www.nature.

com/scientificreports

OPEN IHACRES, GR4J and MISD‑based

multi conceptual‑machine learning
approach for rainfall‑runoff
modeling
1* 2 3
Babak Mohammadi , Mir Jafar Sadegh Safari & Saeed Vazifehkhah
As a complex hydrological problem, rainfall-runoff (RR) modeling is of importance in runoff studies,
water supply, irrigation issues, and environmental management. Among the variety of approaches
for RR modeling, conceptual approaches use physical concepts and are appropriate methods for
representation of the physics of the problem while may fail in competition with their advanced
alternatives. Contrarily, machine learning approaches for RR modeling provide high computation
ability however, they are based on the data characteristics and the physics of the problem cannot
be completely understood. For the sake of overcoming the aforementioned deficiencies, this study
coupled conceptual and machine learning approaches to establish a robust and more reliable RR
model. To this end, three hydrological process-based models namely: IHACRES, GR4J, and MISD are
applied for runoff simulating in a snow-covered basin in Switzerland and then, conceptual models’
outcomes together with more hydro-meteorological variables were incorporated into the model
structure to construct multilayer perceptron (MLP) and support vector machine (SVM) models. At the
final stage of the modeling procedure, the data fusion machine learning approach was implemented
through using the outcomes of MLP and SVM models to develop two evolutionary models of fusion
MLP and hybrid MLP-whale optimization algorithm (MLP-WOA). As a result of conceptual models,
the IHACRES-based model better simulated the RR process in comparison to the GR4J, and MISD
models. The effect of incorporating meteorological variables into the coupled hydrological process-
based and machine learning models was also investigated where precipitation, wind speed, relative
humidity, temperature and snow depth were added separately to each hydrological model. It is found
that incorporating meteorological variables into the hydrological models increased the accuracy
of the models in runoff simulation. Three different learning phases were successfully applied in the
current study for improving runoff peak simulation accuracy. This study proved that phase one (only
hydrological model) has a big error while phase three (coupling hydrological model by machine
learning model) gave a minimum error in runoff estimation in a snow-covered catchment. The
IHACRES-based MLP-WOA model with RMSE of 8.49 ­m3/s improved the performance of the ordinary
IHACRES model by a factor of almost 27%. It can be considered as a satisfactory achievement in
this study for runoff estimation through applying coupled conceptual-ML hydrological models.
Recommended methodology in this study for RR modeling may motivate its application in alternative
hydrological problems.

Rapid climate change is causing significant issues over natural resources as well as human ­beings1. Considering
water resources management, the accurate estimation of accessible water resources and knowledge about the
interactions between the key factors are n ­ ecessary2–4. In this context, runoff gains huge attention which plays
a crucial role in estimating the accessible water resources in the future. It is necessary to predict the runoff for
the quantity and quality of the available water resources and their management, the design capacity of hydraulic
structures, and the associated natural disasters like floods and related environmental i­ ssues5–8.

1
Department of Physical Geography and Ecosystem Science, Lund University, Sölvegatan 12, SE‑223 62 Lund,
Sweden. 2Department of Civil Engineering, Yaşar University, Izmir, Turkey. 3Climate Services, World Meteorological
Organization, Geneva, Switzerland. *email: [email protected]

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 1

Vol.:(0123456789)
 www.nature.com/scientificreports/

Runoff is the main variable for the hydrological analysis from catchment to a continent and global scale
which is in direct interaction with rainfall, groundwater, soil moisture, humidity and snow. In addition, various
meteorological and climatological variables like temperature, evaporation, humidity, and air pressure provoke
the volume of ­runoff9. The nature of hydrological systems could be monitored through the various types of
hydrological models which give a deeper insight into the physical interaction between the various parameters
and their response to each ­other10–13.
The precise estimation of rainfall-runoff (RR) interactions is a major topic among hydrologists since it can
lead managers to have an adequate estimation of the available runoff in the rivers and avoid the negative conse-
quences in the existing hydraulic f­ acilities14–16. The heterogeneous pattern of the hydrological components over
the basins and their nonlinear behavior make the RR process a complex phenomenon. For this reason, numer-
ous RR models are being developed which aim to increase the precision in predicting the runoff in different
spatial and temporal resolutions. Generally, these models could be classified into two fundamental types; the
physically-based or white-box and, the machine learning (ML) or the black-box models; where each type comes
with different advantages and l­ imitations11. For instance, the white-box models require more variables and data
(e.g., soil characteristics, topography, land use, etc.) compared to the black-box models which can assist with a
few data types. However, the morphological and physical features are presented in different levels with white-box
models but are masked and not considered with the black-box models.
In this context, a wide variety of RR models have been employed for modeling runoff in Switzerland. For
instance, Antonetti et al.17 introduced the revised version of the PREVAH hydrological model which is capable
of simulating heavy rainfall events more realistically compared to the traditional version. Antonetti and ­Zappa18
examined the effect of various expert knowledge levels on the accuracy of conceptual hydrological models in
Emme catchment, Switzerland. The result augments the better accuracy of more complex models compared to
the less expert knowledge. Muelchi et al.19 studied the impact of climate change on the runoff regimes in Swit-
zerland using different regional climate models’ precipitation data and a semi-distributed hydrological model.
They revealed the runoff decreased in summer and autumn and increased in the winter however, the annual
mean was projected to decrease in many catchments of Switzerland. Recently, Rottler et al.20 examined the sea-
sonality of flood events in the Rhine river under various climate models’ data for the future. They indicated that
the temperature controls the total runoff at the Basel station which is the closest station upstream of the Rhine
river. They also showed the change of many snowfalls to the rainfall events which eventually increased the total
annual runoff values with the maximum increase in winter.
To solve a variety of environmental problems in the basin, hydrological models can be implemented which
are commonly used tools for the design and planning of water resources systems. The distributed white-box
hydrological models can display the spatial variation of the process by considering the physics of the problem.
On the other hand, black-box models which are also known as empirical approaches are established on the data
without considering the physics of the ­problem21. However, due to the complexity of some hydrological modeling
such as RR modeling, ML application has attracted the interest of many researchers. The complexity of the RR
modeling can be linked to the non-stationary characteristics of the parameters including trend, jump, seasonal-
ity, and most importantly non-linearity of the problem. ML approaches can approximate a nonlinear function
established on the data to determine a certain relationship among system variables without having information
about the physics of the RR ­process8,22,23. It is conducted by using some hydro-meteorological variables and
mostly incorporating the observed rainfall and runoff d ­ ata24. As examples from the literature, the outperformance
of the wavelet-gene expression programming (W-GEP) model to GEP was documented by Shoaib et al.25 using
several datasets collected from different regions. The satisfactory performance of an emotional artificial neural
network (EANN) for RR modeling in comparison to the artificial neural network (ANN) was presented by
Nourani et al.26. Chang et al.27 applied a self-adaptive fuzzy inference network (SaFIN) for RR modeling in
different basins. Nournai et al.28 applied the wavelet-M5 model tree for the same purpose and found out that
multilinear models may give reliable results for catchments having regular rainfall ­patterns29,30. Tikhamarine
et al.23 optimized ANN, least squares support vector machine (LSSVM), and multiple linear regression model
(MLR) using Harris Hawks Optimization (HHO) and particle swarm optimization (PSO) and showed the higher
accuracy of the LSSVM-HHO model for RR modeling. Safari et al.24 recommended the regression in the repro-
ducing kernel Hilbert space (RRKHS) approach for RR modeling and demonstrated its accuracy in capturing
peak runoff values in contrast to the radial basis function artificial neural network (RBFNN) and multivariate
adaptive regression splines (MARS) benchmarks. Morales et al.31 introduced a self-identification neuro-fuzzy
inference model (SINFIM) for RR modeling in a Chilean watershed where the rainfall and runoff lags and the
number of membership functions were determined through the modeling procedure. Better performance of
SINFIM was illustrated in comparison to the ANN, adaptive neuro-Fuzzy inference system (ANFIS), and Long
Short-Term Memory (LSTM) methods.
The aforementioned studies only investigated the RR process either by with-box or black-box models.
Although both approaches have certain advantages where the former gives insights into the physics of the
problem and the latter has robust computation ability, their main limitations respectively are lower computation
precision and neglecting the physics of the RR process. In this study, in order to overcome the deficiencies of
white-box and black-box models, a coupling approach is implemented to consider the physics of the problem
through modeling the RR process utilizing white-box hydrological models together with the application of robust
ML techniques. We aim to model the RR interaction on a process-based methodology through.

(i) Using several common white-box models.

(ii) Applying the black-box models with some extra variables which were not considered at the previous
stage.

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 2

Vol:.(1234567890)
www.nature.com/scientificreports/

(iii) Application of optimization algorithms over the output of the previous stage.
(iv) Introducing a new strategy for improving the ability of ordinary hydrological models in a snow-covered
basin.

Materials and methods

White‑box hydrological models. GR4J. The Génie Rural à 4 paramètres Journalier (GR4J) model can
be used for hydrological models such as runoff modeling and flood f­ orecasting32. This model considers the varia-
bles of precipitation, evapotranspiration, and transpiration as flow data for the runoff ­simulation33. If the amount
of precipitation (P) is more than the amount of evapotranspiration (E), then the net precipitation (Pn) is equal to:
Pn = P − E (1)

En = 0 (2)

x1 (1 − ( xs1 )2 )tanh( Px1n )

PS = s Pn (3)
1+ x1 tanh( x1 )

where x1 is the maximum capacity of the soil moisture accounting (mm), P the net rainfall (mm), s the actual
amount of storage and Ps precipitation in the level of s. If the amount of precipitation is less than the amount of
evapotranspiration then, net evapotranspiration (En), net precipitation (Pn) and, potential evapotranspiration
of the storage (ES) as part of the En can be calculated as follows
En = E − P (4)

Pn = 0 (5)
   
s · 2 − xs1 · tanh Ex1n
ES =     (6)
1 + 1 − xs1 · tanh Ex1n

Equations (7)–(11) show the Perc as the amount of infiltration, Pr is a part of precipitation (routing store) and,
it was divided into two parts (Q1 and Q9) also, Pu is the updated level of production store. Q constitutes 10% of
direct runoff, which is obtained through the hydrograph of unit Hu2 with 2 x4 base time [x4 is base time in unit
hydrograph UH1 (days)]. The Q9 is another part of 90% of the runoff (delay runoff) which is obtained through
the hydrograph of unit Hu1 with X4 base time.
su = s + PS − ES (7)
 
� � �4 �−1/4
4su
Perc = su 1 − 1 +  (8)
9x1

Pr = Pn − Ps + Perc (9)

m

Q1 (i) = 0.1 × HU2 (k) × Pr (i − k + 1) (10)
k=1

l

Q9 (i) = 0.9 × HU1 (k) × Pr (i − k + 1) (11)
k=1

Equation (12) shows F as groundwater exchange and Eqs. (13)–(16) indicate R as the routine moisture stor-
age, Q the final runoff, Qr and Qd runoff in the outlet and direct runoff, ­respectively10.
R 7/2
F = x2 ( ) (12)
x3

R = max(0; R + Q9 + F) (13)
 
� � �4 �−1/4
R
Qr = R1 − 1 +  (14)
x3

Qd = max(0; Q1 + F) (15)

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 3

Vol.:(0123456789)
 www.nature.com/scientificreports/

Q = Qr + Qd (16)
where x2 and x3 are the coefficients of groundwater exchange (mm) and the maximum routing store capacity
one day ahead (mm), respectively.

IHACRES. The ­IHACRES34 is an integrated conceptual RR model whose main purpose is to describe the
hydrological behavior of the basin using the lowest possible ­parameter35. This model requires a time series of
precipitation and air temperature variables as model inputs to simulate the flow as well as the observational flow
data for the model calibration and the accuracy check. The basis of this model is based on two non-linear reduc-
tion models and a linear hydrograph model where the non-linear reduction model converts precipitation into
effective rainfall by considering the infiltration and evaporation ratio. For the effective rainfall estimation, the
basin moisture index and basin saturation index are calculated for each time step. Equations (17–18) are related
to effective rainfall (uk ) and SM index (φk ), ­respectively36.
uk = [c(�k − l)]p × rk (17)

1
�k = rk + (1 − )�k−1 (18)
τk
where c is the equilibrium coefficient of rainfall, τk the drying rate, l threshold for SM index, p the non-linear
(q) (s)
response terms, and rk the observed rainfall. The combination of fast flow ( xk ) and slow flow ( xk ) components
lead to runoff generation ( xk ) (k shows the time) as follows:
q
Xk = Xk + Xks (19)

(q) (q)
Xk = −αq Xk−1 + βq uk (20)

(s) (s)
Xk = −αs Xk−1 + βs uk (21)

−�
τq =
ln(−αq ) (22)

−�
τs = (23)
ln(−αs )

βs
vq = 1 − vs = 1 − (24)
1 + αs
where αq and βq are constant time parameters for fast flow, and αs and βs are constant time parameters for slow
flow. The  is a time interval, τq and τs constant time slides of fast and slow daily currents, respectively; vq the
ratio of fast flow to total flow (1 − vs ), and vs the relative volume of slow ­flow37.

MISD. The MISD is a semi-distributed and lumped RR model (depending on the implemented type) that was
first developed by Brocca et al.38 to predict flood events in the Upper Tiber River in central Italy. This model
mostly focuses on the SM module which is shown to affect the storage capacity and its associated effect on RR
modeling. In this study, we applied the lumped version of MISD with the daily rainfall and air temperature data
as inputs at the basin level which simulate the gradual changes of soil water into two independent states. Water
exits the first layer by evaporation and transpiration, which is calculated through a linear function between
potential evaporation and saturated soil however, the infiltration from the soil surface to the root area is cal-
culated using the non-linear ­relationship39. Three different components cause runoff generation in the MISD
model, including surface excess saturation, the second soil layer, and the subsurface runoff components. The
first two are collected by the instantaneous geomorphological unit hydrograph (IGUH) and routed to the outlet,
while the groundwater runoff is transferred to the outlet by a linear reservoir method. The applied MISD model
in this study uses the Curve Number method to investigate losses. The IGUH and linear reservoirs are used to
track precipitation in sub-basins and areas that discharge directly into the main waterway, respectively. Finally,
routing along the main waterway is estimated through a linear broadcast approach.

Black‑box models. Multilayer perceptron (MLP). The artificial neural network (ANN) has been widely
used for modeling and classification in different engineering fields. Recently different types of ANN were im-
plemented for different aims. The multilayer perceptron (MLP) is one of the widespread ANN methods which
is successfully applied in water science in many ­cases40–42. The current study used MLP as an ANN model for
the modeling aim. There are input layers, an output layer, and hidden layers in the structure of all types of ANN.
Through the MLP modeling process, input variables by some preprocessing are considered as the first stages
(input layer). The neuron(s) transfers information from the input layer to the hidden layer (by considering input
weight and bias unit), and in the hidden layer, the MLP applies some learning algorithm to the data. Finally, the

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 4

Vol:.(1234567890)
www.nature.com/scientificreports/

result transfers to the output layer. The number of weights and bias can be calculated by the summation function
which is given as follows:
n

uk = wki xi + bk (25)
(i=1)

where xi denotes the input variable and n shows the number of inputs, bk is a bias term, and wki shows the con-
nection weight. The summation function analysis information is based on the activation function of the MLP
model. One of the most common types of MLP activation function is the sigmoid function which is given below:

fk = 1/(1 + e(−uk ) ) (26)

The final output of the neuron k can therefore be obtained by
 
� n
yk = fk  wki xi + bk  (27)
(i=1)

The MLP learning process is based on connecting the various network nodes via optimal weights, and then
neurons transfer for output of the above equation to the next step (layer) via selected optimum w ­ eights43. The
Levenberg–Marquardt algorithm was used as a training function, and the number of optimal neurons was
selected by a trial and error method. Also, the Logarithm of the sigmoid function and the linear function were
used as the activation functions from the input layer to the hidden layer and from the hidden layer to the output
layer, respectively.

Support vector machine (SVM). The support vector machine (SVM) is one of the supervised learning methods,
which originates from statistical learning theory and is used in classification, pattern recognition, and regres-
sion ­issues44. To classify linearly inseparable vectors, various kernel functions can be used for multidimensional
cartographers viewed in higher-dimensional spaces, including hierarchical polynomials, radial basis function
(RBF), or hyperbolic t­angents44. The RBF kernel function was used in the current study. This method is one
of the new methods that have shown good efficiency in hydrological studies in recent years. The basis of this
method is a linear classification of data based on the intended vectors to choose a more reliable margin. One of
the important features of this method is that, it simultaneously minimizes the experimental classification error
and maximizes the geometric margins. Dibike et al. 45 suggested SVM for the first time in hydrological studies
by applying SVM for runoff modeling, successfully. It is an efficient training system, which is based on the finite
optimization theory and uses the principle of minimizing structural errors and making them ­optimal45. The con-
nection between dependent and independent variables is supposed to be defined by an algebraic function ( f (x))
plus some noises (ε) in a SVM model.

f (x) = W T × �(x) + b (28)

y = f (x) + noise (29)

where W and b are coefficient vector and constant coefficient, respectively, and they are the components of the
SVM function, and ∅ is the kernel function.

Whale optimization algorithm (WOA). The (whale optimization algorithm) WOA is a nature-based optimiza-
tion algorithm that is inspired by the social behavior of whales in nature introduced by Mirjalili and ­Lewis46. The
WOA works with a set of random solutions as a starting step, and their position can be updated in each iteration
using the algorithm’s operators. Initially, WOA considers that the best solution is bait, and after the best search
agent is identified, other search agents update their location relative to the best search agent. This behavior is
described as follows:
−
→ −
→∗ −
→ − →− →∗ −
→ 
 
X (t + 1) = X (t) − A . C . X (t) − X (t) (30)
−
→
where t describes the running iteration, X the condition vectors of the whale, X* the condition vector of the
best solution and it can be updated if there exists a better solution. If a better answer is available, then X* must
be updated in each iteration. The variables A and C can be calculated as ­follows46:
−
→ (31)
A = 2−
→
a .−
→
r −−
→
a

−
→ (32)
C = 2.−
→
r
where a is a constant that decreases linearly from 2 to 0 during iterations (in both exploration and extraction
stages) and r a random vector at a distance of 0 to 1. The ML models implemented in the MATLAB 2020b envi-
ronment and the optimal parameters of WOA are listed in Table 1.

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 5

Vol.:(0123456789)
 www.nature.com/scientificreports/

Quantity Value
Maximum number of iterations 500
Number of whales 40
The minimum limit for generating unit 2
Total losses 0.3
Total load demand 0.05

Table 1.  The optimal parameters of WOA.

Figure 1.  The flowchart of the applied methodology.

Improving white‑box hydrological models via machine learning strategy. The current study rec-
ommended the improved RR modeling via hydrological models enhanced by ML approaches. For this aim, three
levels were considered for the modeling procedure including level 1: focusing on conceptual runoff modeling via
the white-box models (IHACRES, GR4J, and MISD) using daily temperature, precipitation, and evapotranspira-
tion data; Level 2: improving the accuracy of hydrological models via ordinary ML approaches (MLP and SVM)
and in level 3: improving runoff modeling outcomes from level 2 was considered as input of MLP model coupled
by the WOA (MLP-WOA). The main aim of these processes is to improve the ability of worldwide hydrological
models by data-fusion and ML approaches for runoff modeling in a snow-covered basin.
The values of the daily runoff time series of the selected study area in Switzerland were firstly simulated
through the three white-box hydrological models (IHACRES, GR4J and MISD). For improving the ability of
the mentioned classical hydrological model, we applied the ML approaches (MLP and SVM). In the final step of
the modeling process, we used a nature optimization algorithm for boosting the ability of hydrological and ML
models in runoff simulation. The hybrid nature-inspired model was then proposed to improve the daily runoff
simulation using the hybridization of classical MLP with WOA (MLP-WOA). As shown in Fig. 1, the current
study combined white-box and black-box approaches via three separated levels as follows; Level 1: temperature,
precipitation, and evapotranspiration parameters were considered as inputs of GR4J, IHACRES, and MISD
models and the output is simulated runoff via white-box models. Level 2: simulated runoff by GR4J (from level
1), simulated runoff by IHACRES (from level 1), simulated runoff by MISD (from level 1), temperature, pre-
cipitation, evapotranspiration, relative humidity, and snow depth were considered as inputs of MLP and SVM
models via various scenarios, and the output of level 2 is simulated runoff via black-box models. Level 3: the best
simulated runoff by MLP (MLP5 from level 2) and the best simulated runoff by SVM at level 2 were considered

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 6

Vol:.(1234567890)
www.nature.com/scientificreports/

as inputs of MLP and MLP-WOA models, and the output of level 3 is simulated runoff via data-fusion tasks. The
flowchart of the applied methodology is shown in Fig. 1.

Evaluation criteria. The data used in this study is divided into calibration and validation. Among the entire
data, 70% and 30% of data were considered for calibrating and validating models, respectively for running white-
box and black-box models. The calibration phase was selected from 1 January 1981 to 31 December 2010, and
the validation phase was selected from 1 January 2011 to 26 March 2021. Numerous statistical metrics were used
for the evaluation of RR models. In this study, the mean absolute error (MAE), root mean square error (RMSE)
and Pearson correlation coefficient (r) are used as statistical metrics. In addition, the Nash–Sutcliffe efficiency
(NSE47) and Kling-Gupta efficiency (KGE48) were utilized, which are based on the goodness-of-fit approach and
are among the most common metrics in the hydrological model’s evaluation. The aforementioned evaluation
criteria can be computed as follows
N
1 
MAE = (|Si − Oi |) (33)
N
i=1



1 N
RMSE =  (Si − Oi )2 (34)
N
i=1

n
i=1 (Oi − O).(Si − S)
r=  (35)
n 2 n 2
i=1 (Oi − O) . i=1 (Si − S)

n
(Si − Oi )2
NSE = 1 − i=1
n 2 (36)
i=1 (Oi − O)


µs 2 CVs 2
KGE = 1 − (r − 1)2 + ( − 1) + ( − 1) (37)
µo CVo

where Si and Oi ( S and O ) denote simulated and observed (mean) daily runoff, respectively, and N the number of
observed values used to train and test the models, separately, CV o (CV s ) the coefficient of variation for observed
(simulated) values of the daily runoff. The r is the correlation coefficient between observed and simulated values
of the daily runoff, µo the standard deviation of observed daily runoff values, and µs the standard deviation of
simulated daily runoff values.

The study area and data. The Emme catchment (shown in Fig. 2) was chosen as a pilot study area which
is located in central Switzerland mainly in the Canton of Bern with an approximate area of 924 ­km2. The moun-
tainous Pre-Alps region with 2150 m height around the Augstmatthorn and Tannhorn peaks is the source of
82 km Emme river which drains to the Aare river near the city of Solothurn at 430 m that eventually drains the
Rhine river. The catchment’s mean altitude is around 860 m.
The daily precipitation (P), mean, minimum and maximum temperature (T), relative humidity (RH), evapo-
transpiration (ET), and snow (S) data for the period January 1974–March 2021 are obtained from the MeteoSwiss
for the Langnau station which is located around the central regions of the basin. The daily runoff (Q) measure-
ment for the same period was gathered for the Emme, Wiler Limpach Estuary hydrometric station from the
Swiss Federal Office for the Environment (FOEN). The brief geographic and statistical details of the applied data
are presented in Table 2, and the time series of the precipitation and runoff data are shown in Fig. 3. Dataset is
categorized into the warm-up section (7 years: 1st January 1974 to 31st December 1980), calibration section
(30 years: 1st January 1981 to 31st December 2010), and validation section 10-years: 1st January 2011 to 26th
March 2021) for implementing conceptual RR models.

Results
Conceptual rainfall‑runoff modeling using the white‑box models. Calibration processes. The
dataset is categorized into warm-up Sect. (7 years: 1st January 1974 to 31st December 1980), calibration
Sect. (30 years: 1st January 1981 to 31st December 2010), and validation Sect. (10 years: 1st January 2011 to 26th
March 2021) for implementing of conceptual RR ­models50. For this aim, a numerical optimization method of
derivative-free search method (Pattern Search) was used for calibrating the GR4J model. The optimal param-
eters of the GR4J model are given in Table 3.
A modified SCE-UA (shuffled complex evolution method developed at The University of Arizona), as the
global optimization technique was employed for calibrating IHACRES conceptual ­model51,52. The optimal values
by the SCE-UA method for calibrating the IHACRES model are provided in Table 4. For calibrating the MISD
model a trial-and-error method based on expecting all parameters to be in a monotonic space, and according to
the Kling Gupta Efficiency was ­implemented38. The MISD calibration results are listed in Table 5.

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 7

Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 2.  The geographical map of the Emme watershed in central Switzerland. Figure was created using the
open-source and free QGIS V3.18. The open-access European Digital Elevation Model (EU-DEM), version 1.1
file was retrieved from Copernicus Land Monitoring Service (https://​land.​coper​nicus.​eu/​image​ry-​in-​situ/​eu-​
dem/​eu-​dem-​v1.1?​tab=​metad​ata). The catchment shape file and the river network geospatial data were retrieved
from the open-access Hydrological Atlas of Switzerland (https://​hydro​maps.​ch/)49.

Variable/Stat Mean Min Max Med Std. Dev Kurtosis Skewness Sample size
Precipitation (mm) 3.74 0 95.5 0 7.53 18.07 3.49 17,252
Temperature (°c) 8.15 − 19.8 26 8.4 7.42 − 0.77 − 0.12 17,252
Relative Humidity (%) 81.75 19.6 100 83.6 11.06 − 0.03 − 0.63 17,252
Evapotranspiration (mm) 2.28 0.05 14.45 2.19 0.74 8.13 1.41 17,252
Snow depth (m) 0.34 0 42 0 1.88 92.8 8.4 17,252
Runoff ­(m3/s) 9.51 0.54 305.72 3.67 16.55 50.05 5.42 17,252

Table 2.  The statistical characteristics of the applied data.

Figure 3.  The plot of the applied observed precipitation and runoff time series.

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 8

Vol:.(1234567890)
www.nature.com/scientificreports/

Parameters Description Optimal value

X1 Maximum capacity of the production store (mm) 24
X2 Groundwater exchange coefficient (mm) − 24.32
X3 One day ahead maximum capacity of the routing store (mm) 120
X4 The time base of unit hydrograph UH1 (days) 1.34

Table 3.  Parameters setting by pattern search approach for GR4J model.

Parameters Description Optimal value Typical range

τs Time constant governing rate of recession of slow-flow (day) 59.27 10–350
τq Time constant governing the rate of recession of quick flow (day) 2.95 0.5–10
d Flow threshold (mm) 78.58 50–550
Vs The proportion of slow flow to total flow (proportion) 0.78 0–1
f Plant stress threshold factor (dimensionless) 0.37 0.01–3
e Temperature to potential evapotranspiration conversion factor (dimensionless) 0.25 0.01–1.5

Table 4.  Parameters setting by SCE-UA approach for IHACRES model.

Parameters Description Considered range Optimal value

W_max Fixed water capacity 1st layer 150 150
W_max2 Total water capacity of 2nd layer 100–3000 935.50
W_p Initial conditions (fraction of W_max) 0.1–0.9 0.1
m2 The exponent of drainage for 1st layer 2–1.0 6.75
Ks Hydraulic conductivity for 1st layer 0.1–40 4.30
gamma1 Coefficient lag-time relationship 0.5–3.5 1.71
Kc Parameter of potential evapotranspiration 0.4–3 2.99
alpha Exponent runoff 0.1–15 2.48
Cm Snow module parameter degree-day 0.1/24–3 2.29
m22 An exponent of drainage for 2nd layer 5–3.5 26.24
Ks2 Hydraulic conductivity for 2nd layer 0.01–65 24.09

Table 5.  Parameters setting for implementing MISD model.

Model Phase MAE RMSE NSE r KGE

Calibration 6.53 12.76 0.44 0.66 0.527
IHACRES
Validation 5.83 11.61 0.41 0.64 0.489
Calibration 6.11 11.9 0.51 0.74 0.570
GR4J
Validation 5.72 11.5 0.42 0.68 0.466
Calibration 6.44 12.83 0.43 0.7 0.689
MISD
Validation 5.93 11.85 0.39 0.64 0.584

Table 6.  The results of the applied metrics over different white-box models.

Conceptual rainfall‑runoff models evaluation. The results of applied metrics over the white-box models are
shown in Table 6. Not surprisingly, the acquired values differ from the applied models where in general, the GR4J
shows better performance. The minimum difference between the calibration and validation phase is related to
the GR4J model by RMSE of 0.4 m ­ 3/s, and the maximum difference is related to the IHACRES model by RMSE
of 1.15 ­m /s. In general, the NSE and KGE outcomes explain the acceptable performance of the applied models.
3

Although it differs in the calibration and validation phases overall, the GR4J acquires the least MAE and RMSE
as well as the highest for the NSE, KGE, and r. Considering the plot for the measured and simulated, the GR4J
and MISD models show better performance over the peak values compared to the IHACRES.

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 9

Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 4.  Scatter plots of measured vs simulated runoff using the GR4J, IHACRES and MISD.

Figure 5.  The plot of measured vs simulated runoff time series using GR4J, IHACRES, and MISD models.

The scatter plots of the measured vs simulated runoff using the applied white-box models on their calibration
and validation stages are illustrated in Fig. 4. It is shown that IHACRES underestimates the simulation for higher
values (over 100 ­m3/s) whereas GR4J captures the higher values with lower deviations. For the same category
(less than 100 m­ 3/s), several overestimations are derived from the MISD simulations however, the higher values
are captured much better.
Figure 4 illustrates that the GR4J model performed the best capability during the calibration (RMSE = 11.9
­m3/s and r = 0.74) and validation (RMSE = 11.5 ­m3/s and r = 0.68) phases. There can be a potential for enhancing
the GR4J model ability if we match a proportional fraction of the soil moisture however, this hypothesis requires
to be investigated in future research. The fraction of soil moisture in the GR4J model is considered the difference
between the available soil moisture and the field capacity. Indeed, in nature, soil moisture obtains from satura-
tion level, and this condition occurs between 2 to 4 days, then soil moisture obtains its field capacity after the
drainage process of the soil water. Whereas the GR4J model does not require the upper limit of soil moisture as
saturation soil moisture, which can be another reason for the capability of the GR4J model in RR simulation.
The measured hydrograph in the outlet of the catchment and simulated hydrograph by the hydrological
models are shown in Fig. 5. Calibration and validation phases were selected according to the time series goals
It shows all white-box models have unsatisfactory results in runoff simulation at snow-covered catchments.
Although GR4J and IHACRES models detect some extreme values, the MISD provides poor results in both
calibration and validation phases.

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 10

Vol:.(1234567890)
www.nature.com/scientificreports/

Inputs
Models (machine
No. The input from phase 1 Meteorological variables learning) Output
1 P MLP1 SVM1 Runoff
2 P+T MLP2 SVM2 Runoff
3 The output runoff from IHACRES P + T + ET MLP3 SVM3 Runoff
4 P + T + ET + RH MLP4 SVM4 Runoff
5 P + T + ET + RH + S MLP5 SVM5 Runoff
6 P MLP6 SVM6 Runoff
7 P+T MLP7 SVM7 Runoff
8 The output runoff from GR4J P + T + ET MLP8 SVM8 Runoff
9 P + T + ET + RH MLP9 SVM9 Runoff
10 P + T + ET + RH + S MLP10 SVM10 Runoff
11 P MLP11 SVM11 Runoff
12 P+T MLP12 SVM12 Runoff
13 The output runoff from MISD P + T + ET MLP13 SVM13 Runoff
14 P + T + ET + RH MLP14 SVM14 Runoff
15 P + T + ET + RH + S MLP15 SVM15 Runoff

Table 7.  Intended scenarios for the implementation of coupled scenarios (machine learning via hydrological
models).

Runoff simulation using coupled hydrological models via machine learning approaches. Sce‑
nario definition. The results of applied metrics for conceptual models were not in a satisfactory domain, so it
was decided to improve the accuracy of the acquired output from RR models by applying two widespread ML
methods of MLP and SVM, separately. The various scenarios (explained in Table 7) were proposed by applying
the ML methods using the runoff output obtained by white-box models in the previous stage coupled with dif-
ferent extra ­variables53.

Rainfall‑runoff modeling via black‑box models. According to Table 7, the current study considered combining
the output of each hydrological model via each meteorological variable separately as input of ML models. To
this end, based on the scenario defined, the simulated runoff by IHACRES and precipitation were considered as
input of ML models (MLP1 and SVM1); the simulated runoff by MISD, precipitation, temperature, evapotran-
spiration, relative humidity, and snow depth was considered as input of ML models (MLP15 and SVM15), and
so on. Based on the results, adding meteorological variables helped all hydrological models for having a better
runoff simulation. Adding precipitation, temperature, evapotranspiration, relative humidity, and snow depth
increased the ability of the IHACRES model. Therefore, the fifth scenario (MLP5) boosted ability of IHACRES
model in runoff simulation by MAE = 4.94 ­(m3/s), RMSE = 9.43 ­(m3/s), NSE = 0.61, and KGE = 0.61 in validation
phase (according to Table 8).
Adding precipitation, temperature, and evapotranspiration were the most useful variables for enhancing
the ability of the GR4J model in runoff simulation; such a way that runoff was simulated by MAE = 5.16 ­(m3/s),
RMSE = 9.65 ­(m3/s), NSE = 0.59, and KGE = 0.62 under the frame of MLP8 at the validation phase. Performance
of MISD model was improved by applying precipitation, temperature, evapotranspiration, and relative humid-
ity to MLP model (MLP14); which simulated amount of runoff by MAE = 5.38 ­(m3/s), RMSE = 10.24 ­(m3/s),
NSE = 0.54, and KGE = 0.57 in validation phase. Figure 6 exhibits the scatter plots of measured vs simulated runoff
for the calibration and validation phases through the effect of adding meteorological variables to the hydrological
models (based on the MLP model).
The ability of the SVM model is evaluated as a boosting tool for combining each meteorological variable with
hydrological models. As shown in Table 9 and Fig. 7, the best performance is related to combining precipitation,
temperature, evapotranspiration, relative humidity, and snow depth by the IHACRES model (SVM5). The SVM
scenarios simulated runoff via MAE = 5.62 and 5.02 (­ m3/s) and, r = 0.78 and 0.77 for calibration and validation
phases, respectively. Application of GR4J proved combining precipitation, temperature, and evapotranspiration
by this model can have more accurate result in runoff simulation by MAE = 5.31 ­(m3/s), RMSE = 9.82 ­(m3/s),
NSE = 0.58, and KGE = 0.60 in validation phase. Adding precipitation, temperature, evapotranspiration, rela-
tive humidity, and snow depth (SVM15) to the MISD model can make an more accurate approach for runoff
simulating by result of MAE = 5.50 ­(m3/s), RMSE = 10.32 ­(m3/s), NSE = 0.53, and KGE = 0.59 in validation phase.
The measured and simulated hydrographs by best-proposed models including combining IHACRES models
via meteorological variables reproduced by MLP and SVM models are shown in Fig. 8. Both applied ML tech-
niques (MLP and SVM) proved adding meteorological variables in a parallel situation can increase the perfor-
mance of hydrological models in snow-covered catchments. However, there are divergences between the ability
of MLP5 a and SMV5 for adding meteorological variables to IHACRES models. As the simulated hydrograph
(Fig. 8) shows, MLP5 can have a better simulation of the peak flow (maximum events) in comparison with
SVM5. Also, both MLP5 and SVM5 can be nominated as capable tools for adding extra meteorological variables
to hydrological process-based models.

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 11

Vol.:(0123456789)
 www.nature.com/scientificreports/

Model Phase MAE RMSE NSE r KGE

Calibration 6.28 11.78 0.52 0.72 0.59
MLP1
Validation 5.69 10.91 0.48 0.69 0.54
Calibration 5.59 10.74 0.6 0.78 0.68
MLP2
Validation 5.09 9.85 0.58 0.77 0.58
Calibration 5.56 10.47 0.62 0.79 0.69
MLP3
Validation 4.99 9.52 0.6 0.79 0.60
Calibration 5.56 10.45 0.62 0.79 0.68
MLP4
Validation 5.02 9.54 0.6 0.78 0.61
Calibration 5.53 10.32 0.63 0.79 0.70
MLP5
Validation 4.94 9.43 0.61 0.79 0.61
Calibration 6.35 11.36 0.55 0.74 0.63
MLP6
Validation 6.07 10.82 0.49 0.7 0.56
Calibration 5.63 10.29 0.63 0.8 0.70
MLP7
Validation 5.31 9.86 0.58 0.76 0.59
Calibration 5.53 10.13 0.64 0.8 0.72
MLP8
Validation 5.16 9.65 0.59 0.77 0.62
Calibration 5.54 10 0.65 0.81 0.71
MLP9
Validation 5.21 9.65 0.59 0.77 0.61
Calibration 5.65 10.13 0.64 0.8 0.71
MLP10
Validation 5.28 9.61 0.6 0.78 0.62
Calibration 6.3 11.41 0.55 0.74 0.62
MLP11
Validation 5.84 10.82 0.49 0.7 0.53
Calibration 5.82 10.58 0.61 0.78 0.68
MLP12
Validation 5.43 10.26 0.54 0.74 0.57
Calibration 5.78 10.57 0.61 0.78 0.69
MLP13
Validation 5.46 10.23 0.54 0.74 0.58
Calibration 5.77 10.64 0.61 0.78 0.68
MLP14
Validation 5.38 10.24 0.54 0.74 0.57
Calibration 5.73 10.32 0.63 0.79 0.70
MLP15
Validation 5.42 10.25 0.54 0.74 0.57

Table 8.  The results of the applied metrics on the calibration and validation phases through different MLP
scenarios.

The accuracy improvement by coupled MLP‑WOA model. Although the accuracy of the proposed
scenarios is slightly better than the RR models implemented in the first stage, it was decided to enumerate a data-
fusion model to see whether the accuracy improves. To this aim, the MLP5 and SVM5 which showed the highest
performances were selected as inputs of the model in the third phase. Then, the derived runoff from the men-
tioned models was chosen as input for the application of MLP and MLP-WOA models. In the third phase, firstly
simulated runoff by MLP5 and SVM5 were considered as input of ordinary MLP model, which the result of MLP
showed the application of the data-fusion approach can improve the accuracy of ordinary MLP and SVM. Then,
the WOA optimizer approach was applied for improving the ability of MLP’s training. The outputs of MLP5
and SVM5 are considered as the inputs of the MLP-WOA model. For this aim, the third phase (according to
Table 10) benefits from the advantages of physically-based models in the first phase, and advantages of the ML
process in the second phase and the advantages of the bio-inspired optimization algorithm in the third phase.
Results of the third learning phase of the current study are provided in Table 11 and Fig. 9. The main goal of
the third stage is the application of an advanced method for coupling the best result of the previous stage (MLP5
and SVM5) to reach a high accuracy in runoff simulation. Then, WOA coupled with MLP (namely MLP-WOA)
was considered as an advanced approach for this aim. Two aims were fulfilled in this stage: (i) simulated runoff via
MLP5 and SVM5 were considered as inputs of the model in the third stage. In this way, the final model can benefit
from sages 1 and 2, which means, the final model of the third stage (MLP-WOA) has advantages of black-box and
white-box models at the same time. (ii) for reaching maximum efficiency, this stage employed a high-performance
predictor tool by combining a nature-inspired optimization algorithm via an ordinary ML. Therefore, for the
evaluation of the mentioned combined model (MLP-WOA), its performance is evaluated by standalone MLP.
The MLP-WOA simulated runoff by result of MAE = 5.14 ­(m3/s), RMSE = 9.07 ­(m3/s), NSE = 0.71, r = 0.85, and
KGE = 0.78 in training phase, and MAE = 4.56 ­(m3/s), RMSE = 8.49 ­(m3/s), NSE = 0.68, r = 0.84, and KGE = 0.66
in testing period. In addition, the evaluation of the optimal model (MLP-WOA) with an ordinary model (MLP)
showed that WOA improved the ability of ordinary ML for runoff simulation in the snow-covered catchment.
The scatter plot of MLP-WOA showed that most of the data have fallen close to the best fit line.

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 12

Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 6.  Scatter plots of measured vs simulated runoff for the calibration and validation phases through the
different MLP scenarios.

Hydrographs of the third phase are provided in Fig. 10 and it shows that the third phase is much more accu-
rate than the second and first phases. According to Fig. 10, the time-series graph of MLP-WOA detected the
maximum flow better than the other used strategies. It was successful in reproducing simulated hydrographs for
both training and testing phases from 1981 to 2020.

Runoff peak flow simulation analysis. Maximum events of each model were analyzed by the Taylor
diagram (Figs. 11 and 12). According to peak flow analysis for the top 5% and top 10% of peak flows, considered
learning phases improved the ability of models for peak flow estimation. The ordinary hydrological models
are located at the farthest point in both diagrams, and they have weak correlation and far standard deviation
according to observed peak flow values. Second phase learning (adding meteorological variables by ML models)
improved a little bit performance of peak flow estimation in all models. Then, the third learning phase (data-
fusion: coupled hydrological models via ML models) dramatically improved the peak flow simulation. As it is
shown in the diagram, the blue and red points are the results of the third learning phase, in which the blue point
(MLP-WOA) has less error in peak flow simulation. It refers to the ability of hydrological models and ML models

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 13

Vol.:(0123456789)
 www.nature.com/scientificreports/

Model Phase MAE RMSE NSE r KGE

Calibration 6.33 11.86 0.51 0.72 0.58
SVM1
Validation 5.79 10.93 0.48 0.69 0.53
Calibration 5.8 10.96 0.58 0.76 0.65
SVM2
Validation 5.23 10.17 0.55 0.75 0.56
Calibration 5.68 10.70 0.60 0.78 0.67
SVM3
Validation 5.07 9.79 0.58 0.77 0.58
Calibration 5.66 10.63 0.61 0.78 0.67
SVM4
Validation 5.04 9.82 0.58 0.77 0.57
Calibration 5.65 10.57 0.61 0.78 0.68
SVM5
Validation 5.02 9.79 0.58 0.77 0.57
Calibration 6.32 11.33 0.56 0.75 0.62
SVM6
Validation 6.08 10.87 0.48 0.70 0.56
Calibration 5.74 10.41 0.62 0.79 0.69
SVM7
Validation 5.42 10.13 0.55 0.75 0.59
Calibration 5.67 10.2 0.64 0.80 0.70
SVM8
Validation 5.31 9.82 0.58 0.77 0.60
Calibration 5.66 10.16 0.64 0.80 0.70
SVM9
Validation 5.31 9.85 0.58 0.76 0.60
Calibration 5.64 10.08 0.65 0.81 0.71
SVM10
Validation 5.32 9.84 0.58 0.76 0.60
Calibration 6.33 11.45 0.55 0.74 0.62
SVM11
Validation 5.93 10.97 0.47 0.69 0.53
Calibration 5.98 10.88 0.59 0.77 0.66
SVM12
Validation 5.57 10.44 0.52 0.73 0.55
Calibration 5.96 10.79 0.60 0.77 0.67
SVM13
Validation 5.57 10.38 0.53 0.73 0.56
Calibration 5.91 10.75 0.60 0.77 0.66
SVM14
Validation 5.51 10.34 0.53 0.73 0.56
Calibration 5.82 10.54 0.62 0.79 0.67
SVM15
Validation 5.50 10.32 0.53 0.73 0.59

Table 9.  The results of the applied metrics on the calibration and validation phases through different SVM
scenarios.

together in parallel conditions to have better runoff simulation in snow-covered basins. The statistical param-
eters given in Taylor diagrams (RMSE, r, and SD) for the top 5% and 10% of peak flow are listed in Table 12.

Discussion
Since rainfall-runoff (RR) is a non-linear and complex hydrological phenomenon, a variety of approaches such
as conceptual and empirical have been implemented for runoff estimation. Conceptual approaches which are
known as physically-based or white-box models incorporate morphological and physical features of the prob-
lem and are quite useful for understanding the physics of the problem. However, in terms of accuracy, they
may fail to generate satisfactory results. Alternatively, machine learning (ML) or black-box models have higher
computational ability, while may fail in the physical justification of the problem. Therefore, the main research
question of the study is to develop a methodology to merge the advantages of both aforementioned approaches
to establish a robust-physically based model for runoff estimation. Three conceptual models as the IHACRES,
GR4J, and MISD are developed in a snow-covered basin in Switzerland and then through using these models’
outcomes and a variety of hydro-meteorological parameters, the ML models of SVM, MLP, fusion MLP, and
MLP-WOA are developed.
Results of conceptual models illustrate that the IHACRES, GR4J, and MISD models give almost the same
results while GR4J provides slightly better results in contrast to the IHACRES and MISD models. Evaluation of
the conceptual models’ performances in terms of computing the peak runoff values, IHACRES fails to an accurate
estimation where it underestimates the peak values. Contrary, the MISD model overestimates several peak values.
From a general point of view, conceptual models’ results are not satisfactory and it was the main motivation to
couple these models with ML models for the Emme catchment. To overcome this issue, a variety of scenarios are
defined to develop the IHACRES, GR4J, and MISD-based MLP and SVM models. For this purpose, five different
scenarios consisting of precipitation, temperature, evapotranspiration, relative humidity, and snow depth are
considered. The incorporation of hydro-meteorological variables into the models promotes the accuracy of the
models developed in the first stage where the result of the IHACRES model with RMSE of 11.61 is improved to
9.43 and 9.79 in IHACRES-based MLP and SVM models, respectively. It shows almost 20% improvement in the

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 14

Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 7.  Scatter plots of measured vs simulated runoff of the calibration and validation phases through the
different SVM scenarios.

IHACRES-based MLP model in contrast to the IHACRES model. This improvement is found at almost 14% for
GR4J and MISD-based MLP models. The better performances of conceptual-based ML models can be linked
not only to the robustness of the ML techniques but also to incorporating the variety of hydro-meteorological
parameters of precipitation, temperature, evapotranspiration, relative humidity, and snow depth into the models.
In order to further improve the accuracy of the applied models, a fusion and robust ML model based on the
WOA are implemented. For this aim, the best results which are obtained by the IHACRES-based MLP and SVM
models are used as model inputs. The IHACRES-based MLP-WOA model with an RMSE of 8.49 m ­ 3/s improved
the performance of the ordinary IHACRES model by a factor of almost 27%. It can be considered a satisfactory
achievement for runoff estimation by applying coupled conceptual-ML hydrological models.
The GR4J, IHACRES, and MISD have been applied to various RR studies. For instance, Shin and ­Kim5 tried to
improve the IHACRES and GR4J models by testing multiple component combinations and eventually achieved
the NSEs ranging from 0.5 to o.8. Recently, the subject of increasing the hydrological model accuracy using
ML models gained a huge interest. Tikhamarine et al.23 showed the superiority of using a Least Square Support
Vector Machine (LSSVM) compared to the MLP coupled with the optimization models (PSO and HHO) in RR
modeling with NSE values of 0.4 to o.8 however, they did not apply any hydrological model. In another work,

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 15

Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 8.  The plot of measured vs simulated runoff time series using MLP5 and SVM5 models.

No. Inputs Output Models

1 MLP5 + SVM5 Q MLP MLP-WOA

Table 10.  Intended scenarios for the implementation of MLP and MLP-WOA models.

Model Phase MAE RMSE NSE r KGE

Training 5.33 9.61 0.68 0.82 0.75
MLP
Testing 4.93 9.49 0.61 0.79 0.61
Training 5.14 9.07 0.71 0.85 0.78
MLP-WOA
Testing 4.56 8.49 0.68 0.84 0.67

Table 11.  The results of the applied metrics on the calibration and validation phases of the MLP and coupled
MLP-WOA models.

Figure 9.  Scatter plots of measured vs simulated runoff on the calibration and validation phases through the
MLP and coupled MLP-WOA models.

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 16

Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 10.  Time series plot related to the result of MLP and MLP-WOA models.

Figure 11.  Taylor diagram of top 5% peak flow during the study period (1981–2021).

Lees et al.22 applied the LSTM to four different conceptual models over the entire UK and achieved an average
NSE of 0.7 to 0.8.
Esmaeili-Gisavandani et al.54 employed the Soil & Water Assessment Tool (SWAT), Hydrologiska Byråns
Vattenbalansavdelning (HBV), IHACRES, Australian water balance mode (AWBM), and Soil Moisture Account-
ing (SMA) models for RR modeling in the Hablehroud basin (in Iran). They coupled outputs of hydrological
models with a black-box model (GEP) and the result of the coupled model showed that the black-box model
can improve the ability of the white-box model for RR modeling. Their coupled model accuracy was reported
by NSE = 0.56 at the validation phase, while the coupled model of the current study was reported as NSE = 0.68
at the validation phase. Ahmadi et al.53 applied SWAT, IHACRES, and ANN in the Kan watershed (Iran). They
reported RMSE equal to 3.3 ­(m3/s) and 3.7 ­(m3/s) for the calibration phase of the SWAT and IHACRES models,
respectively, and also they reported RMSE equal to 2.2 ­(m3/s) for the testing phase of ANN ­model53. Their study

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 17

Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 12.  Taylor diagram of top 10% peak flow during the study period (1981–2021).

Metrics IHACRES GR4J MISD MLP5 SVM5 MLP MLP-WOA

Top 5% of peak flow
RMSE 43.11 41.3 41.88 34.95 36.47 32.46 29.14
r 0.57 0.67 0.67 0.69 0.67 0.78 0.84
SD 21.53 38.76 44.68 30.97 30.12 35.43 36.51
Top 10% of peak flow
RMSE 32.6 32.27 33.25 27.24 28.14 25.4 23.15
r 0.58 0.68 0.64 0.69 0.67 0.77 0.81
SD 18.85 30.87 35.83 26.08 25.17 28.64 29.56

Table 12.  RMSE, r, and SD of the top 5% and 10% of measured peak flow and simulated peak flow.

was implemented in a semi-dry climate zone and the models reported have acceptable accuracy. However, due
to the role of snow in cold regions, RR modeling in snow-covered areas is expected with more errors compared
with dry and semi-dry regions. In some other research, the authors considered streamflow lag-times as input of
black-box models for runoff modeling a­ ims55–58 and they reached higher accuracies in runoff modeling. While
the current study focused on the RR modeling concept by considering all inputs at the t (same) time and by
conserving several meteorological variables as input of black-box models to have interpretation meaning for
RR modeling. Previous researches confirm the result of the current study, for example, Ditthakit et al.13 used
the black-box model to increase the efficiency of the white-box model in Thailand. This means that the method
presented in this study can be expanded by other models and also can generalize the implementation of this
method in other regions with different climates.
Several studies focused on the RR modeling over the Emme catchment and the surrounding areas. In terms
of comparison of the model accuracy with the previous studies in the region, Antonetti et al.59 assessed the flash
flood modeling between May to July 2016 using a chain of hydrological, meteorological, and process-based
runoff generation modules and obtained 0.1 to 0.8 and 0.5 to 0.8 for the NSE and KGE, respectively. Sikorska-
Senoner and Q ­ uilty60 achieved 16–29% improvements by applying various data-driven models to the conven-
tional hydrological model for the streamflow simulation on Klein Emme catchment (a neighbor catchment to

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 18

Vol:.(1234567890)
www.nature.com/scientificreports/

the Emme catchment). They recommended the use of extreme variant boosting and Random Forest models as
they demonstrated the best performance.
Uncertainty in hydrological modeling has been a major challenge. The current study like other hydrological
modeling studies was affected by two main uncertainties. (I) Uncertainty in models’ input: conceptual hydro-
logical models showed a significant uncertainty based on models’ input. Affecting global warming (even only
1 °C) can have a significant effect on the results of the IHACRES, GR4J, and MISD models. By increasing the
temperature, snow cover, and evapotranspiration, the amount of runoff can significantly change. However, the
current study tried to reduce this uncertainty by applying meteorological variables as extra inputs besides the
required inputs of conceptual models. (II) Uncertainty in hydrological models’ parameters can be considered
as another limitation of the current study. The current study tried to use some of the famous optimization
methods for calibration of conceptual hydrological models’ parameters, but still, the models are sensitive to any
unexpected or extreme event in a new climate area. That means, a short-term heavy rainfall or a cold season
can have a significant effect on the conceptual models’ calibration process, and the results of models can vary in
different climates. Also, due to the role of snow (and glaciers) runoff modeling in snow-covered basins always
has more complexity. Then, the current study selected a snow-cover basin for providing a solution for solving
such a problem. The literature review proved that in basins without snow (less complexity) levels 1 and 2 of the
current study could most probably lead to an acceptable accuracy for runoff modeling. The climatic zone, the
scale of the basin, absence of snow, and data availability are some of the factors which could be mentioned for
the complexity of the runoff modeling, then the current study recommended applying the current framework
to different climate zones.

Conclusions
In this study, three conceptual approaches of IHACRES, GR4J, and MISD are implemented for modeling the
RR process in a snow-covered basin in Switzerland. The Two well-known ML techniques (the SVM and MLP)
are coupled with the conceptual IHACRES, GR4J, and MISD models. It is found that the conceptual models’
accuracies are prompted by a factor of 14–19% in comparison to the ordinary conceptual models. Among
conceptual-based ML models, the IHACRES-based MLP model gives better performance. Incorporating the
hydro-meteorological variables of precipitation, temperature, evapotranspiration, relative humidity, and snow
depth significantly improved the accuracy of developed models. An advanced ML model constructed through
WOA has improved the performance of the MLP-WOA model by a factor of 27% in contrast to the conventional
IHACRES model. Results of this study demonstrate that coupling conceptual and ML models can provide sat-
isfactory outcomes in terms of accurate computation and physical justification of the problem. The developed
methodology overcomes the basic deficiencies of the conceptual and ML methods where the former may fail to
generate accurate results and the latter masked the physics of the problem. The coupled approach of merging the
conceptual and ML models takes advantage of white-box models (e.g., considering the hydrological interpretation
of the catchment) and black-box models (e.g., runoff modeling with explicit and implicit relationships between
data that is out of the ability of white-box models) to construct a more robust and reliable model. Utilizing dif-
ferent calibration methods for overcoming the hydrological models’ parameters uncertainty is recommended as a
future research direction. It is highly recommended to check the ability of the proposed method (three considered
phases) under changing climate conditions. It is recommended to apply machine learning algorithms as feature
selection tools for finding the most effective variables and overcoming models’ input uncertainty. In addition,
three lumped models were considered in the current study and it is recommended to compare the results of the
proposed method with the distributed hydrological models in the snow-covered basins as an extension of the
current study.

Data availability
The datasets used and/or analyzed during the current study are available from the corresponding author on
reasonable request.

Received: 15 April 2022; Accepted: 6 July 2022

References
1. Tian, Y., Xu, Y. P. & Zhang, X. J. Assessment of climate change impacts on river high flows through comparative use of GR4J, HBV
and Xinanjiang models. Water Resour. Manage 27, 2871–2888 (2013).
2. Okkan, U., Ersoy, Z. B., Ali Kumanlioglu, A. & Fistikoglu, O. Embedding machine learning techniques into a conceptual model
to improve monthly runoff simulation: A nested hybrid rainfall-runoff modeling. J. Hydrol. 598, 126433 (2021).
3. Yang, S. et al. A physical process and machine learning combined hydrological model for daily streamflow simulations of large
watersheds with limited observation data. J. Hydrol. 590, 125206 (2020).
4. Nourani, V. An emotional ANN (EANN) approach to modeling rainfall-runoff process. J. Hydrol. 544, 267–277 (2017).
5. Shin, M. J. & Kim, C. S. Component combination test to investigate improvement of the IHACRES and GR4J rainfall–runoff
models. Water 13, 2126 (2021).
6. Perrin, C., Michel, C. & Andréassian, V. Improvement of a parsimonious model for streamflow simulation. J. Hydrol. 279, 275–289
(2003).
7. Ye, W., Bates, B. C., Viney, N. R., Sivapalan, M. & Jakeman, A. J. Performance of conceptual rainfall-runoff models in low-yielding
ephemeral catchments. Water Resour. Res. 33, 153–166 (1997).
8. Mohammadi, B. A review on the applications of machine learning for runoff modeling. Sustain. Water Resour. Manag. 7, 1–11
(2021).
9. Wei, X., Guo, S. & Xiong, L. Improving efficiency of hydrological prediction based on meteorological classification: A case study
of gr4j model. Water 13, 2546 (2021).

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 19

Vol.:(0123456789)
 www.nature.com/scientificreports/

10. Nayak, A. K., Biswal, B. & Sudheer, K. P. Role of hydrological model structure in the assimilation of soil moisture for streamflow
prediction. J. Hydrol. 598, 126465 (2021).
11. Kim, T. et al. Can artificial intelligence and data-driven machine learning models match or even replace process-driven hydrologic
models for streamflow simulation?: A case study of four watersheds with different hydro-climatic regions across the CONUS. J.
Hydrol. 598, 126423 (2021).
12. Young, C.-C. & Liu, W.-C. Prediction and modelling of rainfall–runoff during typhoon events using a physically-based and artificial
neural network hybrid model. Hydrol. Sci. J. 60, 2102–2116 (2015).
13. Ditthakit, P. et al. Using machine learning methods for supporting GR2M model in runoff estimation in an ungauged basin. Sci.
Rep. 11, 1–16 (2021).
14. Humphrey, G. B., Gibbs, M. S., Dandy, G. C. & Maier, H. R. A hybrid approach to monthly streamflow forecasting: Integrating
hydrological model outputs into a Bayesian artificial neural network. J. Hydrol. 540, 623–640 (2016).
15. Borzì, I., Bonaccorso, B. & Fiori, A. A modified IHACRES rainfall-runoff model for predicting the hydrologic response of a river
basin connected with a deep groundwater aquifer. Water 11, 2031 (2019).
16. Mohammadi, B., Moazenzadeh, R., Christian, K. & Duan, Z. Improving streamflow simulation by combining hydrological process-
driven and artificial intelligence-based models. Environ. Sci. Pollut. Res. https://​doi.​org/​10.​1007/​s11356-​021-​15563-1 (2021).
17. Antonetti, M., Scherrer, S., Kienzler, P. M., Margreth, M. & Zappa, M. Process-based hydrological modelling: The potential of a
bottom-up approach for runoff predictions in ungauged catchments. Hydrol. Process. 31, 2902–2920 (2017).
18. Antonetti, M. & Zappa, M. How can expert knowledge increase the realism of conceptual hydrological models? A case study based
on the concept of dominant runoff process in the Swiss Pre-Alps. Hydrol. Earth Syst. Sci. 22, 4425–4447 (2018).
19. Muelchi, R., Rössler, O., Schwanbeck, J., Weingartner, R. & Martius, O. An ensemble of daily simulated runoff data (1981–2099)
under climate change conditions for 93 catchments in Switzerland (Hydro-CH2018-Runoff ensemble). Geosci. Data J. https://​doi.​
org/​10.​1002/​gdj3.​117 (2021).
20. Rottler, E., Bronstert, A., Bürger, G. & Rakovec, O. Projected changes in Rhine River flood seasonality under global warming.
Hydrol. Earth Syst. Sci. 25, 2353–2371 (2021).
21. Legesse, D., Vallet-Coulomb, C. & Gasse, F. Hydrological response of a catchment to climate and land use changes in Tropical
Africa: Case study south central Ethiopia. J. Hydrol. 275, 67–85 (2003).
22. Lees, T. et al. Benchmarking data-driven rainfall-runoff models in Great Britain: A comparison of long short-term memory
(LSTM)-based models with four lumped conceptual models. Hydrol. Earth Syst. Sci. 25, 5517–5534 (2021).
23. Tikhamarine, Y. et al. Rainfall-runoff modelling using improved machine learning methods: Harris hawks optimizer vs particle
swarm optimization. Hydrol. Earth Syst. Sci. https://​doi.​org/​10.​1016/j.​jhydr​ol.​2020.​125133 (2020).
24. Safari, M. J. S., Rahimzadeh Arashloo, S. & Danandeh Mehr, A. Rainfall-runoff modeling through regression in the reproducing
kernel Hilbert space algorithm. J. Hydrol. 587, 125014 (2020).
25. Shoaib, M., Shamseldin, A. Y., Melville, B. W. & Khan, M. M. Runoff forecasting using hybrid wavelet gene expression program-
ming (WGEP) approach. J. Hydrol. 527, 326–344 (2015).
26. Nourani, V., Molajou, A., Najafi, H. & Danandeh Mehr, A. Emotional ANN (EANN): A New Generation of Neural Networks for
Hydrological Modeling in IoT (Springer, 2019). https://​doi.​org/​10.​1007/​978-3-​030-​04110-6_3.
27. Chang, T. K., Talei, A., Quek, C. & Pauwels, V. R. N. Rainfall-runoff modelling using a self-reliant fuzzy inference network with
flexible structure. J. Hydrol. 564, 1179–1193 (2018).
28. Nourani, V., Davanlou Tajbakhsh, A., Molajou, A. & Gokcekus, H. Hybrid wavelet-m5 model tree for rainfall-runoff modeling. J.
Hydrol. Eng. 24, 04019012 (2019).
29. Nourani, V., Molajou, A., Tajbakhsh, A. D. & Najafi, H. A wavelet based data mining technique for suspended sediment load
modeling. Water Resour. Manage 33, 1769–1784 (2019).
30. Nourani, V., Tajbakhsh, A. D. & Molajou, A. Data mining based on wavelet and decision tree for rainfall-runoff simulation. Hydrol.
Res. 50, 75–84 (2019).
31. Morales, Y., Querales, M., Rosas, H., Allende-Cid, H. & Salas, R. A self-identification Neuro-Fuzzy inference framework for
modeling rainfall-runoff in a Chilean watershed. J. Hydrol. 594, 125910 (2021).
32. Perrin, C., Michel, C. & Andréassian, V. Modèles Hydrologiques du Génie Rural (GR) (Springer, 2007).
33. Perrin, C., Michel, C. & Andréassian, V. Does a large number of parameters enhance model performance? Comparative assessment
of common catchment model structures on 429 catchments. J. Hydrol. 242, 275–301 (2001).
34. Jakeman, A. J. & Hornberger, G. M. How much complexity is warranted in a rainfall-runoff model?. Water Resour. Res. 29,
2637–2649 (1993).
35. Croke, B. F. W. & Jakeman, A. J. A catchment moisture deficit module for the IHACRES rainfall-runoff model. Environ. Model.
Softw. 19, 1–5 (2004).
36. Carcano, E. C., Bartolini, P., Muselli, M. & Piroddi, L. Jordan recurrent neural network versus IHACRES in modelling daily
streamflows. J. Hydrol. 362, 291–307 (2008).
37. Abushandi, E. & Merkel, B. Modelling rainfall runoff relations using HEC-HMS and IHACRES for a single rain event in an arid
region of Jordan. Water Resour. Manage 27, 2391–2409 (2013).
38. Brocca, L., Melone, F. & Moramarco, T. Distributed rainfall-runoff modelling for flood frequency estimation and flood forecasting.
Hydrol. Process. 25, 2801–2813 (2011).
39. Masseroni, D., Cislaghi, A., Camici, S., Massari, C. & Brocca, L. A reliable rainfall-runoff model for flood forecasting: Review and
application to a semi-urbanized watershed at high flood risk in Italy. Hydrol. Res. 48, 726–740 (2017).
40. Noori, R., Deng, Z., Kiaghadi, A. & Kachoosangi, F. T. How reliable are ANN, ANFIS, and SVM techniques for predicting longi-
tudinal dispersion coefficient in natural rivers?. J. Hydraul. Eng. 142, 04015039 (2016).
41. Noori, R., Karbassi, A. R., Mehdizadeh, H., Vesali-Naseh, M. & Sabahi, M. S. A framework development for predicting the longi-
tudinal dispersion coefficient in natural streams using an artificial neural network. Environ. Prog. Sustainable Energy 30, 439–449
(2011).
42. Kişi, Ö. Streamflow forecasting using different artificial neural network algorithms. J. Hydrol. Eng. 12, 532–539 (2007).
43. Bhattacharya, B., Price, R. K. & Solomatine, D. P. Machine learning approach to modeling sediment transport. J. Hydraul. Eng.
133, 440–450 (2007).
44. Vapnik, V., Golowich, S. E. & Smola, A. Support vector method for function approximation, regression estimation, and signal
processing. in Advances in Neural Information Processing Systems (1997).
45. Dibike, Y. B., Velickov, S., Solomatine, D. & Abbott, M. B. Model induction with support vector machines: Introduction and
applications. J. Comput. Civ. Eng. 15, 208–216 (2001).
46. Mirjalili, S. & Lewis, A. The Whale optimization algorithm. Adv. Eng. Softw. 95, 51–67 (2016).
47. Nash, J. E. & Sutcliffe, J. V. River flow forecasting through conceptual models part I: A discussion of principles. J. Hydrol. 10,
282–290 (1970).
48. Gupta, H. V., Kling, H., Yilmaz, K. K. & Martinez, G. F. Decomposition of the mean squared error and NSE performance criteria:
Implications for improving hydrological modelling. J. Hydrol. 377, 80–91 (2009).
49. Bühlmann, A. & Schwanbeck, J. A: Fundamentals 1/2 Catchment Classification: Medium-Scale, Large-Scale, and Similar-Size Catch‑
ments. www.​hydro​logic​alatl​as.​ch.

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 20

Vol:.(1234567890)
www.nature.com/scientificreports/

50. Kim, K. B., Kwon, H. H. & Han, D. Exploration of warm-up period in conceptual hydrological modelling. J. Hydrol. 556, 194–210
(2018).
51. Duan, Q. Y., Gupta, V. K. & Sorooshian, S. Shuffled complex evolution approach for effective and efficient global minimization. J.
Optim. Theory Appl. 76, 501–521 (1993).
52. Chu, W., Gao, X. & Sorooshian, S. A new evolutionary search strategy for global optimization of high-dimensional problems. Inf.
Sci. 181, 4909–4927 (2011).
53. Ahmadi, M., Moeini, A., Ahmadi, H., Motamedvaziri, B. & Zehtabiyan, G. R. Comparison of the performance of SWAT, IHACRES
and artificial neural networks models in rainfall-runoff simulation (case study: Kan watershed, Iran). Phys. Chem. Earth 111, 65–77
(2019).
54. Esmaeili-Gisavandani, H., Lotfirad, M., Sofla, M. S. D. & Ashrafzadeh, A. Improving the performance of rainfall-runoff models
using the gene expression programming approach. J. Water Clim. Change 12, 3308–3329 (2021).
55. Adnan, R. M. et al. Improving streamflow prediction using a new hybrid ELM model combined with hybrid particle swarm opti-
mization and grey wolf optimization. Knowl. Based Syst. 230, 107379 (2021).
56. Bajirao, T. S., Elbeltagi, A., Kumar, M. & Pham, Q. B. Applicability of machine learning techniques for multi-time step ahead runoff
forecasting. Acta Geophys. 70, 757–776 (2022).
57. Khan, M. T. et al. Application of machine learning techniques in rainfall–runoff modelling of the soan river basin, Pakistan. Water
13, 3528 (2021).
58. Khodakhah, H., Aghelpour, P. & Hamedi, Z. Comparing linear and non-linear data-driven approaches in monthly river flow
prediction, based on the models SARIMA, LSSVM, ANFIS, and GMDH. Environ. Sci. Pollut. Res. 29, 21935–21954 (2022).
59. Antonetti, M., Horat, C., Sideris, I. V. & Zappa, M. Ensemble flood forecasting considering dominant runoff processes—Part 1:
Set-up and application to nested basins (Emme, Switzerland). Nat. Hazards Earth Syst. Sci. 19, 19–40 (2019).
60. Sikorska-Senoner, A. E. & Quilty, J. M. A novel ensemble-based conceptual-data-driven approach for improved streamflow simula-
tions. Environ. Model. Softw. 143, 105094 (2021).

Author contributions
B.M.: Conceptualization, Data curation, Formal Analysis, Investigation, Methodology, Resources, Software,
Visualization, Writing—original draft, Writing—review & editing. M.J.S.S.: Conceptualization, Investigation,
Methodology, Resources, Validation, Writing—original draft, Writing—review & editing. S.V.: Data curation,
Investigation, Methodology, Resources, Validation, Writing—original draft, Writing—review & editing.

Funding
Open access funding provided by Lund University.

Competing interests
The authors declare no competing interests.

Additional information
Correspondence and requests for materials should be addressed to B.M.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International
License, which permits use, sharing, adaptation, distribution and reproduction in any medium or
format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the
Creative Commons licence, and indicate if changes were made. The images or other third party material in this
article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the
material. If material is not included in the article’s Creative Commons licence and your intended use is not
permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from
the copyright holder. To view a copy of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/.

© The Author(s) 2022

Scientific Reports | (2022) 12:12096 | https://doi.org/10.1038/s41598-022-16215-1 21

Vol.:(0123456789)