You are on page 1of 20

Stochastic Environmental Research and Risk Assessment

https://doi.org/10.1007/s00477-020-01910-0
(0123456789().,-volV)(0123456789().
,- volV)

ORIGINAL PAPER

Short term rainfall-runoff modelling using several machine learning


methods and a conceptual event-based model
Rana Muhammad Adnan1 • Andrea Petroselli2 • Salim Heddam3 • Celso Augusto Guimarães Santos4 •

Ozgur Kisi5,6

Accepted: 16 October 2020


 Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract
The applicability of four machine learning (ML) methods, ANFIS-PSO, ANFIS-FCM, MARS and M5Tree, together with
multi model simple averaging (MM-SA) ensemble method, is investigated in rainfall-runoff modeling at hourly timescale.
The results are compared with the conceptual EBA4SUB model using rainfall and runoff data from Samoggia River basin,
Italy. The capability of the methods is measured using five statistics, Nash–Sutcliffe efficiency, root mean squared error,
mean absolute error, scatter index, and adjusted index of agreement. Comparison of single ML reveals that the ANFIS-
PSO, ANFIS-FCM and MARS produce similar accuracy which is better than the M5Tree model. MM-SA ensemble model
improves the accuracy of ANFIS-PSO, ANFIS-FCM, MARS and M5Tree models with respect to RMSE by 8.5%, 5%,
7.4% and 28.8%, respectively. Comparison with the conceptual event-based method indicates that the ML methods
generally performs superior to the EBA4SUB; however, latter method provides better accuracy than the M5Tree and
MARS in some cases.

& Andrea Petroselli 3


Agronomy Department, Faculty of Science, Hydraulics
petro@unitus.it Division University, 20 Août 1955, Route El Hadaik, BP 26,
Skikda, Algeria
& Ozgur Kisi
4
ozgur.kisi@iliauni.edu.ge; ozgurkisi@duytan.edu.vn Department of Civil and Environmental Engineering, Federal
University of Paraı́ba, Paraı́ba, Brazil
Rana Muhammad Adnan
5
rana@hhu.edu.cn Civil Engineering Department, Ilia State University, Tbilisi,
Georgia
Salim Heddam
6
s.heddam@univ-skikda.dz Institute of Research and Development, Duy Tan University,
Da Nang 550000, Vietnam
Celso Augusto Guimarães Santos
celso@ct.ufpb.br
1
State Key Laboratory of Hydrology-Water Resources and
Hydraulic Engineering, Hohai University, Nanjing 210098,
China
2
Department of Economy, Engineering, Society and Business
(DEIM), University of Tuscia, Viterbo, Italy

123
Stochastic Environmental Research and Risk Assessment

Graphic abstract

Keywords EBA4SUB  Conceptual event-based method  Machine learning  Rainfall-runoff modeling

1 Introduction stationary process that is difficult to be captured by these


conventional autoregressive black box models (Nayak et al.
Accurate rainfall-runoff modelling has always been one of 2013). For modelling such complex nonlinear and non-
the hottest topics for researchers in hydrology, because it stationary process, machine learning models, such as,
has a critical role in water resources management, hydro- neural network (ANN) based, fuzzy based and regression
power development, urban planning, irrigation and other based machine learning models have been applied suc-
agro-hydrological/meteorological activities’ planning (Al- cessfully in recent years.
izadeh et al. 2017). With the increasing depletion of In the last two decades, ANN based machine learning
available water resources throughout the world, accurate models have been applied for modeling rainfall-runoff
estimation of runoff becomes more vital and demands process (Kasiviswanathan and Sudheer 2013,2017; Partal
always more accurate methods for modelling the rainfall- et al. 2015; Papacharalampous et al. 2019; Aghelpour and
runoff process. However, accurate estimation of runoff is a Varshavian 2020). Birikundavyi et al. (2002), de Vos and
challenging task due to the complicated relationship Rientjes (2005), Antar et al. (2006) and Chen et al. (2014)
between rainfall and runoff. River runoff changes not only used the multilayer perceptron (MLP) neural network
are dependent on rainfall but also on many other meteo- model for predicting runoff of Mistassibi River, Geer
rological parameters such as evapotranspiration, solar catchment, Blue Nile Basin, and Jinsha River, respectively,
radiation, wind speed, air temperature, etc., and also and Senthil Kumar et al. (2004) applied radial basis func-
catchment specific characteristics such as topography, tion (RBF) neural network model in comparison of MLP
shape, slope, altitude, soil type, land cover, moisture neural network model for rainfall-runoff modeling. They
holding capacity of soil, and so on. For successful running found RBF performed better than MLP neural network
of rainfall-runoff models, the data about all these param- model, and Jain et al. (2004) utilized the back propagation
eters are usually required and this is a difficult and chal- (BP) neural network model for daily runoff modeling of
lenging task especially in developing countries and in Kentucky River basin. Kisi (2007) applied Levenberg–
scarcely gauged catchments (Wu and Chau 2011). Marquardt (LM) neural network in comparison of BP
If reliable data of the mentioned climatic and site neural network model and he found that the LM performed
specific factors are not available, black box models are better than the BP neural network model. Kisi (2008) used
preferred over conceptual and physically based models for generalized regression neural network (GRNN) model in
modelling rainfall-runoff process due to their low input comparison of RBF and feed forward neural network
data requirements (Nourani et al. 2009). In black box (FFNN) models. He found that the GRNN provided better
models category, autoregressive models were used for results in comparison of other neural network models.
runoff modelling in the past (Unal et al. 2004; Niedzielski Recently, long short-term memory neural network
2007; Moeeni and Bonakdari 2017; Fathian et al. 2019; (LSTMNN) model has received a lot of attention in runoff
Sun et al. 2019). However, the effects of the above-men- modeling due to its great non-linear modeling ability (Yuan
tioned parameters on runoff make it a nonlinear and non- et al. 2018; Zhu et al. 2020). In addition to modeling

123
Stochastic Environmental Research and Risk Assessment

precisely rainfall-runoff relationship, ANN models esti- essential. For this purpose, particle swarm optimization
mated rainfall time series alone very accurately (Singh and (PSO) heuristic algorithm was adopted to tuning the
Borah 2013). Aksoy and Dahamsheh (2009) utilized the parameters of ANFIS models (Zhang et al. 2018).
two artificial neural networks (FFN and RBF) models in Among the regression based machine learning models,
comparison of multiple linear regression (MLR) model to used for decades, multivariate adaptive regression splines
predict the monthly rainfall data of three climatic stations model (MARS) and M5 model tree (M5Tree) have been
in Jordan. utilized in many hydrological applications (Yaseen et al.
Although the ANN models were successfully applied in 2016; Adnan et al. 2018, 2019; 2020; Yin et al. 2018; Al-
literature for rainfall-runoff modelling, however, these Sudani et al. 2019; Rezaie-Balf et al. 2019; Mehdizadeh
models could not capture the uncertainties in the rainfall- et al. 2019; Kisi et al. 2019; Nourani et al. 2019a, b). Al-
runoff data very well. To capture these uncertainties, fuzzy Sudani et al. (2019) applied MARS model for modeling
models can be used due to their ability to handle the runoff of Tigris River in comparison to the least square
uncertain data (Chen et al. 2019). Therefore, a hybrid support vector regression (LSSVR) model. Adnan et al.
combination of neural network and fuzzy system namely (2019) utilized the MARS models in predicting monthly
ANFIS (Adaptive Neuro-Fuzzy Inference System) would river flows of Chakdara and Kalam stations. They also used
be more useful to precisely model rainfall-runoff process the LSSVR and the optimally pruned extreme learning
(Nourani and Komasi 2013). Nayak et al. (2004) utilized machine (OP-ELM) models. In the results, they found that
the ANFIS model in comparison to the ANN for modeling both MARS and LSSVR provided better results for the
the runoff of Baitarani River in India. They found that the both stations than the OP-ELM model. Rezaie-Balf et al.
ANFIS performed better than the ANN models. Remesan (2019) applied the MARS models coupling with complete
et al. (2008) applied ANFIS model for rainfall-runoff ensemble empirical mode decomposition with adaptive
modeling and compared the results with neural network noise (CEEMDAN) denoising technique, for streamflow
auto regressive with exogenous input (NNARX) model. prediction of High Dam. They evaluated CEEMDAN-
Jothiprakash et al. (2009) used ANFIS for modelling runoff MARS with NN-CEEMDAN and M5-CEEDMAN models
of an intermittent river. They trained ANFIS models with and found that the MARS model combined with CEEM-
different number of membership functions and in results DAN gave superior performance compared to the other
they found that the ANFIS model with three membership models. Mehdizadeh et al. (2019) used the MARS models
functions provided the best results. Talei et al. (2010) to model monthly runoff of two hydro-meteorological
applied ANFIS for event-based rainfall-runoff modeling in stations in Iran and two stations in Canada. They evaluated
comparison of storm water management model (SWMM). MARS models with other autoregressive models using
In the results, they found that the ANFIS estimated peak local and external data and found that the MARS models
flow more precisely than the SWMM model. Shiri and Kisi outperformed the autoregressive models. The M5Tree was
(2010) developed a new wavelet ANFIS model for esti- used to predict long term runoff of four different stations,
mation of short term and long term runoff. Further devel- namely, Besiri and Malabadi located in Turkey, Hit and
opment in the research field of the ANFIS model Baghdad located in Iraq using the periodicity as input by
application for runoff prediction were ANFIS-grid parti- Yaseen et al. (2016); to forecast monthly and daily river
tioning (ANFIS-GP) and ANFIS-subtractive clustering flow of Hunza River basin using nearby climatic data by
(ANFIS-SC) which has been applied by Sanikhani and Kisi Adnan et al. (2018); to model runoff over the Mediter-
(2012). They used the monthly flow data of the Besiri and ranean region incorporating climate signal information in
the Baykan Station on the Bitlis Stream in Turkey. Their comparison to the LSSVR and MARS models by Kisi et al.
study results demonstrated that the ANFIS-SC model fits (2019); to model the rainfall-runoff using three different
better than the ANFIS-GP in runoff prediction. The results data division strategies using wavelet M5tree model by
of the research done by Adnan et al. (2017) also validates Nourani et al. (2019a, b).
the outcome of Sanikhani and Kisi (2012) study. They In this paper, the recently developed conceptual hydro-
found ANFIS-SC was better in predicting river runoff of logical model named Event-Based Approach for Small and
Gilgit Basin in comparison of ANFIS-GP, FFNN, RBNN Ungauged Basins (EBA4SUB) is also evaluated for rain-
and GRNN models. Kisi et al. (2018) developed the fall-runoff modeling. Few applications related to the
ANFIS-fuzzy C-Means clustering (ANFIS-FCM) for long EBA4SUB model can be found in the literature. Grimaldi
term runoff prediction of two hydrometric stations in USA. and Petroselli (2015) firstly utilized this model in an Italian
They found that the ANFIS-FCM provided better results basin for peak discharge estimation in comparison of
than the other two ANFIS models. To reduce the ANFIS rational formula method. Then, Piscopia et al. (2015)
model’s uncertainty and improve the model prediction applied this method successfully and found satisfactory
performance, optimal selection of parameters is very results compared to rational formula. Then, Petroselli and

123
Stochastic Environmental Research and Risk Assessment

Grimaldi (2018) evaluated the model results with observed resolution was provided by the Italian Geographic Military
discharges of five catchments and found satisfactory Institute, land cover data were retrieved from the Coordi-
results. To test the model capability in areas which have nation of Information on Environment (CORINE) Database
less information and data, recently Młyński et al. (2018), website, and soil data has been derived from the soil map
Petroselli et al. (2019a, b; 2020) and Vojtek et al. (2019) provided by the local administration. The elevations of the
applied the model for determining design peak flow in investigated area range within the interval 51–883 m, the
ungauged basins of Bosnia Herzegovina, Iran, Poland and total contributing area is 178.5 km2, and the basin average
Slovakia in comparison to other methods and found slope is approximately 19%. Regarding the land use, the
promising results using EBA4SUB for these catchments. selected case study is characterized by mountain areas
In all of the aforementioned studies related to EBA4- covered mostly by broadleaved woods, and by bottom
SUB, the model has been utilized only for design flood valleys that are mainly floodplains covered by farmlands
estimation comparing its results with other rainfall-runoff and urbanized. Regarding the soil data, the case study can
hydrological models. However, EBA4SUB has not been be classified as a mix between loamy sand and sandy loam.
compared with the machine learning methods in rainfall- For the selected case study, three years (2014, 2015 and
runoff modeling yet. Hence, the main contribution of the 2016) of synchronized hourly rainfall and runoff data have
present study will be the first time application of EBA4- been selected and used in the present work. Regarding the
SUB hydrological model in comparison of ANFIS-FCM, rainfall data, the rainfall information has been averaged
ANFIS-PSO, MARS and M5Tree machine learning models using the Thiessen polygon method and considering three
and improving the results by applying multi model simple raingauges present in the basin area.
averaging ensemble method (Nourani et al. 2019a, b).
2.2 EBA4SUB

2 Materials and methods Event-Based Approach for Small and Ungauged Basins
(EBA4SUB) estimates the design hydrograph and its peak
2.1 Case study discharge related to a given return period and is based on
an event-based approach that can be used in ‘‘small’’ and
The selected case study, shown in Fig. 1, is the Samoggia ‘‘ungauged’’ watersheds. For ‘‘small’’ we mean having a
River basin, tributary of the Reno River and located in total contributing area smaller than few hundreds of square
central Italy. The Digital Elevation Model (DEM) at 20-m kilometers, so that the assumption of spatially homogenous

Fig. 1 Study area. Localization (left), DEM (center) and land cover (right)

123
Stochastic Environmental Research and Risk Assessment

8
rainfall is still acceptable, while for ‘‘ungauged’’ we mean <  iðtÞ for  t\tpond
without discharge measurements. The EBA4SUB frame- q0 ð t Þ ¼ N s ð2Þ
work consists of three steps: (a) gross rainfall identification : KS 1 þ for t [ tpond
F ðt Þ
(design hyetograph based on Depth-Duration-Frequency—
DDF—curves or observed rainfall data can be used), where q0(t) is the infiltration rate, tpond is the ponding time,
(b) excess rainfall estimation, and (c) design hydrograph F(t) is the cumulative infiltration, Ks is the saturated
calculation. Other than rainfall information, the only input hydraulic conductivity, and Ns is the moisture-tension
data needed by EBA4SUB are the DEM, the land use and parameter. This equation assumes that tpond is reached
the soil type information. when P(t) equals to Ia, while Ks is automatically calibrated
so that the total cumulative net runoff obtained by the SCS-
2.2.1 Gross rainfall CN procedure equals to the corresponding one computed
by the Green-Ampt method. The other Green-Ampt equa-
If DDF curves based on assumed return periods are available, tion parameters are determined starting from the selected
EBA4SUB can estimate the gross rainfall hyetograph case study soil type based on look up tables.
selecting the desired rainfall duration (usually equal to the
basin concentration time, that is estimated from DEM using 2.2.3 Runoff hydrograph
empirical formulas) and the desired rainfall pattern (rectan-
gular, triangular, or Chicago). In this work, anyway, since we The design hydrograph is determined using the DEM and
have available three years of continuous rainfall at hourly the width function based instantaneous unit hydrograph
timescale, we followed a different approach. First, we (WFIUH). DEM at standard resolution is appropriate for
identified single independent rainfall events that are defined lumped modeling, but higher resolution is preferred if
as rainfall sequences preceded and followed by at least 24 available (Petroselli 2012).
consecutive hours without rain thus assuming the physical The following steps are performed to pre-process the
hypothesis that the initial abstraction phenomena return to DEM: (1) a Physical Erosion Model for PIT removal
the pre-event condition in one day (Grimaldi et al. 2012). (PEM4PIT) is used to remove the pits and flat areas
Then, we modeled the rainfall-runoff relationship according (Santini et al. 2009), (2) the drop analysis is used to extract
what expressed in the following. the river network (Tarboton et al. 1991), and (3) an opti-
mized single flow direction algorithm is used to define the
2.2.2 Excess rainfall flow path based on Nardi et al. (2008). The WFIUH is
based on the following formula (Grimaldi et al. 2012):
Excess rainfall is determined based on the observed gross Lc ð xÞ Lh ð xÞ
WFIUH ðtÞ ¼ þ ð3Þ
rainfall hyetograph using the procedure suggested by Gri- vc ð xÞ vh ð xÞ
maldi et al. (2013), named Curve Number for Green-Ampt
(CN4GA). CN4GA is a mixed approach combining the two where Lc and Lh are respectively the length of the path for
methods. First, the SCS-CN procedure (NRCS 2008) is channel and hillslope cell x of the DEM, which are deter-
used to determine the ponding time and the cumulative mined based on optimized flow direction, and vc and vh are
excess rainfall, as follows: respectively surface flow velocity for channel and hillslope
8 cell, which represent parameters affecting the shape of the
< ðP  IaÞ2 WFIUH. Then, after the WFIUH is defined, the hydrograph
Pn ¼ P þ S  Ia for P  Ia ð1Þ Q(t) is obtained according to:
:
0 for P\Ia
Zt
where Pn is the cumulative excess rainfall (mm), P is the QðtÞ ¼ A WFIUH ðt  sÞPn ðsÞds ð4Þ
cumulative gross precipitation (mm), S is the maximum 0
potential basin retention (mm), and Ia is the initial
abstraction due to the interception, infiltration and surface where A is the basin area (km2), t is the rainfall duration
storage (mm). In the present analysis, Ia is assumed as the (h), s is the time step, and Pn(s) is the CN4GA excess
cumulative gross rainfall between the event start and the rainfall (mm/h). Regarding the estimation of surface flow
first moment when observed runoff is [ 0, while S is velocities, vh can be linked according empirical formulas to
derived from Eq. (1) computing cumulative rainfall and local slope and land cover data (Grimaldi et al. 2012),
runoff from observed data. The next step is to determine while vc represents the only parameter to be determined,
the temporal distribution of the excess rainfall based on the and its value can be calibrated for each event so that the
Green-Ampt equation (Santos et al. 2003; Green and Ampt modeled peak discharge can be similar to the observed one.
1911):

123
Stochastic Environmental Research and Risk Assessment

2.3 ANFIS following values were set: population size = 25, max-it-
eration = 500 and 1000.
ANFIS is a type of ANN based on fuzzy inference system
(Jang 1993), in which the ANN and fuzzy logic comple- 2.3.2 FCM
ment each other. ANFIS integrates both ANN and fuzzy
logic features; thus, it is able to take advantage of both in a The fuzzy C-means clustering (FCM), a method developed
single system. It is able to model the uncertain scenarios by Dunn (1973) and improved by Bezdek (1981), is used
using fuzzy logic and ANN learns such a model (Jang et al. here in the ANFIS for clustering. It allows one piece of
1997; Jalalkamali 2015). Two main types of fuzzy infer- data to belong to two or more clusters. Further details about
ence systems are well known, i.e. Mamdani and Sugeno the method can be obtained in the aforementioned refer-
systems (Kennedy and Eberhart 1948; Mamdani and ences. However, it is based on minimization of an objective
Assilian, 1975). In this study, the Sugeno fuzzy inference function, the fuzzy partitioning is carried out by an iterative
system was chosen because it (1) is computationally effi- optimization of such function and the iteration stops when
cient; (2) works well with linear techniques; (3) works well a termination criterion is reached. Thus, data are bound to
with optimization and adaptive techniques; (4) guarantees each cluster by means of a membership function, which
output surface continuity; and (5) is well-suited to mathe- represents the fuzzy behavior of such procedure. Therefore,
matical analysis. Further details can be found in Jang a matrix is built, whose factors are numbers between 0 and
(1993); and Jang et al. (1997). The dataset was divided into 1 and represent the degree of membership between data
three equal parts, and a cross validation was performed, in and centers of clusters.
which two parts were used for training and one for testing;
however, each method was tested using all data. In the 2.4 MARS
presented study, two ANFIS methods were implemented,
one is ANFIS optimized with particle swarm optimization Multivariate adaptive regression splines (MARS) is a non-
(PSO) and other is ANFIS with fuzzy C-means clustering linear and non-parametric method proposed by Friedman
(FCM). PSO and FCM are briefly explained in the fol- (1991), which is able to predict a continuous dependent
lowing sections. parameter in high-dimensional data through a forward–
backward procedure. Several basis functions are added to
2.3.1 PSO the model during the forward stepwise stage to initially
overfit the data. It is worth noting that the basis functions
The concept for the optimization of nonlinear functions are dependent on a spline function which split the dataset
using particle swarm methodology was firstly introduced into piecewise linear segments. The MARS-based models
by Kenndy and Eberhart (1948), when they called such a are described as follows:
methodology as particle swarm optimization (PSO). Since X
M
then, PSO has been widely applied in non-linear problems. f ð x Þ ¼ d0 þ dm hm ð X Þ ð5Þ
It is an evolutionary algorithm firstly inspired by the m¼1
behavior of birds, but now it is also based on biological and where hm(x) are spline functions, d0 and dm are coefficients
sociological behavior of other organisms such as ants, bees that can be computed by the minimum sum of squared
and fishes (Bashir and El-Hawary 2009; Santos et al. errors from splines functions, and M is the number of
2010, 2011). One can note that a school of fish or a swarm functions. The unnecessary basis functions are eliminated
of birds or insets searches for protection or food in a very during the backward stage, which is called as pruning
typical manner, which could be applied to a manner to stage. The functions are eliminated by the generalized
search for a global minimum of an objective function in a cross-validation criterion. Thus, this procedure improves
feasible space. Every member of the swarm searches for the prediction accuracy and M is determined (Kisi and
the best in its locality, learns from its own experience and Parmar 2016). More detailed about the MARS algorithm is
shares with other members. The PSO mimics such a available in Friedman (1994).
behavior, in which each individual of the swarm is a par-
ticle that has a position and a velocity. Thus, the member of 2.5 M5 model tree
a swarm will communicate the good seen positions to each
other and adjust their own position and velocity. Further M5 model tree (M5Tree) is a regression tree, which is
details can be found in Kennedy and Eberhart (1948), and based on two steps: (1) creating a decision tree using a
Santos et al. (2010, 2011). Based on several tests, the splitting criterion, and (2) designing the model tree by

123
Stochastic Environmental Research and Risk Assessment

pruning the overgrown tree and replace the sub-trees with PN


½ðQ0 Þi  ðQc Þi 2
linear regression functions (Singh et al. 2010). NSE ¼ 1  Pi¼1
N
; 1\NES  1 ð9Þ
i¼1 ½ðQ0 Þi  Q 0 2
The M5Tree uses the same concept of a binary decision
tree to split the data, but instead of the class labels, it uses SI ¼ RMSE=Q0 ð10Þ
linear regression functions at the nodes, which predict PN
ðQ0 Þi  ðQc Þi
continuous numerical attributes. The splitting criterion is d ¼ 1  PN  i¼1   ð11Þ
ðQc Þ  Q0 ðQo Þ  Q0 
based on treating the standard deviation (sd) of the class i¼1 i i
values that reach a leaf as a measure of error at that leaf. in which, N is the number of data, Qo, Qc, Q0 are the
The expected reduction in this error is calculated as a result measured, calculated, mean measured, hourly streamflow,
of testing each attribute at that leaf. The standard deviation respectively.
reduction (SDR) is computed as: The statistical parameters of the applied data are sum-
X jN i j marized in Table 1. As seen from the table, whole data
SDR ¼ sd ðN Þ  sd ðNi Þ ð6Þ
jN j have three years, so they first were split into three equal
parts. Then, each model was trained using two parts and
where sd is the standard deviation, N is the examples that
tested with remaining part. At each step, training and
reach the leaf and Ni is the subset of examples with the ith
testing parts were changed. Thus, all the models were
outcome of the potential set.
tested with the whole data. It is apparent from Table 1 that
One may note that data in child nodes present less sd
the precipitation and streamflow data have considerably
than those at the parent node, which makes it purer. Thus,
high skewed distributions (Skewness ranges 9.76–10.4 for
after all split possibilities, M5 selects the one that maxi-
precipitation and 11.9–19.7 for streamflow).
mizes the expected error reduction, which may lead to a
Results for hourly streamflow prediction using the
large tree-like structure inducing to an over-fitting. Then,
ANFIS-PSO with different combinations of streamflow
the aforementioned step (2) takes place when, for example,
(Q) and precipitation are summarized in Table 2. For the
a sub-tree is replaced by a leaf (Pal and Deswal 2009).
employment of ANFIS-PSO, population was set to 25 and
two iterations were applied as 500 and 1000. As Table 2
shows, for the combinations of Q and precipitation used to
3 Results and discussion predict the Q, this study yielded diverse results. Using only
previous precipitation as input variable (combination 1–4),
In the presented study, we applied five models using different
the ANFIS-PSO performed relatively worse, with mean
input combination for forecasting hourly streamflows using
RMSE between 4.39 m3/s and 6.56 m3/s, and mean MAE
the cross-validation technique. The proposed models were:
in the range of 1.33–1.66 m3/s, and the high NSE values
(1) ANFIS-PSO, (2) ANFIS-FCM, (3) MARS, (4) M5Tree
and (5) the multi model with simple averaging called MM-
SA. MM-SA is commonly used in the literature (Nourani
Table 1 The statistical parameters of the applied data
et al. 2019a, b) and its results are calculated by simply aver-
aging the results of the single implemented methods (e.g., Whole data M1 M2 M3
ANFIS-PSO, ANFIS-FCM, MARS, M5Tree and EBA4- Precipitation, mm
SUB). Results of the different models and modelling strate- Mean 0.113 0.110 0.107 0.123
gies with different input combinations were assessed and Min 0 0 0 0
compared, and then further validated by comparison with the Max 16.5 12.2 14.1 16.5
EBA4SUB model. As a basis for comparison, models were Median 0 0 0 0
validated using five statistics indices: Nash–Sutcliffe effi-
Skewness 10.3 9.76 10.4 10.1
ciency (NSE), root mean squared error (RMSE), mean
Std. dev 0.582 0.501 0.578 0.658
absolute error (MAE), scatter index (SI) and adjusted index of
Streamflow, m3/s
agreement (d) calculated as follows:
Mean 1.406 1.732 1.078 1.410
1X
N
Min 0 0 0 0
MAE ¼ jðQ0 Þi  ðQc Þi j ð7Þ
N i¼1 Max 298 167 104 298
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Median 0.310 0.590 0.197 0
u N
u1 X Skewness 19.7 14.2 11.9 19.4
RMSE ¼ t ½ðQ0 Þi  ðQc Þi 2 ð8Þ Std. dev 6.63 6.36 3.90 8.71
N i¼1
M1 = 2016, M2 = 2015, M3 = 2014

123
Stochastic Environmental Research and Risk Assessment

Table 2 Test statistics of the ANFIS-PSO using different combinations of streamflow and precipitation for hourly streamflow prediction
Statistics Cross validation Test data set Input combination
(i) (ii) (iii) (iv) (v) (vi) (vii) (viii)

RMSE M1 2016 4.58 4.51 3.28 7.55 1.40 1.83 1.37 1.57
M2 2015 4.17 5.19 3.81 4.67 1.01 0.96 0.90 0.88
M3 2014 9.43 5.27 6.07 7.45 2.97 3.02 2.69 2.69
Mean 6.06 4.99 4.39 6.56 1.79 1.94 1.65 1.71
MAE M1 2016 1.48 1.41 1.23 1.66 0.26 0.29 0.27 0.28
M2 2015 1.32 1.43 1.40 1.71 0.25 0.26 0.23 0.23
M3 2014 1.46 1.16 1.40 1.60 0.29 0.29 0.28 0.27
Mean 1.42 1.33 1.34 1.66 0.27 0.28 0.26 0.26
NSE M1 2016 0.483 0.499 0.735 - 0.408 0.951 0.917 0.954 0.939
M2 2015 - 0.140 - 0.767 0.049 - 0.427 0.934 0.940 0.947 0.949
M3 2014 0.457 0.635 0.516 0.271 0.884 0.880 0.905 0.905
Mean 0.360 0.600 0.433 0.369 0.923 0.912 0.935 0.930
In the table, input combinations are (i) Pt-1, Pt-2, …, Pt-10, (ii) Pt-1, Pt-2, …, Pt-20, (iii) Pt-1, Pt-2, …, Pt-30, (iv) Pt-1, Pt-2, …, Pt-40, (v) Pt-1,
Pt-2, …, Pt-30, Qt-1, (vi) Pt-1, Pt-2, …, Pt-30, Qt-1, Qt-2, (vii) Pt-1, Pt-2, …, Pt-30, Qt-1, Qt-2, Qt-3 and (viii) (v) Pt-1, Pt-2, …, Pt-30, Qt-1, Qt-
2, Qt-3, Qt-4. Among the P inputs, input combination (iii) provided the best accuracy for the M1 and M2 data sets and therefore, after the 4th
combination, streamflows were added to this combination. For the M3 set, however, input combination (ii) provided the best accuracy

obtained by the model was 0.735 using the combination 3 provided the best accuracy. As can be seen by referring to
for the M1 dataset. Differences between the three datasets the statistical indices reported in Table 3, all ANFIS-FCM
(M1–M3) were slight, though overall performances models using the M1–M3 subset are not improved in terms
declined slightly as the models included more precipitation of RMSE, MAE and NSE values. Comparing the results
data (combination 1–4). However, in regard to NSE values, using only the precipitation as input variables (combination
it is clear from Table 2 that the lowest accuracy was 1–4), shows the deficiency of the ANFIS-FCM to capture
obtained using the M2 dataset. Inclusion of previous the high variation across the entire dataset. In one hand, the
Q contributed to increasing the accuracy of the three performances of the ANFIS-FCM appear to have slightly
datasets (M1–M3), where the improvement was most sig- improved for the input combination 1, 2 and 4, using the
nificant in case of M1 and M2 datasets. In general, the M1 subset, presenting NSE values that are slightly higher
reduction of RMSE and MAE occurred most effectively for than those using the ANFIS-PSO: 0.519 versus 0.483,
those subsets (M1 and M2), significantly more than the M3 0.509 versus 0.499, and 0.560 versus - 0.408, respec-
subset. All in all, the most accurate models were those tively. In the other hand, for the combination 4, it can be
having the previous Q as input variables (combination clearly observed that the ANFIS-PSO was more accurate
5–8), yet the most significant improvement was achieved than the ANFIS-FCM with 17.8% improvement in the NSE
using the model with the combination 7, for which the best value. No differences in model performance (as indicated
performance was obtained, with mean NSE value of 0.935 by the NSE) are apparent between the input combinations
and mean RMSE and MAE of 1.65 m3/s and 0.26 m3/s, 1–4 for the M2 subset. The lowest NSE value (0.445) was
respectively. In addition, the least accurate models were obtained using the combination 2. In addition, for the M3
those featuring only the previous precipitations as predic- subset, the NSE values in both input combinations (1–4) is
tors without the addition of any Q lag, which were not very small, reflecting the limitation of the models to cor-
improved in any case as more previous precipitation were rectly predict the Q using only the precipitation as input
included. Hence, it was demonstrated that the inclusion of variables, leading to the necessity to use previous stream-
previous Q for the input variables significantly helped in flow record as input. Comparisons between the results in
reducing the RMSE and MAE value, yet it did introduce a Table 3 are informative. They suggest that models based on
remarkable improvement of the models’ performances and previous Q values (combination 5–8) perform as well as
yielded the best accuracies. those with only the precipitation for all the three subsets
Results using the ANFIS-FCM models are summarized (M1–M3), but markedly high using the M1 subset. Models
in Table 3. For this method, iteration was set to 100 and for subset M1 all perform well for all combinations (5–8),
cluster number from 2 to 8 was tried and 2 clusters with NSE values in excess of 0.950, small RMSE and MAE

123
Stochastic Environmental Research and Risk Assessment

Table 3 Test statistics of the ANFIS-FCM using different combinations of streamflow and precipitation for hourly streamflow prediction
Statistics Cross validation Test data set Input combination
(i) (ii) (iii) (iv) (v) (vi) (vii) (viii)

RMSE M1 2016 4.42 4.46 4.24 4.23 1.41 1.32 1.32 1.33
M2 2015 11.4 39.2 4.34 4.33 1.01 1.32 0.89 0.89
M3 2014 6.45 6.50 6.35 6.34 2.97 2.68 2.55 2.56
Mean 7.42 16.72 4.98 4.97 1.80 1.77 1.59 1.59
MAE M1 2016 1.41 1.37 1.41 1.40 0.30 0.31 0.29 0.29
M2 2015 1.67 2.18 1.71 1.71 0.31 0.31 0.28 0.29
M3 2014 1.48 1.34 1.31 1.30 0.31 0.28 0.28 0.28
Mean 1.52 1.63 1.48 1.47 0.31 0.30 0.28 0.29
NSE M1 2016 0.519 0.509 0.557 0.560 0.951 0.957 0.957 0.957
M2 2015 - 7.55 - 99.6 - 0.232 - 0.230 0.933 0.957 0.948 0.948
M3 2014 0.454 0.445 0.470 0.471 0.884 0.905 0.914 0.914
Mean 2.841 33.5 0.420 0.420 0.923 0.940 0.940 0.940
In the table, input combinations are (i) Pt-1, Pt-2, …, Pt-10, (ii) Pt-1, Pt-2, …, Pt-20, (iii) Pt-1, Pt-2, …, Pt-30, (iv) Pt-1, Pt-2, …, Pt-40, (v) Pt-1,
Pt-2, …, Pt-30, Qt-1, (vi) Pt-1, Pt-2, …, Pt-30, Qt-1, Qt-2, (vii) Pt-1, Pt-2, …, Pt-30, Qt-1, Qt-2, Qt-3 and (viii) (v) Pt-1, Pt-2, …, Pt-30, Qt-1, Qt-
2, Qt-3, Qt-4. Among the P inputs, input combination (iii) provided the best accuracy and therefore, after the 4th combination, streamflows were
added to this combination

compared to the mean value (Table 4). For M3 subset, the 0.957 for combination 6. The model performed slightly less
models are relatively lower than for M1 subset, with NSE well compared to the M1 dataset and significantly higher
in the range of 0.884–0.914 and RMSE substantially in than the M3, and similar accuracies were found for the
excess of the average values. Regarding the M2 subset, the combination 5 with NSE of 0.957. Not surprisingly,
developed ANFIS-FCM was able to successfully predict accuracy as measured by mean NSE, RMSE and MAE was
the streamflow in one case for the four-input combination higher for the combination 7 with values of 0.940, 1.59 m3/
(5–8). The mapping accuracy of the M2 dataset was good s and 0.28 m3/s, respectively.
with NSE values ranging from 0.933 for combination 5 to

Table 4 Test statistics of the MARS models using different combinations of streamflow and precipitation for hourly streamflow prediction
Statistics Cross validation Test data set Input combination
(i) (ii) (iii) (iv) (v) (vi) (vii) (viii)

RMSE M1 2016 4.61 4.33 4.34 4.34 1.43 1.76 1.64 1.80
M2 2015 5.08 4.85 4.82 4.82 1.19 0.93 0.87 0.90
M3 2014 6.47 6.24 6.16 6.16 2.87 2.50 2.39 2.39
Mean 5.39 5.14 5.11 5.11 1.83 1.73 1.63 1.70
MAE M1 2016 1.51 1.42 1.46 1.46 0.26 0.25 0.24 0.25
M2 2015 1.61 1.65 1.72 1.72 0.22 0.20 0.18 0.18
M3 2014 1.45 1.32 1.30 1.30 0.28 0.25 0.24 0.24
Mean 1.52 1.52 1.52 1.49 0.25 0.23 0.22 0.22
NSE M1 2016 0.477 0.539 0.536 0.536 0.950 0.923 0.934 0.920
M2 2015 - 0.688 - 0.540 - 0.522 - 0.522 0.907 0.943 0.950 0.947
M3 2014 0.450 0.489 0.501 0.501 0.892 0.918 0.925 0.925
Mean 0.538 0.538 0.538 0.520 0.916 0.928 0.936 0.931
In the table, input combinations are (i) Pt-1, Pt-2, …, Pt-10, (ii) Pt-1, Pt-2, …, Pt-20, (iii) Pt-1, Pt-2, …, Pt-30, (iv) Pt-1, Pt-2, …, Pt-40, (v) Pt-1,
Pt-2, …, Pt-20, Qt-1, (vi) Pt-1, Pt-2, …, Pt-20, Qt-1, Qt-2, (vii) Pt-1, Pt-2, …, Pt-20, Qt-1, Qt-2, Qt-3 and (viii) (v) Pt-1, Pt-2, …, Pt-20, Qt-1, Qt-
2, Qt-3, Qt-4. Among the P inputs, input combination (ii) provided the best accuracy and therefore, after the 4th combination, streamflows were
added to this combination

123
Stochastic Environmental Research and Risk Assessment

Table 5 Test statistics of the M5Tree using different combinations of streamflow and precipitation for hourly streamflow prediction
Statistics Cross validation Test data set Input combination
(i) (ii) (iii) (iv) (v) (vi) (vii) (viii)

RMSE M1 2016 5.02 4.63 4.73 4.54 2.61 2.60 2.60 2.60
M2 2015 4.42 4.34 4.88 4.76 1.17 2.58 2.58 2.58
M3 2014 6.69 4.89 5.63 5.91 2.59 2.55 2.48 2.52
Mean 5.38 4.62 5.08 5.07 2.12 2.58 2.55 2.57
MAE M1 2016 1.51 1.44 1.46 1.40 0.25 0.24 0.24 0.24
M2 2015 1.46 1.38 1.42 1.40 0.18 0.24 0.24 0.24
M3 2014 1.55 1.30 1.31 1.34 0.265 0.24 0.23 0.24
Mean 1.51 1.37 1.40 1.38 0.23 0.24 0.24 0.24
NSE M1 2016 0.379 0.472 0.449 0.491 0.832 0.833 0.833 0.833
M2 2015 - 0.279 - 0.233 - 0.560 - 0.484 0.910 0.836 0.836 0.836
M3 2014 0.411 0.686 0.584 0.541 0.912 0.915 0.919 0.917
Mean 0.356 0.500 0.531 0.505 0.885 0.861 0.863 0.862
In the table, input combinations are (i) Pt-1, Pt-2, …, Pt-10, (ii) Pt-1, Pt-2, …, Pt-20, (iii) Pt-1, Pt-2, …, Pt-30, (iv) Pt-1, Pt-2, …, Pt-40, (v) Pt-1,
Pt-2, …, Pt-30, Qt-1, (vi) Pt-1, Pt-2, …, Pt-30, Qt-1, Qt-2, (vii) Pt-1, Pt-2, …, Pt-30, Qt-1, Qt-2, Qt-3 and (viii) (v) Pt-1, Pt-2, …, Pt-30, Qt-1, Qt-
2, Qt-3, Qt-4. Among the P inputs, input combination (iv) provided the best accuracy for the M2 and M3 data sets. For the M1 set, however, input
combination (iv) provided the best accuracy and therefore, after the 4th combination, streamflows were added to this combination

Table 6 Test statistics of the


Statistics Cross validation Test data set ANFIS-PSO ANFIS-FCM MARS M5Tree MM-SA
best ANFIS-PSO, ANFIS-FCM,
MARS, M5Tree and MM-SA in RMSE M1 2016 1.37 1.32 1.64 2.61 1.37
hourly streamflow prediction
M2 2015 0.90 0.89 0.87 1.17 0.79
M3 2014 2.69 2.55 2.39 2.59 2.38
Mean 1.65 1.59 1.63 2.12 1.51
MAE M1 2016 0.27 0.29 0.24 0.25 0.23
M2 2015 0.23 0.28 0.18 0.18 0.19
M3 2014 0.28 0.28 0.24 0.265 0.24
Mean 0.26 0.28 0.22 0.23 0.22
NSE M1 2016 0.954 0.957 0.934 0.832 0.954
M2 2015 0.947 0.948 0.950 0.910 0.959
M3 2014 0.905 0.914 0.925 0.912 0.926
Mean 0.935 0.940 0.936 0.885 0.946
SI M1 2016 0.802 0.770 0.959 1.527 0.800
M2 2015 0.837 0.825 0.812 1.090 0.736
M3 2014 1.928 1.831 1.712 1.856 1.705
Mean 1.189 1.142 1.161 1.491 1.080
d M1 2016 0.932 0.928 0.942 0.937 0.943
M2 2015 0.918 0.901 0.937 0.937 0.931
M3 2014 0.829 0.930 0.939 0.932 0.939
Mean 0.893 0.920 0.939 0.935 0.938
MM-SA is multi model with simple averaging. The estimates of this model were simply calculated by
averaging the estimates of the optimal ANFIS-PSO, ANFIS-FCM, MARS and M5Tree models

The generated overall accuracies using the MARS 5.11 to 5.39 m3/s, mean MAE ranging from 1.49 to
model are displayed in Table 4. The application of MARS 1.52 m3/s, and mean NSE value ranging from 0.520 to
model for the four first input combinations (i.e. 1–4) pro- 0.530, respectively. The M1 subset produced a better result
duced a lower accuracy with mean RMSE ranging from with NSE ranging from 0.477 to 0.539, compared to the

123
Stochastic Environmental Research and Risk Assessment

Fig. 2 The scatterplot ANFIS-PSO ANFIS-FCM


comparison of observed and 200 200
predicted runoff by machine y = 0.9409x + 0.0659
learning methods—M1 data set y = 0.9748x + 0.0611 R² = 0.9575
150 R² = 0.954 150

Model, m3/s

Model, m3/s
100 100

50 50

0 0
0 50 100 150 200 0 50 100 150 200
Observed, m3/s Observed, m3/s

MARS
M5Tree
250 y = 1.0748x - 0.1195 300
R² = 0.95
200 250 y = 1.1164x - 0.1648
Model, m3/s

Model, m3/s
200 R² = 0.8895
150
150
100
100
50 50

0 0
0 50 100 150 200 250 0 50 100 150 200 250 300
Observed, m3/s Observed, m3/s

MM-SA
250
y = 1.0267x - 0.0394
200 R² = 0.9586
Model, m3/s

150

100

50

0
0 50 100 150 200 250
Observed, m3/s

M2 and M3 subsets. The statistical indices illustrate that RMSE and MAE of 1.63 m3/s, 0.220 m3/s, respectively. In
the results from these input combinations having only the addition, the NSE values, however, indicate that the MARS
precipitation as input variables are the poorest. Comparison performances for the M1 subset produced a better result
of the results from MARS method with the result listed in only for combination 5. For numerical comparison, the
Tables 3 and 4 using ANFIS-FCM and ANFIS-PSO NSE value of M1 subset improves the M2 and M3 subsets
models, illustrated that none of applied models was able to by 4.3% and 5.8%, respectively. Overall accuracies
achieve the best accuracy using only precipitation as input regarding the RMSE, MAE, and NSE values were pro-
variable. When the Q was included as input (combination duced using the M2 subset for the combination 7, with
5–8), statistical indices for the three subsets (M1–M3) values equal to 0.87 m3/s, 0.18 m3/s and 0.950,
indicate that the MARS model using the combination 7 respectively.
generated significantly better result than the other as sup- The M5 model tree (M5Tree) was also used to explore
ported by the higher mean NSE value of 0.936 and mean the relation between hourly streamflow and precipitation

123
Stochastic Environmental Research and Risk Assessment

Fig. 3 The scatterplot ANFIS-PSO ANFIS-FCM


comparison of observed and 125 125
predicted runoff by machine y = 0.8889x + 0.1178 y = 0.9538x + 0.0998
learning methods—M2 data set 100 R² = 0.9508 100 R² = 0.9485

Model, m3/s

Model, m3/s
75 75

50 50

25 25

0 0
0 25 50 75 100 125 0 25 50 75 100 125
Observed, m3/s Observed, m3/s

MARS M5Tree
125 125
y = 1.0015x + 0.0279 y = 0.9081x + 0.0764
100 R² = 0.9525 100 R² = 0.9098
Model, m3/s

Model, m3/s
75 75

50 50

25 25

0 0
0 25 50 75 100 125 0 25 50 75 100 125
Observed, m3/s Observed, m3/s

MM-SA
125
y = 0.9381x + 0.0805
100 R² = 0.9593
Model, m3/s

75

50

25

0
0 25 50 75 100 125
Observed, m3/s

using several input combinations. Summaries of the accu- NSE value of 0.686 was obtained using the combination 2
racy of M5Tree model using the M1–M3 subsets are shown for the M3 subset, while they had dropped drastically for
in Table 5, where the RMSE, MAE, and the NSE of the the M1 and M2 subsets and become less than 0.50, rep-
eight input combinations are reported, together with the resenting very low accuracies. Accordingly, with only
mean value calculated for the M1–M3 subset. These show precipitation as input, the M5Tree did not show significant
that similar to the ANFIS-PSO, ANFIS-FCM, and MARS improvement despite increased of the number of the pre-
models, the M5Tree was not able to correctly predict the vious precipitation for combination 1–4. Modifying the
Q using only the precipitation as input variable, and the structure of the input variables by adding the previous
models worked poorly with high RMSE and MAE values, streamflow (combination 5–8) may result in a significant
and the lowest NSE values. Table 5 shows that there are improvement of the models’ performances for the three
greater variations between the accuracies of the three subsets. According to Table 5, for the M1 subset, the
subsets (M1–M3), and the best accuracy with the highest performances of the M5Tree for the four input

123
Stochastic Environmental Research and Risk Assessment

Fig. 4 The scatterplot ANFIS-PSO ANFIS-FCM


comparison of observed and
350 350
predicted runoff by machine y = 0.9672x + 0.0719 y = 0.9502x + 0.0843
learning methods—M3 data set 300 R² = 0.9087 300 R² = 0.9156
250 250

Model, m3/s

Model, m3/s
200 200
150 150
100 100
50 50
0 0
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
Observed, m3/s Observed, m3/s

M5Tree
MARS 350
350 y = 0.8881x + 0.1515
y = 0.9172x + 0.1157 300
300 R² = 0.9126
R² = 0.9252 250

Model, m3/s
250
Model, m3/s

200
200
150 150

100 100
50 50
0 0
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
Observed, m3/s Observed, m3/s

MM-SA
350
y = 0.9307x + 0.1059
300 R² = 0.9257
250
Model, m3/s

200
150
100
50
0
0 50 100 150 200 250 300 350
Observed, m3/s

combinations (5–8) remained very close, with nearly equal from the Table 5 that the accuracy of the model becomes
RMSE (& 2.60 m3/s), MAE (& 0.23 m3/s) and NSE constant with equal NSE values of 0.836.
(& 0.83). The same conclusion was made with respect to In Table 6, we compared the results obtained using the
the M3 subset for which we observed that the M5Tree best models according to the lowest mean values of the
models possess approximately the same RMSE RMSE, MAE, SI and the highest NSE and d with those
(& 2.50 m3/s), MAE (& 0.24 m3/s) and NSE (& 0.915) obtained using the multi model with simple averaging
values. In addition, for the M2 subset, the M5Tree using (MM-SA). Replacing the individual models with the MM-
the combination 5 appeared to be the most accurate model SA model for forecasting streamflow slightly increased the
with high NSE value above 0.91, RMSE and MAE below overall accuracy of the ANFIS-FCM for which the NSE
1.20 m3/s and 0.20 m3/s, respectively. Moreover, it is clear and d values slightly increased (NSE from 0.940 to 0.946
and d from 0.920 to 0.938), the mean RMSE and SI

123
Stochastic Environmental Research and Risk Assessment

Table 7 Comparison of
Method RMSE MAE NSE SI d RMSE MAE NSE SI D
machine learning and
EBA4SUB methods in Event 1 Event 2
modeling rainfall-runoff of
various events ANFIS-PSO 16.2 4.89 0.892 0.665 0.921 7.23 2.20 0.948 0.485 0.945
ANFIS-FCM 16.4 4.89 1.000 0.674 0.921 6.86 2.34 1.000 0.460 0.941
MARS 15.1 4.45 0.905 0.623 0.927 9.21 2.79 0.916 0.618 0.933
M5Tree 15.7 5.00 0.898 0.646 0.917 15.8 4.11 0.753 1.057 0.903
MM-SA 15.1 4.38 0.906 0.621 0.928 7.49 2.37 0.944 0.502 0.942
EBA4SUB 20.3 7.23 0.830 0.834 0.886 7.20 2.70 0.949 0.842 0.934
Event 3 Event 4
ANFIS-PSO 4.54 1.77 0.946 0.452 0.927 3.55 0.75 0.952 0.501 0.959
ANFIS-FCM 3.20 1.41 1.000 0.319 0.945 3.62 0.83 1.000 0.511 0.955
MARS 3.80 1.44 0.962 0.378 0.944 3.46 0.81 0.954 0.489 0.956
M5Tree 6.48 2.30 0.891 0.645 0.908 6.60 1.82 0.833 0.931 0.901
MM-SA 3.60 1.34 0.966 0.358 0.947 3.88 0.94 0.942 0.547 0.949
EBA4SUB 3.90 1.61 0.960 0.388 0.938 7.06 2.09 0.809 0.996 0.889
Event 5 Event 6
ANFIS-PSO 2.44 0.73 0.773 0.683 0.900 0.84 0.44 0.985 0.144 0.955
ANFIS-FCM 2.45 0.75 1.000 0.685 0.897 1.19 0.53 1.000 0.203 0.944
MARS 2.53 0.59 0.757 0.708 0.920 1.10 0.44 0.974 0.188 0.955
M5Tree 2.23 0.50 0.812 0.623 0.932 1.24 0.50 0.967 0.211 0.949
MM-SA 2.30 0.59 0.799 0.644 0.920 0.99 0.41 0.979 0.169 0.957
EBA4SUB 5.54 2.59 -0.163 1.548 0.668 3.26 1.89 0.774 0.556 0.810

Fig. 5 Comparison of machine learning and EBA4SUB methods in modeling rainfall-runoff of event 1

123
Stochastic Environmental Research and Risk Assessment

Fig. 6 Comparison of machine learning and EBA4SUB methods in modeling rainfall-runoff of event 2

marginally decreased (RMSE from 1.59 to 1.51 m3/s and EBA4SUB model for simulating six extreme streamflow
SI from 1.142 to 1.080), and with negligible improvement events during the period under study. The six events are
of the MAE value, which dropped from 0.28 to 0.22 m3/s. designated as Event 1 to Event 6. For Event 1 as seen in
While it can be clearly observed that ANFIS-PSO, ANFIS- Table 7 and Fig. 5, performance metrics show that MM-SA
FCM, and MARS models perform significantly better than is better than the other models according to the RMSE,
the M5Tree, and an improvement is observed when MAE, SI and d values, while the NSE value (& 0.906) of
including the previously measured Q as input, it is also the aforementioned model is below that of the ANFIS-
clear that these models cannot surpass the MM-SA model FCM (NSE = 1.00) model, slightly higher than that of the
in terms of overall accuracies. Comparisons of the mea- MARS (NSE = 0.905) and significantly higher than those
sured and calculated streamflow by all developed models obtained using ANFIS-PSO (NSE = 0.892), M5Tree
revealed that the MM-SA model (NSE = 0.946) had a (NSE = 0.898) and EBA4SUB (NSE = 0.898). In overall,
superior performance in comparison to the ANFIS-PSO it can be concluded that the EBA4SUB has the poorest
(NSE = 0.935), ANFIS-FCM (NSE = 0.940), MARS accuracies among all proposed models with high RMSE
(NSE = 0.936) and M5Tree (NSE = 0.885) models. The and MAE of 20.3 m3/s and 5 m3/s, respectively. This is not
MM-SA model decreased the RMSE value of ANFIS-PSO surprising since EBA4SUB has been mainly developed to
by 8.48%, 5.03% for ANFIS-FCM, 7.36% for MARS, and be a parsimonious model, in terms of input parameters, to
28.77% for M5Tree. To sum up significant improvement in be used for design hydrograph estimation using the same
Q forecasting by the MM-SA in comparison with M5Tree information of the rational formula. The MM-SA has a
model is observed in NSE value which increased from marginally lower MAE, RMSE and SI (referring to
& 0.885 to & 0.946 as well as the RMSE that decreased Table 7) compared to MARS model, but overall the dif-
from 2.12 to 1.51 m3/s. The scatterplot of observed and ference between the two is negligible. In conclusion, our
predicted Q using the machines learning methods for M1– results confirm that the EBA4SUB has limitations in cor-
M3 subset are showed in Figs. 2, 3 and 4. rectly simulating the streamflow when several extreme
Finally, we compared in Table 7 the results obtained values (peaks) are presented, whereas the machines learn-
using the developed models with those obtained using the ing models were notably able to accurately simulate the

123
Stochastic Environmental Research and Risk Assessment

Fig. 7 Comparison of machine learning and EBA4SUB methods in modeling rainfall-runoff of event 3

extreme events. For Event 2 as seen in Table 7 and Fig. 6, models were very close together. However, the MM-SA
the ANFIS-FCM performed better than all other models in provided slightly better performances, according to the
terms of RMSE (& 6.86 m3/s), SI (0.460) and NSE RMSE, MAE, SI, NSE and d. For Event 4 (Table 7 and
(& 1.00) values; in general ANFIS-FCM tended to cor- Fig. 8), it is clear that, the comparative results between the
rectly predict the peaks values more accurate than all the proposed models reveal that the best accuracy was obtained
other models. The superiority of the ANFIS-FCM was by the MARS model in accord to the RMSE (3.46 m3/s),
more evident when compared to the M5Tree, for which an MAE (0.81 m3/s) and SI (& 0.489) values, slightly lower
improvement of 56.58% and 43.06% in terms of RMSE than the ANFIS-PSO model (RMSE = 3.55 m3/s;
and MAE was achieved. The results of the M5Tree model MAE = 0.75 m3/s; SI = 0.501). In addition, the lowest
were worse than those obtained using all the other models; accuracy was obtained using the EBA4SUB with higher
in addition, the aforementioned model tends to poorly RMSE, MAE and SI values of 7.06 m3/s, 2.09 m3/s and
estimate the streamflow peaks values, with large RMSE 0.996 and lower NSE and d values of 0.809 and 0.889,
(& 15.80 m3/s), MAE (& 4.11 m3/s), SI (& 1.057), and respectively. For Event 5 (not shown for brevity), The
low NSE (& 0.753) and d (0.903) values. If the accuracy M5Tree provided the best ability with the lowest RSME
prediction of the EBA4SUB is analyzed, in addition to (2.23 m3/s), MAE (0.50 m3/s), SI (0.623) and the highest d
relatively high NSE (& 0.949) the RMSE value was (0.932). According to the NSE, however, the ANFIS-FCM
slightly smaller ((& 7.20 m3/s), when compared to the (NSE = 1.000) has better efficiency than the M5Tree
ANFIS-PSO (RMSE = 7.23 m3/s; NSE = 0.948) and MM- (NSE = 0.812). The accuracy of M5Tree model is closely
SA (RMSE = 7.49 m3/s; NSE = 0.944). Results for Event followed by the MM-SA model with respect to all five
3 (Table 7 and Fig. 7) showed that the best accuracy was comparison criteria. Finally, for Event 6 (not shown for
obtained using the ANFIS-FCM with the lowest RMSE of brevity), the ANFIS-PSO got the 1st rank in modeling of
3.20 m3/s and SI of 0.319 and the highest NSE of 1.00. In this rainfall runoff event with the lowest RMSE (0.84 m3/
addition, MARS, MM-SA and EBA4SUB showed higher s), MAE (0.44 m3/s) and SI (0.144). This model is closely
NSE values as compared to ANFIS-PSO. The differences followed by the MM-SA model. In addition, the ANFIS-
were the largest for M5Tree, and the three aforementioned

123
Stochastic Environmental Research and Risk Assessment

Fig. 8 Comparison of machine learning and EBA4SUB methods in modeling rainfall-runoff of event 4

FCM model has the higher NSE compared to ANFIS-PSO • Inclusion of streamflow information considerably
and other models. improved the models accurcies by remarkably reducing
root means square error and mean absolute error and
increasing Nash–Sutcliffe efficiency.
4 Conclusions • Among the single machine learning methods, ANFIS-
PSO, ANFIS-FCM and MARS were found to have
The study investigated the applicability of four machine similar accuracy and they performed superior to the
learning methods, ANFIS-PSO, ANFIS-FCM, MARS and M5Tree in rainfall-runoff modeling.
M5Tree, in short term rainfall-runoff modeling and results • MM-SA ensemble model was found to be useful
were compared with those of the physically event-based method and it improved the single models’ accuracy;
method, EBA4SUB. First, four methods were implemented increments in RMSE of ANFIS-PSO, ANFIS-FCM,
considering several input scenarios including previous MARS and M5Tree are 8.5%, 5%, 7.4% and 28.8%,
hourly precipitation and streamflow data obtained from respectively.
Samoggia River basin, Italy and cross validation was • Comparison of the methods in event based prediction
applied to get more robust predictions. Then, four rainfall- indicated that the machine learning methods generally
runoff events were selected from the data and the EBA4- outperformed the physically event-based method,
SUB model was applied to predict runoff from rainfall and EBA4SUB, whereas, in some cases it provided better
its results were compared with the machine learning accuracy than the M5Tree (Event 2 and Event 3) and
methods. Finally, multi model simple averaging ensemble MARS (only Event 2).
method was applied to the results and compared with the
The study revealed that the developed ML methods
single methods. Comparison of the results produced the
(e.g., ANFIS-PSO, ANFIS-FCM, MARS) which use lim-
following conclusions:
ited number of parameters can be useful in ungauged
• It was observed that the least accurate models were basins where runoff information, needed for calibrating
found from the models having the previous precipitations advanced physically based models, are not available. In
as inputs only without the addition of any streamflow. this study, the developed ML methods were compared and

123
Stochastic Environmental Research and Risk Assessment

assessed using rainfall-runoff data from of one case study Fathian F, Fakheri-Fard A, Ouarda TB, Dinpashoh Y, Nadoushani
in Italy. For future works, more data from different basins SSM (2019) Multiple streamflow time series modeling using
VAR–MGARCH approach. Stoch Env Res Risk Assess
can be utilized for evaluating the proposed methods. 33(2):407–425
Friedman JH (1991) Multivariate adaptive regression splines. Ann
Acknowledgements The authors are deeply grateful to Attilio Stat 19:1–67
Castellarin, University di Bologna, DICAM Department, Italy, for Green WH, Ampt GA (1911) Studies on soil physics. J Agric Sci
having provided the watershed data used in the present manuscript. 4(1):1–24
Grimaldi S, Petroselli A (2015) Do we still need the rational formula?
An alternative empirical procedure for peak discharge estimation
References in small and ungauged basins. Hydrol Sci J 60:66–67
Grimaldi S, Petroselli A, Nardi F (2012) A parsimonious geomor-
phological unit hydrograph for rainfall–runoff modeling in small
Adnan RM et al (2017) Application of soft computing models in ungauged basins. Hydrol Sci J 57(1):73–83
streamflow forecasting. In: Proceedings of the institution of civil Grimaldi S, Petroselli A, Romano N (2013) Curve-number/green-
engineers-water management, Vol 172, No 3. Thomas Telford ampt mixed procedure for streamflow predictions in ungauged
Ltd. 123–134 October 2017 basins: parameter sensitivity analysis. Hydrol Process
Adnan RM, Yuan X, Kisi O, Adnan M, Mehmood A (2018) Stream 27(8):1265–1275
flow forecasting of poorly gauged mountainous watershed by Jain A, Sudheer KP, Srinivasulu S (2004) Identification of physical
least square support vector machine, fuzzy genetic algorithm and processes inherent in artificial neural network rainfall–runoff
M5 model tree using climatic data from nearby station. Water models. Hydrol Process 18:571–581
Resour Manage 32(14):4469–4486 Jalalkamali A (2015) Using of hybrid fuzzy models to predict
Adnan RM, Liang Z, Trajkovic S, Zounemat-Kermani M, Li B, Kisi spatiotemporal groundwater quality parameters. Earth Sci Inf
O (2019) Daily streamflow prediction using optimally pruned 8:885–894
extreme learning machine. Journal of Hydrology 577:123981 Jang JSR (1993) ANFIS: adaptive-network-based fuzzy inference
Adnan, R. M., Liang, Z., Heddam, S., Zounemat-Kermani, M., Kisi, system. IEEE Trans Syst Man Cybern 23:665–685
O., & Li, B. (2020). Least square support vector machine and Jang JSR, Sun CT, Mizutani E (1997) Neuro-fuzzy and soft
multivariate adaptive regression splines for streamflow predic- computing: a computational approach to learning and machine
tion in mountainous basin using hydro-meteorological data as intelligence. IEEE Trans Autom Control 42(10):1482–1484
inputs. Journal of Hydrology, 586, 124371. Jothiprakash V, Magar RB, Kalkutki S (2009) Rainfall–runoff models
Aghelpour P, Varshavian V (2020) Evaluation of stochastic and using adaptive neuro–fuzzy inference system (ANFIS) for an
artificial intelligence models in modeling and predicting of river intermittent river. Int J Artif Intell 3:1–23
daily flow time series. Stoch Env Res Risk Assess 34(1):33–50 Kasiviswanathan KS, Sudheer KP (2013) Quantification of the
Aksoy H, Dahamsheh A (2009) Artificial neural network models for predictive uncertainty of artificial neural network based river
forecasting monthly precipitation in Jordan. Stoch Environ Res flow forecast models. Stoch Env Res Risk Assess 27(1):137–146
Risk Assess 23(7):917–931 Kasiviswanathan KS, Sudheer KP (2017) Methods used for quanti-
Alizadeh MJ, Kavianpour MR, Kisi O, Nourani V (2017) A new fying the prediction uncertainty of artificial neural network based
approach for simulating and forecasting the rainfall-runoff hydrologic models. Stoch Env Res Risk Assess 31(7):1659–1670
process within the next two months. J Hydrol 548:588–597 Kennedy J, Eberhart R, (1948) Particle swarm optimization. In:
Al-Sudani ZA, Salih SQ, Yaseen ZM (2019) Development of Proceedings of the IEEE international conference on neural
multivariate adaptive regression spline integrated with differen- networks Vol 4, 1942–1948
tial evoltion model for streamflow simulation. J Hydrol 573:1–12 Kisi O (2007) Streamflow forecasting using different artificial neural
Antar MA, Elassiouti I, Alam MN (2006) Rainfall–runoff modeling network algorithms. J Hydrol Eng 12(5):532–539
using artificial neural networks technique: a Blue Nile catchment Kisi O (2008) River flow forecasting and estimation using different
case study. Hydrol Process 20(5):1201–1216 artificial neural network techniques. Hydrol Res 39(1):27–40
Bashir ZA, El-Hawary ME (2009) Applying wavelets to short-term Kisi O, Parmar KS (2016) Application of least square support vector
load forecasting using PSO-based neural networks. IEEE Trans machine and multivariate adaptive regression spline models in
Power Syst 24:20–27 long term prediction of river water pollution. J Hydrol
Bezdek JC (1981) Pattern recognition with fuzzy objective function 534:104–112
algoritms. Plenum Press, New York Kisi O, Shiri J, Karimi S, Adnan RM (2018) Three different adaptive
Birikundavyi S, Labib R, Trung HT, Rousselle J (2002) Performance neuro fuzzy computing techniques for forecasting long-period
of neural networks in daily streamflow forecasting. J Hydrol Eng daily streamflows. In: Roy S, Samui P, Deo R, Ntalampiras S
7(5):392–398 (eds) Big data in engineering applications. Studies in Big Data,
Chen L, Singh VP, Guo S, Zhou J, Ye L (2014) Copula entropy vol 44. Springer, Singapore, pp 303–321. https://doi.org/10.
coupled with artificial neural network for rainfall–runoff simu- 1007/978-981-10-8476-8_15
lation. Stoch Env Res Risk Assess 28(7):1755–1767 Kisi O, Choubin B, Deo RC, Yaseen ZM (2019) Incorporating
Chen CS, Jhong YD, Wu WZ, Chen ST (2019) Fuzzy time series for synoptic-scale climate signals for streamflow modelling over the
real-time flood forecasting. Stoch Env Res Risk Assess Mediterranean region using machine learning models. Hydrol
33(3):645–656 Sci J 64(10):1240–1252
de Vos NJ, Rientjes THM (2005) Constraints of artificial neural Mamdani EH, Assilian S (1975) An experiment in linguistic synthesis
networks for rainfall–runoff modeling: trade-offs in hydrological with a fuzzy logic controller. Int J Man Mach Stud 7:1–13
state representation and model evaluation. Hydrol Earth Syst Sci Mehdizadeh S, Fathian F, Safari MJS, Adamowski JF (2019)
9:111–126 Comparative assessment of time series and artificial intelligence
Dunn JC (1973) A fuzzy Relative of the ISODATA process and its models to estimate monthly streamflow: a local and external data
use in detecting compact well-separated clusters. J Cybern analysis approach. J Hydrol 579:124225
3:32–57

123
Stochastic Environmental Research and Risk Assessment

Młyński D, Petroselli A, Wałe˛ga A (2018) Flood frequency analysis Rezaie-Balf M, Nagann SR, Kisi O, El-Shafie A (2019) Enhancing
by an event-based rainfall-runoff model in selected catchments streamflow forecasting using the augmenting ensemble proce-
of southern Poland. Soil Water Res 13:170–176 dure coupled machine learning models: case study of Aswan
Moeeni H, Bonakdari H (2017) Forecasting monthly inflow with High Dam. Hydrol Sci J 64(13):1629–1646
extreme seasonal variation using the hybrid SARIMA-ANN Sanikhani H, Kisi O (2012) River flow estimation and forecasting by
model. Stoch Env Res Risk Assess 31(8):1997–2010 using two different adaptive neuro-fuzzy approaches. Water
Nardi F, Grimaldi S, Santini M, Petroselli A, Ubertini L (2008) Resour Manage 26(6):1715–1729
Hydrogeomorphic properties of simulated drainage patterns Santini M, Grimaldi S, Nardi F, Petroselli A, Rulli MC (2009)
using DEMs: the flat area issue. Hydrol Sci J 53(6):1176–1193 Preprocessing algorithms and landslide modelling on remotely
Natural Resources Conservation Service (NRCS) (2008) Hydrology, sensed DEMs. Geomorphology 113(1–2):110–125
national engineering handbook. Washington DC: US Department Santos CAG, Srinivasan VS, Suzuki K, Watanabe M (2003)
of Agriculture, part 630 Application of an optimization technique to a physically based
Nayak PC, Sudheer KP, Rangan DM, Ramasastri KS (2004) A neuro- erosion model. Hydrol Process 17(5):989–1003. https://doi.org/
fuzzy computing technique for modeling hydrological time 10.1002/hyp.1176
series. J Hydrol 291(1–2):52–66 Santos CAG, Pinto LEM, Freire PKMM, Mishra SK (2010)
Nayak PC, Venkatesh B, Krishna B, Jain SK (2013) Rainfall-runoff Application of a particle swarm optimization to a physically-
modeling using conceptual, data driven, and wavelet based based erosion model. Wars Univ Life Sci—SGGW Ann Land
computing approach. J Hydrol 493:57–67 Reclam 42(1):39–49
Niedzielski T (2007) A data-based regional scale autoregressive Santos CAG, Freire PKMM, Mishra SK, Soares Júnior A (2011)
rainfall-runoff model: a study from the Odra River. Stoch Env Application of a particle swarm optimization to the tank model.
Res Risk Assess 21(6):649–664 IAHS Publ 347:114–120
Nourani V, Komasi M (2013) A geomorphology-based ANFIS model Senthil Kumar AR, Sudheer KP, Jain SK, Agarwal PK (2004)
for multi-station modeling of rainfall–runoff process. J Hydrol Rainfall–runoff modeling using artificial neural network: com-
490:41–55 parison of networks types. Hydrol Process 19(6):1277–1291
Nourani V, Komasi M, Mano A (2009) A multivariate ANN-wavelet Shiri J, Kisi O (2010) Short-term and long-term streamflow forecast-
approach for rainfall–runoff modeling. Water Resour Manage ing using a wavelet and neuro-fuzzy conjunction model. J Hydrol
23(14):2877 394(3–4):486–493
Nourani V, Elkiran G, Abdullahi Tahsin A (2019) Multi-region Singh P, Borah B (2013) Indian summer monsoon rainfall prediction
modeling of daily global solar radiation with artificial intelli- using artificial neural network. Stoch Env Res Risk Assess
gence ensemble. Nat Resour Res 28(4):1217–1238 27(7):1585–1599
Nourani V, Davanlou Tajbakhsh A, Molajou A, Gokcekus H (2019) Singh KK, Pal M, Singh VP (2010) Estimation of mean annual flood
Hybrid Wavelet-M5 Model tree for rainfall-runoff modeling. in Indian catchments using backpropagation neural network and
J Hydrol Eng 24(5):04019012 M5 model tree. Water Resour Manage 24(10):2007–2019
Pal M, Deswal S (2009) M5 model tree based modeling of reference Sun Y, Niu J, Sivakumar B (2019) A comparative study of models for
evapotranspiration. Hydrol Process Int J 23(10):1437–1443 short-term streamflow forecasting with emphasis on wavelet-
Papacharalampous G, Tyralis H, Koutsoyiannis D (2019) Comparison based approach. Stoch Env Res Risk Assess 33(10):1875–1891
of stochastic and machine learning methods for multi-step ahead Talei A, Chua LHC, Quek C (2010) A novel application of a neuro-
forecasting of hydrological processes. Stoch Env Res Risk fuzzy computational technique in event-based rainfall–runoff
Assess 33(2):481–514 modeling. Expert Syst Appl 37(12):7456–7468
Partal T, Cigizoglu HK, Kahya E (2015) Daily precipitation Tarboton DG, Bras RL, Rodriguez-Iturbe I (1991) On the extraction
predictions using three different wavelet neural network algo- of channel networks from digital elevation data. Hydrol Process
rithms by meteorological data. Stoch Env Res Risk Assess 5(1):81–100
29(5):1317–1329 Ünal NE, Aksoy H, Akar T (2004) Annual and monthly rainfall data
Petroselli A (2012) LIDAR data and hydrological applications at the generation schemes. Stoch Env Res Risk Assess 18(4):245–257
basin scale. GISci Remote Sens 49(1):139–162 Vojtek M, Petroselli A, Vojteková J, Ashgarynia S (2019) Flood
Petroselli A, Grimaldi S (2018) Design hydrograph estimation in inundation mapping in small and ungauged basins: sensitivity
small and fully ungauged basin: a preliminary assessment of the analysis using the EBA4SUB and HEC-RAS modeling
EBA4SUB framework. J Flood Risk Manage 11:197–210 approach. Hydrol Res 50(4):1002–1019
Petroselli A, Mulaomerović-Šeta A, Lozančić Ž (2019a) Comparison Wu CL, Chau KW (2011) Rainfall–runoff modeling using artificial
of methodologies for design peak discharge estimation in neural network coupled with singular spectrum analysis. J Hydrol
selected catchments of Bosnia and Herzegovina. 399(3–4):394–409
GRAÐEVINAR 71(9):729–739 Yaseen ZM, Kisi O, Demir V (2016) Enhancing long-term streamflow
Petroselli A, Vojtek M, Vojteková J (2019b) Flood mapping in small forecasting and predicting using periodicity data component:
ungauged basins: a comparison of different approaches for two application of artificial intelligence. Water Resour Manage
case studies in Slovakia. Hydrol Res 50(1):379–392 30(12):4125–4151
Petroselli A, Asgharinia S, Sabzevari T, Saghafian B (2020) Yin Z, Feng Q, Wen X, Deo RC, Yang L, Si J, He Z (2018) Design
Comparison of design peak flow estimation methods for and evaluation of SVR, MARS and M5Tree models for 1, 2 and
ungauged basins in Iran. Hydrol Sci J 65(1):127–137 3-day lead time forecasting of river flow data in a semiarid
Piscopia R, Petroselli A, Grimaldi S (2015) A software package for mountainous catchment. Stoch Env Res Risk Assess
the prediction of design flood hydrograph in small and ungauged 32(9):2457–2476
basins. J Agric Eng XLV I(432):74–84 Yuan X, Chen C, Lei X, Yuan Y, Adnan RM (2018) Monthly runoff
Remesan R, Shamim MA, Han D, Mathew J (2008) ANFIS and forecasting based on LSTM–ALO model. Stoch Env Res Risk
NNARX based rainfall-runoff modeling. In 2008 IEEE Interna- Assess 32(8):2199–2212
tional Conference on Systems, Man and Cybernetics (pp. Zhang Z, Zhang Q, Singh VP, Shi P (2018) River flow modelling:
1454–1459). IEEE comparison of performance and evaluation of uncertainty using

123
Stochastic Environmental Research and Risk Assessment

data-driven models and conceptual hydrological model. Stoch Publisher’s Note Springer Nature remains neutral with regard to
Env Res Risk Assess 32(9):2667–2682 jurisdictional claims in published maps and institutional affiliations.
Zhu S, Luo X, Yuan X, Xu Z (2020) An improved long short-term
memory network for streamflow forecasting in the upper
Yangtze River. Stoch Environ Res Assess 34:1–17

123

You might also like