Journal of Hydrology 597 (2021) 125722

Contents lists available at ScienceDirect

Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol

Research papers

Analysis of spatio-temporal dependence of inflow time series through

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

This manuscript was handled by Sally Elizabeth This paper aims to assess fully the spatio-temporal dependence dimensions of inflow across two adjacent and
Thompson, Editor-in-Chief, with the assistance parallel basins and among different time steps through Causality. This is addressed from the perspective of Causal
of Carlos Oroza, Associate Editor Reasoning, supported by Bayesian modelling, under a novel framework named Bayesian Causal Modelling
(BCM). This is applied, through a “concept-proof”, to the Jucar River Basin (the second largest basin of Eastern
Keywords:
Spain, characterized by long and severe drought conditions). In this “concept-proof” a double goal is evaluated;
Causality
first dedicated to a lumped analysis of dependence and second a specific one over dry periods focused on time-
Causal reasoning
Bayesian spatio-temporal dependence horizon of the Jucar basin typical drought (3 years). These challenges comprise the development of two fully
Stochastic hydrology connected Bayesian Networks (BNs), one for each challenge populated/trained from historical-inflow records.
Jucar river basin BNs were designed at a season-scale and consequently, time was upscaled and grouped into Irrigation and Non-
Historical inflow time series Irrigation periods, according to Jucar River Basin Authority operational practices. Results achieved showed that
BCM framework satisfactorily captured the spatio-temporal dependencies of systems. Furthermore, BCM is able
to answer to some key questions over interdependencies between adjacent and parallel subbasins. Those ques­
tions may comprise, the amount of spatial dependences among time series, the temporarily conditionality among
subbasins and the spatio-temporal dependence among basins. This provides a relevant insight on the intrinsic
spatio-temporal dependence structure of inflow time series in complex basins systems. This approach could be
very valuable for water resources planning and management, due to its application power for predicting extreme
events (e.g. droughts) as well as improving and optimizing the reservoirs operation rules.

1. Introduction that inherently underlies in the runoff time series (Molina et al., 2019).
The meteorological-hydrological basin response is captured by means of
The inherent nonlinear nature of hydrologic series and their associ­ this internal structure (Wang et al., 2009; Hao and Singh, 2016; Molina
ated processes is widely known (Wei et al., 2012; Molina et al., 2016; and Zazo, 2018; Zazo et al., 2019). Furthermore, the spatio-temporal
Peng et al., 2017). Furthermore, the nonstationarity of records (annual changes in weather patterns are likely to further aggravate the appear­
or even larger time scale) is rapidly growing worldwide over the last ance and persistence of drought events (Mishra and Singh, 2010).
years due to modifications/alterations in the climatic patterns (Allan Likewise, the dependence of hydrological processes is a traditional
and Soden, 2008; Trenberth, 2011; Kalra et al., 2013; Chang et al., 2015; studied matter, given its importance to assess, model and forecast the
Wasko and Sharma, 2015; Donat et al., 2016) as result of the influence of behavior and availability of water resources, especially in a context of
climate change into small-scale meteorological processes (Marotzke droughts (Mishra and Singh, 2011). Classic ways to evaluate these de­
et al., 2017; Pfahl et al., 2017). This global situation is materialized in pendences are mainly related to statistical procedures, including
the increasing occurrence, intensity and persistence of extreme events regression (e.g. Hao and Singh, 2016), time series modeling (e.g. Salas
(Jyrkama and Sykes, 2007; Marcos-Garcia et al., 2017; Molina and Zazo, et al., 1980; Hipel and McLeod, 1994; Mishra and Desai, 2005), recur­
2018) including droughts, which can produce significant economic rence analysis (e.g. Estrela et al., 2000), and principal component
losses (Gil et al., 2011; Kahil et al., 2015; Freire-González et al., 2017; analysis (e.g. Vicente-Serrano et al., 2004; Vicente-Serrano and Cuadrat-
Lopez-Nicolas et al., 2017). In order to reduce these negative impacts, it Prats, 2007). For a deeper explication of these techniques, the reader is
is necessary to better understand the internal structure of dependencies referred to the work Mishra and Singh (2011).

https://doi.org/10.1016/j.jhydrol.2020.125722
Received 16 August 2020; Received in revised form 16 October 2020; Accepted 29 October 2020
Available online 5 November 2020
0022-1694/© 2020 Elsevier B.V. All rights reserved.
H. Macian-Sorribes et al. Journal of Hydrology 597 (2021) 125722

Regarding temporal dependence of hydrological time series, this has and characterize the logical and non-trivial structure of dependencies
traditionally received a more attention by the scientific community, for that inherently underlies historical records (Zazo, 2017; Molina and
example by means of Persistence (inherent statistical property of time Zazo, 2018; Molina et al., 2019; Zazo et al., 2019). Consequently, there
series), which is strongly related to the measure of the time series long- is a general clear necessity of strengthen the spatio-temporal depen­
term memory through Hurst coefficient (Hurst, 1951), or storage and dence studies on water systems (Holmström et al., 2015).
drought statistics (Salas et al., 1980), which are currently benchmarks; This study has been applied in Jucar River Basin (the second largest
or other ones such us Copula applications (De Michele and Salvadori, basin on Mediterranean side of Spain). This river basin is characterized
2003; Dehghani et al., 2019; Li et al., 2019; Saghafian and Sanginabadi, by an irregular hydrology, with long drought episodes, an intense irri­
2020), the usage of multivariate distributions in modeling different gation activity and the existence of large reservoirs able to provide
dependence structures (Sarabia-Alzaga and Gómez-Déniz, 2008; Tsou­ carryover storage (Marcos et al., 2017). In order to achieve an efficient
kalas et al., 2019; Eum et al., 2020), or novel analysis strategies based on system operation, within a context of sustainable and safe water re­
Artificial Intelligence (AI), such as Artificial neural networks (ANNs; sources management of a basin, it is necessary to adequately charac­
(Ochoa-Rivera, 2008), Bayesian Networks (BN; Molina and Zazo, 2017; terize, both in space and time, the hydrological discharge to its main
Zazo et al., 2019), BNs applied to inflow predictions (Kim et al., 2018; reservoirs (Alarcon and Contreras) and their evolution. Furthermore, it
Jia et al., 2020) or Dynamic Bayesian networks (DBNs; Molina et al., is also necessary to deepen in the knowledge of their spatio-temporal
2013; Roozbahani et al., 2018; Ropero et al., 2018). However, spatial behavior through the discovery and analysis of the non-trivial and
and spatio-temporal dependence for hydrologic science and engineering logical structure of interdependence, which intrinsically underlies the
is much poorer studied (Holmström et al., 2015) and even more through historical records and that define of its behavior.
a Bayesian perspective (Wikle et al., 1998; Jona Lasinio et al., 2005). This paper is mainly aimed to provide a new methodological
This is because of: (a) complexity of characterizing and differentiate approach, named Bayesian Causal Modelling (BCM). This is done by
water sub-systems, (b) scarcity of spatial data availability and (c) diffi­ means of an innovative “concept-proof”, focused on Causality, here
culties on the application of spatial statistical methods, among others. addressed by Causal Reasoning and supported by Bayesian modelling.
On the other side, in the last years, the emergence of AI and Infor­ This methodological approach is mainly based on special features that
mation Theory techniques have allowed tackling new ways of BNs offer such as: (1) their inherent ability to explicitly model uncer­
approaching, characterizing and quantifying dependences (Wang et al., tainty (Vogel et al., 2014) and/or (2) their unsupervised structural
2009; Hejazi and Cai, 2011; Singh, 2011; Wu et al., 2015; Lima et al., learning (Molina and Zazo, 2017). This provides not only temporal
2016; Molina et al., 2016; Kousari et al., 2017; Yang et al., 2016, 2017a, dependence behavior but also the spatial and spatio-temporal de­
2017b). These techniques are useful for exploring the evolution of pendencies behavior of inflow across river basins. This knowledge is
meteorological and hydrological dynamics in order to find suitable practically unexplored in the field of water engineering and science and
drought predictors at different spatio-temporal scales (Molina et al., that is what the novelty is about. In this sense, the application of BCM
2020). both to the prediction of droughts and to operation rules improvement
On the other hand, Causality in hydrological records has not been of reservoirs, by means of this approach seems quite straightforward in
deeply studied and it could be done by means of the joint use of different the case of in parallel interconnected basins.
forms of reasoning patterns. These forms are Causal Reasoning (CR), On the other side, this work, although addressed through a “concept-
Evidential Reasoning (ER) and Intercausal Reasoning (IR) (Koller and proof”, is a further step in the research initiated with Molina et al. (2016)
Friedman, 2009; Pearl, 2009). CR is used when the approach is done and Zazo (2017). This research activity is largely characterized by the
from top to bottom. In this sense, the analysis is focused on the cause and application of BNs, in a dynamic and stochastically way, to increase
the objective comprises the prediction of the effect or consequence. knowledge over water resources behavior of basins.
Consequently, the queries in form of conditional probability, where the This paper is organized as follows: after this Introduction, next Sec­
“downstream” effects of various factors are predicted, are instances of tion 2 covers a description of the applied materials and methods for this
causal reasoning or prediction. Classical Stochastic Hydrology has been research. Then, Section 3 is devoted to case studies description and
very focused along several decades on the characterization of temporal available hydrological historical records. General methodology is shown
behavior for runoff or streamflow in different basins worldwide. In this and explained in depth in Section 4. After that, Section 5 presents the
sense, CR approach becomes especially useful when hydrological main experimental results drawn for the research. In Section 5, the re­
dependence analysis takes place because it is about an evaluation of the sults are discussed in detail. Finally, Section 6 addresses a Discussion
influence of past events on posterior hydrological processes. ER com­ and Conclusions from the study, where the most important aspects
prises bottom-up reasoning, so the analysis is focused on the conse­ learnt from this research are shown.
quence (effect) and the cause is inferred (Bayesian Inference). This
approach would be appropriate to the analysis of catastrophic events 2. Materials and methods
such as flooding where the most important thing to be studied is the
consequence. IR is probably the hardest concept to understand. It 2.1. Bayesian networks
comprises the interaction of different causes for the same effect. This
type of reasoning is very useful in hydrology, where a consequence can Bayesian Networks (BN), also known as Bayesian Belief Networks or
be generated or explained from several causes. This approach would be Belief Networks, belong to the AI methods, such as artificial neural
very appropriate for developing Decision Support Systems based on networks, fuzzy logic systems and decision trees (Molina et al., 2016).
Causality and Causal Modelling (Molina et al., 2010) where there is a They are Probabilistic Graphical Models (PGM) in which a visual rep­
high heterogeneity of variableś nature and high complexity of causal resentation of a reasoning problem is performed to infer a new knowl­
relationships. Furthermore, one of the most exciting prospects in recent edge into an uncertainty context (Cabañas de Paz, 2017). Furthermore,
years has been the possibility of using the theory of BNs to discover PGM are considered as convenient and intuitive formalisms to capture
causal structures in raw data (historical runoff record) (Pearl et al., the probabilistic independent information of a domain (Madsen et al.,
2014). This is done in this research through the usage of historical inflow 2003; Pearl, 1988; Vogel et al., 2018).
data to train and populate the BN model. Consequently, AI techniques BNs combine probability theory with graph theory (Vogel et al.,
such as CR and ER and/or IR provide new horizons for this type of 2014). In this sense, they are based on: 1) a graphical representation
studies. comprising nodes and links, of a given set of random variables with their
Recently, the potential of BNs to find temporal causal structures into conditional interdependencies; and 2) a probabilistic model consisting
raw statistical data (Spirtes, 2010)has begun to be applied to discover of probabilistic expressions that describe the probability distributions/

2
H. Macian-Sorribes et al. Journal of Hydrology 597 (2021) 125722

functions of the variables involved and the relationships between them 2.2. Spatio-temporal dependence analysis through Causality
depicted in the links (Cain, 2001; Castelletti and Soncini-Sessa, 2007;
Sperotto et al., 2017). Furthermore, BNs aim to model the joint proba­ Spatio-temporal dependence analyses in hydrology are mainly
bility distribution of all considered variables. This is performed by related to the use of statistics and time series modeling procedures,
means of propagation of (in)dependencies between the variables which require them as a preliminary step (Salas et al., 1980; Hipel and
throughout the whole graphical structure, which can be seen as a gen­ McLeod, 1994). A crucial part of these approaches is the postulation of
eral description of the behavior of a system (Vogel et al., 2018). In other the stochastic model to be used (e.g. AR 1, ARMA 1,1, AR 2), and the
words, PGMs offer a compact representation of the joint probability choice of the statistical distribution followed by the variables considered
distribution over sets of random variables (Said, 2006; Castelletti and by the model. Both decisions constrain the modeling of spatio-temporal
Soncini-Sessa, 2007). dependences, since they assume that the analyzed variables are struc­
On the other side, each node in the network represents a variable. tured in a specific way and that this structure does not change over time,
The input variables to a BN are graphically represented through the or that it changes following a specific pathway. In fact, the goal of these
parent nodes (nodes without arriving links) and the output variables are choices is to look for a spatio-temporal pattern close enough to the one
computed following the links between the root nodes and the leaf nodes shown by the available data records of the modeled variables. AI models,
(nodes without departing links, which represent the output variables in on the other hand, do not need to make assumptions on the data, as they
the network). For each link, the probabilistic expressions representing are able to find out the properties and spatio-temporal patterns shown
the relationship between the departing and arriving nodes are applied. by it, even if they could not be identified before the analysis (Molina
The input variables can be provided as single values or as probability et al., 2016; Molina and Zazo, 2017). Although, AI approaches require
functions, while the output variables are obtained as probability func­ large amount of data that not always is easy to obtain (Zazo et al., 2019;
tions. This probabilistic assessment is the main distinctive feature of a Molina et al., 2020).
BN and enables the possibility of characterizing uncertainty, perform At this point, it is worth highlighting that BNs are able to discover,
risk analyses and support decision-making processes (Castelletti and extract and quantify the logical and non-trivial time-dependency re­
Soncini-Sessa, 2007; Vogel et al., 2013, 2012; Sperotto et al., 2017). lationships existing among variables by CR (Molina and Zazo, 2018;
Formally, BN N = (G, P) consists of a Direct Acyclic Graph (DAG; Molina et al., 2019). This research comprises an advance over the pre­
Lappenschaar et al., 2012) denoted by G = (V, E) and a set of probability vious ones as it is able to discover and extract both spatial and spatio-
distributions P (Molina et al., 2013), therefore, a BN N = (G, P) defines a temporal dependencies. Furthermore, once the BN is built, it possible
joint probability distribution over all the analyzed variables. The DAG is to determine how strong are the strength of relationships between var­
defined by a set of nodes (or variables) V and a set of links (or edges) E. iables by modifying the probability distribution of one of them and
The edges join the variables and are oriented, indicating the causal seeing how it affects the probability distributions of the remaining ones
dependence between the connected variables (A|B indicates A causes B (Molina et al., 2016). This modification may be successfully addressed
or B is the effect of A) (Madsen et al., 2003). The set of probability by means of maximizing its highest interval, and its impact is measured
distributions P includes “a priori” probabilities and a set of conditional quantifying the change in the expected value of the other variables (the
probabilities (expressed in form of Conditional Probabilities Tables; average) found when the first one is maximized (Molina et al., 2016).
CPTs). Theorem of Bayes is performed as updater of “a priori” proba­ Under this approach, small changes mean that the relationship is weak,
bilities adding the observed evidence and providing “a posteriori” while large modifications refer to strong bonds between the variables
probabilities, mathematically expressed e as (Bayes, 1763): represented in those nodes. If the BN represents the values of a given
variable for different locations and time stages, the relationships ob­
P(A|B)Â⋅P(B) tained from the analysis are spatial, temporal and spatio-temporal.
P(B|A) =
P(A)
2.3. Potential applications of BN for spatio-temporal analysis
where P(B|A) is the conditional probability of B for a given state of
variable A; P(B|A) the other-way-round conditional probability; P(B) the
The use of BNs to perform spatio-temporal analyses shows a main
probability of B; and P(A) the probability of A.
advantage in comparison with classic statistics as the spatio-temporal
Specifically applied to “concept-proof” over BCM of this research, the
dynamics of these relationships can be easily assessed. Their potential
nodes in this case represent the hydrological space and time variables,
applications include:
which summarize the response of a basin to the hydrological cycle
(Zazo, 2017) due to the fact that the inflow values intrinsically en­
1. Situations in which it is desired to explore if two or more variables
compasses the influence of key factors such as rainfall-temperature
show a causal relationship (e.g. between dry or wet cycles and cli­
variability, orography, fluvial morphodynamics, and anthropic factor
matic indicators such as the ENSO).
(e.g. land-use change) as well as river-aquifer interaction that condition
2. Situations in which the spatio-temporal pattern of the relationship
of the water flow in a point (Molina and Zazo, 2018), in this case inflow
cannot be not known in advance or cannot be adequately modelled
at a reservoir.
with traditional statistics (because stochastic models able to fit the
On the other hand, BNs present relevant advantages such as: 1) the
given pattern would be very complex and thus, less parsimonious).
explicit graphical representation, that lets any user see the variables and
3. Situations in which fitting a probability distribution to the variables
the relationships analyzed by them; and 2) the propagation of uncer­
cannot be reliably done (because no probability function is found
tainty, something very valuable in risk assessment and decision-making.
adequate enough or the chosen one would be very complex).
They are also able to quantify the interdependencies existing between
4. Situations in which the spatio-temporal dependencies vary over time
variables and they allow defining, in an easy and automatically way,
(e.g. river basins subject to changing climatic patterns).
relationships into complex systems (Molina et al., 2010)and proving
more accurate results in the modeling of natural processes (See and
3. Case study: the Jucar river basin in Spain
Openshaw, 2000; Jain and Kumar, 2007). In contrast, their data re­
quirements are big, their complexity grows are the systems and pro­
3.1. Description
cesses being modeled increase, and the fact that it is difficult to perform
a quantitative validation on the results (Molina et al., 2010; Sperotto
The case study is located at Jucar River Basin, a Mediterranean basin
et al., 2017).
of 22,261 km2 in Eastern Spain; which is focused on Alarcon and Con­
treras sub-basins, both belonging to Upper Jucar River Basin (Fig. 1).

3
H. Macian-Sorribes et al. Journal of Hydrology 597 (2021) 125722

Fig. 1. Jucar River Basin location.

River Jucar flows through 497 km between the Iberian Mountains and
the Mediterranean Sea. Following a typical Mediterranean pattern
(peaks in autumn and very little values during summer), the annual
rainfall spans from 309 to 717 mm, with an average value of 473 mm. Its
mean annual resource is 1605 mm3/year, in which 70% comes from
Fig. 2. Hydrogeological maps of Alarcon and Contreras sub-basin. Source:
groundwater outflow through springs and stream-aquifer interaction IGME (IGME, 2020b).
(CHJ, 2015). Its main reservoirs are Alarcon (1088 mm3 useful capac­
ity), Contreras (429 mm3), they are the most one’s capacity of Jucar
Velazquez, 2017). In this sense, once environmental requirements are
River Basin. Alarcón and Contreras are placed at the outlets of the upper
met, the current operating rules give priority to urban uses over agri­
basins of the Jucar and Cabriel rivers, while Tous is located at the limit
cultural areas. However, it should be noted that the main legal
of the Jucar floodplain, in which most urban and agricultural demands
constraint in the Jucar River system is the Alarcon Agreement (CHJ,
get concentrated.
2015), signed between the Spanish Ministry of the Environment (on
In lithological terms, Alarcon and Contreras sub-basins are charac­
behalf of the CHJ) and a users’ association, “Unidad Sindical de Usuarios
terized by existing of important limestone formations, dolomites, con­
del Jucar” (USUJ), which gathers together the users with the most senior
glomerates, sandstones and Tertiary detritus (IGME, 2020a).
water rights owning the Alarcon reservoir (Macian-Sorribes and Pulido-
Hydrogeology speaking, Alarcon sub-basin presents impermeable or
Velazquez, 2017). Among the latter, the users with elder rights from the
very low permeability formations, together with extensive, discontin­
lower basin are given the highest surface priority. The remaining de­
uous and local aquifers of moderate permeability and production. In
mands, with access to both surface and ground waters, can only use
contrast, Contreras sub-basin is characterized by very permeable aqui­
surface water if the Alarcon storage is higher than the limits established
fers, generally extensive and productive with other ones impermeable or
by the Alarcon Agreement (CHJ, 2015).
very low permeability zones (IGME, 2020b). Moreover, both sub-basins
Finally, both sub-basins (Alarcon and Contreras) are considered as
show a piezometric gradient towards the coastal area, from West to East
lumped units, in line with the assumptions made by the Jucar River
(CHJ, 2006, 2020) (Fig. 2).
Basin Authority (CHJ, 2020). Furthermore, and according to operational
From a management point of view for the Jucar river basin, the
practices of the Jucar River Basin Authority (CHJ, 2020) two different
annual consumptive demand of the Jucar river system is 1557 Mm3 for
periods are considered in this study. First one, the Irrigation Period (to
the 2015–2021 period (CHJ, 2015). The major consumptive use is
May from September) corresponds to the irrigation season, and second
agriculture (90%), followed by urban (8%) and industrial (2%) uses. The
one, the Non-Irrigation Period covers to October from April of next year.
most important urban districts correspond to Valencia, Albacete and
To a complete description of Management of the Jucar River Basin, it
Sagunto. Surface irrigation demands concentrate on the lower basin
refers the reader to Macian-Sorribes and Pulido-Velazquez (Macian-
(Riberas del Jucar and Canal Jucar-Turia), while the main groundwater-
Sorribes and Pulido-Velazquez, 2017)
irrigated area is the Mancha Oriental, in the middle basin. The aquifer
overdraft from this demand has caused an inversion of the stream-
aquifer interaction found between the Mancha Oriental aquifer and 3.2. Available and processed data
the Jucar river, moving from gaining to losing river. Furthermore,
minimum flows are set on 18 streams by environmental reasons (CHJ, The BCM analysis of spatio-temporal inflow dependence has been
2015). applied to the Alarcon and Contreras sub-basins, which correspond to
Regarding the planning and operation of the Jucar River Basin is the the upper part of the Jucar River Basin and are the main contributors to
main responsibility of the Jucar River Basin Authority (in Spanish the water discharge that can be regulated through the Alarcon and
Confederacion Hidrográfica del Júcar, CHJ), which is subject to physical, Contreras reservoirs (see Fig. 1). The available data consists of monthly
environmental, and legal constraints (Macian-Sorribes and Pulido- inflow time series to the reservoirs for the 1940–2012 period (72 hy­
drological years; please note: in Spanish hydrological years the first

4
H. Macian-Sorribes et al. Journal of Hydrology 597 (2021) 125722

month is October and the last is September of following year), obtained Table 1
from the Spanish Ministry of Agriculture, Fishery, Food and Environ­ Alarcon and Contreras monthly time series. Main statistic parameters.
ment (http://ceh-flumen64.cedex.es/anuarioaforos/default.asp). Given Statistic parameters Sub-basins
that the outlets of both sub-basins coincide with reservoirs (Alarcon and
Alarcon Contreras
Contreras reservoirs respectively) whose inflow, storage and release is
measured, it is not necessary to consider further gauging stations to Monthly mean (Hm3) 33.01 28.51
Annual mean (Hm3) 396.09 342.12
calculate its total hydrological discharge. Minimum (Hm3) 3.86 5.52
Data records in both reservoirs were restored to natural regime by Maximum (Hm3) 262.14 211.70
computing the evaporation and seepage losses in Alarcon and Contreras Range (Hm3) 258.28 206.18
using evaporation rates, storage-surface curves and infiltration rates Variance 1099.39 532.31
Standard deviation (Hm3) 33.16 23.07
provided by CHJ; and then adding both loss terms to the actual inflows
Variation coefficient (%) 100.5 80.9
recorded at both reservoirs. Because of water use and abstraction is Skewness coefficient (*) 2.78 2.98
considered negligible in both Alarcon and Contreras headwaters in Positive skewness Positive skewness
comparison to its discharge, no further adjustments were needed to Kurtosis 10.12 14.30
restore the recorded inflows to natural regime. This procedure is
equivalent to the procedure followed by CHJ in computing the natural
4.1. Step-1. Data analysis
discharges of both basins. Fig. 3 and Table 1 display the monthly time
series considered and their main statistic parameters respectively, as
The key factors of this step are the origin of the data (historical re­
well as the probability distribution functions of both time series and
cords, outputs from hydrological modeling, results obtained by sto­
their correlograms are presented in Fig. 4.
chastic models and so on), its spatial coverage (e.g. the entire
The similarity exposed by their probability distributions is the result
hydrological basin as a lumped unit, the entire basin divided into sub-
of its spatial proximity and their resemblance in physical, meteorolog­
basins, certain sub-basins, …) and its temporal coverage (period of
ical and hydrological properties. Their correlograms show a strong
analysis, time scale of the data, etc.). They should be shaped according
spatial and spatio-temporal dependence. The temporal dependence of
to the objective of the analysis. For example, the evaluation of climate
Contreras runoff is higher than in Alarcon. On the contrary, the spatio-
change impacts would require the obtaining of future inflow scenarios
temporal correlograms show a higher influence of Alarcon on Contreras
through hydrological modeling, while analyses based on the current
inflow values than in the opposite. The main reason behind these dif­
hydrology may be done employing historical records. One aspect that
ferences seems to be the role of groundwater discharge (CHJ, 2006,
should also be considered is the data needs of a BN, which may require to
2020). These relationships will be explored more deeply with the BCM
extend the available records using modelling techniques (e.g. Molina
analysis.
et al., 2016).
In the case of the Jucar river basin, the analysis has been developed
4. Methodology
at the seasonal scale, dividing the monthly inflows into two different
periods per year. This division has been made in accordance with the
This research was articulated in three consecutive and interrelated
operational practices of the Jucar River Basin Authority (CHJ). The
sequential steps (Fig. 5). Step-1 comprises the treatment and analysis of
Irrigation Period (IP; to May from September) corresponds to the irri­
the raw inflow data, in order to adapt it to the subsequent steps and to
gation season, in which agricultural demands concentrate and rainfall is
the goals pursued by the study. After that, Step-2 corresponds to the
reduced, so reservoirs are lowered to satisfy them. The Non-Irrigation
building of the BN_PGM models applied to perform the spatio-temporal
Period (NIP; to October from April of next year) is characterized by
BCM analysis; and they models are populated and trained by the treated
low agricultural demands and higher rainfalls than the summer one. The
data previously. Finally, Step-3 is fully entirely devoted to calculating
Jucar river system operation by the CHJ Operation Office is supported
and analyzing the causal relationships through “concept-proof” BCM.
by projections on inflows during the irrigation period based on the hy­
This last step is crucial because of it will allow discovering, extracting
drological discharge on the past months (Macian-Sorribes and Pulido-
and quantifying of the logical and non-trivial time-dependency structure
Velazquez, 2017). Based on them, deliveries from the reservoirs to the
that inherently underlies into hydrological series and that define of the
farmers are scheduled for the upcoming irrigation season. Consequently,
behavior of the adjacent sub-basins. Finally, this “concept-proof” is
this division aligns the analysis with the practices of the system
validated through a sensitivity analysis, based on Information Theory.
operators.

Fig. 3. Alarcon and Contreras historical monthly series after removing effect of evaporation and seepage losses, according to Spanish hydrological years (first month
October and the last one is September of following year).

5
H. Macian-Sorribes et al. Journal of Hydrology 597 (2021) 125722

Fig. 4. Preliminary assessment of the spatial and temporal dependencies of the Alarcon and Contreras sub-basins. Upper: Probability distribution functions. Bottom:
Temporal and Spatio-temporal correlograms with Anderson probability limits for an independent series (95 and 99 percent probability level). Please note that the
Anderson limits define independence area of a correlogram.

4.2. Step-2. Causal model building 4.3. Step-3. Bayesian causal modelling

This step involves developing BN-PGM model by means of other four This is a key step in this research and consists in calculating and
sub-steps:1) Preliminary design; 2) training; 3) design refining; and 4) analyzing the causal relationships using the BCM developed in the last
validation and sensitivity analysis. At the beginning of the process, a first step, following the process explained in Section 2.2. The estimation of
design of the PGM is proposed. A training process on this draft is per­ the relationships embedded in each BN-PGM consisted in alternatively
formed to refine the design of the BN-PGM model and make sure it modifying (maximizing) one node of the model and evaluate how the
captures the spatio-temporal relationships shown by the data. Once the remaining nodes change in response (Molina et al., 2016). A remarkable
model is set, it is necessary to validate and to perform a sensitivity change in one particular node associated with a change in another
analysis, based on the perspective of the Theory of Information. would imply a strong causal relationship between them, whereas a little
Given the novelty of application of this BCM methodology, there one would mean the absence of a bond. The analysis of these relation­
were not previous references on how to proceed regarding the applica­ ships consists in comparing and sorting the changes previously calcu­
tion of the PGMs for the triple dimension of analysis pursued by this lated, considering the type of relationship that each pair of nodes has (e.
research, so it was built in an iterative way. After a period of analysis g. spatial, temporal, spatio-temporal).
within the preliminary design phase, an initial draft was developed. This step was also tackled from a triple dimension that comprises
Once the network design was set, supported by a powerful training spatial, temporal and spatio-temporal analysis. Drawn results from these
through Learning Wizard (HUGIN® version 7.3), a fully connected PGM dimensions show important differences among them. Furthermore, due
was manually built for wholly capturing and modelling the double to the flexibility of design and the back propagation of the probability,
dimension (spatio-temporal) of both sub-basins. This is consistent with the BN-PGM was able to evaluate the influence of both sub-basins each
the strong relationships found in the previous data analysis (please see other in both ways. This is a relevant characteristic, which is essential
Fig. 4 bottom, correlograms). Finally, model validation was developed for developing an accurate and rigorous water rivers basins manage­
through the calculation of some important parameters belonging to In­ ment. This is highlighted through two different analysis. First one
formation Theory discipline. In particular, it has been used Total En­ dedicated to a lumped analysis of dependence, and the second one by
tropy, Conditional Entropy and Mutual information. In the Jucar river means of a specific analysis over dry periods focused on time-horizon of
basin two different BN-PGM models have been built: one describing the the Jucar basin typical drought (3 years; CHJ, 2007). Both analyses
average and lumped behavior of the upper sub-basins and another comprise the development of two fully connected Bayesian Networks
specific one used to obtain the spatio-temporal relationships during a (BNs), designed at a season-scale and grouped in two different periods,
dry period. Temporal Irrigation and Non-Irrigation ones, according to Jucar River

6
H. Macian-Sorribes et al. Journal of Hydrology 597 (2021) 125722

Fig. 5. General methodology.

Basin Authority operational practices. Furthermore, it was considered a (spatial, temporal and spatio-temporal) at seasonal scale. This index is
year as dry year if either Alarcon or Contreras runoff during the winter obtained as the division between the expected value of the probability
season falls below the percentile 33 of the whole historical period. function after maximization of the probability intervals from designed
A novel variation ratio (VR) has been proposed that allows a BN-PGM models and the one from before. In order to facilitate its
dependence analysis between the 2 subbasins at the three dimensions compression, the outcomes were normalized, in this way that variation

Fig. 6. BCM conceptual frameworks to upper Jucar river sub-basins. Direct Acyclic Graph considered between one season and the subsequent one. Upper lumped
analysis. Bottom specific analysis over dry cycle. Abbreviations used: Non-irrigation period (NIP). Irrigation period (IP). Contreras reservoir (Con). Alarcon reservoir
(Ala). Please note that the numbers indicate the year of the dry cycle analyzed.

7
H. Macian-Sorribes et al. Journal of Hydrology 597 (2021) 125722

ratio had the same physical sense as a classical correlation coefficient. about the true state of X (Molina et al., 2016; Zazo, 2017). The entropy
That is, the higher the value departs from zero (0), the stronger the measure therefore enables an assessment of the additional information
dependence is. In contrast, values lower than 0 indicate an inverse required to specify a particular alternative. Shannon’s measure of
dependence, and 0 value displays non-dependence. mutual information is used to assess the effect of collecting information
about one variable (Y) in reducing the total uncertainty about variable X
4.4. BCM concept-proof. Applied conceptual frameworks using:
I(Y.X) = H(Y) − H(Y|X)
Due to the absence of previous references on this type of application,
it is essential to describe the applied conceptual frameworks (Fig. 6). where I(Y.X) is the mutual information between variables. This measure
Each node of the BN represents the runoff of an upper Jucar sub-basin reports the expected degree to which the joint probability of X and Y
for a given season. Alarcon NIP is the runoff of the Alarcon sub-basin diverges from what it would be if X were independent of Y. If I(Y.X) = 0,
for the non-irrigation period; Contreras NIP is the runoff of Contreras X and Y are mutually independent. H(Y∣X) is conditional entropy which
for the non-irrigation period; Alarcon IP corresponds to the runoff from means the uncertainty that remains about Y when X is known to be x.
the Alarcon sub-basin during the irrigation season; and Contreras IP According to the entropy approach presented, mutual information
represents the runoff from Contreras for the irrigation period. values higher than 0 would imply some dependence between the vari­
On the other side, the data used to populate and train this BN cor­ ables analyzed, while the opposite would mean that they are indepen­
responds to the available monthly historical records presented in Section dent from each other (Pearl, 1988). Therefore, this analysis represents
3.2., obtained from the Jucar River Basin Authority (CHJ), upscaled to an alternative way to characterize the dependences between variables.
seasonal periods (months from October to April added to form the non- In this research, this validation is useful because its results can be
irrigation runoff and months from May to September to form the irri­ compared with the temporal dependence analysis carried out.
gation runoff). Assessments of the entropy associated to each variable, the condi­
Consequently, in order to propagate the Bayes’ Theorem in both tional entropy and mutual information for each variable and connection
space and time the Fig. 6 summarizes the applied conceptual frame­ were performed supported by HUGIN Expert® software.
works both for lumped analysis and specific ones focused on dry cycle
analysis. In order to assure that the model structure faithfully represents 5. Results
the reality, the structure of BN was done manually. Furthermore, CPTs
were automatically built from a database containing inflow time series 5.1. Average spatio-temporal dependences
data for Alarcon and Contreras.
The BCM was built using the Learning Wizard of HUGIN Expert® (V The average spatio-temporal dependence analysis has been done by
7.3), which generated probabilistic distributions of the variables asso­ maximizing the probability distribution of one variable of the BN each
ciated with each node, as well as a logic structure according to the in­ time and measuring how this perturbation affects the probability dis­
ternal dependences and relationships detected. The Learning Wizard tributions of the rest of the network. This dependence has been evalu­
result was a fully connected network (each node interacts with the ated using a novel and normalized developed variation ratio, for each
remaining three), which is in accordance to the strong dependences remaining node of the network, as the division between the expected
shown by the statistical analysis performed before (please see Fig. 4 value of the probability function after the perturbation and the one from
bottom, correlograms). Therefore, each node of the network shows a before (Fig. 7). Moreover, the dependences obtained have been divided
causal relationship with the remaining three. Another reason for this into spatial (same season, different sub-basin), temporal (different sea­
fully connected design is that it is the best way for assuring the complete son, same sub-basin) and spatio-temporal (different season, different
capture of the runoff dependence for the double dimension: space and sub-basin). The considered seasons are non-irrigation (NIP) and irriga­
time. tion (IP); while the sub-basins correspond to Alarcon (A) and Contreras
Finally, although this conceptual frameworks, by itself could be (C).
understood as results of the research, it has been considered appropriate It can be observed that spatial dependences (blue bars) are the
to show them as part of the methodological section, given the novelty of strongest ones, with values distinctly higher than 0 and higher than the
these approaches, in this way the reader is facilitated to understand of temporal and the spatio-temporal dependences. This bond is in accor­
the developed process. dance with the geographical proximity of both sub-basins. Furthermore,
dependences are higher for the non-irrigation season than during irri­
4.5. Concept-proof. Sensitivity analysis and validation of results gation. This may be caused by the higher rainfalls in the non-irrigation
season, since its geographic proximity drives similar rainfall events,
Sensitivity analysis and validation of results is addressed through the strengthening the bond between them. Another interesting aspect (also
development of an Entropy approach, based on the perspective of the noticed in the correlograms, see Fig. 4 bottom) is the fact that the in­
Theory of Information. In this sense, Conditional Entropy, Mutual in­ fluence of Alarcon subbasin on Contreras’ inflows is higher than in the
formation is calculated for the Bayesian causal models. opposite way, for both seasons. This asymmetry in their dependences
Sensitivity analysis can be performed using two types of measures; may be caused by groundwater flows (please see Fig. 2), which show an
entropy and Shannon’s measure of mutual information (Pearl, 1988). underground discharge from Alarcon sub-basin to Contreras sub-basin,
The entropy measure assumes that the uncertainty or randomness of a according to piezometric levels as it is shown CHJ (2006), CHJ (2020).
variable X, characterized by probability distribution P(x), can be rep­ Considering the temporal dependences, Contreras shows a higher
resented by the entropy function H(X) (Molina et al., 2016): bond than Alarcon, given the higher variation ratio noticed in Contreras.
H(X) = −
∑
P(x)logP(x) This agrees with the correlograms, in which this influence is clearly
x∕
∈X shown by the significant statistical evidence (see Fig. 4 bottom; Alarcon
& Contreras correlogram displays higher correlation coefficients, rk,
Entropy of a probability distribution can be defined as a measure of than Contreras & Alarcon correlogram in all analyzed monthly time
the associated uncertainty to that random process that this distribution lags). The reason behind this may be the geological differences found
describes. Consequently, a score of uncertainty/certainty level of events between both sub-basins, which can be explored using the information
can be made attending to this entropy, H(X). provided by the Geological Survey of Spain (Instituto Geológico y
Reducing H(X) by collecting information in addition to the current Minero de España, IGME, http://www.igme.es/default.asp). In this
knowledge about variable X is interpreted as reducing the uncertainty

8
H. Macian-Sorribes et al. Journal of Hydrology 597 (2021) 125722

Fig. 7. Average spatio-temporal dependences on the upper Jucar river basin.

sense, the groundwater bodies under the Alarcon sub-basin present a the other hand, spatial dependences are similar to the average ones. This
significant portion of detrital sedimentary rocks, while the aquifers seems to be caused by the fact that rainfall during the irrigation season is
related with the Contreras sub-basin are mainly chemical sedimentary low even during wet years, so the differences among years are less
rocks. Given that detrital rocks are associated with quicker hydrological relevant than the ones found for the non-irrigation season. Another
responses than chemical ones, the geological differences noticed may change between normal and dry years is that Contreras plays more in­
cause a slower hydrological response (and thus with longer time spans) fluence on Alarcon than Alarcon on Contreras during dry years (C (IP) &
in Contreras. A (IP) is greater than A (IP) & C (IP)). This inversion is also shown in the
Regarding the spatio-temporal dependences, the prevalence pattern spatio-temporal relationships (C (NIP) & A (IP) is greater than A (NIP) &
observed for the spatial ones (more influence on Alarcon in Contreras C (IP)). The temporal relationships do not suffer significant modifica­
than the opposite) is maintained from the non-irrigation to the irrigation tions, although Contreras does not show higher temporal dependences
season. In fact, Contreras runoff on non-irrigation does not seem to play than Alarcon.
any influence on the Alarcon runoff during irrigation, as the variation The causal relationships observed after a dry year (which in most
ratio is approximately 0. The reason behind this behavior may be the cases corresponds to a dry year) present a remarkable increase in tem­
maintenance across time of the bond observed in the spatial analysis. poral and spatio-temporal relationships with respect to the average
The lack of influence of Contreras on Alarcon seems to be caused by the dependences. This increase in the relationships with temporal compo­
limited groundwater flow from the Contreras sub-basin to the Alarcon nent may be caused by groundwater discharge, since its importance
one. It is also interesting notice that the spatio-temporal dependence grows during long periods without rainfall. Spatial relationships show
from Alarcon to Contreras (A (NIP) & C (IP)) is similar to the temporal higher levels than the ones found for the first dry year, but do not reach
dependence of Contreras (C (NIP) & C (IP)). the levels shown on average. This seems to point that groundwater ex­
Overall, results show a relevant asymmetry on the spatial and spatio- changes between sub-basins increase after a dry year, may be because
temporal influence of both basins each other. This may be due to the fact the geological differences between both sub-basins cause an unbalance
that Alarcon sub-basin plays more influence on Contreras than the between piezometric heads again (due to different hydrological
opposite, probably by the existence of groundwater flows following response times).
mainly the Alarcon-Contreras direction (piezometric gradient towards Two years after the start of the dry spell (which except for the most
the coastal area, from West to East (CHJ, 2006, 2020)). severe droughts found, is not dry) the temporal and spatio-temporal
dependences are, on a broader view, similar to the average. This is
5.2. Spatio-temporal dependence on dry years consistent with the ending of a dry period. However, spatial de­
pendences are lower than the average while showing similar levels
Given that dry spells in the Jucar river basin last more than one year, regardless of the sub-basin and season. This may be caused by a rise in
a new BCM has been developed extending its temporal-horizon up to 3 precipitations, since its influence in the bond between sub-basins is the
years (6 seasons, see Fig. 6 Bottom). According to dry year definition a same regardless of the direction of the relationship (from Alarcon to
set of 27 years were considered. In the most of them, both Alarcon and Contreras or from Contreras to Alarcon).
Contreras inflow during winter fall below the threshold, as well as
during the summer season. 5.3. Sensitivity and validation analysis
Fig. 8 shows the results of the causal relationship analysis done for
dry periods, starting at a dry year and spanning to 2 years after a dry one. Calibration of causal reasoning process is done through a compara­
Most of the years after a dry year were dry too, but a 3-year dry spell was tive analysis with traditional techniques such as correlation. In this
only found for the case of an extreme drought. sense, the results obtained from the causal reasoning analysis for the
The main difference between the average dependences and the ones average BN-PGM model resemble the ones provided by correlation
in a dry year is the absence of significant spatial relationships during the (please see Fig. 4 Bottom), as explained previously.
non-irrigation season, as the variation ratios are similar to 0 (please see On the other side, assessments of the entropy associated to each
Fig. 8 upper; A (NIP) & C (NIP) and C (NIP & A (NIP)). This may be variable, the conditional entropy and mutual information for each var­
caused by the absence of rainfall and the lowering of groundwater ta­ iable and connection were done for the average BN-PGM model (please
bles. The decrease in heads is steeper in the Alarcon sub-basin due to see Fig. 6 Upper, BCM conceptual framework for lumped analysis). In
having more detrital rock areas than Contreras, causing the interruption this sense, Fig. 9 summarizes the main achieved results. These outcomes
of the groundwater flow from Alarcon to Contreras. During irrigation, on of the entropy analysis agree with the causal relationship analysis (see

9
H. Macian-Sorribes et al. Journal of Hydrology 597 (2021) 125722

Fig. 8. Spatio-temporal dependences on the upper Jucar river basin during dry periods.

Fig. 7). In particular, the same prevalence pattern is observed: spatial causal relationship found between both nodes in comparison to the rest
dependences show the highest values of mutual information, then of spatio-temporal dependences. In spite of the similarity between them,
temporal and finally spatio-temporal dependences, whose values are the there exist some differences. The most important is that the results of the
lowest. mutual information between two nodes are the same regardless of the
Regarding the mutual information values of the spatial dependences direction considered (e.g. A (NIP) & C (NIP) has the same value than C
are higher in the non-irrigation period, as the causal relationships. (NIP) & A (NIP)), while the causal relationship analysis is sensitive to it.
Similarly, mutual information values on the temporal dependences of Moreover, the remarkable differences found between the spatial and the
Contreras are higher than the ones in Alarcon. Regarding the spatio- remaining causal relationships do not appear in the entropy analysis, in
temporal dependences, the one between A (NIP) and C (IP) has the which only slight modifications are found.
highest mutual information value, in accordance with the strongest On the other hand, Fig. 10 shows an entropy analysis of the BN-PGM

10
H. Macian-Sorribes et al. Journal of Hydrology 597 (2021) 125722

Fig. 9. Sensitivity analysis results for the average behavior of Alarcon (Ala) and Contreras (Con) sub-basins.

Fig. 10. Entropy analysis over dry cycle analysis performed over time-horizon of 3 year (6 seasons).

model for a dry period, which is also supported through HUGIN Expert® are higher in Contreras than in Alarcon (although with slight increases).
software. The entropy results clearly reflect the temporal mitigation of On a broader view, the results of the entropy analysis agree with the
dependences, since the mutual information values decrease when causal relationship analysis. The mutual information values for the first
moving forward or backward in time. The mutual information values year of a dry cycle are lower than on average, in accordance with the
distinctly decrease when moving more than one year apart from the causal relationship analysis, as well as the increase noticed in the mutual
hypothesis node. This is in line with the correlograms (see Fig. 4), which information values after the first year. Similarly, temporal mutual in­
show dependence for 12 months. The fact that Contreras has more formation values after a dry year increase, as causal relationships do.
temporal dependence than Alarcon is also shown in the entropy analysis, However, there are also points in which both analyses depart. For
since mutual information values associated with temporal dependences example, the mutual information values for the third year are in general

11
H. Macian-Sorribes et al. Journal of Hydrology 597 (2021) 125722

similar to the second, while the causal relationships showed a decrease system. If the influence of each variable can be assessed and ranked with
between those years. respect to the others, one can determine where investments in reducing
With respect to the spatio-temporal relationships, the mutual infor­ uncertainty (e.g. increasing the monitoring or the quality of the mea­
mation values for the non-irrigation season are higher in Alarcon than in surements) would be efficient. The fact of getting much more informa­
Contreras, while during the irrigation season they are similar except for tion on the interaction between adjacent basins (in series and/or in
the second year, in which Contreras shows higher values. This is parallel) is crucial for implementing an effective water resources man­
consistent with the causal relationship analysis (see Fig. 8) except for the agement largely due to predictive and management reasons.
first dry year, in which they disagree (entropy points at a higher influ­ BN models developed in this research have been populated, trained
ence from Alarcon on Contreras while the causal relationship shows the and validated using historical data. This could be done due to the exis­
opposite). In spite of the differences found, most of the pattern changes tence of enough records (72 years for the whole historical period and 27
noticed between the average behavior and a dry year found by the causal dry years). However, they would exist situations in which historical
relationship analysis can be drawn from the entropy analysis. Given this, records would not reach the minimum size to ensure a safe training. In
the BN-PGM for dry years can be considered valid. this case, suitable alternatives to extend the available time series would
consist in using stochastic processes like Monte Carlo simulation or
5.4. Comparison between BCM and traditional statistical analysis stochastic models, as done by Molina et al. (2016). However, these
techniques assume a specific spatio-temporal dependence structure, so
Although their distinct differences in terms of setup and interpreta­ they would influence the analysis of the causal relationship performed,
tion, BCM results (Fig. 7) agree in general terms with the ones obtained given that the generated data would follow a pattern that may not be in
from a traditional statistical correlation analysis (Fig. 4). In particular, line with the one intrinsic to the historical records.
both procedures find out that Contreras shows higher temporal depen­ Another alternative to extend the available runoff data would be the
dence than Alarcon (the correlogram of Contreras takes higher values use of hydrological models, whose assumptions are less tight in terms of
than Alarcon, as well as the temporal variation ratios), that the depen­ spatio-temporal dependences. However, these models would require
dence from Alarcon to Contreras is higher than from Contreras to enough historical records of meteorological variables (e.g. temperature
Alarcon (the correlogram between Alarcon and Contreras takes higher and rainfall) and information about the hydrological properties of the
values than the one from Contreras to Alarcon, as the spatial and spatio- sub-basins to be analyzed. Furthermore, hydrological models usually
temporal variations ratios); and that these dependences hold for the need more time to be set up than stochastic models.
whole hydrological year (correlograms rise above the Anderson limits The changes found out by BCM at the beginning of the dry period
while variation ratios depart from zero). An advantage of statistical open the possibility of predicting droughts from examining the evolu­
analysis is their well-developed theory, which is able to clearly identify tion of these dependences. In fact, BNs’ assessment of dependence is able
when a relationship is significant (using the Anderson limits), in contrast to determine how each variable depends on previous measurements,
to the use of variation ratios by BCM, since they are unbounded and no enabling its use as a predictor (Molina and Zazo, 2017). For the Jucar
precise threshold can be set to infer its significance. Although both river basin in particular, the absence of spatial dependence between
methodologies are sensitive to the amount of available data, BCM is Alarcon and Contreras sub-basins can be used to foresee water scarcity
more data-demanding than statistical analyses in general. periods and start activating drought management measures in advance.
However, BCM outperforms statistical analysis in other aspects. First Using the BCM defined in combination with water resource man­
of all, it is able to offer context-dependent information in a straight­ agement models, one can automatically trigger these measures accord­
forward way (e.g. how the relationships vary between normal and dry ing to the evolution of the causal relationship. These possibilities of
years, or between the irrigation season and the rest of the year), being predictive potential will be explored in upcoming research.
also possible to determine when this general behavior changes better From the methodology developed in this research and its application
than statistical analysis. Moreover, BCM does not require prior data to the case study, the following conclusions can be drawn:
transformations or standardizations before and after running the anal­
ysis, being also not bound to a particular distribution function or set of • Causal reasoning has been successfully applied to the upper Jucar
distribution functions. Although not explored in this research, BCM can river basin. The BN_PGM model developed has been able to identify
also generate probability functions conditioned to previous observations the dependences existing between runoff variables in both time (in-
of related variables in an efficient way compared to statistical year seasons) and space (sub-basins). These dependences are similar
procedures. to the ones provided by an entropy analysis and agree with the ones
found by a previous statistical analysis.
6. Discussion and conclusions • Causal reasoning identified how dependences change between
average and dry years for the upper Jucar river basin. They were also
Bayesian Causal Modelling, evaluated from Bayesian Networks able to characterize how these dependences evolve during a dry
perspective, has demonstrated in this “concept-proof” to be able to period. This ability to determine the entrance and exit of a drought
identify which relationships between variables are important, as well as enables the use of BNs as drought forecasting tool or as drought
how they change between an average behavior and a dry year. scenarios generation technique.
Compared with classic statistical analysis, neither prior knowledge nor
assumptions on the probability distribution of the variables involved is Additionally, this research demonstrates the suitability of BCM
required. On the contrary, Bayesian Networks learn it by themselves, methodology for spatio-temporal analysis of water resource behavior of
being also able to recursively add more information as soon as it is one or several basins, in a dynamic and stochastic way. In this sense BCM
available (Molina et al., 2016). is able to reveal a logical, hidden and non-trivial structure of interde­
The spatio-temporal dependences in the upper Jucar river sub-basins pendence underlying in the hydrological historical records.
analyzed through causal reasoning have been compared with an entropy On the other hand, the spatial approach proposed in this “concept-
analysis. Although some differences were found, both of them agreed in proof” based on BCM, and focused on dry periods (droughts), may be a
most of the main patterns found in the data, as well as the modifications suitable starting point for its application to other types of extreme events
suffered by these for a dry year. However, CR considers the direction of such as floods, within a context of flood damage prediction.
each relationship. This advantage is a distinct feature of causal Finally, this “concept-proof” reinforces that past information will
reasoning, making it able to determine which variables are playing the provide prior knowledge of the future, particularly useful in water re­
highest influence on the spatio-temporal relationships found in the sources planning and management of highly dependent basins.

12
H. Macian-Sorribes et al. Journal of Hydrology 597 (2021) 125722

CRediT authorship contribution statement IGME, 2020b. Hydrogeological map of Spain scale 1/1000.000. https://igme.maps.
arcgis.com/home/item.html?id=036292dc5b8946bd979a7dc47d2f8561 (Accessed
01.03.20).
Hector Macian-Sorribes: Conceptualization, Methodology. Jose- Jain, A., Kumar, A.M., 2007. Hybrid neural network models for hydrologic time series
Luis Molina: Conceptualization, Data curation, Software, Writing - forecasting. Appl. Soft Comput. 7, 585–592.
original draft. Santiago Zazo: Visualization, Investigation. Manuel Jia, Q.M., Li, Y.P., Li, Y.F., Huang, G.H., 2020. Analyzing variation of inflow from the Syr
Darya to the Aral Sea: a Bayesian-neural-network-based factorial analysis method.
Pulido-Velázquez: Supervision, Validation. J. Hydrol. 587, 124976.
Jona Lasinio, G., Sahu, S.K., Mardia, K.V., 2005. Modeling Rainfall Data using a Bayesian
Kriged-Kalman Model. Southhampton, UK.
Declaration of Competing Interest Jyrkama, M.I., Sykes, J.F., 2007. The impact of climate change on spatially varying
groundwater recharge in the Grand River Watershed (Ontario). J. Hydrol. 338 (3–4),
237–250.
The authors declare that they have no known competing financial Kahil, M.T., Dinar, A., Albiac, J., 2015. Modeling water scarcity and droughts for policy
interests or personal relationships that could have appeared to influence adaptation to climate change in arid and semiarid regions. J. Hydrol. 522, 95–109.
Kalra, A., Li, L., Ahmad, S., 2013. Improving streamflow forecast lead time using oceanic-
the work reported in this paper. atmospheric oscillations for Kaidu River Basin, Xinjiang, China. J. Hydrol. Eng. 18
(8), 1031–1040.
Kim, K., Lee, S., Jin, Y., 2018. Forecasting quarterly inflow to reservoirs combining a
Acknowledgements
copula-based Bayesian network method with drought forecasting. Water 10, 233.
Koller, D., Friedman, N., 2009. Probabilistic Graphical Models: Principles and
This work was supported by the by the IMPADAPT project Techniques. The MIT Press, Cambridge, Massachusetts, USA.
Kousari, M.R., Hosseini, M.E., Ahani, H., Hakimelahi, H., 2017. Introducing an
(CGL2013-48424-C2-1-R) with Spanish MINECO (Ministerio de Econo­
operational method to forecast long-term regional drought based on the application
mía y Competitividad) and European FEDER funds; and the European of artificial intelligence capabilities. Theor. Appl. Climatol. 127 (1–2), 361–380.
Union’s Horizon 2020 research and innovation programme under the Lappenschaar, M., Hommersom, A., Lucas, P.J.F., 2012. Qualitative chain graphs and
IMPREX project (GA 641.811). their use in medicine. Proceedings of the Sixth European Workshop on Probabilistic
Graphical Models. Proceedings of the Sixth European Workshop on Probabilistic
Graphical Models, Granada, Spain, 2012, pp. 179–185.
References Li, H., Wang, D., Singh, V.P., Wang, Y., Wu, J., Wu, J., et al., 2019. Non-stationary
frequency analysis of annual extreme rainfall volume and intensity using
Archimedean copulas: a case study in eastern China. J. Hydrol. 571, 114–131.
Allan, R.P., Soden, B.J., 2008. Atmospheric warming and the amplification of
Lima, A.R., Cannon, A., Hsieh, W.W., 2016. Forecasting daily streamflow using online
precipitation extremes. Science 321 (5895), 1481–1484.
sequential extreme learning machines. J. Hydrol. 537, 431–443.
Bayes, T., 1763. An essay towards solving a problem in the doctrine of chances. Philos.
Lopez-Nicolas, A., Pulido-Velazquez, M., Macian-Sorribes, H., 2017. Economic risk
Trans. Roy. Soc. Lond. 53, 370–418.
assessment of drought impacts on irrigated agriculture. J. Hydrol. 550, 580–589.
Cabañas de Paz, R., 2017. New Methods and Data Structures for Evaluating Influence
Macian-Sorribes, H., Pulido-Velazquez, M., 2017. Integrating historical operating
Diagrams. Ph.D. Thesis. University of Granada, Granada, Spain, p. 360 pp..
decisions and expert criteria into a DSS for the management of a multireservoir
Cain, J., 2001. Guidelines for using Bayesian Networks to Support the Planning and
system. J. Water Resour. Plann. Manage. Am. Soc. Civil Eng. 143 (1).
Management of Development Programmes in the Water Sector and Beyond. Centre
Madsen, A.L., Lang, M., Kjaerulff, U.B., Jensen, F., 2003. The Hugin tool for learning
for Ecology & Hydrology Wallingford, Wallingford, UK.
Bayesian networks. Symbol. Quant. Approaches Reason. Uncertain., Proc. 2711,
Castelletti, A., Soncini-Sessa, R., 2007. Bayesian Networks and participatory modelling in
594–605.
water resource management. Environ. Modell. Software 22, 1075–1088.
Marcos, P., Lopez-Nicolas, A., Pulido-Velazquez, M., 2017. Analysis of climate change
Chang, B., Guan, J., Aral, M.M., 2015. Scientific discourse: climate change and sea-level
impact on meteorological and hydrological droughts through relative standardized
rise. J. Hydrol. Eng. 20 (1), A4014003.
indices. EGUGA, 1391.
CHJ, 2006. Mapa piezométrico general de la confederación hidrográfica del Júcar. In
Marcos-Garcia, P., Lopez-Nicolas, A., Pulido-Velazquez, M., 2017. Combined use of
“Comprobación y evaluación en la Cuenca Piloto del río Júcar de las Guías
relative drought indices to analyze climate change impact on meteorological and
desarrolladas en el marco de la Estrategia común para la implementación de la
hydrological droughts in a Mediterranean basin. J. Hydrol. 554, 292–305.
Directiva Marco del Agua (in S. Valencia).
Mishra, A.K., Desai, V.R., 2005. Drought forecasting using stochastic models. Stoch.
CHJ, 2007. Plan Especial de Alerta y Eventual Sequia en la Confederación Hidrográfica
Environ. Res. Risk Assess. 19 (5), 326–339.
del Júcar (in Spanish). Valencia.
Mishra, A.K., Singh, V.P., 2010. A review of drought concepts. J. Hydrol. 391 (1–2),
CHJ, 2015. Plan Hidrológico de la Demarcación Hidrográfica del Júcar. Valencia.
202–216.
CHJ, 2020. Sistema de Información del Agua (SIA) de la Confederación Hidrográfica del
Mishra, A.K., Singh, V.P., 2011. Drought modeling – a review. J. Hydrol. 403 (1–2),
Júcar.Red Piezométrica Operativa.
157–175.
Dehghani, M., Saghafian, B., Zargar, M., 2019. Probabilistic hydrological drought index
Molina, J.L., Bromley, J., García-Aróstegui, J.L., Sullivan, C., Benavente, J., 2010.
forecasting based on meteorological drought index using Archimedean copulas.
Integrated water resources management of overexploited hydrogeological systems
Hydrol. Res. 50, 1230–1250.
using Object-Oriented Bayesian Networks. Environ. Modell. Software 25 (4),
De Michele, C., Salvadori, G., 2003. A Generalized Pareto intensity-duration model of
383–397.
storm rainfall exploiting 2-Copulas. J. Geophys. Res. Atmos. 108, 4067.
Molina, J.L., Pulido-Velazquez, D., Luis Garcia-Arostegui, J., Pulido-Velazquez, M., 2013.
Donat, M.G., Lowry, A.L., Alexander, L.V., O’Gorman, P.A., Maher, N., 2016. More
Dynamic Bayesian networks as a decision support tool for assessing climate change
extreme precipitation in the World’s dry and wet regions. Nat. Clim. Change 6 (5),
impacts on highly stressed groundwater systems. J. Hydrol. 479, 113–129.
508–513.
Molina, J.L., Zazo, S., Rodríguez-Gonzálvez, P., González-Aguilera, D., 2016. Innovative
Eum, H., Gupta, A., Dibike, Y., 2020. Effects of univariate and multivariate statistical
analysis of runoff temporal behavior through Bayesian networks. Water.
downscaling methods on climatic and hydrologic indicators for Alberta, Canada.
Molina, J.L., Zazo, S., 2017. Causal reasoning for the analysis of rivers runoff temporal
J. Hydrol. 588, 125065.
behavior. Water Resour. Manage. 1–13.
Estrela, M.J., Peñarrocha, D., Millán, M., 2000. Multi-annual drought episodes in the
Molina, J.L., Zazo, S., 2018. Assessment of temporally conditioned runoff fractions in
Mediterranean (Valencia region) from 1950–1996. A spatio-temporal analysis. Int. J.
unregulated rivers. J. Hydrol. Eng. 23, 04018015.
Climatol., John Wiley & Sons Ltd 20 (13), 1599–1618.
Molina, J.L., Zazo, S., Martín-Casado, A.M., 2019. Causal reasoning: towards dynamic
Freire-González, J., Decker, C., Hall, J.W., 2017. The economic impacts of droughts: a
predictive models for runoff temporal behavior of high dependence rivers. Water 11,
framework for analysis. Ecol. Econ. 132, 196–204.
877.
Gil, M., Garrido, A., Gómez-Ramos, A., 2011. Economic analysis of drought risk: an
Molina, J.L., Zazo, S., Martín-Casado, A., Patino-Alonso, M., 2020. Rivers’ temporal
application for irrigated agriculture in Spain. Agric. Water Manag. 98 (5), 823–833.
sustainability through the evaluation of predictive runoff methods. Sustainability 12,
Hao, Z., Singh, V.P., 2016. Review of dependence modeling in hydrology and water
1720.
resources. Prog. Phys. Geogr. 40 (4), 549–578.
Ochoa-Rivera, J.C., 2008. Prospecting droughts with stochastic artificial neural
Hejazi, M.I., Cai, X., 2011. Building more realistic reservoir optimization models using
networks. J. Hydrol. 352, 174–180.
data mining – a case study of Shelbyville Reservoir. Adv. Water Resour. 34 (6),
Pearl, J., 1988. Reasoning in Intelligent Systems: Networks of Plausible Inference.
701–717.
Morgan Kaufmann, San Francisco, USA.
Hipel, K.W., McLeod, A.I., 1994. Time Series Modelling of Water Resources and
Pearl, J., 2009. Causality: Models, Reasoning and Inference. Cambridge University Press,
Environmental Systems. Elsevier, Amsterdam, The Netherlands.
New York City, NY, USA.
Holmström, L., Ilvonen, L., Seppä, H., Veski, S., 2015. A Bayesian spatiotemporal model
Pearl, J., 2014. Graphical models for probabilistic and causal reasoning. In: Gonzalez, T.,
for reconstructing climate from multiple pollen records. Inst. Math. Stat. Ann. Appl.
Diaz-Herrera, J., Tucker, A. (Eds.), Computing Handbook, Third Edition: Computer
Stat. 9 (3), 1194–1225.
Science and Software Engineering, Chapman & Hall/CRC, Boca Raton, Florida, USA.
Hurst, H.E., 1951. Long-term storage capacity of reservoirs. Trans. Am. Soc. Civil Eng.
Peng, T., Zhou, J., Zhang, C., Fu, W., 2017. Streamflow forecasting using empirical
(ASCE) 116, 770–808.
wavelet transform and artificial neural networks. Water 9, 406.
IGME, 2020a. Map of Permeabilities of Spain scale 1/1000.000. https://igme.maps.
arcgis.com/home/webmap/viewer.html?
webmap=da8eb570845b41bbb5548c8266eaed0d (Accessed 01.03.20).

13
H. Macian-Sorribes et al. Journal of Hydrology 597 (2021) 125722

Roozbahani, A., Ebrahimi, E., Banihabib, M.E., 2018. A framework for ground water In: Deodatis, G., Ellingwood, B.R., Frangopol, D.M. (Eds.), Safety, Reliability, Risk
management based on bayesian network and MCDM techniques. Water Resour. and Life-cycle Performance of Structures and Infrastructures. CRC Press, London,
Manage. 32, 4985–5005. pp. 3123–3130.
Ropero, R.F., Nicholson, A.E., Aguilera, P.A., Rumi, R., 2018. Learning and inference Vogel, K., Riggelsen, C., Korup, O., Scherbaum, F., 2014. Bayesian network learning for
methodologies for hybrid dynamic Bayesian networks: a case study for a water natural hazard analyses. Nat. Haz. Earth Syst. Sci. 14 (9), 2605–2626. https://doi.
reservoir system in Andalusia, Spain. Stoch. Environ. Res. Risk Assess. 32, org/10.5194/nhess-14-2605-2014.
3117–3135. Vogel, K., Weise, L., Schröeter, K., Thieken, A.H., 2018. Identifying driving factors in
Said, A., 2006. The implementation of a Bayesian network for watershed management flood-damaging processes using graphical models. Water Resour. Res. 54,
decisions. Water Resour. Manage. 20, 591–605. 8864–8889.
Saghafian, B., Sanginabadi, H., 2020. Multivariate groundwater drought analysis using Wang, W., Chau, K., Cheng, C., Qiu, L., 2009. A comparison of performance of several
copulas. Hydrol. Res. 51, 666–685. artificial intelligence methods for forecasting monthly discharge time series.
Salas, J.D., Delleur, J.W., Yevjevich, V.M., Lane, W.L., 1980. Applied Modeling of J. Hydrol. 374 (3–4), 294–306.
Hydrologic Time Series. Water Resources Publications, Littleton, Colorado, USA. Wasko, C., Sharma, A., 2015. Steeper temporal distribution of rain intensity at higher
Sarabia-Alzaga, J., Gómez-Déniz, E., 2008. Construction of multivariate distributions: a temperatures within Australian storms. Nat. Geosci. 8 (7), 527–U166.
review of some recent results. Stat. Oper. Res. Trans. 32, 3–36. Wei, S., Zuo, D., Song, J., 2012. Improving prediction accuracy of river discharge time
See, L., Openshaw, S., 2000. A hybrid multi-model approach to river level forecasting. series using a Wavelet-NAR artificial neural network. J. Hydroinf. 14, 974–991.
Hydrol. Sci. J. J. Des Sci. Hydrol. 45, 523–536. Wikle, C.K., Berliner, L.M., Cressie, N., 1998. Hierarchical Bayesian space-time models.
Singh, V.P., 2011. Hydrologic synthesis using entropy theory: review. J. Hydrol. Eng. 16 Environ. Ecol. Stat. 5, 117–154.
(5), 421–433. Wu, F., Zhan, J., Su, H., Yan, H., Ma, E., 2015. Scenario-based impact assessment of land
Sperotto, A., Molina, J.L., Torresan, S., Critto, A., Marcomini, A., 2017. Reviewing use/cover and climate changes on watershed hydrology in Heihe River Basin of
Bayesian Networks potentials for climate change impacts assessment and Northwest China. Advances in Meteorology 410198.
management: a multi-risk perspective. J. Environ. Manage., Academic Press 202, Yang, T., Asanjan, A.A., Faridzad, M., Hayatbini, N., Gao, X., Sorooshian, S., 2017a. An
320–331. enhanced artificial neural network with a shuffled complex evolutionary global
Spirtes, P., 2010. Introduction to causal inference. J. Mach. Learn. Res. 11, 1643–1662. optimization with principal component analysis. Inf. Sci. 418, 302–316.
Trenberth, K., 2011. Changes in precipitation with climate change. Clim. Res. 47 (1–2), Yang, T., Asanjan, A.A., Welles, E., Gao, X., Sorooshian, S., Liu, X., 2017b. Developing
123–138. reservoir monthly inflow forecasts using artificial intelligence and climate
Tsoukalas, I., Efstratiadis, A., Makropoulos, C., 2019. Building a puzzle to solve a riddle: phenomenon information. Water Resour. Res. 53 (4), 2786–2812.
a multi-scale disaggregation approach for multivariate stochastic processes with any Yang, T., Gao, X., Sorooshian, S., Li, X., 2016. Simulating California reservoir operation
marginal distribution and correlation structure. J. Hydrol. 575, 354–380. using the classification and regression-tree algorithm combined with a shuffled
Vicente-Serrano, S.M., Cuadrat-Prats, J.M., 2007. Trends in drought intensity and cross-valid+ation scheme. Water Resour. Res. 52 (3), 1626–1651.
variability in the middle Ebro valley (NE of the Iberian peninsula) during the second Zazo, S., Macian-Sorribes, H., Sena-Fael, C.M., Martín-Casado, A.M., Molina, J.L., Pulido-
half of the twentieth century. Theor. Appl. Climatol. 88 (3), 247–258. Velazquez, M., 2019. Qualitative approach for assessing runoff temporal dependence
Vicente-Serrano, S.M., Gonzálezlez-Hidalgo, J.C., de Luis, M., Ravents, J., 2004. Drought through geometrical symmetr. (Contributed Paper). Proceedings Internacional
patterns in the Mediterranean area. Clim. Res., Inter-Res. Sci. Center 26 (1), 5–15. Congress on Engineering. Engineering for Evolution (ICEUBI2019). 27–29 November
Vogel, K., Riggelsen, C., Merz, B., Kreibich, H., Scherbaum, F., 2012. In: Flood damage 2019, Covilhã, Portugal.
and influencing factors: A Bayesian network perspective. University of Granada, Zazo, S., 2017. Analysis of the Hydrodynamic Fluvial Behaviour through Causal
Granada, Spain, pp. 347–354. Reasoning and Artificial Vision. Ph.D. Thesis. University of Salamanca, Ávila, Spain,
Vogel, K., Riggelsen, C., Scherbaum, F., Schröter, K., Kreibich, H., Merz, B., 2013. p. 159.
Challenges for Bayesian network learning in a flood damage assessment application.

14