all other
In this study, we found parameter sets that optimized PBIAS objective functions (Tables 4 and 5).
to its best solution (value = −0.1) in calibrating the SWAT+.
However, further inspection of the observed and simulated
3.4 Objective functions and their connection to the
streamflow hydrographs as well as the statistical solutions for
streamflow hydrograph
the different hydrograph components showed significantly high
deviation for the peak and low flows (Figs. 7 and 8). The NSE is one of the most frequently used objective functions in
parameter set resulting in the highest NSE had an acceptable calibrating hydrological models. The NSE is calculated based
(absolute) PBIAS value. Because the bias is included in the KGE on the observed mean and standard deviation, and the best
 HYDROLOGICAL SCIENCES JOURNAL 1211

Figure 10. Scatter plot showing the correlation of different objective functions. The triangles represent the best optimized values for each pair of objective functions
with the target objective function shown on the vertical axis; the square points represent the best values achieved based on the target objective function shown on the
horizontal axis.

calibration solution found in this study was 0.64 with values the 5th percentile flow that corresponded to the flow excee­
of 0.62 and −9.3, for KGE and PBIAS, respectively. If model dance probability of 95% ranged from 0 to 26 m3/s, while the
parameters were optimized using KGE as the objective func­ observed value was 18 m3/s. When considering NSE as the
tion, the best calibration solution resulted in a KGE of 0.74 primary objective function, the simulation output based on
with corresponding values of 0.57 and −2.5 for NSE and parameter sets fitted for the low flow segment provided the
PBIAS, respectively. Statistical solutions differed when the best representation of the observed value for the 5th percentile.
model was calibrated based on different objective functions In this case, the simulated value was overestimated by only 2%.
and hydrograph segments, which indicates that the improve­ Likewise, the parameter sets fitted to the average hydrograph
ment of one calibration criterion was achieved at the expense underestimated flows by 7%. Conversely, the simulated flow
of other calibration criteria and flow segments (Table 4). The rates using parameter sets fitted for the middle flow (stream­
optimum parameter set based on KGE and PBIAS reduced flow volume corresponding to the flow exceedance probability
the NSE values, highlighting the potential differences in the between 20% and 80%), and high flows (flow exceedance
model parameter selections and the part of hydrographs probability between 5% and 20%) segments resulted in under­
represented by each of the objective functions. Moreover, estimations of the flows by 18% and 34%, respectively.
NSE and KGE metrics did not correlate linearly and the Regarding the flow volume at the 80% flow exceedance
NSE metric did not relate well to the PBIAS and KGE probability, the model overestimated this value by 39%, 13%,
metrics. However, for a perfect model simulation both and 1% when using parameter sets fitted to the high, middle,
KGE and NSE should have a value of 1. When using NSE and low flow segments, respectively. The parameter values
as the main objective function the model performance is fitted for the extremely high and low flow segments led to an
evaluated based on a benchmark value, which is the mean underestimation and overestimation of the volume corre­
value of the observation, while the Euclidean distance from sponding to the 80% flow exceedance probability by 93% and
the point of ideal model performance is considered in the 35%, respectively. Simulations using fitted parameters for the
case of KGE performance (Equation 2). average hydrograph and middle flows segment resulted in a
better representation of the median flow. The observed median
flow was underestimated by about 2% and 0.3% when using
3.5 Comparison of simulated percentiles under different
models calibrated to the average hydrograph and middle flow
objective functions
segment, respectively. Similarly, the third quantile (a flow
For each objective function and calibrated model parameter volume corresponding to the 25% flow exceedance probabil­
set, the simulated streamflow was compared to the observed ity) was better estimated by the model calibrated to the average
streamflow for both the calibration and validation periods hydrograph, whereas the flow volume at the 5% exceedance
using threshold values based on observed flow percentiles probability was best estimated by the high flow model, with
(Figs. 11 and 12). Variations in the fitted model parameters only a 0.7% underestimation. The middle flow and average
and objective functions led to significant differences in flow flow models overestimated the flow volume at the 5% excee­
percentiles. For instance, the simulated flow volume value for dance probability by 5% and 4%, respectively. These results are
1212 T. B. TIGABU ET AL.

Figure 11. Scatter plot in the log10 scale showing the observed managed streamflow at Oroville gauging station against SWAT+ simulated streamflow at the same
gauging station with parameter sets optimized to the best of NSE, KGE, PBIAS, high flow, middle flow, and low flow for the water years between 2010 and 2020
(calibration period). The red broken lines indicate threshold values of low and high flows.

Figure 12. Scatter plot in the log10 scale showing the observed managed streamflow at Oroville gauging station against SWAT+ simulated streamflow at the same
gauging station with parameter sets optimized to the best of NSE, KGE, PBIAS, high flow, middle flow, and low flow for the water years between 2000 and 2009
(validation period).

consistent with the corresponding model performance values probabilities. In contrast, a model calibrated to the optimal
(see Tables 4 and 5). solution of KGE resulted in an underestimation of the low flow
Furthermore, the model parameters that led to optimal (80 to 100% exceedance probabilities) and overestimation of
values of NSE, KGE, and PBIAS for the average hydrograph the median flow. A model calibrated to the optimal value of
estimated different flow volumes with various uncertainty PBIAS led to an overestimation for all flow volumes corre­
levels. A model fitted to the optimal value of NSE resulted in sponding to exceedance probabilities of 5 to 100% and under­
overestimations of the low, high, and very high flow volumes, estimation of flow volumes at 0 and 5% exceedance
and underestimation at the 50%, 25%, and 5% exceedance probabilities.
 HYDROLOGICAL SCIENCES JOURNAL 1213

Overall, there were clear differences among the calibrated While combining flow segments for model calibration can
models for the optimal solutions of different objective func­ yield a plausible model, the hydrograph and FDCs can incor­
tions and flow segments in mimicking the observed flow porate both poorly and well-performing segments.
volumes. Thus, a single objective function and parameter set Consequently, the overall performance of the model may be
might not appropriately represent the hydrological processes influenced more by either the good or poor performing seg­
of various flow stages. This could be due to the complex water ments. In our study, we obtained credible model performance
abstractions, heterogeneity of land use/cover, high seasonal for the entire streamflow calibration, at least considering the
differences, and high variability of model parameters to ade­ employed objective function. Based on Moriasi et al. (2007)
quately capture the dynamic hydrological processes over dif­ and Towner et al. (2019), the statistical values for the objective
ferent seasons. Consequently, it is worth calibrating the SWAT functions fell within an acceptable range. However, significant
+ model separately for different flow segments using multi- deviations were observed for extreme high and low flows
objective functions to achieve a more accurate representation compared to the observed streamflow (Figs. 6 and 7). The
of streamflow and flow volumes. observed discrepancy in the high flow segment may be
ascribed to a limitation of the curve number method. The
soil conservation service (SCS)-curve number method used
4 Discussion
in the SWAT model, as emphasized by Nie et al. (2011), does
Although the new SWAT+ was used to simulate managed not consider the duration and intensity of rainfall. Moreover,
streamflow in a complex hydrological system, characterized the intricate water delivery cascading reservoir system and the
by high seasonal differences and frequent extreme flows, cali­ complex climate of California further contribute to the discre­
brating a comprehensive SWAT+ model for Feather River pancy in the high flow segment. Therefore, independent para­
watershed posed challenges. meterization for different flow segments can contribute to
In our study, we independently parameterized the SWAT+ constructing plausible models, increasing confidence in the
model for low flow, middle flow, and high flow segments of the use of calibrated models for various purposes. The improve­
hydrograph, leading to significant improvements in model per­ ments in the predictive power of the SWAT+ model achieved
formance. These improvements were achieved by appropriately through calibration for different flow segments were also evi­
representing sensitive model parameters. Among the 19 sensi­ dent in the percentile flow estimates.
tive parameters, CN2 stood out as one of the most influential, Furthermore, the independently calibrated SWAT+ model
causing overestimation or underestimation of flow segments. Its can predict future streamflow for high, low, and middle flows
value was adjusted differently for each flow segment. by incorporating projected precipitation and temperature data
Initially, when simulating the entire streamflow hydrograph based on various emission scenarios. The differences observed
using a calibrated SWAT+ model, underestimation occurred in parameter selection and model performance across different
for high and middle flows, while overestimation was observed hydrograph segments indicate that employing multiple para­
for low flow segments compared to observed streamflow. To meter sets can significantly enhance the accuracy and reliabil­
mitigate this, we increased the CN2 value by 13.5, reducing the ity of predictions for high and low flows in the future.
underestimation from 10.2% to 1.4% for the high flow seg­ Consequently, the integration of a multi-parameter and
ment. Similarly, increasing the CN2 value by 9.8 improved the multi-segment calibration in the SWAT+ model will provide
simulation of middle flows and enhanced the goodness of fit valuable insights into the most probable future conditions
for the hydrograph, which was previously underestimated. In within the study catchment. Specially, the calibrated SWAT+
contrast, for the low flow segment, which was overestimated model will be utilized to explore the potential incidences of
during the average simulation, we decreased the CN2 value by peak flows and low flow events. This investigation will involve
4.6 to achieve an adequate curve fitting and improved model applying parameter sets tailored for high and low flow seg­
performance. Likewise, the values of other parameters tailored ments, considering future climate conditions as outlined by
to fit the average hydrographs were adjusted to new values that the shared socioeconomic pathway (SSP) and representative
led to a satisfactory simulation of various flow segments when concentration pathway (RCP) scenarios. Implementation of
employing multi-objective functions and parameterization. multi-segment parameterization and calibration approaches
Various studies worldwide have demonstrated that becomes crucial in regions characterized by extreme climate
updating model parameter values and employing multi- conditions, such as California. These approaches are instru­
objective functions for different flow segments can enhance mental in simulating extremes, including droughts and floods,
the predictive capabilities of hydrological models. Tegegne under diverse climate change scenarios, such as SSPs and
et al. (2019) calibrated a SWAT model for two hydro- RCPs.
geographically distinct catchments, one well managed and When considering the connections between objective func­
the other poorly managed, and found that using different tions and flow segments, optimizing the NSE and RSR resulted
CN2 values for different flow components improved the in similar improvements of different parts of the flow seg­
model’s ability to predict various streamflow stages. ments. Maximizing NSE for the entire streamflow led to
Pfannerstill et al. (2014) reported that calibrating the underestimation of high and very low flow segments, while
SWAT model based on multiple flow segments improves overestimating middle flows. Similarly, maximizing KGE sig­
its performance across different parts of the streamflow nificantly underestimated the low flow segment. In contrast,
hydrograph, including the overall flow hydrograph and optimizing PBIAS resulted in an overestimation of the low
very low flows. flow segment. By calibrating the model using a combination
1214 T. B. TIGABU ET AL.

of all objective functions, the middle flow segment was better segments. Consequently, the overall performance of the
simulated, with the lowest standardized RMSE and highest model may be influenced more by the good segment or the
NSE value. Moreover, all statistical values for the other flow poorly performing segment. Therefore, it is valuable to
segments were acceptable, indicating that combining multiple account for multi-objective functions simultaneously to
objective functions improves the overall model performance. obtain a credible model that can balance the trade-off
However, each individual objective function exhibited slight between different objective functions.
deterioration compared to the optimal values obtained
through individual calibration. This finding aligns with In general, this study emphasizes the significance of indepen­
Gupta et al. (2009), and Garcia et al. (2017) who also reported dent flow segment calibration using multi-objective functions
improved simulation of seasonal and annual mean streamflow to accurately represent flow conditions during wet, average
in TOPMODEL when using combined objective functions. flow, and dry periods. Consequently, parameterizing hydro­
Comparing the parameter values fitted to the optimal solu­ logical models based on different flow conditions is crucial for
tions of NSE, KGE, and PBIAS (Table 4) with those fitted to constructing reliable hydrological models. As a result, this
the high, middle, and low segments, slight differences in the study provides valuable insights for water managers and
Pearson correlation coefficient (r) were observed. In both the researchers in effectively managing water resources during
NSE and KGE cases, parameter values fitted to the middle high and low flow seasons.
segment exhibited the highest correlation coefficient
(r = 0.97), while the low flow parameter values showed the
highest correlation coefficient when PBIAS (r = 0.91) was used Acknowledgements
as the objective function. When comparing the parameter We acknowledge the California Department of Water Resources for
correlation between NSE and KGE, NSE and PBIAS, and providing the PRISM climate data sources and Lawrence Livermore
KGE and PBIAS, NSE and KGE demonstrated the highest National Laboratory under contract DE-AC52-07NA27344 for the fund­
ing support under Project 22-SI-008.
correlation (r = 0.96), followed by NSE and PBIAS (r = 0.95),
indicating that NSE and KGE have a similar effect on different
flow segments. Disclosure statement
No potential conflict of interest was reported by the author(s).
5 Conclusion
This study aimed to enhance the simulation of managed Funding
streamflow in the Feather River Watershed, Sierra Nevada,
California, United States, by exploring multi-objective func­ Part of this work was performed under the auspices of the US Department
tions and multi-segment calibration approaches for the SWAT of Energy by Lawrence Livermore National Laboratory under contract
DE-AC52-07NA27344 and was supported by the LLNL-LDRD Program
+ model. The study investigated how model parameters varied under Project 22-SI-00. The U.S. Department of Energy through
concerning different objective functions and distinct stream­ Lawrence Livermore National Laboratory.
flow segments of the hydrograph (e.g. high, middle, and low
flow segments). The model evaluation criteria NSE, KGE,
PBIAS, and RSR, were employed to assess the simulated man­ ORCID
aged streamflow. The study’s findings led to the following Tibebe B. Tigabu http://orcid.org/0000-0003-0104-5924
conclusions:

(1) Parameterizing hydrological models based on different References

flow segments (e.g. high flow, middle flow, and low Abbaspour, K.C., et al., 2015. A continental-scale hydrology and water
flow) of the hydrograph improved the prediction of quality model for Europe: calibration and uncertainty of a high-resolu­
streamflow compared to using a model calibrated to tion large-scale SWAT model. Journal of Hydrology, 524, 733–752.
the average hydrograph. doi:10.1016/j.jhydrol.2015.03.027.
(2) When a single model calibration against the entire Abbaspour, K.C., Vaghefi, S.A., and Srinivasan, R., 2017. A guideline for
successful calibration and uncertainty analysis for soil and water
streamflow hydrograph was used the middle flow seg­ assessment: a review of papers from the 2016 international SWAT
ment of the hydrograph was simulated best, while low conference. Water, 10 (1), 6. doi:10.3390/w10010006.
flow and high flow segments were underestimated in Alemayehu, T., et al., 2022. On the calibration of spatially distributed
the simulation. Independently fitting parameters for hydrologic models for poorly gauged basins: exploiting information
different flow segments significantly enhanced the from streamflow signatures and remote sensing-based evapotranspira­
tion data. Water, 14 (8), 1252. doi:10.3390/w14081252.
model’s performance for each segment. This suggests Aliye, M.A., et al., 2020. Evaluating the performance of HEC-HMS and
that proposing multiple sets of model parameters can SWAT hydrological models in simulating the rainfall-runoff process for
increase confidence in constructing a reliable model for data scarce Region of Ethiopian Rift Valley Lake Basin. Open Journal of
managed streamflow predictions. Modern Hydrology, 10 (4), 105. doi:10.4236/ojmh.2020.104007.
(3) Optimizing model parameters for a single objective func­ Arnold, J.G., et al., 1998. Large area hydrologic modeling and assessment part
I: model development 1. JAWRA Journal of the American Water Resources
tion to its optimal value may lead to a deterioration in Association, 34 (1), 73–89. doi:10.1111/j.1752-1688.1998.tb05961.x.
another objective function. However, the hydrograph and Arnold, J.G., et al., 2018. Use of decision tables to simulate management
FDCs can encompass both poorly and well-performing in SWAT+. Water, 10 (6), 713. doi:10.3390/w10060713.
 HYDROLOGICAL SCIENCES JOURNAL 1215

Arnold, J.G. et al., 2012. SWAT: Model use, calibration, and validation. Hanak, E. and Lund, J.R., 2012. Adapting California’s water management
Transactions of the ASABE, 55 (4), 1491–1508. to climate change. Climatic Change, 111, 17–44. doi:10.1007/s10584-
Arnold, J.G. and Fohrer, N., 2005. SWAT2000: current capabilities and 011-0241-3.
research opportunities in applied watershed modelling. Hydrological Hargreaves, G.H. and Samani, Z.A., 1985. Reference crop evapotranspira­
Processes: An International Journal, 19 (3), 563–572. doi:10.1002/hyp. tion from temperature. Applied Engineering in Agriculture, 1 (2), 96–
5611. 99. doi:10.13031/2013.26773.
Avanzi, F., 2018. Feather river hydrologic observatory: improving snowpack Hedden-Nicely, D.R., 2022. Climate change and the future of western US
forecasting for hydropower generation using intelligent information water governance. Nature Climate Change, 12 (2), 108–110. doi:10.
systems: a report for California’s fourth climate change assessment. 1038/s41558-021-01141-3.
Sacramento, CA, USA, California Energy Commission. Homer, C., et al., 2004. Development of a 2001 national land cover
Bailey, R.T., et al., 2023. Augmenting the national agroecosystem model database for the United States. Photogrammetric Engineering &
with physically based spatially distributed groundwater modeling. Remote Sensing, 70 (7), 829–840. doi:10.14358/PERS.70.7.829.
Environmental Modelling and Software, 160, 105589. doi:10.1016/j. Huang, G., Kadir, T., and Chung, F., 2012. Hydrological response to
envsoft.2022.105589. climate warming: the upper feather river watershed. Journal of
Barnett, T.P., Adam, J.C., and Lettenmaier, D.P., 2005. Potential impacts Hydrology, 426, 138–150.
Immerzeel, W.W., et al., 2020. Importance and vulnerability of the
of a warming climate on water availability in snow-dominated regions.
world’s water towers. Nature, 577 (7790), 364–369. doi:10.1038/
Nature, 438 (7066), 303–309. doi:10.1038/nature04141.
s41586-019-1822-y.
Bieger, K., et al., 2017. Introduction to SWAT+, a completely restructured
Immerzeel, W.W., Van Beek, L.P., and Bierkens, M.F., 2010. Climate
version of the soil and water assessment tool. JAWRA Journal of the
change will affect the Asian water towers. Science, 328 (5984), 1382–
American Water Resources Association, 53 (1), 115–130. doi:10.1111/
1385. doi:10.1126/science.1183188.
1752-1688.12482.
Jahn, R., et al., 2006. Guidelines for soil description. Vol. 1, Rome, Italy:
Casado-Rodríguez, J. and Del Jesus, M., 2022. Hydrograph separation for Food and Agriculture Organization of the United Nations.
tackling equifinality in conceptual hydrological models. Journal of Kannan, N., et al., 2019. Some challenges in hydrologic model calibration
Hydrology, 610, 127816. for large-scale studies: a case study of SWAT model application to
Chen, Y. et al., 2023. Evaluation and uncertainty assessment of weather Mississippi-Atchafalaya River Basin. Hydrology, 6 (1), 17. doi:10.3390/
data and model calibration on daily streamflow simulation in a large- hydrology6010017.
scale regulated and snow-dominated river basin. Journal of Hydrology, Knoben, W.J., Freer, J.E., and Woods, R.A., 2019. Inherent benchmark or
617, 129103. not? Comparing Nash–Sutcliffe and Kling–Gupta efficiency scores.
Daggupati, P., et al., 2015. Impact of model development, calibration, and Hydrology and Earth System Sciences, 23 (10), 4323–4331. doi:10.
validation decisions on hydrological simulations in West Lake Erie 5194/hess-23-4323-2019.
Basin. Hydrological Processes, 29 (26), 5307–5320. doi:10.1002/hyp. Koczot, K.M., et al., 2004. Precipitation-runoff processes in the Feather
10536. River Basin, northeastern California, with prospects for streamflow
Dal Molin, M., et al., 2023. Exploring signature‐based model calibration predictability, water years 1971–97. US Geological Survey Scientific
for streamflow prediction in ungauged basins. Water Resources Investigations Rep, 5202, 82.
Research, 59 (7), e2022WR031929. doi:10.1029/2022WR031929. Koczot, K.M., Markstrom, S.L., and Hay, L.E., 2012. Watershed scale
Donnelly, C., Andersson, J.C., and Arheimer, B., 2016. Using flow signa­ response to climate change—Feather River Basin, California (No.
tures and catchment similarities to evaluate the E-HYPE multi-basin 2011-3125). US Geological Survey.
model across Europe. Hydrological Sciences Journal, 61 (2), 255–273. Lakshmi, G., and Sudheer, K.P., 2021. Parameterization in hydrological
doi:10.1080/02626667.2015.1027710. models through clustering of the simulation time period and multi-
DWR (California Department of Water Resources). 2022. Pre-existing objective optimization based calibration. Environmental Modelling &
SWAT2009 model and daily weather data. Software, 138, 104981.
DWR (California Department of Water Resources). 2023. The California Lindström, G., et al., 2010. Development and testing of the HYPE (hydro­
State Water Project. https://water.ca.gov/programs/state-water-pro logical predictions for the environment) water quality model for dif­
ject# [Accessed 17 March 2023]. ferent spatial scales. Hydrology Research, 41 (3–4), 295–319. doi:10.
Garcia, F., Folton, N., and Oudin, L., 2017. Which objective function to 2166/nh.2010.007.
calibrate rainfall–runoff models for low-flow index simulations? Liu, Z., et al., 2021. Identifying climate change impacts on surface
Hydrological Sciences Journal, 62 (7), 1149–1166. doi:10.1080/ water supply in the Southern Central Valley, California. Science of
02626667.2017.1308511. the Total Environment, 759, 143429. doi:10.1016/j.scitotenv.2020.
Gassman, P., et al., 2009. The agricultural policy environmental extender 143429 .
(APEX) model: an emerging tool for landscape and watershed envir­ Mahmoodi, N., et al., 2021. Spatially distributed impacts of climate
change and groundwater demand on the water resources in a wadi
onmental analyses.
system. Hydrology and Earth System Sciences, 25 (9), 5065–5081.
Golmohammadi, G., et al., 2014. Evaluating three hydrological distributed
doi:10.5194/hess-25-5065-2021.
watershed models: MIKE-SHE, APEX, SWAT. Hydrology, 1 (1), 20–39.
Mengistu, A.G., van Rensburg, L.D., and Woyessa, Y.E., 2019. Techniques
doi:10.3390/hydrology1010020.
for calibration and validation of SWAT model in data scarce arid and
Gupta, H.V., et al., 2009. Decomposition of the mean squared error and
semi-arid catchments in South Africa. Journal of Hydrology: Regional
NSE performance criteria: implications for improving hydrological
Studies, 25, 100621.
modelling. Journal of Hydrology, 377 (1–2), 80–91. doi:10.1016/j.jhy Moriasi, D.N., et al., 2007. Model evaluation guidelines for systematic
drol.2009.08.003. quantification of accuracy in watershed simulations. Transactions of
Gupta, H.V., Sorooshian, S., and Yapo, P.O., 1999. Status of automatic the ASABE, 50 (3), 885–900. doi:10.13031/2013.23153.
calibration for hydrologic models: comparison with multilevel expert Nash, J.E. and Sutcliffe, J.V., 1970. River flow forecasting through con­
calibration. Journal of Hydrologic Engineering, 4 (2), 135–143. ceptual models’ part I—A discussion of principles. Journal of
Guse, B., et al., 2020. Improving information extraction from simulated Hydrology, 10 (3), 282–290.
discharge using sensitivity‐weighted performance criteria. Water Nelson, T., et al., 2016. Reservoir operating rule optimization for
Resources Research, 56 (9), e2019WR025605. doi:10.1029/ California’s Sacramento Valley. San Francisco Estuary and Watershed
2019WR025605. Science, 14 (1). doi:10.15447/sfews.2016v14iss1art6.
Haas, M.B. et al., 2016. A joined multi-metric calibration of river dis­ Nie, W., et al., 2011. Assessing impacts of landuse and landcover changes
charge and nitrate loads with different performance measures. Journal on hydrology for the upper San Pedro watershed. Journal of Hydrology,
of Hydrology, 536, 534–545. 407 (1–4), 105–114. doi:10.1016/j.jhydrol.2011.07.012.
1216 T. B. TIGABU ET AL.

Pfannerstill, M. et al., 2017. How to constrain multi‐objective calibrations Tigabu, T.B., et al., 2020. Modeling the spatio-temporal flow dynamics of
of the SWAT model using water balance components. JAWRA Journal groundwater-surface water interactions of the Lake Tana Basin, Upper
of the American Water Resources Association, 53 (3), 532–546. Blue Nile, Ethiopia. Hydrology Research, 51 (6), 1537–1559. doi:10.
Pfannerstill, M., Guse, B., and Fohrer, N., 2014. Smart low flow signature 2166/nh.2020.046.
metrics for an improved overall performance evaluation of hydrological Tigabu, T.B., et al., 2023. Pitfalls in hydrologic model calibration in a data
models. Journal of Hydrology, 510, 447–458. doi:10.1016/j.jhydrol.2013.12. scarce environment with a strong seasonality: experience from the
044. Adyar catchment, India. Environmental Earth Sciences, 82 (15), 367.
Pierce, D.W., Cayan, D.R., and Dehann, L., 2016. Creating climate projec­ doi:10.1007/s12665-023-11047-2.
tions to support the 4th California climate assessment. La Jolla, CA, Towner, J., et al., 2019. Assessing the performance of global hydrological
USA: University of California at San Diego, Scripps Institution of models for capturing peak river flows in the Amazon basin. Hydrology
Oceanography. and Earth System Sciences, 23 (7), 3057–3080. doi:10.5194/hess-23-
PRISM Climate Group, Oregon State University, 4 February 2014. https:// 3057-2019.
prism.oregonstate.edu/terms/. [Accessed 28 July 2022]. United States Geological Survey (USGS). 2022. USGS water data for the
Refsgaard, J.C. and Storm, B., 1995. Mike She. In: V.P. Singh, ed. nation [Online]. National Water Information System: Web Interface.
Computer models of watershed hydrology. Colorado: Water Resources https://waterdata.usgs.gov/nwis [Accessed 27 July 2022].
Publication, 809–847. Wagner, P.D., et al., 2022. Representation of hydrological processes in a rural
Schürz, C., 2019. Development of tools and interfaces for the implemen­ lowland catchment in Northern Germany using SWAT and SWAT+.
tation, pre-processing, and sensitivity analysis of the SWAT model in Hydrological Processes, 36 (5), e14589. doi:10.1002/hyp.14589.
an R programming environment. Westerberg, I.K., et al., 2011. Calibration of hydrological models using
flow-duration curves. Hydrology and Earth System Sciences, 15 (7),
Shafii, M. and Tolson, B.A., 2015. Optimizing hydrological consistency by
2205–2227. doi:10.5194/hess-15-2205-2011.
incorporating hydrological signatures into model calibration objec­
White, M.J., et al., 2022. Development of a field scale SWAT+ modeling
tives. Water Resources Research, 51 (5), 3796–3814.
framework for the contiguous US. JAWRA Journal of the American
Siirila-Woodburn, E.R., et al., 2021. A low-to-no snow future and its impacts
Water Resources Association, 58 (6), 1545–1560. doi:10.1111/1752-
on water resources in the western United States. Nature Reviews Earth &
1688.13056.
Environment, 2 (11), 800–819. doi:10.1038/s43017-021-00219-y.
Wu, H., et al., 2021. An improved calibration and uncertainty analysis
Soetaert, K. and Petzoldt, T., 2010. Inverse modelling, sensitivity and
approach using a multicriteria sequential algorithm for hydrological mod­
monte carlo analysis in R using package FME. Journal of Statistical eling. Scientific Reports, 11 (1), 16954. doi:10.1038/s41598-021-96250-6.
Software, 33, 1–28. Wu, J., et al., 2020. Development of reservoir operation functions in
Tegegne, G., et al., 2019. Hydrological modelling uncertainty analysis for SWAT+ for national environmental assessments. Journal of
different flow percentiles: a case study in two hydro-geographically Hydrology, 583, 12455. doi:10.1016/j.jhydrol.2020.124556.
different watersheds. Hydrological Sciences Journal, 64 (4), 473–489. Zambrano-Bigiarini, M., 2017. Package ‘hydroGOF’. Goodness-of-fit
doi:10.1080/02626667.2019.1587562. functions for comparison of simulated and observed.

Appendix

Table A1. Parameter value ranges that were used for LHS resulted in 500, 1000, 1000, and 500 parameter sets during the four phases of automatic calibration. The first
phase of 500 iterations was carried out to determine the minimum and maximum value ranges for 19 sensitive parameters; the next phase of 1000 iterations was
carried out to calibrate snowmelt, snowfall, temperature, and precipitation lapse rate parameters; and the next 1000 and 500 iterations were used to calibrate other
parameters.
First 500 runs Fourth 500 runs
Allowable range (preliminary phase) Second 1000 runs Third 1000 runs (final phase)

Parameter Default Absolute min Absolute max Min Max Min Max Min Max Min Max Change method
snomelt_tmp 1 −5 5 −5.00 5.00 −5.00 5.00 Calibrated independently in previous step Replace
snofall_tmp 0.5 −5 5 −5.00 5.00 −5.00 3.75 Replace
snomelt_max 4.5 0 10 0.00 10.00 −5.00 9.00 Replace
snomelt_min 4.5 0 10 0.00 10.00 0.00 9.00 Replace
snomelt_lag 1 1 1 0.00 1.00 0.00 1.00 Replace
plaps 0 0 200 0.00 200.00 0.00 75.00 Replace
tlaps −10 −10 10 −10.00 10.00 −10.00 10.00 Replace
cn2 Variable 35.00 95.00 −20.00 20.00 Default Default −15.00 15.00 −5.00 15.00 Relative
[35,98]
esco 0.95 0.00 1.00 0.00 1.00 0.95 0.95 0.00 0.60 0.00 0.60 Replace
epco 1.00 0.00 1.00 0.01 1.00 1.00 1.00 0.30 0.80 0.30 0.80 Replace
lat_ttime 0.00 0.50 180.00 0.00 180.00 0.00 0.00 1.00 75.00 1.00 75.00 Replace
awc Variable 0.01 1.00 0.01 0.50 Default Default 0.06 0.48 0.06 0.48 Relative change
[0,1]
perco 0.90 0.00 1.00 0.00 1.00 0.90 0.90 0.13 0.21 0.13 0.21 Replace
alpha 0.05 0.00 1.00 0.00 1.00 0.05 0.05 0.04 0.86 0.04 0.86 Replace
flo_min 3.00 0.00 50.00 0.00 50.00 3.00 3.00 13.50 40.00 13.50 38.00 Replace
revap 0.02 0.02 0.20 0.02 0.20 0.02 0.02 0.02 0.12 0.02 0.12 Replace
rchg_dp 0.05 0.00 1.00 0.00 1.00 0.05 0.0.5 0.13 0.80 0.13 0.80 Replace
revap_min 5.00 0.00 500.00 0.00 500.00 5.00 5.00 0.00 32.00 0.00 32.00 Replace
latq_co 0.01 0.00 1.00 0.00 1.00 0.01 0.01 0.00 0.90 0.00 0.90 Relative
Long name for model parameters: lat_ttime: exponential of the lateral flow travel time; esco: soil evaporation compensation factor; epco: plant water uptake
compensation factor; snomelt_tmp: snowmelt temperature; snofall_tmp: snow fall temperature; snomelt_max: maximum snowmelt temperature; snomelt_min:
minimum snowmelt temperature; snomelt_lag: snow pack temperature lag factor; perco: percolation coefficient; awc: available water capacity of soil layer; alpha:
alpha factor for groundwater recession curve; CN2: condition II curve number; revap: threshold depth of water in shallow aquifer required to allow revap to occur;
rchg_dp: recharge to deep aquifer (the fraction of root zone percolation that reaches the deep aquifer); revap_min: water table depth for revap to occur water table
depth for revap to occur; flo_min: water table depth for return flow to occur; plaps: precipitation lapse rate (mm per km of elevation difference); tlaps: temperature
lapse rate; deg C per km of elevation difference, latq_co: Plant ET curve number coefficient; replace: absolute value; relative change: adding/subtracting the value to
the default one. The change methods relative indicate add/subtract the value on the default and replace indicates replace the default value by the absolute value of
the new number.
 HYDROLOGICAL SCIENCES JOURNAL 1217

Table A2. Optimization matrix designed to facilitate the integration of multi-parameters and multi-segment calibration, resulting in final parameter sets for running the
model for the intended purpose. Each run encompasses multiple sets of parameter values and statistical indices for both the average and different flow segments. For
brevity, this table excludes RSR and PBIAS computations for the average and other flow segments, considering space constraints.
Flow segments separated based on historical streamflow threshold values

Average hydrograph NSE for the different segments KGE for the different segments

Run Objective function NSE PBIAS KGE RSR Very high High Mid Low Very low Very high High Mid Low Very low
121 NSE 0.64 −9.30 0.62 0.60 0.45 0.72 0.98 0.56 −2.73 0.35 0.79 0.91 0.36 0.50
56 PBIAS 0.57 −0.10 0.65 0.66 0.56 0.97 0.94 −2.79 −20.89 0.39 0.95 0.86 0.51 0.38
454 KGE 0.57 −2.50 0.74 0.65 0.77 0.76 0.98 −3.63 −29.22 0.54 0.79 0.91 0.51 0.23
121 RSR 0.64 −9.30 0.62 0.60 0.45 0.72 0.98 0.56 −2.73 0.35 0.79 0.91 0.36 0.50
408 Very_high_NSE 0.31 −44.70 0.43 0.83 0.88 −2.64 −1.44 −24.53 −78.71 0.70 0.46 −0.04 −0.18 NA
340 High_NSE 0.60 −2.60 0.66 0.63 0.58 0.99 0.97 −4.18 −9.30 0.41 0.96 0.89 0.09 0.45
136 Mid_NSE 0.57 −4.00 0.68 0.66 0.66 0.98 0.99 0.42 −1.32 0.45 0.94 0.98 0.74 0.54
194 Low_NSE 0.31 −49.60 0.17 0.83 −0.29 −6.44 −0.05 0.99 0.82 0.10 0.26 0.21 0.98 0.94
303 Very_low_NSE 0.42 −29.50 0.39 0.76 0.25 −1.73 0.66 0.73 0.96 0.28 0.58 0.51 0.53 0.91
408 Very_high_KGE 0.31 −44.70 0.43 0.83 0.88 −2.64 −1.44 −24.53 −78.71 0.70 0.46 −0.04 −0.18 NA
102 High_KGE 0.60 −7.50 0.65 0.63 0.57 0.97 0.99 0.32 −5.86 0.40 0.97 0.94 0.29 0.70
136 Mid_KGE 0.57 −4.00 0.68 0.66 0.66 0.98 1.00 0.42 −1.32 0.45 0.94 0.98 0.74 0.54
194 Low_KGE 0.31 −49.60 0.17 0.83 −0.29 −6.44 −0.05 1.00 0.82 0.10 0.26 0.21 0.98 0.94
281 Very_low_KGE 0.46 −32.7. 0.37 0.73 0.10 −2.16 0.62 0.84 0.89 0.22 0.51 0.50 0.66 0.96
Long name or meanings: NSE: Nash-Sutcliff Efficiency; KGE: Kling–Gupta efficiency; PBIAS: Percent bias; RSR: standardized root means square error; Very_high, High,
MID, Low, and Very_Low: very high flow, high, middle, low, and very low flow segment.

Table A3. Displays how changing the default value of a parameter affects water yield. The comparison is between the simulation using the default parameters and
simulations done by changing the value of a single parameter at a time.
% change on the % change in
Parameter Default value Minimum Maximum Change method Changed value default para value water yield
CN2 Variable 35 98 Relative −15.00 Variable −14.29
tlaps 6.5 0 10 Replace 10.00 53.85 1.22
latq_co 0.01 0 1 Replace 0.50 98.00 102.86
awc Variable 0 1 Relative 0.50 Variable 11.43
snomelt_max 4.5 0 10 Replace 10.00 122.22 0.12
perco 1 0 1 Replace 0.90 10.00 27.96
plaps 0 −25 25 Replace 10.00 Infinity 77.55

Figure A1. Daily volume of released and stored water from/in the Oroville Reservoir as simulated using calibrated SWAT+ models for high, middle, and low flow
simulation. The release and storage patterns reflect the historical wet and dry years. As the calibration work did not focus on release and storage volumes specifically,
the uncertainty levels could be higher than expected. The figure keys high, middle, and low refer to simulations under the parametrization of high, middle, and low
flow segments.