Atmospheric Research 255 (2021) 105531

Contents lists available at ScienceDirect

Atmospheric Research
journal homepage: www.elsevier.com/locate/atmosres

Forecasting rainfall using transfer entropy coupled directed–weighted

complex networks
Hakan Tongal a, *, Bellie Sivakumar b
a
Department of Civil Engineering, Engineering Faculty, Süleyman Demirel University, 32260 Isparta, Turkey
b
Department of Civil Engineering, Indian Institute of Technology Bombay, Powai, Mumbai 400076, India

A R T I C L E I N F O A B S T R A C T

Keywords: In rainfall forecasting, selection of an appropriate gauging station to forecast and identification of the appro­
Rainfall forecasting priate explanatory variables are nontrivial tasks due to the complexity of the physical processes involved and
Directed–weighted complex network spatiotemporal variability. In this study, for the first time, the concept of transfer entropy is coupled with the
Transfer entropy
complex network analysis for rainfall forecasting by determining the nonlinear directional relationship between
Directional relationship
Artificial neural networks
the stations. The proposed methodology involves determining the directional relationship between the stations in
Turkey a given basin, and defining the most influenced station by using the node strength and directed clustering co­
efficient that are specifically developed for directed-weighted networks. Further, the obtained information flow
within the basin is utilized to produce current monthly forecasts of the most influenced station. The methodology
is implemented for rainfall forecasting in the Western Mediterranean Basin, Turkey, considering monthly total
rainfall data from seven stations. The results indicate that the proposed methodology is useful and effective in the
identification of the appropriate station to forecast and the relevant explanatory variables to serve as inputs for
the artificial neural networks (ANNs) model. The results from the proposed methodology are compared with
those from two widely employed input determination approaches (i.e. using rainfall from the most-correlated
station in the basin and using rainfall from all the other stations). It is found that the proposed methodology
significantly improves the forecasting performance of the ANN model. The results obtained in this study have
broad implications for designing optimal rain gauge density, identification of the complexity of the rain gauge
network structure, and interpolation (or extrapolation) of hydrological data for ungauged locations.

1. Introduction 2020; Shi et al., 2020); statistical models (Villarini and Serinaldi, 2012;
Zoglat et al., 2018; Qian et al., 2020); and machine learning models,
Rainfall is a key element of the hydrological cycle. It forms a sig­ such as artificial neural networks (Zahmatkesh and Goharian, 2018;
nificant input for a variety of applications in hydrology and water re­ Parviz and Rasouli, 2019; Hossain et al., 2020; Johny et al., 2020). In
sources, including streamflow modeling, flood forecasting, soil moisture recent years, applications of machine learning models for rainfall fore­
estimation, and sediment transport. Changing rainfall patterns, casting have been gaining momentum. Machine learning models are
including those associated with global warming, introduce significant efficient in modeling complex (hydro-meteorological) processes and
stress on surface water resources (Barnett et al., 2005). Therefore, nonlinear relationships between different sub-processes. Specifically,
adequate understanding of rainfall patterns and reliable rainfall fore­ artificial neural networks (ANNs) have received significant attention in
casts are important for proper planning and management of our water rainfall forecasting, due to their high nonlinearity, flexibility, and data-
resources. However, rainfall modeling and forecasting are highly chal­ driven learning without any prior knowledge about the underlying
lenging, due to the complexity of the processes involved and spatio­ dynamical processes (e.g., Aksoy and Dahamsheh, 2009; Hung et al.,
temporal variability (Hu et al., 2019). 2009; Abbot and Marohasy, 2014; Hamidi et al., 2015; Solgi et al., 2017;
There exists a variety of models for rainfall forecasting, including Dash et al., 2018; Xiang et al., 2018; Pham et al., 2020). Such studies
weather precipitation models and remote sensing observations (Yu have revealed that ANNs could successfully forecast rainfall amounts at
et al., 2017; Chakraborty et al., 2020; Chen et al., 2020; Newman et al., various timescales in advance in arid to semi-arid to humid areas.

* Corresponding author.
E-mail addresses: [email protected] (H. Tongal), [email protected] (B. Sivakumar).

https://doi.org/10.1016/j.atmosres.2021.105531
Received 3 October 2020; Received in revised form 31 January 2021; Accepted 16 February 2021
Available online 18 February 2021
0169-8095/© 2021 Elsevier B.V. All rights reserved.
H. Tongal and B. Sivakumar Atmospheric Research 255 (2021) 105531

However, as indicated by Winchell et al. (1998) and Smith et al. (2004), The objectives of this study are (1) to determine the most influenced
a basin is mainly characterized by the spatio-temporal variability of rainfall station in a basin by considering the information flow among the
rainfall. Therefore, the performance of models for rainfall forecasting is stations by analysing the transfer entropy coupled complex network, and
highly sensitive to the spatio-temporal variability of rainfall. (2) to develop a methodology for rainfall forecasting of the most influ­
Due to the complex nonlinear relationship of rainfall between the enced station by considering its transmitter stations. To these objectives,
different stations in a basin, selection of a station for rainfall forecasting monthly total rainfall data observed over the period 1979–2018 at seven
directly affects the rainfall forecasts and the resulting performance of the stations located in the Western Mediterranean Basin, Turkey are ana­
machine learning models. Therefore, a scientifically-sound selection of a lysed using the concepts of complex networks, especially node strength
station with particular consideration of the information flow among the and clustering coefficient. To obtain the directed-weighted network of
stations within a basin would be more useful than selecting a station the stations, the connections, directions, and weights are determined
arbitrarily. To this aim, by considering the information flow between the using the transfer entropy methodology. The rainfall network is ana­
stations, the dominant influencing stations as well as the stations that lysed with in- and out-clustering coefficient methods that have been
are influenced within the basin could be selected for various forecasting specifically developed for the directed-weighted networks. Further, the
purposes. For instance, if the forecasted values of the most influenced strength of the relations is analysed with in- and out-strength co­
station are sufficient, then the gauging station could be moved to a more efficients. These coefficients are especially useful for determining the
needed place having fewer gauge stations. Therefore, the information dominant influencing stations as well as the stations that are influenced
flow among the stations can be utilized to analyse the potential effects of within the basin, besides depicting the information flow between the
decreasing or increasing the density of available meteorological stations stations. The results of the complex network analysis are further utilized
that eventually could provide insights into potentially improve the to develop an ANN model to forecast the current monthly rainfall values
performance of hydrological models. of the most appropriate station. The explanatory variables for fore­
Another limitation in the above-mentioned studies is related to the casting the current monthly rainfall values are identified by considering
explanatory variables selected to forecast rainfall values at any selected the information flow among the stations.
station(s). Since, in general, the station for forecast is selected arbi­ The rest of this paper is organised as follows. Section 2 introduces the
trarily, the observed values of the other stations are assumed as concepts of directed–weighted complex networks, the transfer entropy,
explanatory variables. In most cases, to determine the significant and construction of ANN model including the methodology for selecting
explanatory variables, cross-correlation, autocorrelation, and mutual most relevant input variables. Section 3 presents the details of the study
information functions are used (Corzo and Solomatine, 2007; Chua and area and the rainfall data considered. The results and discussion are
Wong, 2010; Wu et al., 2010; Babel et al., 2015; Li et al., 2015). While presented in Section 4, followed by conclusions and implications in
these functions are useful for determining the degree of relationship Section 5.
between the variables, they often do not provide insight into the di­
rection of the relationship. Recently, transfer entropy has been demon­ 2. Methodology
strated to be a useful tool in quantifying the statistical coherence
between variables in dynamical systems (Schreiber, 2000; Marschinski 2.1. Complex networks
and Kantz, 2002; Jizba et al., 2012). Further, since transfer entropy is
quite efficient in quantifying the flow of information between two time A set of nodes (or vertices) interconnected with links (or edges) forms
series in a non-parametric and in an explicitly non-symmetric way (Jizba a network or a graph. Formally, a network is defined by an ordered pair
et al., 2012), it is useful for analysing non-linear complex systems where D = (V, A), where V = {V1 , V2 , …, VN } is a set of nodes and A is a set of
sub-processes are interlinked (Lee et al., 2012), as is the case in hydro­ links {i, j} that are 2-element subsets of V (Agarwal et al., 2018). Ac­
logical systems. Therefore, transfer entropy could be useful for depicting cording to the nature of the connections between the nodes, several
directionality, relative strength, and statistical significance of informa­ network typologies can be obtained, such as undirected–unweighted,
tion flow among the stations in rainfall forecasting studies. It can also directed–unweighted, undirected–weighted, and directed–weighted. In
play a particularly important role when used in other approaches that this study, the directed–weighted network is used, since it could provide
address the connections in systems in other ways, such as network better representation of the spatio-temporal variability of processes.
theory, which is the focus of this study. While there are several metrics, such as clustering coefficient, degree
A network or graph can be defined as a group or collection of com­ distribution, degree centrality, size and density, and average path length
ponents that are meaningfully related. The concepts of complex net­ and diameter for analysing undirected–unweighted networks, there are
works, which are recent developments in network theory, have been only a few metrics for analysing directed–weighted networks, such as
applied to several hydrological problems, such as determining connec­ node strength and directed clustering coefficients (Clemente and Grassi,
tions in streamflow networks (Sivakumar and Woldemeskel, 2014; 2018). Therefore, in this study, the node strength and clustering coef­
Halverson and Fleming, 2015; Han et al., 2018; Yasmin and Sivakumar, ficient are used to analyse the dynamics within the network.
2018; Rodríguez-Alarcón and Lozano, 2019), drought analysis (Kona­ For each link (i, j) in a network, a weight wij > 0 can be associated, so
pala and Mishra, 2017; Ciemer et al., 2018, and dynamics of rainfall that a weighted–directed network is obtained. The adjacency relation­
connections (Sivakumar and Woldemeskel, 2015; Jha and Sivakumar, ship between the nodes of the network D and the weights on the pairs (or
2017; Naufan et al., 2018). It is appropriate to note, at this point, that the arcs) can be described with the weighted adjacency matrix, W. The
networks considered in the above studies, especially those in rainfall weighted adjacency matrix is a non-negative real n-square matrix; for
studies, were undirected–unweighted complex networks. This type of the unweighted case, it can be denoted by A whose generic element aij =
networks, however, has some shortcomings, such as: (1) the selected 1 if and only if there is an edge connecting nodes i and j (i.e., if they are
correlation threshold value directly affects the structure of the network; neighbours) and zero otherwise. The in-degree (din i ) and out-degree
and (2) employing linear correlations may be insufficient to resolve
(douti ) of a node i can be defined as (Clemente and Grassi, 2018):
complex relationships among the rain gauge stations. These problems
∑
may be overcome, at least to a certain extent, by determining the di­ dini = aji = ATi U
rection of the connections or direction of information flow in the rainfall j∕
=i
∑ (1)
dynamics. To this end, the present study, for the first time to our dout = aij = Ai U
knowledge, proposes to use directed-weighted complex networks
i
j∕
=i
coupled with transfer entropy method to reveal the nature of informa­
tion flow within a basin for the purpose of forecasting rainfall. where U shows the unit column vector of n elements; and Ai and ATi are

2
H. Tongal and B. Sivakumar Atmospheric Research 255 (2021) 105531

the i-th row of A and of its transpose, respectively. For the weighted are selected as one in the interest of the computational cost involved
network, these definitions can be updated by considering the weights for (Ruddell and Kumar, 2009; Lee et al., 2012). The calculation of transfer
obtaining in-strength (sin
i ) and out-strength (si ) as:
out entropies from joint probability distributions is based on the dis­
∑ ( ) cretisation of the time series. Further, it should be noted that the transfer
sini = aji ⋅wji = AT W ii = WTi U entropy is explicitly non-symmetric (directional), (i.e.,
= TEY→X ), and if there is any information flow, then TEX→Y
j∕
=i
∑ ( ) (2) TEX→Y ∕
sout
i = aij ⋅wij = AWT ii = Wi U should be greater than zero. To check whether the calculated transfer
j∕
=i
entropy between two time series (i.e. rainfall at two stations in this case)
The total strength (stot is statistically significant, the effective transfer entropy is employed, as
i ) then can be given as the total of in- and out-
strengths. introduced by Marschinski and Kantz (2002).
The local clustering coefficient was formulated by Watts and Strogatz In this study, the significance level is adopted at 5%, and the arcs
(1998) as: with p-values lower than this level are assumed as statistically signifi­
cant. To eliminate any undesirable effects of finite data length, the
2|t(i) | A3ii effective transfer entropy is used. To this end, shuffling is repeated 100
Ci (A) = = (3)
di (di − 1) di (di − 1) times and the average of the resulting shuffled transfer entropy esti­
mates across all replications is subtracted from the transfer entropy.
where A3ii counts twice the number of triangles in which a node i par­ Therefore, the effective transfer entropy can be thought of as a bias-
ticipates (Clemente and Grassi, 2018). This local clustering coefficient is corrected transfer entropy. The statistical significance of transfer en­
the fraction of the number of triangles t(i) connected to node i and the tropy is calculated using the Markov block bootstrap method, as sug­
potential triangles. This coefficient is for undirected and unweighted gested by Dimpfl and Peter (2013). Contrary to shuffling, this method
networks. An extension of this formula was proposed by Clemente and preserves the dependencies within each time series but destroys the
Grassi (2018) to consider directed and weighted links. This general statistical dependency between X and Y. Under the null hypothesis of no
directed–weighted clustering coefficient (CClGr
i ) is calculated as: information transfer between the nodes, the distribution of the transfer
[( )( )2 ] entropy estimate is obtained by bootstrapping the samples 300 times (as
0.5 W + WT A + AT we find that bootstrapping the samples higher than 300 times makes no
CClGr
i (W) = ↔
ii
(4) difference to the results). The weights of the arcs are the values of the
stot tot
i (si − 1) − 2si
transfer entropy calculated for each node (i.e., rainfall station in this
where si↔ is the strength of the bilateral arcs given by: case) with this approach.
∑( ) ( )
si↔ = 0.5 aij wij + aji wji = 0.5 AWT + AT W ii (5)
j∕
=i
2.3. Artificial neural networks
In this study, two types of triangles in which node i (or rainfall sta­
tion) can take part are considered: in-clustering (Cini ) and out-clustering Artificial neural networks (ANNs) are mathematical structures hav­
(Cout ing interconnected neurons to map the complex non-linear input-output
i ). More details about these coefficients can be found in Clemente
and Grassi (2018). relationships. With these models, relationships among the input and
output variables involved in a physical process can be analysed without
considering the underlying physical processes. In recent decades, ANN
2.2. Transfer entropy
models have been increasingly applied to various hydrological processes
and problems (Badrzadeh et al., 2013; Nastos et al., 2014; Johny et al.,
Transfer entropy is a measure of information flow between two time
2020; Pham et al., 2020). The main reason behind this could be the
series that is transferred only between certain parts of probability dis­
simplicity of the model structure, its ease of applicability and robustness
tributions. Let p(x) and p(y) be the marginal probability density func­
in recognizing complex dynamical relationships between the input and
tions of two discrete time series X = {x1 , x2 , …, xN } and Y =
output.
{y1 , y2 , …, yN }, respectively. The average information contained in the
Multi-layer perceptron feed-forward neural networks (MLP-FFNNs)
source X about the next state of the destination Y that was not already
are one of the most widely employed and promising ANNs (Singh et al.,
contained in the destination’s past can be estimated from the transfer
2009; Lohani et al., 2011). In these networks, each neuron is connected
entropy (Sandoval, 2014). The transfer entropy from the variable X to
to all neurons in the subsequent layer, and the output of each layer is
the variable Y is calculated as:
taken as an input for the subsequent layer. It is a common practice to
( ⃒ )
⃒ (l) (k) have three layers (input, output, and one hidden layer) in such networks.
∑ ( ) p yi ⃒yi− t , xi− τ
TEX→Y =
(l) (k)
p yi , yi− t , xi− τ log2 ( ⃒ ) (6) As indicated by Hagan and Menhaj (1994) and Aqil et al. (2007b),
⃒ (l)
(l) (k)
yi ,y ,x p yi ⃒yi− t Newton or quasi-Newton optimization algorithms, such as Levenberg-
Marquart and Broyden-Fletcher-Goldfarb-Shanno algorithms, are fast
i− t i− τ

where i indicates a given point in time, and k and l are the block lengths in convergence time and being less easily trapped in local minima.
of past values in X and Y, respectively (Lee et al., 2012). The conditional Further, they have better generalization capabilities than the descent
probabilities in Eq. (6) are conditioned on the past values of xi− τ =
(k) gradient back-propagation algorithm (Aqil et al., 2007a) and provide
{ } enough complexity to capture the complex non-linear relationships in
{xi− τ− k+1 , xi− τ− k+2 , …, xi− τ } and yi− t = yi− t− l+1 , yi− t− l+2 , …, yi− t where
(l) (l) (l) (l)
hydrological processes (Tongal, 2020).
τ and t are the delay times. Several studies on rainfall forecasting have considered two hidden
There are no clear-cut guidelines as to the selection of the delay times layer neural networks. The main reason for this is that the rainfall
τand t. However, setting the delay times as one is generally considered to process is a complex non-linear process where several sub-processes play
be reasonable to a degree, since, in a Markovian process, the highest significant roles. However, as indicated by De Villiers and Barnard
information is obtained from the immediate history of the considered (1993), data constraints, such as only a limited number of available
variable (Ruddell and Kumar, 2009). Further, the block lengths k and l observations, could lead to poor performance when two hidden layers

3
H. Tongal and B. Sivakumar Atmospheric Research 255 (2021) 105531

are used when compared to its single layer counterpart. Further, having Further, to eliminate the variables that are redundant due to their
additional hidden layers with a considerably high number of neurons inherent correlations, Min-Redundancy (R) is introduced as:
could lead to over-fitting that decreases the generalizability of the
1 ∑ ( )
network. Therefore, the three-layer feed-forward neural network (i.e. R= I xi ; xj (9)
|S|2 xi ,xj ∈S
only one input, hidden, and output layers) is used in this study. As
indicated by Zadeh et al. (2010) and Maier and Dandy (1998), the hy­ The mRMR combines the above two constraints and maximizes the
perbolic tangent function provides better predictions than the logistic following condition by optimizing D and R simultaneously as:
sigmoid function and enables faster training of the network. Further, a
combination of a hyperbolic tangent function and a linear function in maxϕ(D, R), ϕ = D − R (10)
the hidden and output layers, respectively, is effective when extrapo­
lating beyond the range of training data (Maier and Dandy, 2000). Thus,
2.5. Description of the proposed forecasting methodology
this combination is utilized in this study.
In this study, to improve monthly forecasting of rainfall, transfer
2.4. Minimum redundancy maximum relevance for feature selection entropy coupled complex network analysis is proposed (Fig. 1). In this
proposed methodology, it is aimed to improve the current monthly
Feature selection is the process of obtaining a subset from the orig­ forecasts of the most influenced station in the basin. To this aim, the
inal feature set by selecting relevant features of the dataset and most influenced station in the basin is determined from the transfer
removing the irrelevant and redundant ones. While feature selection entropy coupled complex network analysis using in-clustering, out-
methods can be categorized into supervised, semi-supervised, and un­ clustering, in-strength, and out-strength values. The first-degree trans­
supervised, the supervised feature selection method has been widely mitter stations are the source stations of information flow to the most
employed for regression problems. In the supervised feature selection, a influenced station. Similarly, the second-degree transmitter stations are
subset of features is selected by measuring the importance and relevance the source stations of information flow to the first-degree transmitter
of the features using a target class. stations. To utilize the information flow, the current monthly rainfall
Numerous supervised feature selection methods have been devel­ values of the first-degree transmitter stations are predicted from the
oped, including robust feature selection, discriminative least squares second-degree transmitter stations. The most significant variables are
regression, global redundancy minimization, ReliefF, and correlation- determined from the mRMR analysis. To eliminate any information loss,
based feature selection. However, the methods that employ mutual in­ each station’s significant lagged values are determined individually. The
formation criterion have been reported to yield promising results. The outputs of the optimal ANN models constructed for the current monthly
mutual information of two variables is a measure of information ob­ rainfall at the first-degree transmitter stations are aggregated with a
tained about one variable by observing the other variable using proba­ combiner ANN model to capture different rainfall dynamics. Then, the
bility distributions. Therefore, the mutual information can be used to current monthly forecasts of rainfall values at the first-degree trans­
quantify the non-linear dependency between the variables, which, in mitter stations are joined to the significant lagged values of these sta­
general, can be multi-dimensional (Bugata and Drotar, 2020). Proposed tions. These are then utilized to obtain the current monthly forecasts of
by Ding and Peng (2005), the minimum redundancy maximum rele­ the most influenced station.
vance (mRMR), which employs mutual information as a dependency
measure, is one of the most widely employed feature selection methods. 2.6. Performance indices
The mRMR method is a powerful method for selecting a subset of fea­
tures from the feature space (Abdourahamane et al., 2019). It selects the In this study, four performance indices, i.e. root mean square error
optimal features by considering both minimum redundancy and (RMSE), Nash-Sutcliffe efficiency (NSE), Kling-Gupta efficiency (KGE),
maximum relevance with feature class labels among features using and volumetric error (VE), are used to evaluate the model performance.
maximal statistical dependency based on mutual information (Bugata These indices are widely used in hydrological studies, due to their reli­
and Drotar, 2020). Let X and Y be two discrete variables. Then, their ability and easy interpretability. Further, they are recommended for
mutual information is defined as: evaluating the generalizability and reliability of hydrological fore­
∑∑ (
p(x, y)
) casting models (Legates and McCabe, 1999; Wang et al., 2009; Patel and
I(X; Y) = p(x, y)ln (7) Ramachandran, 2015; Doycheva et al., 2017). The four performance
p(x)p(y)
y∈Y x∈X
indices are briefly described below.
The RMSE is one of the most widely employed criteria in assessing
where p(x) and p(y) are the marginal probabilities of x and y, respec­
model efficiency. Since, this criterion is sensitive to large forecast errors
tively, and p(x, y) is the joint probability. Feature selection aims to find a
(i.e., errors amplified with square), it is biased towards errors in the
feature subset S with m features {xi }, which jointly have the largest
forecasting of high values (Mekanik et al., 2016). As indicated by Das
dependency on the target class c, called as ‘Max-Dependency.’ However,
and Nanduri (2018), the NSE (Nash and Sutcliffe, 1970) assesses the
in the multivariable case, the calculation of mutual information becomes
model’s capability to mimic the output data away from the observed
complex (and sometimes even unsolvable) due to the requirement of
mean value, and so it is commonly used in the evaluation of perfor­
calculation of multivariate density functions. Ding and Peng (2005)
mances of hydrological models. To overcome a potential systematic
showed that the Maximum-Relevance criterion could be useful for
underestimation of peaks and variability in the NSE, Gupta et al. (2009)
approximating Max-Dependency with the mean value of all mutual in­
proposed the Kling-Gupta efficiency (KGE), where correlation, bias, and
formation values between individual features xi in the current feature
variability are considered jointly. Further, to eliminate some problems
subset S and the target class c (Cai et al., 2019). The Max-Dependency
in the NSE, such as overemphasizing high values relative to other ob­
(D) is expressed as:
servations due to the squared deviations, the volumetric efficiency (VE)
1 ∑ was proposed by Criss and Winston (2008).
D= I(xi ; c) (8)
|S| xi ∈S The equations for calculation of the four performance indices are as
follows:

4
H. Tongal and B. Sivakumar Atmospheric Research 255 (2021) 105531

Fig. 1. Methodology flowchart for the developed forecasting methodology. Information flows from first degree transmitter stations to the most influenced station and
from second degree transmitter stations to first degree transmitter stations.

i. Root mean square error (RMSE): influential over the region. Further, the Mediterranean climate has ef­
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ fects over the inner regions to a degree, due to the perpendicular
√
√ ∑ N ( )2 mountains to the Aegean Sea. There are 10 natural lakes and 12 dams in
RMSE = √N− 1 Pi − P ̂i (11)
the basin. The dams in the basin are used mainly for domestic water
i=1
supply, power generation, irrigation, and flood mitigation.
The available rainfall data is comprised of time series of daily rainfall
measured at a number of stations distributed throughout the basin. In
ii. Nash-Sutcliffe efficiency (NSE): this study, seven rain gauge stations are selected, in such a way that they
are scattered throughout the basin to obtain representative basin-wide
∑N ( ̂
)2
information flow among the stations (Fig. 2). Further, these stations
i=1 Pi − P i
NSE = 1 − ∑ ( )2 (12) have long-term daily observations with high quality from 1979 to date.
N
i=1 Pi − P
These daily rainfall records are obtained from the Turkish State Mete­
orological Service in Turkey. The quality of data is firmly controlled
where N is the length of the data set, P is the mean of observed value, Pi before release. Table 1 offers some useful insights on the monthly total
rainfall statistics.
and P
̂ i are observed and forecasted values, respectively.
For the present analyses, monthly total rainfall values from the above
stations for the period 1979–2018 are used. The monthly scale is far
iii. Kling-Gupta efficiency (KGE):
more appropriate than daily or annual scale for medium-term (from few
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
years to few decades) water resources planning and management
KGE = 1 − (r − 1)2 + (α − 1)2 + (β − 1)2 (13)
(Sivakumar and Woldemeskel, 2015). Therefore, in the present study,
the ideas of complex networks are applied at a monthly scale to reveal
where r is the correlation coefficient, α represents variability and is the
the spatial dynamic characteristics of rainfall for its forecasting.
ratio of the standard deviation of simulated and observed time series,
It is appropriate to note, at this point, that several other hydro-
and β is the bias and obtained by dividing the mean of simulated time
climatic variables could also be included in the analyses. However, in
series by the mean of observed time series.
the present study, the focus is on the proposed methodology, and so it is
decided to keep the analyses simple with only a single meteorological
iv. Volumetric efficiency (VE):
variable. With rainfall being the primary meteorological input for hy­
∑N ⃒⃒ ̂ ⃒
⃒ drological studies, only rainfall data are considered, as has been done in
i=1 ⃒ P i − Pi ⃒
VE = 1 − ∑N (14) many hydrological studies (Lohani et al., 2011; Wu and Chau, 2011; Deo
i=1 Pi and Şahin, 2016). Further, the main aim of this study is to investigate
RMSE measures the goodness-of-fit relevant to high flow values. whether the proposed methodology improves the rainfall forecasts,
While NSE values between 0 and 1 show an acceptable model perfor­ rather than to obtain the optimal forecasting model. Therefore,
mance, an NSE value lower than zero means that the mean value of the including other hydro-climatic variables would significantly complicate
series is a better estimation than the constructed model (Moriasi et al., the analysis and interpretations, since one cannot be sure whether the
2007). Similar to NSE, the KGE ranges between − ∞ and 1 (perfect fit). differences in the results actually come from the developed methodology
The VE varies between − ∞ and 1, with the latter representing the per­ or the input structure. It should also be noted that using spatially uni­
fect fit. If the VE is less than zero, the model is no better than a predictor formly and densely distributed stations are not comparative for the
using zeros (Cheng et al., 2016). proposed methodology. To reveal the most dominant or most influenced
station, there should be at least three stations that are linearly and/or
3. Study area and data nonlinearly connected in a basin for the proposed methodology. While it
is possible to use gridded data set, additional uncertainties due to
The Western Mediterranean Basin is located in the southwestern part downscaling or extrapolation could be introduced into the relationships
of Turkey. The basin area is about 20,953 km2, which is about 2.7% of between the stations. Finally, at-site measurements still play important
the total area of Turkey. Since the basin is located near to the Medi­ roles in meteorological observations and hydrological studies (e.g.,
terranean Sea and the Aegean Sea, the Mediterranean climate is there are over 800 rainfall stations in the Lake Eyre Basin in central

5
H. Tongal and B. Sivakumar Atmospheric Research 255 (2021) 105531

Fig. 2. Locations of the seven rainfall stations across the Western Mediterranean Basin considered in this study with the spatial distribution of mean annual
precipitation.

Table 1
Descriptive statistics of monthly total rainfall data from seven selected rain gauge stations in the Western Mediterranean Basin for the period 1979–2018.
Station Station No.* Mean (mm) Std. Dev. (mm) Min. (mm) Max. (mm) Skewness Kurtosis Coefficient of Variation Percentage of Zeros (%)

Elmalı–17,952 St-01 40.08 42.79 0 354.1 2.22 8.21 1.07 5.42

Milas–17,884 St-02 59.94 67.36 0 371.2 1.30 1.29 1.12 17.29
Finike–17,375 St-03 78.90 104.02 0 653.3 2.13 6.06 1.32 15.83
Datça–17,297 St-04 55.93 71.98 0 378.3 1.67 2.84 1.29 26.67
Köyceğiz–17,924 St-05 92.01 108.84 0 624.8 1.56 2.54 1.18 15.63
Fethiye–17,296 St-06 70.52 87.30 0 474.5 1.59 2.36 1.24 19.58
Acıpayam–17,890 St-07 44.51 44.84 0 327.6 1.76 4.66 1.01 4.79
*
The station number is taken from Fig. 2, and a similar format is used in the other parts of the text.

Australia). Therefore, observations from gauge stations are utilized for mean values (i.e. coefficient of variation), the stations in the coastal
the present study. regions (i.e. the Finike, Datça, and Fethiye stations) have higher un­
The coefficient of variation is calculated by dividing the standard certainty around the mean values (see Fig. 2 for the locations of the
deviation with the mean value. Based on the statistics presented in stations). Similarly, the northern and inner regions (i.e. the Elmalı and
Table 1, it may be inferred that while the northern and inner regions (i.e. Acıpayam stations) have lower percentage of zeros, while the other re­
the Acıpayam and Elmalı stations) have lower uncertainty around the gions have relatively higher percentage of zeros (about 3 and 5 times

Fig. 3. Complex network structure for monthly rainfall dynamics in the Western Mediterranean Basin. The weights obtained from the transfer entropy analysis show
the strength of directional relationships between the stations.

6
H. Tongal and B. Sivakumar Atmospheric Research 255 (2021) 105531

higher). This indicates that the Mediterranean climate is more influen­ stations to form a cluster. For instance, in Fig. 4 (a), the highest in-
tial over the coastal regions than over the inner and northern regions of clustering coefficient (1.0) is obtained for Station 2. However, as indi­
the basin. cated earlier, Station 2 receives the lowest number of in-coming con­
nections with just two connections, i.e. from Stations 3 and 4. Since
4. Results and discussion there are also connections between Stations 3 and 4, these form a cluster
where significant information flow exists between the nodes. On the
4.1. Complex network of rainfall in the Western Mediterranean Basin other hand, while Station 3 has four in-coming connections (i.e. from
Stations 2, 4, 6, and 7), it has the lowest in-clustering value (0.42). This
The complex network structure of rainfall with the weights obtained is because, Stations 2, 4, 6, and 7 do not form a closed cluster among
from the transfer entropy analysis in the Western Mediterranean Basin is themselves. There are mutual connections between Station 2 and Station
shown in Fig. 3. The network indicates that information flow involves 4 and between Station 4 and Station 7. Station 6 receives information
every station in the basin. However, it is clear that while some stations only from Station 4. This can be seen from Fig. 3 and Table 2. As a result,
have strong connections with other stations, some others have weak there is no continuous information flow between the connected stations
connections with others. For instance, the strongest connection is from that are linked to Station 3.
Station 4 (Datça) to Station 6 (Fethiye) (0.396 bits), and the weakest The out-clustering value is the highest for Stations 1 and 7 (1.0)
connection is from Station 1 (Elmalı) to Station 7 (Acıpayam) (0.222 (Fig. 4), while the lowest values are obtained for Stations 4 and 6 (0.49
bits). This finding is interesting, since geographically nearby stations (e. and 0.50, respectively). There are out-going connections from Station 1
g. Stations 5 and 6, or Stations 1 and 3) do not always have the strongest to Stations 4 and 7, and there is a mutual connection between Stations 4
connections. Further, the farthest stations, such as Stations 2 and 3, do and 7. Similarly, for Station 7, there are out-going connections to Sta­
not have the weakest connection. Therefore, transfer entropy could be tions 3 and 4, and there is a bilateral connection between Stations 3 and
useful for depicting significant information flow based on probability 4. Therefore, the out-clustering coefficient is the highest (1.0) for Sta­
distributions. tions 1 and 7. On the other hand, Stations 4 and 6 have the lowest out-
As can be seen from Fig. 3, the number of in-coming and out-going clustering values. As indicated earlier, Station 4 has the highest number
connections also differs for the stations. For instance, the highest num­ of in-coming and out-going connections. However, as shown by the out-
ber of out-going connections is observed for Stations 3 and 4 with six clustering coefficient, there are fewer clusters between the stations on
connections each, and the lowest number of out-going connections is which Station 4 has an influence. Station 6 has influence on Stations 1, 3,
observed for Stations 1 and 7. It is relevant to note that Stations 3 and 4 and 5, and it receives information from Stations 3, 4, and 5. While
are located in the coastal region of the basin and, seemingly, these are Station 6 has the second lowest out-clustering value (0.50), it has the
significantly influential over the region. However, Stations 1 and 7 are second-highest in-clustering value (0.84). This can be an indication that
located in the inner regions and far away from the coast. The highest while Station 6 has less impact on a group of stations that have relations
number of in-coming connections is observed for Stations 1 and 4 with with it, it is more influenced by a group of stations that have in­
five connections each, and the lowest number of in-coming connections terconnections. In addition to the clusters formed between the stations,
is observed for Station 2, with just two connections (notably Station 2 another important criterion for the information flow is the strength of
receives significant information from Station 3, the farthest station from the connections.
Station 2). Interestingly, Station 4 has the highest number of both in- As shown in Fig. 5, Station 4 has the highest in- and out-strength
coming and out-going connections, indicating a reciprocal relationship values. While Station 1 has the second-highest in-strength value, its
between this and the other stations. Being located in the coastal region, out-strength value is the lowest among the stations. Therefore, it can be
it is tightly connected to the other stations in the basin. However, this concluded that Station 1 is one of the most influenced stations in the
station is not only influential over the region but also influenced by the basin. However, Station 1 has the highest out-clustering value. This in­
other stations. While Station 1 has the highest number of in-coming dicates that the out-strength and out-clustering coefficients show
connections, it has the lowest number of out-going connections. It is different aspects of the information flow between the stations. While the
clear that Station 1 is one of the most influenced stations and having out-strength indicates the connection strength, the out-clustering coef­
fewer connections with the other stations in the basin. ficient gives an insight into the nature of information flow.
The different numbers of in-coming and out-going connections for
the stations indicate that while some stations are more dominant (or 4.2. Construction of the input structure of ANN model of Station 6
transmitter), the others could be more affected (or receiver). To un­ (Fethiye)
derstand the behaviour of clustering between the stations, in- and out-
clustering coefficients could be useful. The in- and out-clustering co­ To demonstrate the usefulness of the constructed complex network
efficients calculated for the seven stations are shown in Fig. 4. for improving the rainfall forecasts, it is important to select a proper
From Fig. 4, it can be seen that the stations have different (and station in the basin. This is because, randomly selecting a station whose
significantly different in some cases) in- and out-clustering coefficients. values are to be forecasted could be inefficient, especially when there
It should be noted that these coefficients reflect the inclination of the exists a complex relationship between the stations. For instance, since

Fig. 4. (a) In-clustering, and (b) out-clustering coefficients of seven rainfall stations in the Western Mediterranean Basin.

7
H. Tongal and B. Sivakumar Atmospheric Research 255 (2021) 105531

Table 2
Established connections with the complex network analysis and their weights between the stations.
Source Target Weight (bits) Source Target Weight (bits) Source Target Weight (bits)

St-01 St-04 0.283 St-03 St-06 0.339 St-05 St-04 0.388

St-01 St-07 0.222 St-03 St-07 0.374 St-05 St-06 0.356
St-02 St-01 0.316 St-04 St-01 0.286 St-05 St-07 0.329
St-02 St-03 0.318 St-04 St-02 0.375 St-06 St-01 0.264
St-02 St-04 0.362 St-04 St-03 0.332 St-06 St-03 0.323
St-03 St-01 0.291 St-04 St-05 0.373 St-06 St-05 0.355
St-03 St-02 0.349 St-04 St-06 0.396 St-07 St-03 0.292
St-03 St-04 0.341 St-04 St-07 0.371 St-07 St-04 0.321
St-03 St-05 0.334 St-05 St-01 0.309

Fig. 5. Normalized (a) in-strength and (b) out-strength values of the seven stations in the Western Mediterranean Basin.

Station 4 has the highest number of in-coming and out-going connec­ information. However, in the second input structure, the optimal num­
tions, it is an indication of a complex and reciprocal relationship where ber of input variables is determined using the mRMR method from all
forecasting of rainfall values for this station could be difficult. Therefore, stations in the basin. The mRMR method can significantly reduce the
rather than arbitrarily selecting a station, the network structure ob­ redundancy by eliminating irrelevant variables and, therefore, all the
tained, such as the above, could be used for revealing the relationships. significant information in the basin could be utilized to forecast the
In particular, a station that has a relatively low out-clustering value and rainfall without any information loss or redundancy.
relatively high in-clustering value with high in-strength connections In the proposed methodology, in addition to the statistically signif­
could be a candidate to this end. This is because, the in-clustering and in- icant lagged observations of the stations connected to Station 6 (i.e.,
strength coefficients could be an indication of an affected station more Stations 3, 4, and 5), the forecasts in the current month of Stations 3, 4,
than a dominant one. Therefore, Station 6 is selected here for the and 5 are considered. In this method, the transmitter stations of Stations
analysis, since it has the second lowest out-clustering value (it is 0.50, 3, 4, and 5 are used for forecasting rainfall in the current month for these
and the lowest out-clustering value is 0.49) while having the second- stations. For instance, the current monthly rainfall of Station 3 is fore­
highest in-clustering value with relatively strong connections (Fig. 4). casted by using Station 2, Station 4, and Station 7, separately. Since each
Station 6 receives information from Stations 3, 4, and 5 (Fig. 3). forecast represents different aspects of rainfall dynamics between the
Therefore, the rainfall forecasting model of Station 6 should include the stations, the current monthly forecasts are utilized through another
observations from these three stations. In this study, we develop a combiner ANN model to obtain a more representative single forecast of
forecasting methodology that considers the information flow between current monthly rainfall of Station 3. This combiner system improves the
the stations. To evaluate the performance of the methodology, the re­ forecasts of the connected stations when the base models produce a
sults are also compared with those obtained from two different input certain output combination. Finally, the current monthly forecasts of
structures. In the first conventional input structure, only the highest- rainfall values at Stations 3, 4, and 5 are joined to the significant lagged
correlated station with Station 6 is considered, i.e., Station 5. In the values of these stations (Table 3) to forecast the current monthly rainfall
second conventional approach, all the other stations are used for fore­ at Station 6. The data set is divided into two parts, i.e. training and
casting. The Pearson correlation coefficient in the monthly total rainfall testing data sets. About the first 80% of the data is used for training the
values between Station 5 and Station 6 is 0.89. It should be noted that models, while the remaining 20% is used for testing.
the proximity of the stations could result in a linear correlation between Table 3 presents the significantly-correlated lagged variables of other
the precipitation values of the stations. For instance, the Pearson cor­ stations to Station 6, obtained from the mRMR analysis.
relation coefficient between the monthly total rainfall values from the In Table 3, the lagged values obtained are ordered from left to right
adjacent stations 1 and 3 is 0.84. However, in the transfer entropy by their significance. In the conventional input structures, the variables
analysis, the probability distributions of the rainfall values of different in Table 3 are used to obtain the current monthly rainfall value for
stations are taken into account. The obtained transfer entropy value Station 6. In the context of the methodology developed in this study, the
between the stations 1 and 3 is 0.291 bits, which is the fifth-lowest value current forecasts of Stations 3, 4, and 5 (i.e., the transmitter stations of
among the 26 values in Table 2. Therefore, it can be construed that while
the proximity of the stations could lead to higher linear correlations, the Table 3
stations far from each other can also be related (e.g. stations 4 and 7). Most important features of other stations related to the current monthly rainfall
Specifically, as the topography and climate manifest themselves in the of Station 6 (Fethiye) determined by the mRMR analysis.
probability distribution of rainfall to a certain degree, the stations far
Station Correlated variables with Station 6
from each other can form triangles that information (or influences/af­
Station 1 Pt-12 Pt-5 Pt-1 Pt-11 Pt-6 Pt-4 Pt-10 Pt-8 Pt-7
fects) flows within.
Station 2 Pt-12 Pt-4 Pt-1 Pt-11 Pt-6 Pt-5 Pt-7
We are aware that these two input structures reflect two extremes Station 3 Pt-12 Pt-4 Pt-2 Pt-6 Pt-11 Pt-7
and there are also several methods that can be used to optimize the Station 4 Pt-12 Pt-1 Pt-5 Pt-7
number of stations. While the first input structure could lead to some Station 5 Pt-12 Pt-9 Pt-6 Pt-1 Pt-5 Pt-7
information loss, the second approach could lead to some redundancy of Station 7 Pt-12 Pt-4 Pt-6 Pt-1 Pt-11 Pt-5 Pt-10 Pt-7 Pt-9

8
H. Tongal and B. Sivakumar Atmospheric Research 255 (2021) 105531

Station 6) are obtained from their transmitter stations. It should be noted around 100 mm. Both approaches fail in capturing the peak (or high)
that Stations 3, 4, and 5 are determined by the complex network anal­ values, i.e. any value above 200 mm, but the first approach gives better
ysis. To show the usefulness of the proposed methodology, the observed results than the second approach. This is reasonable, since, in the second
values of Station 6 are not considered in the constructed model struc­ approach, considering basin-wide stations could decrease the forecasted
tures. Further, rather than finding the optimal features by using all peak values (because of the ‘averaging’ effect, leading to underestima­
lagged values, each station’s significant lagged values are determined tion). Further, using basin-wide observations lead to more scattered
individually to avoid information loss. values around the 45-degree line (especially with greater underestima­
tion) by introducing more uncertainty. As can be seen from the right
4.3. Forecasting performances of different input structures panels of Fig. 6 (a) and Fig. 6 (b), the fitted red line for the first approach
is closer to the 45-degree line.
4.3.1. Forecasting rainfall at Station 6 (Fethiye) with conventional input
structures 4.3.2. Forecasting rainfall at Station 6 (Fethiye) with input structure
In the conventional model structures for forecasting rainfall at Sta­ obtained from complex network
tion 6, two scenarios are considered: (1) only the most correlated station With the complex network-based analysis, the current monthly
with Station 6 (i.e., Station 5); and (2) all the other stations in the basin. rainfall value at Station 6 is obtained with its predictor (or transmitter)
The number of hidden neurons is determined by a trial-and-error pro­ stations (i.e., Stations 3, 4 and 5). To utilize the information flow, the
cedure by varying the number of hidden neurons from 2 to 10. As a current monthly rainfall values of Stations 3, 4, and 5 are predicted from
result, the obtained model structures are 6-9-1 and 41-4-1, respectively. their transmitter stations (Fig. 7). Then, the forecasts obtained for the
The resulting performance indices for these approaches for the current monthly rainfall of Stations 3, 4, and 5 are used to forecast the
testing period are as follows: with only Station 5, RMSE is 54.14 (mm), current monthly rainfall at Station 6.
NSE is 0.52, KGE is 0.72, and VE is 0.41. With all the other stations As can be seen from Fig. 7, Station 3 receives information from
RMSE is 50.76 (mm), NSE is 0.58, KGE is 0.70, and VE is 0.46. It is clear Stations 2, 4, and 7 (Fig. 7 (a)), Station 4 receives information from
that using all the other stations gives better results (except the KGE Stations 1, 2, 3, 5, and 7 (Fig. 7 (b)), and Station 5 receives information
value) than using only the single most correlated station. However, from Stations 3 and 4 (Fig. 7 (c)). The correlated variables of Stations 3,
interestingly, the difference between the two approaches is very small. 4, and 5 with their transmitter stations are determined with the mRMR
Using 41 input variables from all the other stations does not significantly analysis (not shown). It should be noted that rather than finding the
improve the model performance when compared to using only six input optimal features by using all lagged values of all transmitter stations (e.
variables. The main reason for this could be the complex relationship g., Stations 2, 4, and 7 for Station 3), each station’s significant lagged
between the stations and the non-linear nature of the rainfall process. values are determined individually to avoid information loss.
The monthly rainfall forecasts obtained from the two approaches are The outputs of the optimal ANN models constructed for the current
shown in Fig. 6. monthly rainfall at Stations 3, 4, and 5 are aggregated with a combiner
As can be seen from Fig. 6, both approaches yield an average per­ ANN model to capture different rainfall characteristics of the stations
formance in forecasting monthly rainfall at Station 6. Several strengths that reflect the local features of topography and climate. Subsequently,
and weaknesses of the two approaches can be seen from this figure. For the current monthly forecasts of rainfall values at Stations 3, 4, and 5 are
instance, the probability density obtained using the first approach is joined to the significant lagged values of these stations (
more similar to the observed density (Fig. 6 (a), right panel). However, Table 3) to forecast the current monthly rainfall at Station 6. To
the second approach is more successful in forecasting values up to avoid over-fitting and simplify the process, we do not use another

Fig. 6. Monthly rainfall forecasts of Station 6 by using the observations from (a) the most correlated station (i.e. Station 5); and (b) all the other stations.

9
H. Tongal and B. Sivakumar Atmospheric Research 255 (2021) 105531

Fig. 7. Information flow of transmitter stations of Station 6 from other stations to (a) Station 3; (b) Station 4; and (c) Station 5.

10
H. Tongal and B. Sivakumar Atmospheric Research 255 (2021) 105531

Fig. 8. Monthly rainfall forecasts obtained for Station 6 using the proposed methodology that considers information flow between the stations.

aggregator ANN model between Station 6 and Stations 3, 4, and 5. The directed–weighted complex network of rainfall monitoring stations in a
monthly rainfall forecasts obtained with the proposed methodology for basin for rainfall forecasting, with implementation in the Western
Station 6 are given in Fig. 8. Mediterranean Basin, Turkey. The results indicated that the geographic
The monthly rainfall forecasts obtained indicate that the proposed proximity between the stations does not necessarily imply a stronger
methodology gives noticeably better results than the two conventional relationship. For instance, it was found that neither the nearest stations
input structures employed in this study for the purpose of comparison. have the strongest connections nor the farthest stations have the weakest
The forecasts from the proposed methodology follow the patterns and connections. Therefore, the complex network analysis coupled with the
changes in the observed values closely, and there is no significant dif­ transfer entropy method is useful for determining the nonlinear direc­
ference between the observed and forecasted values. Interestingly, tional relationship between the stations.
contrary to the conventional approaches, the proposed methodology In rainfall studies, it is a common practice to arbitrarily select a
yields similar, or even higher, rainfall values than the observed peak station for forecasting. However, as the results from the proposed
values. Furthermore, the probability density obtained is closer to the complex networks-based methodology in this study indicate, several
observed density. The forecasted values are less scattered around the 45- stations could be more dominant, while some others could be more
degree line, which indicates lower uncertainty in the forecasts. influenced. Therefore, the complex network approach is useful for
The performance indices obtained for this method are 38.39 mm, determining the nature of information flow among the stations and,
0.76, 0.86, and 0.61 for RMSE, NSE, KGE, and VE, respectively. The hence, in selecting an appropriate station for forecasting. Further, the
performance indices obtained indicate that the proposed methodology insights gained from the nature of information flow were utilized to
significantly improves the performance of rainfall forecasting at Station improve the performance of ANN model to forecast the current monthly
6. For instance, for the proposed methodology, there is about 29–24% total rainfall values. Comparison of the forecasting results from the
decrease in RMSE, 46–31% increase in NSE, 19–23% increase in KGE, networks-based approach with those from two widely employed con­
and 49–33% increase in VE, when compared to the first (i.e. using only ventional input structures (the most correlated station and all the other
Station 5) and second (i.e., using all the other stations) conventional stations in the basin) indicated that the proposed methodology signifi­
input structure, respectively. These improvements can be better seen cantly improved the performance of the ANN model.
from Fig. 9. It should be noted that, in this figure, NRMSE indicates the It is worth noting that the proposed methodology can also be used to
normalized RMSE values, i.e. RMSE values are divided by the highest investigate the spatial connections in rainfall at different temporal scales
RMSE value. and network sizes. It is indeed possible to unravel the hidden causal
structure at different temporal scales. Specifically, establishing spatial
5. Conclusions and future directions of research connections among the stations across different temporal scales has
certain implications for scaling-based studies and regionalization
Identifying and utilizing information on the non-linear relationships studies. For instance, to reveal the temporal dynamics within a basin,
in rainfall dynamics between monitoring stations in a basin is crucial for instead of using the annual mean or accumulated data, each year can be
various hydrological purposes, such as rainfall forecasting, rainfall data considered as a node that consists of daily observations (see Jha and
interpolation, catchment classification, and predictions in ungauged Sivakumar, 2017; Han et al., 2018). Rather than using a pre-defined
stations. For the first time, this study examined the usefulness of the threshold value to establish a connection between the nodes, the pro­
posed methodology provides statistically significant directed relation­
ships among the nodes. The proposed methodology can also be useful in
deciding on the installation of any new monitoring stations or removal
of any existing monitoring stations in a region. For instance, additional
monitoring stations will not offer any significant benefits for the regions
that have already high in-clustering values. Indeed, removing one or
more monitoring stations from such a region to the regions with low in-
clustering values can be more beneficial. It is also possible to improve
drought forecasts in a station with high in-clustering and in-strength
coefficients with the help of other stations that are connected to this
station due to information flow.
The results presented in this study are also useful for streamflow
forecasting. For instance, information flow from other stations to the
station of interest can be utilized to forecast streamflow or to fill the
missing values. The information flow could be useful in establishing
Fig. 9. Comparison of the performance indices of the rainfall forecasts based on
the input structures developed by considering rainfall values from the single
connections between streamflow or depending upon the situation (e.g.,
station (i.e. Station 5), all stations within the basin, and the stations determined purpose and problem), between catchment characteristics, hydro-
by the complex network analysis. climatic variables, and model parameters. Further, since these

11
H. Tongal and B. Sivakumar Atmospheric Research 255 (2021) 105531

connections differ in space, time, and space-time, the predictions in Ciemer, C., Boers, N., Barbosa, H.M., Kurths, J., Rammig, A., 2018. Temporal evolution
of the spatial covariability of rainfall in South America. Clim. Dyn. 51 (1–2),
ungauged basins (PUB) could be improved by utilizing the actual
371–382.
neighbours rather than just the geographically-nearest neighbours Clemente, G.P., Grassi, R., 2018. Directed clustering in weighted networks: a new
(Sivakumar and Woldemeskel, 2014). perspective. Chaos, Solitons Fractals 107, 26–38.
Finally, it is important to note that the present study attempted Corzo, G., Solomatine, D., 2007. Baseflow separation techniques for modular artificial
neural network modelling in flow forecasting. Hydrol. Sci. J. 52 (3), 491–507.
rainfall forecasting only based on rainfall values. Use of other hydro- Criss, R.E., Winston, W.E., 2008. Do nash values have value? Discussion and alternate
meteorologic variables (e.g. temperature, wind velocity, pressure, hu­ proposals. Hydrol. Process. 22 (14), 2723–2725.
midity) would certainly yield better outcomes, which deserves further Das, J., Nanduri, U.V., 2018. Assessment and evaluation of potential climate change
impact on monsoon flows using machine learning technique over Wainganga River
investigation. Further, in this study, ANN was preferred as a forecaster Basin, India. Hydrol. Sci. J. 63 (7), 1020–1046.
model, since it is widely used due to its simplicity and accuracy. How­ Dash, Y., Mishra, S.K., Panigrahi, B.K., 2018. Rainfall prediction for the Kerala State of
ever, other machine learning/data mining models, such as support India using artificial intelligence approaches. Comput. Electr. Eng. 70, 66–73.
De Villiers, J., Barnard, E., 1993. Backpropagation neural nets with one and two hidden
vector regression (SVR), random forest (RF), and chaos theory-based layers. Neural Netw. IEEE Trans. 4 (1), 136–141.
models, can also be used. These will be investigated in future studies. Deo, R.C., Şahin, M., 2016. An extreme learning machine model for the simulation of
monthly mean streamflow water level in eastern Queensland. Environ. Monit.
Assess. 188 (2), 90.
Dimpfl, T., Peter, F.J., 2013. Using transfer entropy to measure information flows
Declaration of Competing Interest between financial markets. Stud. Nonlinear Dyn. Econometr. 17 (1), 85–102.
Ding, C., Peng, H., 2005. Minimum redundancy feature selection from microarray gene
The authors declare that they have no known competing financial expression data. J. Bioinforma. Comput. Biol. 03 (02), 185–205.
Doycheva, K., Horn, G., Koch, C., Schumann, A., König, M., 2017. Assessment and
interests or personal relationships that could have appeared to influence weighting of meteorological ensemble forecast members based on supervised
the work reported in this paper. machine learning with application to runoff simulations and flood warning. Adv.
Eng. Inform. 33, 427–439.
Gupta, H.V., Kling, H., Yilmaz, K.K., Martinez, G.F., 2009. Decomposition of the mean
Acknowledgements squared error and nse performance criteria: implications for improving hydrological
modelling. J. Hydrol. 377 (1), 80–91.
Hagan, M.T., Menhaj, M.B., 1994. Training feedforward networks with the Marquardt
The first author would thank the Turkish State Meteorological Ser­
algorithm. Neural Netw. IEEE Trans. 5 (6), 989–993.
vice in Turkey for providing the rainfall data. Bellie Sivakumar ac­ Halverson, M.J., Fleming, S.W., 2015. Complex network theory, streamflow, and
knowledges the support from the IIT Bombay seed grant (RD/0519- hydrometric monitoring system design. Hydrol. Earth Syst. Sci. 19 (7), 3301–3318.
IRCCSH0-027). The authors thank the editor and two reviewers for their Hamidi, O., Poorolajal, J., Sadeghifar, M., Abbasi, H., Maryanaji, Z., Faridi, H.R.,
Tapak, L., 2015. A comparative study of support vector machines and artificial
constructive comments and useful suggestions on an earlier version of neural networks for predicting precipitation in Iran. Theor. Appl. Climatol. 119
the manuscript, which led to significant improvements to the quality (3–4), 723–731.
and presentation of the work. Han, X., Sivakumar, B., Woldmeskel, F.M., Guerra de Aguilar, M., 2018. Temporal
dynamics of streamflow: application of complex networks. Geosci. Lett. 5 (1), 10.
Hossain, I., Rasel, H., Imteaz, M.A., Mekanik, F., 2020. Long-term seasonal rainfall
References forecasting using linear and non-linear modelling approaches: a case study for
Western Australia. Meteorog. Atmos. Phys. 132 (1), 131–141.
Hu, J., Liu, Y., Sang, Y.-F., 2019. Precipitation complexity and its spatial difference in the
Abbot, J., Marohasy, J., 2014. Input selection and optimisation for monthly rainfall
Taihu Lake Basin, China. Entropy 21 (1), 48.
forecasting in Queensland, Australia, using artificial neural networks. Atmos. Res.
Hung, N.Q., Babel, M.S., Weesakul, S., Tripathi, N., 2009. An artificial neural network
138, 166–178.
model for rainfall forecasting in Bangkok, Thailand. Hydrol. Earth Syst. Sci. 13 (8).
Abdourahamane, Z.S., Acar, R., Serkan, Ş., 2019. Wavelet–Copula-based mutual
Jha, S.K., Sivakumar, B., 2017. Complex networks for rainfall modeling: spatial
information for rainfall forecasting applications. Hydrol. Process. 33 (7),
connections, temporal scale, and network size. J. Hydrol. 554, 482–489.
1127–1142.
Jizba, P., Kleinert, H., Shefaat, M., 2012. Rényi’s information transfer between financial
Agarwal, A., Marwan, N., Maheswaran, R., Merz, B., Kurths, J., 2018. Quantifying the
time series. Phys. A Stat. Mech. Appl. 391 (10), 2971–2989.
roles of single stations within homogeneous regions using complex network analysis.
Johny, K., Pai, M.L., Adarsh, S., 2020. Adaptive Eemd-Ann hybrid model for Indian
J. Hydrol. 563, 802–810.
summer monsoon rainfall forecasting. Theor. Appl. Climatol. 1–17.
Aksoy, H., Dahamsheh, A., 2009. Artificial neural network models for forecasting
Konapala, G., Mishra, A., 2017. Review of complex networks application in
monthly precipitation in Jordan. Stoch. Env. Res. Risk A. 23 (7), 917–931.
hydroclimatic extremes with an implementation to characterize spatio-temporal
Aqil, M., Kita, I., Yano, A., Nishiyama, S., 2007a. A comparative study of artificial neural
drought propagation in continental USA. J. Hydrol. 555, 600–620.
networks and neuro-fuzzy in continuous modeling of the daily and hourly behaviour
Lee, J., Nemati, S., Silva, I., Edwards, B.A., Butler, J.P., Malhotra, A., 2012. Transfer
of runoff. J. Hydrol. 337 (1–2), 22–34.
entropy estimation and directional coupling change detection in biomedical time
Aqil, M., Kita, I., Yano, A., Nishiyama, S., 2007b. Neural networks for real time
series. Biomed. Eng. Online 11 (1), 19.
catchment flow modeling and prediction. Water Resour. Manag. 21 (10),
Legates, D.R., McCabe, G.J., 1999. Evaluating the use of “goodness-of-fit” measures in
1781–1796.
hydrologic and hydroclimatic model validation. Water Resour. Res. 35 (1), 233–241.
Babel, M.S., Badgujar, G.B., Shinde, V.R., 2015. Using the mutual information technique
Li, X., Maier, H.R., Zecchin, A.C., 2015. Improved Pmi-based input variable selection
to select explanatory variables in artificial neural networks for rainfall forecasting.
approach for artificial neural network and other data driven environmental and
Meteorol. Appl. 22 (3), 610–616.
water resource models. Environ. Model. Softw. 65, 15–29.
Badrzadeh, H., Sarukkalige, R., Jayawardena, A.W., 2013. Impact of multi-resolution
Lohani, A.K., Goel, N., Bhatia, K., 2011. Comparative study of neural network, fuzzy
analysis of artificial intelligence models inputs on multi-step ahead river flow
logic and linear transfer function techniques in daily rainfall-runoff modelling under
forecasting. J. Hydrol. 507, 75–85.
different input domains. Hydrol. Process. 25 (2), 175–193.
Barnett, T.P., Adam, J.C., Lettenmaier, D.P., 2005. Potential impacts of a warming
Maier, H.R., Dandy, G.C., 1998. The effect of internal parameters and geometry on the
climate on water availability in snow-dominated regions. Nature 438 (7066),
performance of back-propagation neural networks: an empirical study. Environ.
303–309.
Model. Softw. 13 (2), 193–209.
Bugata, P., Drotar, P., 2020. On some aspects of minimum redundancy maximum
Maier, H.R., Dandy, G.C., 2000. Neural networks for the prediction and forecasting of
relevance feature selection. SCIENCE CHINA Inf. Sci. 63 (1), 1–15.
water resources variables: a review of modelling issues and applications. Environ.
Cai, Z., Ma, H., Zhang, L., 2019. Feature selection for airborne lidar data filtering: a
Model. Softw. 15 (1), 101–124.
mutual information method with parzon window optimization. GIScience Rem. Sens.
Marschinski, R., Kantz, H., 2002. Analysing the information flow between financial time
1–15.
series. Eur. Phys. J. B Condens. Matter Complex Syst. 30 (2), 275–281.
Chakraborty, P., Sarkar, A., Kumar, S., George, J.P., Rajagopal, E., Bhatla, R., 2020.
Mekanik, F., Imteaz, M., Talei, A., 2016. Seasonal rainfall forecasting by adaptive
Assessment of Ncmrwf global ensemble system with differing ensemble populations
network-based fuzzy inference system (Anfis) using large scale climate signals. Clim.
for tropical cyclone prediction. Atmos. Res. 105077.
Dyn. 46 (9–10), 3097–3111.
Chen, H., Chen, Y., Gao, J., Sun, T., Carlin, J.T., 2020. A radar reflectivity data
Moriasi, D.N., Arnold, J.G., Van Liew, M.W., Bingner, R.L., Harmel, R.D., Veith, T.L.,
assimilation method based on background-dependent hydrometeor retrieval: an
2007. Model evaluation guidelines for systematic quantification of accuracy in
observing system simulation experiment. Atmos. Res. 105022.
watershed simulations. Trans. ASABE 50 (3), 885–900.
Cheng, Q.-B., Reinhardt-Imjela, C., Chen, X., Schulte, A., Ji, X., Li, F.-L., 2016.
Nash, J.E., Sutcliffe, J.V., 1970. River flow forecasting through conceptual models, part I-
Improvement and comparison of the rainfall–runoff methods in swat at the
a discussion of principles. J. Hydrol. 10, 282–290.
monsoonal watershed of Baocun, eastern China. Hydrol. Sci. J. 61 (8), 1460–1476.
Nastos, P.T., Paliatsos, A.G., Koukouletsos, K.V., Larissi, I.K., Moustris, K.P., 2014.
Chua, L.H., Wong, T.S., 2010. Improving event-based rainfall–runoff modeling using a
Artificial neural networks modeling for forecasting the maximum daily total
combined artificial neural network–kinematic wave approach. J. Hydrol. 390 (1–2),
precipitation at Athens, Greece. Atmos. Res. 144, 141–150.
92–107.

12
H. Tongal and B. Sivakumar Atmospheric Research 255 (2021) 105531

Naufan, I., Sivakumar, B., Woldemeskel, F.M., Raghavan, S.V., Vu, M.T., Liong, S.-Y., Solgi, A., Nourani, V., Bagherian Marzouni, M., 2017. Evaluation of nonlinear models for
2018. Spatial connections in regional climate model rainfall outputs at different precipitation forecasting. Hydrol. Sci. J. 62 (16), 2695–2704.
temporal scales: application of network theory. J. Hydrol. 556, 1232–1243. Tongal, H., 2020. Comparison of local and global approximators in multivariate chaotic
Newman, S., Carminati, F., Lawrence, H., Bormann, N., Salonen, K., Bell, W., 2020. forecasting of daily streamflow. Hydrol. Sci. J. 65 (7), 1129–1144.
Assessment of new satellite missions within the framework of numerical weather Villarini, G., Serinaldi, F., 2012. Development of statistical models for at-site
prediction. Remote Sens. 12 (10), 1580. probabilistic seasonal rainfall forecast. Int. J. Climatol. 32 (14), 2197–2212.
Parviz, L., Rasouli, K., 2019. Development of precipitation forecast model based on Wang, W.-C., Chau, K.-W., Cheng, C.-T., Qiu, L., 2009. A comparison of performance of
artificial intelligence and subseasonal clustering. J. Hydrol. Eng. 24 (12), 04019053. several artificial intelligence methods for forecasting monthly discharge time series.
Patel, S., Ramachandran, P., 2015. A comparison of machine learning techniques for J. Hydrol. 374 (3–4), 294–306.
modeling river flow time series: the case of upper Cauvery River Basin. Water Watts, D.J., Strogatz, S.H., 1998. Collective dynamics of ‘small-world’ networks. Nature
Resour. Manag. 29 (2), 589–602. 393 (6684), 440.
Pham, B.T., Le, L.M., Le, T.-T., Bui, K.-T.T., Le, V.M., Ly, H.-B., Prakash, I., 2020. Winchell, M., Gupta, H.V., Sorooshian, S., 1998. On the simulation of infiltration-and
Development of advanced artificial intelligence models for daily rainfall prediction. saturation-excess runoff using radar-based rainfall estimates: effects of algorithm
Atmos. Res. 237, 104845. uncertainty and pixel aggregation. Water Resour. Res. 34 (10), 2655–2670.
Qian, S., Chen, J., Li, X., Xu, C.Y., Guo, S., Chen, H., Wu, X., 2020. Seasonal rainfall Wu, C., Chau, K., 2011. Rainfall–runoff modeling using artificial neural network coupled
forecasting for the Yangtze River Basin using statistical and dynamical models. Int. J. with singular spectrum analysis. J. Hydrol. 399 (3), 394–409.
Climatol. 40 (1), 361–377. Wu, C.L., Chau, K.W., Fan, C., 2010. Prediction of rainfall time series using modular
Rodríguez-Alarcón, R., Lozano, S., 2019. A complex network analysis of Spanish River artificial neural networks coupled with data-preprocessing techniques. J. Hydrol.
Basins. J. Hydrol. 578, 124065. 389 (1–2), 146–167.
Ruddell, B.L., Kumar, P., 2009. Ecohydrologic process networks: 1. Identification. Water Xiang, Y., Gou, L., He, L., Xia, S., Wang, W., 2018. A Svr–Ann combined model based on
Resour. Res. 45 (W03419). ensemble emd for rainfall prediction. Appl. Soft Comput. 73, 874–883.
Sandoval, L., 2014. Structure of a global network of financial companies based on Yasmin, N., Sivakumar, B., 2018. Temporal streamflow analysis: coupling nonlinear
transfer entropy. Entropy 16 (8), 4443–4482. dynamics with complex networks. J. Hydrol. 564, 59–67.
Schreiber, T., 2000. Measuring information transfer. Phys. Rev. Lett. 85 (2), 461. Yu, P.-S., Yang, T.-C., Chen, S.-Y., Kuo, C.-M., Tseng, H.-W., 2017. Comparison of random
Shi, H., Chen, J., Li, T., Wang, G., 2020. A new method for estimation of spatially forests and support vector machine for real-time radar-derived rainfall forecasting.
distributed rainfall through merging satellite observations, Raingauge records, and J. Hydrol. 552, 92–104.
terrain digital elevation model data. J. Hydro Environ. Res. 28, 1–14. Zadeh, M.R., Amin, S., Khalili, D., Singh, V.P., 2010. Daily outflow prediction by multi
Singh, K.P., Basant, A., Malik, A., Jain, G., 2009. Artificial neural network modeling of layer perceptron with logistic sigmoid and tangent sigmoid activation functions.
the river water quality—a case study. Ecol. Model. 220 (6), 888–895. Water Resour. Manag. 24 (11), 2673–2688.
Sivakumar, B., Woldemeskel, F.M., 2014. Complex networks for streamflow dynamics. Zahmatkesh, Z., Goharian, E., 2018. Comparing machine learning and decision making
Hydrol. Earth Syst. Sci. 18 (11), 4565–4578. approaches to forecast long lead monthly rainfall: the city of Vancouver, Canada.
Sivakumar, B., Woldemeskel, F.M., 2015. A Network-based analysis of spatial rainfall Hydrology 5 (1), 10.
connections. Environ. Model. Softw. 69, 55–62. Zoglat, A., Amar, A., Badaoui, F., Hassou, L.A., 2018. A Copulas Approach for Forecasting
Smith, M.B., Koren, V.I., Zhang, Z., Reed, S.M., Pan, J.-J., Moreda, F., 2004. Runoff the Rainfall, International Conference on Advanced Intelligent Systems for
response to spatial variability in precipitation: an analysis of observed data. Sustainable Development. Springer, pp. 236–244.
J. Hydrol. 298 (1–4), 267–286.

13