Hydrogeol J

DOI 10.1007/s10040-017-1636-7

REPORT

Application of Archimedean copulas to the impact assessment

of hydro-climatic variables in semi-arid aquifers of western India
Pawan S. Wable 1 & Madan K. Jha 1

Received: 20 January 2017 / Accepted: 30 June 2017

# Springer-Verlag GmbH Germany 2017

Abstract The effects of rainfall and the El Niño Southern exploring the impacts of rainfall and ENSO on groundwater at
Oscillation (ENSO) on groundwater in a semi-arid basin of basin scales.
India were analyzed using Archimedean copulas considering
17 years of data for monsoon rainfall, post-monsoon ground- Keywords Archimedean copulas . Groundwater
water level (PMGL) and ENSO Index. The evaluated depen- vulnerability . El Niño-Southern Oscillation (ENSO) . Climate
dence among these hydro-climatic variables revealed that change . India
PMGL-Rainfall and PMGL-ENSO Index pairs have signifi-
cant dependence. Hence, these pairs were used for modeling
dependence by employing four types of Archimedean cop- Introduction
ulas: Ali-Mikhail-Haq, Clayton, Gumbel-Hougaard, and
Frank. For the copula modeling, the results of probability Groundwater contributes to one-third of the global freshwater
distributions fitting to these hydro-climatic variables indicated supply, which supports a population of over 2 billion
that the PMGL and rainfall time series are best represented by (Gorelick and Zheng 2015). In India, groundwater is a source
Weibull and lognormal distributions, respectively, while the of water supply for more than 80% of the rural and 50% of the
non-parametric kernel-based normal distribution is the most urban populations, and for 50% of irrigation demand, which
suitable for the ENSO Index. Further, the PMGL-Rainfall pair contributes to 70–80% of irrigated production (Mall et al.
is best modeled by the Clayton copula, and the PMGL-ENSO 2006). There is a consensus among climate scientists that
Index pair is best modeled by the Frank copula. The Clayton global warming will intensify, accelerate, or enhance the water
copula-based conditional probability of PMGL being less than cycle, which will have important consequences for the world’s
or equal to its average value at a given mean rainfall is above freshwater resources (UNESCO 2009). Although the effects
70% for 33% of the study area. In contrast, the spatial varia- of climate change on water resources are already visible
tion of the Frank copula-based probability of PMGL being worldwide, the greatest concern of water experts is its impact
less than or equal to its average value is 35–40% in 23% of on groundwater as this is a more dependable source of water
the study area during El Niño phase, while it is below 15% in supply for domestic, irrigation and industrial sectors (e.g.,
35% of the area during the La Niña phase. This copula-based Holman 2006; Gurdak et al. 2009; Gorelick and Zheng
methodology can be applied under data-scarce conditions for 2015). Groundwater is affected by climate through major hy-
drological processes such as precipitation, evapotranspiration,
and runoff as well as through interaction with surface-water
* Pawan S. Wable
bodies. The extremes of climate (droughts and floods) are
[email protected] often related to the drivers of climate variability, i.e., large-
* Madan K. Jha
scale climatic patterns/oscillations such as the North Atlantic
[email protected] Oscillation (NAO), Arctic Oscillation (AO), Pacific Decadal
Oscillation (PDO), El Niño Southern Oscillation (ENSO), etc.
1
Agricultural and Food Engineering Dept., IIT Kharagpur, These large-scale and long-term climatic cycles can have the
Kharagpur 721 302, India most discernible impacts on groundwater due to slow aquifer
 Hydrogeol J

recharge processes and long aquifer response times (Russo 2–7 years (IRI 2017). During the warm phase (El Niño),
et al. 2014); however, proper assessment of climate-change the sea surface temperature is anomalously warm, while
impacts on groundwater is challenging due to its complex in the cool phase (La Niña), it is anomalously cool. In a
relations with hydro-climatic variables (Mishra and Singh recent study, it was found that 10 out of the past 13
2010). This challenge is further complicated for data-scarce droughts in India had a high correlation with the El
developing countries. Niño phase (Singh 2014). A study conducted in an arid
In the recent past, several researchers ascertained the con- region of Western Rajasthan (India) revealed that the
nection between groundwater and large-scale climate patterns droughts during El Niño phases were more severe than
across the world (e.g., Jones and Banner 2003; Fleming and La Niña phases (Ganguli and Reddy 2013). In another
Quilty 2006; Hanson et al. 2006; Luque-Espinar et al. 2008; study, Reddy and Ganguli (2012b) applied bivariate
Gurdak et al. 2009; Tremblay et al. 2011; Perez-Valdivia et al. copulas and reported that during La Niña (El Niño),
2012). Fleming and Quilty (2006) studied this link in the there was higher (lower) precipitation and shallow
aquifer of southwest British Columbia, Canada, and reported (deeper) groundwater levels in the Manjra River basin
that the groundwater of the area has a significant correlation of western India.
with seasonal ENSO-related precipitation anomalies, i.e., The preceding reviews suggest that the association
groundwater levels are higher during La Niña years and between hydro-climatic variables can be studied using
lower during El Niño years. Using wavelet and coherence emerging tools and techniques. One of such techniques
analysis, Tremblay et al. (2011) investigated the links of cli- is copula functions, which have been extensively used
matic oscillations (NAO, AO, PDO, Pacific-Northern- in hydro-meteorological studies (ICSH 2017). The cop-
America-Pattern, and multivariate ENSO Index) with the ula technique has salient features, which are helpful in
groundwater levels of three unconfined aquifers in Canada hydrological studies: (1) it derives joint distributions
and concluded that the inter-annual cycles observed in large- independent of the marginal, (2) along with the compos-
scale climatic patterns were also found in groundwater levels, ite likelihood approach, it reduces uncertainty in the
thereby suggesting substantial influence of these climatic pat- estimates of frequency distribution parameters, and (3)
terns on groundwater. All of these studies confirm that there it handles non-linearity for modeling dependence be-
exists a definite linkage between groundwater levels and long- tween random variables (Genest and Favre 2007;
term climatic cycles occurring thousands of kilometers away Chowdhary and Singh 2010). Although the copula-
from the area/basin under study. based bivariate, trivariate and quadravariate analyses
India receives around 70–90% of its rainfall from the have been used for the frequency analysis of extreme
southwest monsoon and its interannual variability is events (Salvadori and De Michele 2004; Shiau et al.
mostly influenced by the large-scale climatic pattern of 2007; Kao and Govindaraju 2008; Karmakar and
ENSO (Revadekar et al. 2012). The ENSO is a climatic Simonovic 2009; Wong et al. 2010), the use of bivariate
phenomenon, which affects global climate variability analysis is mostly preferred due to its simplicity (Klein
owing to the interaction between the tropical Pacific et al. 2011).
Ocean and its surrounding atmosphere. It has two It is also apparent from the aforementioned review that
phases, El Niño and La Niña, which are linked to sea to date, only one study (Reddy and Ganguli 2012b) has
surface temperatures and approximately alternates every applied copula models for the risk assessment of changes

Table 1 Copula cumulative distribution functions, their generator function φθ, and the relation of Kendall’s tau (τ) with parameter θ

Family C(u, v) φθ Relation of Kendall’s tau (τ) with θ

Ali-Mikhail-Haq ðu;vÞ

 3θ−2 3θ−2
½1−θð1−uÞð1−vÞ ln ½1−θðt1−tÞ 3θ − 3θ lnð1−θÞ

Clayton 1  −θ  θ
(u−θ + v−θ − 1)−1/θ θ t −1 ðθþ2Þ

Gumbel-Hougaard 1 ðθ−1Þ
expð−½ð−lnuÞ θ þ ð−lnvÞ θ θ Þ (−ln t)−θ θ

Frank    4
1 ðe−θu −1Þðe−θv −1Þ ðe−θt −1Þ 1þ ½D1 ðθÞ−1
− θ ln 1 þ ðe−θ −1Þ
−ln ðe−θ −1Þ θ

x
Note: Dk(x) is Debye function; for any positive integer k, Dk ðxÞ ¼ k=xk ∫0 t k ðet −1Þ dt
Hydrogeol J

Gujarat

China Madhya Pradesh

Pakistan

Chhattisgarh
Nepal

Bangladesh Maharashtra
India

Andhra Pradesh

Bay of Bengal Study Area

Arabian Sea Arabian Sea

Karnataka

Fig. 1 Location of the study basin along with boundaries of administrative ‘blocks’ and locations of gauging sites

in hydro-climatic variables on groundwater; however, this these research gaps and increasing drought incidences in
study considered only one observation well as a represen- India, this study was carried out to address some of the
tative for the entire river basin (143.32 km2), which is not aforementioned research gaps considering the Sina River
practically appropriate because the findings of the study basin as a study area, which is located in the semi-arid
may not be useful for the entire basin. To address this region of Maharashtra, western India, and it comes under
shortcoming, the present study was conceived to explore the ‘chronically drought-prone area’ (PACS 2004). As a
the applicability of the copula technique at a larger scale result, frequent droughts occur and water scarcity is a
(e.g., basin/sub-basin scale) so as to ensure more realistic serious problem in the area (Chary et al. 2010; DTE
findings for the area under study. Also, the studies on 2016; News World India 2016). In addition, to the best
groundwater linkage with the ENSO phenomenon are of the authors’ knowledge, no scientific study has been
very limited in developing countries in general and the conducted so far in the study area and the present study is
Indian subcontinent in particular (Reddy and Ganguli first of its kind in the area. The specific objectives of this
2012b; Susilo et al. 2013; Seeboonruang 2014). Given study are: (1) to evaluate dependence among hydro-
 Hydrogeol J

Fig. 2 Locations of raingauge

stations and areas of their
Thiessen polygons

 
climatic variables, (2) to model joint dependence between where F X 1 ;X 2 ::…X n x1 ; x2 ;…; xn stands for the joint cumula-
hydro-climatic variables using suitable Archimedean cop- tive distribution function (CDF) with continuous marginal
ulas, and (3) to explore the effect of hydro-climatic vari- distributions F X 1 ðx1 Þ; F X 2 ðx2 Þ; :…; F X n of the random vari-
ables (rainfall and ENSO phenomena) on groundwater ables X1, …, Xn and C is a copula, i.e., a CDF whose margins
using copula-based conditional distributions. The method- are uniform on the interval (0, 1).
ology used to fulfill these objectives is an extension of
that reported by Reddy and Ganguli (2012b).
Table 2 Number of observation wells, area and percentage of total area
under the Thiessen polygons of nine raingauge stations
Overview of copulas Name of station No. of observation wells Area (km2) Area (%)

A copula is a multivariate probability distribution having uni- Alni 16 1,827.75 15

form marginal distribution of random variables. It can repre- Chinchondipatil 23 1,814.56 15
sent and model dependence between associated random vari- Jamked 39 2,992.47 24
ables irrespective of their marginal distributions. Sklar’s theo- Kasegaon 18 1,625.90 13
rem (Sklar 1959) states that every joint distribution F can be Tembhurni 21 1,683.57 14
expressed as: Supa 4 501.48 4
  Solapur 5 1,055.50 9
F X 1 ;X 2 ::…X n x1 ; x2 ;…; xn Kolgaon 3 507.98 4
Bandalgi 3 235.23 2
¼ C ½ F X 1 ðx1 Þ; F X 2 ðx2 Þ; :…; F X n  ð1Þ
Hydrogeol J

Table 3 Expressions of probability density function and parameter estimation for parametric and non-parametric distributions used in the study

Distribution Probability density function Parameter estimation

Parametric distributions
Gamma h qffiffiffiffiffiffiffiffiffiffiffiffii
1
f ðxÞ ¼ xα−1 e−x=β ; x > 0 A ¼ lnðxÞ− ∑lnnðxÞ,β ¼ 4A
1
1 þ 1 þ 4A 1
;α ¼ βx
β α ΓðαÞ
α ¼ shape parameter; β ¼ scale parameter
Lognormal 0 1
∑yi
1 B 1 C μy ¼
f X ðxÞ ¼ qffiffiffiffiffiffiffiffiffiffiffiffi exp@− qffiffiffiffiffiffiffiffiffiffiffiffiA; x > 0 n
∑y
x 2πσ y 2
x 2πσ2 y σ2 y ¼ i
n

Y ¼ lnðXÞ and − ∞ < μy < ∞

μy ¼ mean ; σy ¼ standard deviation
Weibull
f X ðxÞ ¼ αxα−1 β −α exp½ð−x=β Þα ; x≥0; α; β > 0 Estimated by iterative procedure
α ¼ shape parameter; β ¼ scale parameter

Nonparametric kernel density-based distributions

Normal
1=5
K ðxÞ ¼ p1ffiffiffiffi expð−x2 =2Þ ; −∞ < x < ∞
2π Optimal bandwidth ¼ σ 4 3n
K(•) = Kernel density function
Quadratic
3ð1−x2 Þ
K ðxÞ ¼ 4 for jxj ≤1; otherwise K ðxÞ ¼ 0

This study focuses on two variables at a time, i.e., n = 2. pseudo-inverse is denoted ϕ−1. Various parametric clas-
These variables are denoted X and Y; their joint distribution ses of Archimedean copulas are listed in Table 1, in
FX,Y can be expressed in terms of their CDFs FX and FY as terms of their CDF, generator function and other prop-
follows: erties. For more information about copulas and their

 application, the interested reader can refer to Salvadori
F X ;Y ðx; yÞ ¼ C F x ðxÞ; F y ðyÞ ¼ C ½u; v ð2Þ and De Michele (2007), Genest and Nešlehová (2012a,
b) or Genest and Chebana (2016).
where C is unique whenever FX and FY are continuous, else
uniquely estimated by range FX × range FY.
Methodology
Archimedean copulas
Study area
In general, the copula C in Eq. (2) is assumed to come
from a parametric class. Archimedean copulas, elliptical
For the present study, the Sina River basin was selected as the
copulas, and extreme-value copulas families have been
study area. This basin is located in Maharashtra, western India
applied in hydrological studies. However, Archimedean
(Fig. 1), between 17° 28′ N and 19° 16′ N latitude, 74° 28′ E
copulas are most frequently used due to their flexibility
and 76° 7′ E longitude. The basin has an area of 12,244 km2,
and simplicity (e.g., Genest and MacKay 1986; Zhang
with the topographic elevation ranging from 420 to 964 m
and Singh 2006; Klein et al. 2011) and, hence, this type
(above mean sea level; MSL). It comprises four districts,
was selected for this study. A bivariate copula C is said
namely Ahmednagar, Beed, Osmanabad and Solapur, but
to be Archimedean if it can be written in the following
the largest portion (42%) of the basin falls in Solapur district.
form:
The 19 smaller subdivisions, i.e., blocks, for these four admin-
C ðu; vÞ ¼ ϕ−1 ðϕðuÞ þ ϕðvÞÞ ð3Þ istrative districts are shown by different colors in Fig. 1. The
average maximum and minimum air temperatures are 40.5 °C
where the generator ϕ is a function ϕ: [0, 1] → [0, ∞] in the month of May and 10.5 °C in the month of December,
which is convex, decreasing and such that ϕ(1) = 0. Its respectively. The rainy season extends from mid-June to the
 Hydrogeol J

Data

Hydro-meteorological data used in this study were collected from

various government organizations/agencies. Daily rainfall data of
nine raingauge stations for the period of 1985–2009 were col-
lected from India Meteorological Department (IMD), Pune and
State Data Storage Center, Hydrology Project (HP), Nashik,
India. It should be noted that the World Meteorological
Organization recommendation of 1 rainfall station per 600–
900 km2 for plain areas could not be met, so the rainfall data
were supplemented by data from stations in the vicinity (outside)
the study area, to better represent the spatial average rainfall
within the study area. Pre-monsoon (May month) and post-
monsoon (October month) groundwater-level data of 132 sites
(observation wells) over the basin for the 1985–2009 period were
also acquired from the Groundwater Survey and Development
Agency (GSDA), Pune, India. These data are from unconfined
aquifers, which are predominant in the study area. The locations
of observation wells and raingauge stations are shown in Fig. 1.
Groundwater-level data for many sites and for some years are
missing from the 1985–2009 dataset; this is a common problem
in most developing nations of the world. As a result, the appli-
cation of time-series analysis techniques under limited-data con-
ditions becomes a challenging task for the researchers of devel-
oping nations; therefore, considering the low availability and
continuity of time-series groundwater-level data in the study area,
the present study was carried out under data-scarce conditions. In
this study, 17 years (1990–2006) of groundwater-level and rain-
fall data have been used to investigate the applicability of the
copula technique at larger scale. Thiessen polygons were created
using the rainfall stations available in the study area (Fig. 2). The
areas of the Thiessen polygons and the number of observation
wells falling within each Thiessen polygon are given in Table 2.
It is worth also mentioning that in some of the recent studies on
Fig. 3 Temporal variation of mean monthly rainfall for a upper, b copula modeling, limited datasets (15–18 years) have been used
middle, c lower part stations of the study area during 1990–2006 period (e.g., Durocher et al. 2016; Reddy and Ganguli 2012a).
The impacts of hydro-climatic factors on groundwater are
reflected in recharge and discharge processes occurring in a
groundwater basin. However, detailed information about these
end of October. The average annual rainfall of the study area is processes are often lacking at a basin scale, especially in the
644 mm; most of the rainfall occurs due to the southwest developing world. Generally, groundwater level is monitored
monsoon. and, hence, it is easily available data compared to other com-
Geologically, the study area is underlain by Deccan basalts, ponents of groundwater. In fact, spatio-temporal variations of
which are composed of vesicular amygdaloidal basalt and groundwater levels in a basin are the outcome of spatially and
fraction jointed basalt (Deolankar 1980). The water-bearing temporally varying recharge and discharge processes occur-
formations are generally shallow unconfined or semi- ring in the basin. Given this fact and the unavailability of other
confined aquifers in the cover of weathered or fractured upper groundwater-related data in the study area, the effects of
portions of Deccan basalts, along with a patch of local alluvi- hydro-climatic factors on groundwater have been explored
um. The depth of weathered/fractures zones under unconfined in this study using groundwater-level data.
conditions ranges from 7.2 to 22.5 m below the ground level. The ENSO phenomenon is well represented by a recent
Specific yield (effective porosity) of the unconfined aquifers index known as ‘multivariate ENSO Index (MEI)’. MEI is
ranges from 0.010 to 0.026, which indicates relatively low defined using the first un-rotated principal component of six
storage capability of the aquifers. observed variables—sea-level pressure, zonal and meridional
Hydrogeol J

Fig. 4 Pre-monsoon and post-

monsoon groundwater-level fluc-
tuations for a upper, b middle, c
lower parts of the study area dur-
ing 1990–2006 period

Fig. 5 Monthly variation of 4

ENSO Index during 1990–2006 3
period 2
ENSO Index

1
0
Jan-90
Jul-90
Jan-91
Jul-91
Jan-92
Jul-92
Jan-93
Jul-93
Jan-94
Jul-94
Jan-95
Jul-95
Jan-96
Jul-96
Jan-97
Jul-97
Jan-98
Jul-98
Jan-99
Jul-99
Jan-00
Jul-00
Jan-01
Jul-01
Jan-02
Jul-02
Jan-03
Jul-03
Jan-04
Jul-04
Jan-05
Jul-05
Jan-06
Jul-06

-1
-2
-3
-4
Time (month-year)
 Hydrogeol J

Fig. 6 Variation of standardized

values of ENSO Index, post-
monsoon groundwater levels
(PMGL) and monsoon rainfall for
the 1900–2006 period at rainfall
stations: a Alni, b
Chinchondipatil, c Jamkhed, and
d Kasegaon

components of the surface wind, sea surface temperature, sur-

face air temperature and total cloudiness fraction of the sky
over the tropical Pacific (Wolter and Timlin 2011). Monthly Table 4 Summary of Spearman’s rho (ρ) values for the pairs of PMGL-
Rainfall, PMGL-ENSO Index, and Rainfall-ENSO Index
MEI values for the study period (1990–2006) were obtained
from the database provided by the National Oceanic and Name of station Spearman’s rho (ρ)
Atmospheric Administration (NOAA 2017).
PMGL- PMGL- Rainfall-
Rainfall ENSO Index ENSO Index
Dependence among the hydro-climatic variables
Alni 0.86 * −0.43 −0.51**
(p ≤ 0.01) (p = 0.09) (p = 0.04)
Before evaluating dependence, the data independency in all the Chinchondipatil 0.78* −0.47 −0.29
time series was checked by an auto-correlation test (Ljung-Box (p ≤ 0.01) (p = 0.06) (p ≥ 0.10)
Jamked 0.68* −0.53** −0.29
Q-test). For evaluating dependence, rank-based (and hence scale- (p ≤ 0.01) (p = 0.03) (p ≥ 0.10)
free) measures of dependence, such as Spearman’s rho (ρ) and Kasegaon 0.70* −0.47 0.68*
(p ≤ 0.01) (p = 0.06) (p ≤ 0.01)
Kendall’s tau (τ) are preferred over Pearson’s correlation coeffi- Tembhurni 0.76* −0.54** −0.40
(p ≤ 0.01) (p = 0.03) (p ≥ 0.10)
cient, given that they do not rely on any assumption of linearity Supa 0.59** −0.28 −0.58**
between the random variables and are not affected by outliers (p = 0.02) (p ≥ 0.10) (p = 0.02)
Solapur 0.63** −0.53** −0.32
(Klein et al. 2011). In this study, Spearman’s ρ was used to (p = 0.01) (p = 0.03) (p ≥ 0.10)
evaluate dependence among hydro-climatic variables at all the Kolgaon 0.86* −0.39 −0.45
(p ≤ 0.01) (p ≥ 0.10) (p = 0.07)
nine raingauge stations for the 17 years period (1990–2006). The Bandalgi 0.33 −0.66** −0.17
elevation of post-monsoon groundwater levels (PMGL) was con- (p ≥ 0.10) (p = 0.05) (p ≥ 0.10)
sidered instead of depth to groundwater below the ground surface Note: Values in brackets show p-value; values with * indicate significant
in order to maintain a common datum for all the groundwater- dependence at 1% level of significance; values with ** indicate signifi-
monitoring sites. cant dependence at 5% level of significance
Hydrogeol J

Fig. 7 Cumulative distribution

function of gamma, lognormal,
and Weibull distributions fitted to
post-monsoon groundwater levels
(PMGL) in the zones of rainfall
stations: a Chinchondipatil and b
Jamkhed

In standard climatology, variables affected by large-scale cli- (GM), lognormal (LN) and Weibull (WB); however, for the
matic patterns should be averaged over the area (Fleming and ENSO Index, non-parametric kernel-density-based normal
Quilty 2006); hence, to study the effect of monsoon rainfall and and quadratic distributions were considered because paramet-
the ENSO phenomenon on groundwater levels, post-monsoon ric distributions do not fit climate indices properly (Reddy and
groundwater levels for the observation wells of a particular Ganguli 2012b). The probability density functions and param-
Thiessen polygon were averaged. To perform this analysis, the eter estimates for the parametric and non-parametric kernel-
cumulative monsoon rainfall and the average of monthly MEI density-based distributions are shown in Table 3. In all cases,
values for the period June to October were used. The presence of the estimates were obtained using the method of maximum
dependence between each pair of hydro-climatic variables was likelihood. The best distribution was selected based on select-
examined at the 1 and 5% levels of significance based on the p- ed univariate statistical indicators—root mean square error
values of the standard two-tailed t-test. It is worth mentioning that (RMSE), Akaike information criterion (AIC), and
to reject the null hypothesis, the p-value should be less than or Kolmogorov-Smirnov (KS) test—and a graphical indicator
equal to the level of significance (α). For a visual illustration, the (cumulative distributive function plot).
variation of hydro-climatic variables over their standardized val-
ue was also plotted.
Archimedean copulas for modeling dependence

Fitting marginal distributions to hydro-climatic variables A priori, the choice of parametric Archimedean copulas fam-
ilies as possible models for the dependence between hydro-
After evaluating the dependence, marginal distributions were climatic variables is guided by the range of association they
fitted to each of the variables. For PMGL and rainfall, the most allow. The Clayton and Gumbel-Hougaard copulas are used if
popular parametric distributions were used, namely gamma the dependence is positive, whereas Ali-Mikhail-Haq and

Table 5 Performance evaluation of different probability distributions fitted to PMGL

Name of station AIC RMSE KS-test

GM LN WB GM LN WB GM LN WB

Alni −57.51 −57.54 −58.09 0.07299 0.06905 0.07817 0.1561 0.1497 0.1748
Chinchondipatil −69.83 −69.85 −68.83 0.07309 0.07020 0.07425 0.1582 0.1541 0.1889
Jamked −63.77 −63.81 −62.72 0.08630 0.08271 0.08328 0.1599 0.1532 0.1604
Kasegaon −64.84 −64.87 −66.04 0.06652 0.06328 0.06643 0.1271 0.1218 0.1235
Tembhurni −70.38 −70.38 −72.75 0.06409 0.06137 0.08117 0.1560 0.1495 0.1545
Supa −69.67 −69.70 −69.89 0.06911 0.06519 0.07549 0.1348 0.1290 0.1364
Solapur −72.67 −72.72 −70.80 0.04160 0.04280 0.05189 0.1159 0.1104 0.1042
Kolgaon −66.80 −66.83 −66.28 0.05566 0.05361 0.05448 0.1584 0.1548 0.1051
Bandalgi −71.60 −71.63 −72.21 0.06895 0.06581 0.08279 0.1699 0.1627 0.1568

Note: KS critical value at 5% level of significance = 0.330; italic value indicate the best value for particular evaluation criteria
 Hydrogeol J

Table 6 Parameters of the probability distributions fitted to PMGL the generator function for each copula family with its deriva-
Name of station GM LN WB tive, together with the relation of Kendall’s τ with copula
parameter θ, are presented in Table 1. Copula modeling was
α β μ σ α β performed using MATLAB software.
Alni 145,269.9 0.0039 6.333 0.0027 432.5 563.8
Goodness-of-fit tests for selecting copulas
Chinchondipatil 87,448.2 0.0072 6.442 0.0035 356.8 628.4
Jamked 99,977.8 0.0056 6.331 0.0033 386.4 562.4
Goodness-of-fit tests can be used to check whether a specific
Kasegaon 70,408.9 0.0069 6.187 0.0039 289.4 487.2
copula family fits the data at hand. In this study, both graphical
Tembhurni 54,545.3 0.0092 6.221 0.0044 243.1 504.5
and statistical indicators were used to assess the fitness of
Supa 95,811.4 0.0068 6.483 0.0033 356.8 654.9
Archimedean copulas.
Solapur 40,973.9 0.0114 6.146 0.0051 250.1 468.1
Kolgaon 99,020.4 0.0062 6.415 0.0033 370.4 611.8 Graphical diagnostics
Bandalgi 40,987.5 0.0110 6.115 0.0051 228.8 453.7
In order to assess the fit of a given Archimedean copulas
family Cθ, 1,000 observations were generated from Cθ after
estimating its parameter. These pseudo-observations were
Frank copulas are applied for modeling both positive and neg- then transformed back into the variables’ original units using
ative dependence. The Frank copula can model the entire the inverses of the marginal distribution FX and FY. The scatter
range of dependence values [−1, +1], whereas the Ali- plot of the resulting pairs was then visualized and compared to
Mikhail-Haq family of copulas is only suitable for weakly the original data. Algorithms to generate random pairs from
dependent variables (Nelsen 2006). There exists a connection different copula families (Cθ) can be found in Whelan (2004)
between a rank-based non-parametric measure of dependence and Genest and Favre (2007).
called Kendall’s τ and Archimedean copulas generators,
which is given as follows (Genest and MacKay 1986): Statistical indicators

Apart from the graphical diagnostics, three statistical indica-

Z tors for bivariate copulas were used in this study, namely
ϕðt Þ
τ ¼1þ4 dt ð4Þ RMSE; AIC; and KS goodness-of-fit test. Detailed descrip-
ϕ0
tions of these statistical indicators can be found in Klein et al.
where ϕ′ denotes the derivative of ϕ with respect to t. This (2011).
relation can be used to estimate the parameter θ of an
Archimedean copula by the method of moments, which con- Effect of rainfall and the ENSO phenomenon
sists of replacing τ by an estimate thereof in Eq. (4) and solv- on groundwater
ing for θ for any given choice of Archimedean copulas; thus,
paired random variables can be modeled through copulas by In order to study the impacts of rainfall and the ENSO phe-
preserving their mutual dependence. In this study, four fami- nomenon on groundwater, the copula-based conditional dis-
lies of Archimedean copulas (Clayton, Gumbel-Hougaard, tribution probabilities of PMGL ≤ PMGLavg for average and
Ali-Mikhail-Haq and Frank) were applied. The expression of non-average monsoon rainfall scenarios as well as for ENSO

Fig. 8 Cumulative distribution

function of gamma, lognormal,
and Weibull distributions fitted to
rainfall at stations: a
Chinchondipatil and b Jamkhed
Hydrogeol J

Table 7 Performance evaluation of different probability distributions fitted to rainfall

Name of station AIC RMSE KS-test

GM LN WB GM LN WB GM LN WB

Alni −231.3 −229.5 −235.9 0.1006 0.08619 0.1238 0.1733 0.1519 0.2200
Chinchondipatil −225.3 −226.8 −225.2 0.05393 0.06218 0.05915 0.1337 0.1694 0.1424
Jamked −221.3 −219.9 −225.0 0.1038 0.09247 0.1139 0.2152 0.2022 0.1998
Kasegaon −231.3 −230.6 −233.7 0.07418 0.06115 0.09350 0.1458 0.1227 0.1805
Tembhurni −221.1 −221.2 −222.3 0.04689 0.04316 0.05545 0.09304 0.1027 0.08620
Supa −205.9 −205.3 −208.5 0.07675 0.07048 0.09537 0.1708 0.1510 0.2057
Solapur −227.8 −226.7 −230.7 0.07545 0.06410 0.09496 0.1422 0.1251 0.1851
Kolgaon −213.5 −214.0 −213.6 0.07007 0.06337 0.07461 0.1387 0.1484 0.1426
Bandalgi −220.3 −219.3 −223.5 0.07467 0.06638 0.08815 0.1379 0.1266 0.1469

Note: KS critical value at 5% level of significance = 0.3298; italic values indicate the best value for particular evaluation criteria

phases were determined from the following equation (Zhang five monsoon months, i.e., from June to October. For all rainfall
and Singh 2006; Reddy and Ganguli 2012b): stations, among monsoon months, the maximum amount of rain-
fall is received in the month of September. In upper, middle, and
C ðu; vÞ
F X j Y ≤ y ¼ CU j V ≤v ¼ ð5Þ lower parts of the study area, the maximum rainfall is received
v for Chinchondipatil, Alni and Solapur stations, respectively. The
Further, the spatial variation of these probabilities over the variation of the monthly rainfall over the period of 25 years for
study area was analyzed by generating probability maps using that particular month is shown by standard error bars in Fig. 3a–
ArcGIS software. c. All stations have the highest standard error in the month of
September, except Kasegaon station, which has it in month
August. Minimum standard errors are found in the months of
January and February at all the stations.
Results and discussion

Preliminary data analysis Groundwater characteristics

Rainfall characteristics The pre- and post-monsoon groundwater-level elevation time

series data for Chinchondipatil, Jamkhed and Kasegaon
The monthly variation of rainfall for stations in upper, middle,
and lower parts of the study area is shown in Fig. 3a–c. 1
Maximum amount of rainfall in the study area is confined to
0.9

0.8
Table 8 Parameters estimated of the probability distributions fitted to
rainfall 0.7

Name of station GM LN WB 0.6

F(Z)

α β μ σ α β 0.5

Alni 8.756 86.162 6.568 0.3359 2.658 848.1 0.4

Chinchondipatil 4.048 111.452 5.983 0.5574 2.235 510.0 0.3
Jamked 10.160 59.220 6.350 0.3153 3.019 673.3
0.2 ENSO Index
Kasegaon 5.571 110.483 6.330 0.4401 2.323 696.7
Normal
Tembhurni 7.303 70.521 6.174 0.3910 2.868 578.5 0.1
Quadratic
Supa 10.338 37.331 5.906 0.3201 3.200 430.5
0
Solapur 7.348 85.465 6.373 0.3765 2.628 707.9 0 1 2
Kolgaon 5.536 65.816 5.805 0.4572 2.645 411.5 ENSO Index
Bandalgi 10.578 56.291 6.341 0.3121 3.150 664.8 Fig. 9 Cumulative distribution function of non-parametric normal and
quadratic distributions fitted to the ENSO Index
 Hydrogeol J

Table 9 Performance evaluation of different probability distributions Evaluating dependence among the hydro-climatic
fitted to ENSO Index
variables
AIC RMSE KS-test
The auto-correlation test revealed that there is no significant
Normal Quadratic Normal Quadratic Normal Quadratic time autocorrelation at the 5% level of significance, thereby
−32.54 −32.44 0.05989 0.06200 0.1249 0.1297 suggesting that each time series is independent during the
study period. However, it is apparent from Fig. 6a–d that all
Note: KS critical value at 5% level of significance = 0.3298; italic values
indicate the best value for particular evaluation criteria the hydro-climatic variables, i.e., PMGL, rainfall and ENSO
Index are cross-correlated to one another. For brevity, the
graphs of four selected stations are shown in Fig. 6a–d as an
example and the dependence measured using Spearman’s ρ is
stations representing, respectively, the upper, middle, and low-
presented in Table 4. The evaluation of dependence indicated
er parts of the study area are plotted in Fig. 4a–c along with the
that there is positive dependence in the PMGL-Rainfall pair,
annual rainfall time series data. The pre-monsoon groundwa-
which means increase in rainfall increases the PMGL. On the
ter-level elevations for Chinchondipatil, Jamkhed and
other hand, PMGL-ENSO Index and Rainfall-ENSO Index
Kasegaon stations are 616.09–626.73, 549.99–566.16,
are negatively associated with each other. These relationships
479.52–493.07 m MSL, and the post-monsoon groundwater-
suggest that there will be a decrease in the PMGL as well as
level elevations are in the range of 622.73–630.39, 558.45–
rainfall with increase in the ENSO Index values.
564.09, 483.38–489.71 m MSL, respectively. These values
For the PMGL-Rainfall pair, a high level of dependence
clearly show the response of post-monsoon groundwater-level
was found to be significant for all the stations, except at the
elevation to the variation of rainfall, i.e., post-monsoon
Bandalgi station, which is located in the downstream portion
groundwater-level elevation increases with increase in rainfall
of the study area. This lower dependence between PMGL and
and vice-versa. In case of pre-monsoon groundwater-level el-
Rainfall at the Bandalgi station could be attributed to concen-
evation, a sudden peak is observed in year 1999 for Jamkhed
trated runoff (overland flow) at the downstream end. For the
and Kasegaon stations, which is attributed to recharge from
PMGL-ENSO Index pair, only Jamkhed, Tembhurni, Solapur
the maximum rainfall in the previous year (i.e., 1998).
and Bandalgi stations, which cover 49% of the study area
(6,185 km2), exhibited statistically significant negative depen-
ENSO Index dence (Table 4). There is no statistically significant depen-
dence for the remaining five stations and, hence, in the areas
The monthly variation of the ENSO index during 1990–2006 covered by these stations, the relationship between PMGL and
period is shown in Fig. 5. For the ENSO phases during the ENSO can only be used for qualitative predication (high or
considered period, the top 30th percentile of ENSO index low PMGL). This insignificant dependence may be attributed
values represents El Niño years, whereas the bottom 30th per- to other climatic oscillations (Jones and Banner 2003). For the
centile of ENSO Index values represents La Niña years, and Rainfall-ENSO Index pair, only three stations (Alni,
the remaining as neutral years (Wolter and Timlin 2011). Kasegaon and Supa) that cover 32% of the study area showed
Accordingly, 1995–1996 and 1998–2000 years indicates La statistically significant negative dependence; hence, this pair
Niña years; 1990 and 2001–2006 years denote neutral years was not considered in subsequent analyses.
and the remaining 5 years, i.e., 1991–1994, 1997, were El
Niño years. Identifying marginal distribution for fitting
hydro-climatic variables
Table 10 Parameters of different copulas fitted to PMGL-Rainfall pairs
The performance evaluation for the distribution fitting of
Name of station Clayton Frank Gumbel-Hougaard
PMGL at all the stations was carried out using cumulative
Alni 4.476 11.019 3.238 distribution function (CDF) plots and statistical indicators as
Chinchondipatil 2.973 7.871 2.486 shown in Fig. 7a,b and Table 5, respectively. Table 6 summa-
Jamked 2.075 5.907 2.038 rizes the estimated parameters of GM, LN, and WB distribu-
Kasegaon 2.387 6.602 2.194 tions. Upon visually assessing CDF fit for the PMGL time
Tembhurni 2.857 7.624 2.429 series (Fig. 7a,b) for different stations and AIC criteria
Supa 2.000 5.736 2.000 (Table 5), it can be seen that the WB distribution provides a
Solapur 2.000 5.736 2.000 better fit than GM and LN distributions. For the rainfall time
Kolgaon 4.182 10.411 3.091 series, CDF plots (Fig. 8a,b) and RMSE values (Table 7) sug-
Bandalgi 0.519 1.920 1.259 gest that it is better represented by the LN distribution com-
pared to WB and GM distributions. The parameter estimates
Hydrogeol J

(a) Alni
568 568 568
τ
τ τ

PMGL (m MSL)
564
564

PMGL (m MSL)
564
PMGL (m MSL)

560 τ = 0.69
560 τ = 0.70 560 τ = 0.70
556
Simulated (Gumbel-
556 Simulated (Clayton) 556 Simulated (Frank) 552 Hougaard)
Observed
Observed Observed

552 552 548

0 1000 2000 3000 0 1000 2000 3000 0 1000 2000 3000
Rainfall (mm) Rainfall (mm) Rainfall (mm)

(b) Chinchondipatil
636 636 636

632 632 632

PMGL (m MSL)

PMGL (m MSL)
628
PMGL (m MSL)

628 628
τ = 0.57
τ = 0.62
624 τ = 0.61 624
624
620 620
620
Simulated (Frank) Simulated (Gumbel-
Simulated (Clayton) 616 Hougaard)
616 Observed 616 Observed
Observed
612
612 612
0 1000 2000 3000
0 1000 2000 3000 0 1000 2000 3000
Rainfall (mm) Rainfall (mm) Rainfall (mm)

(c) Jamkhed
568 568 568
564 564 564
PMGL (m MSL)

PMGL (m MSL)

560 560
PMGL (m MSL)

τ = 0.51 560
τ = 0.54
556 τ = 0.53
556 556
552 Simulated
552
Simulated (Frank) 552 Simulated (Gumbel-
(Clayton) Observed Hougaard)
548 Observed
548
548 Observed
544 544
544
0 1000 2000 0 1000 2000
0 1000 2000
Rainfall (mm) Rainfall (mm) Rainfall (mm)
Fig. 10 Scatter plots of observed (black dots) versus 1,000 simulated (gray dots) samples using Clayton, Frank and Gumbel-Hougaard copulas for the
PMGL-Rainfall pair at a Alni, b Chinchondipatil, and c Jamkhed stations. In this figure, Kendall’s tau (τ) value is shown for simulated samples

for the GM, LN, and WB distributions, fitted rainfall time based normal and quadratic distributions performed nearly
series are given in Table 8. the same; the KS-test also supports both the distributions
Moreover, the CDF plot for the ENSO Index is depicted in (Table 9). However, the statistical evaluation confirmed that
Fig. 9, which reveals that both the non-parametric kernel- the ENSO Index is best fitted by the ‘non-parametric kernel-
 Hydrogeol J

Table 11 Kendall’s tau (τ) values for the PMGL-Rainfall and PMGL- width is the only parameter for the two non-parametric kernel-
ENSO Index pairs
based distributions and its value is estimated as 0.4455.
Name of station Kendall’s tau (τ)

PMGL-Rainfall PMGL-ENSO Index Selecting suitable copula for modeling dependence

Alni 0.6912 −0.2500

As mentioned in the previous section, the PMGL and rainfall
Chinchondipatil 0.5978 −0.3235
time series followed different distributions and hence, the tra-
Jamked 0.5092 −0.3529 ditional bivariate distribution cannot be used for dependence
Kasegaon 0.5441 −0.2941 modeling. Even if in the case of same marginal distribution for
Tembhurni 0.5882 −0.3382 the PMGL and rainfall time series, copula function is pre-
Supa 0.5000 −0.2206 ferred to the traditional bivariate distribution due to its better
Solapur 0.5000 −0.3529 performance (Ganguli and Reddy 2012). Hence, the depen-
Kolgaon 0.6765 −0.2500 dence between PMGL and rainfall is modeled using a copula
Bandalgi 0.2059 −0.4412 function, because it does not need the condition of random
variables to follow the same marginal distribution family. As
the pair of PMGL-Rainfall exhibited highly positive depen-
dence (p < 0.01), an attempt was made to capture their depen-
based normal distribution’ with lower values of AIC (−32.54) dence using Clayton (Cl), Frank (Fr) and Gumbel-Hougaard
and RMSE (0.05989) as shown in Table 9. The optimal band- (GH) copula models. The estimates of copula model parame-

Table 12 Results of goodness-of-fit tests for different the PMGL-Rainfall copula models

Name of station AIC RMSE KS test

Cl Fr GH Cl Fr GH Cl Fr GH

Alni 22.59 18.22 11.97 0.02386 0.02858 0.02981 0.1521 0.1533 0.1575
Chinchondipatil 10.01 16.79 16.11 0.03309 0.03913 0.03875 0.2240 0.2352 0.2477
Jamked 12.40 10.74 11.77 0.01791 0.02699 0.03126 0.2695 0.2708 0.2798
Kasegaon 7.176 13.74 17.43 0.03895 0.04136 0.04255 0.1164 0.1256 0.1366
Tembhurni 13.00 12.94 13.28 0.02709 0.03420 0.03516 0.1569 0.1588 0.1716
Supa 9.875 10.36 11.85 0.02740 0.03018 0.03161 0.1458 0.1681 0.1826
Solapur 13.45 10.46 15.42 0.02820 0.03759 0.03945 0.1832 0.1861 0.1995
Kolgaon 24.61 26.50 23.18 0.02547 0.03718 0.04262 0.1710 0.1792 0.1810
Bandalgi 2.592 4.044 5.986 0.03326 0.03045 0.02741 0.1544 0.1639 0.1736

Note: KS critical value at 5% level of significance = 0.3298; italic values indicate the best values for corresponding criterion

Simulated (Frank) Observed

568 512 476

564 508 472

PMGL (m MSL)

PMGL (m MSL)
PMGL (m MSL)

504 468
560
500 464
556 460
496
τ = - 0.34 τ = -0.36 456 τ = -0.36
552 492
(a) Jamkhed (b) Tembhurni 452
(c) Solapur
548 488 448
-1 0 1 2 3 -2 0 2 4 -1 0 1 2 3
ENSO index ENSO index ENSO index
Fig. 11 Scatter plots of observed (black dots) versus 1,000 simulated (gray dots) samples using Frank copula for the PMGL-ENSO index pair at a
Jamkhed, b Tembhurni, and c Solapur stations
Hydrogeol J

Table 13 Estimated parameters and KS-test values of the Frank copula well intertwined with the observed data (shown as black dots).
for PMGL-ENSO Index pairs
Furthermore, the values of Kendall’s τ for the simulated data
Name of station Parameter estimated KS-test are close to those of the observed data (Fig. 10a–c and
Table 11). However, the Clayton copula better simulates the
Alni −2.372 0.1280 trend of the observed data compared to the other two copulas.
Chinchondipatil −3.189 0.1712 In addition, all the statistical indicators (Table 12) also confirm
Jamked −3.545 0.2193 that the Clayton copula is a better choice among the three
Kasegaon −2.851 0.1535 copula families considered. It should be noted from Fig.
Tembhurni −3.365 0.1636 10a–c that the upper bound appears for groundwater levels
Supa −2.068 0.2631 and rainfall dependence. For this, the estimated non-
Solapur −3.545 0.1136 parametric upper tail dependence coefficient for all stations
Kolgaon −2.372 0.1882 in the study area is found to be varying from 0.30 to 0.72.
Bandalgi −4.758 0.1510 Also, at a certain threshold of high rainfall, very weak depen-
dency exists between PMGL and rainfall. In fact, the rate of
Note: KS critical value at 5% level of significance = 0.3298
recharge is a function of depth to the water table. Hence, when
the water table reaches the threshold value, the recharge rate is
ters for the PMGL-Rainfall pair are shown in Table 10. The drastically reduced, which causes less dependency on rainfall.
scatter plots of observed and simulated data from the three For the PMGL-ENSO Index pair, a negative dependence
fitted copula models are shown in Fig. 10a–c, together with was found (Table 4). Therefore, they were only modeled by
the Kendall’s τ values computed from simulated samples for using the Frank copula, which is applicable to the entire range
the three copula models. It is evident that the random pairs of dependence [−1, +1]. The scatter plots show a good overlap
generated by all three copula models (shown as gray dots) are and close Kendall’s τ values between the observed data and

Fig. 12 Conditional probabilities Rain ≤ 754.39 mm (Avg) Rain ≤ 435.17 mm (5th P) Rain (Avg) ≤ 451.15 mm Rain ≤ 133.46 mm (5th P)
of post-monsoon groundwater Rain ≤ 597.82 mm (25th P) Rain ≤ 667.8 mm (50th P)
Rain≤ 282.75 mm(25th P) Rain ≤ 433.2 mm (50th P)
Rain ≤ 819.95 mm (75th P) Rain ≤ 1503.04 mm (95th P)
levels (PMGL) for given average Rain ≤ 531.98 mm (75th P) Rain ≤ 909.02 mm(95th P)
1
and non-average rainfall scenari- 1
os (5th, 25th, 50th, 75th and 95th
percentiles) at four stations: a 0.8
0.8
Alni, b Chinchondipatil, c
FH/R ≤r

Jamkhed, and d Kasegaon. In this

FH/R≤r

0.6 0.6
figure, P denotes percentile of
rainfall
0.4 0.4
(a) Alni (b) Chinchondipatil
0.2 0.2
PMGLavg = 563 m MSL
PMGLavg = 627.5 m MSL

0 0
560 561 562 563 564 565 566 622 624 626 628 630 632
PMGL (m MSL) PMGL (m MSL)

Rain ≤ 601.67 mm (Avg) Rain ≤ 383.55 mm (5th P) Rain ≤ 615.47 mm (Avg) Rain ≤ 256.57 mm (5th P)

Rain ≤ 467 mm (25th P) Rain ≤ 497 mm (50th P) Rain ≤ 423.28 mm (25th P) Rain ≤ 554.30 mm (50th P)

Rain ≤ 680.65 mm (75th P) Rain ≤ 1069.85 mm (95th P) Rain ≤ 683.90 mm (75th P) Rain ≤ 1293.12 mm (95th P)

1 1

0.8 0.8
FH/R≤r

0.6 0.6 (d) Kasegaon

FH/R≤r

(c) Jamkhed
0.4 0.4

PMGLavg = 561.5 m
0.2 0.2 PMGLavg = 486.0 m

0
0
482 484 486 488 490 492
558 560 562 564 566
PMGL (m MSL) PMGL (m MSL)
 Hydrogeol J

Fig. 13 Probability (P) map of

PMGL ≤ PMGLavg for a given
average rainfall scenario. NS non-
significant dependence

the pseudo-sample generated by the Frank copula for all the by considering Chinchondipatil station as shown in Fig. 12b.
stations (Fig. 11a–c and Table 11). The results of the KS-test The PMGL values of 624 and 628 m are respectively lower
(Table 13) also suggest that the dependence in the PMGL- and higher magnitude in the zone/area represented by this
ENSO Index pair is adequately captured by the Frank copula. station. For a given average rainfall, the probability of
The estimates of the parameters of the Frank copula fitted to PMGL being less than or equal to 624 m (MSL) is 15%,
the PMGL-ENSO Index pair are shown in Table 13. This whereas that of PMGL less than or equal to 628 m (MSL) is
choice of copula model corroborates the earlier study reported nearly 80% (Fig. 12b). These probability values are not sym-
by Reddy and Ganguli (2012b) in which depth-to- metrical when two different values of PMGL are considered.
groundwater data were considered instead of PMGL. However, if only one value of PMGL is considered, then
symmetry (x% and 100–x%) of probability values will ex-
Impacts of rainfall and the ENSO phenomenon ist—for example, the probability of PMGL being less than
on groundwater or equal to 624 m (MSL) is 15%, whereas that of PMGL being
greater than 624 m (MSL) is 85%.
Impacts of rainfall Furthermore, based on Fig. 12a–d, a spatial map of the
probability of PMGL ≤ PMGLavg, i.e., probability of non-
In order to study the effects of rainfall on groundwater, the exceedance for a given average rainfall scenario, is generated
graphs of the Clayton-copula-based conditional distribution as shown in Fig. 13. It can be seen from Fig. 13 that for a given
probabilities of PMGL for given average and non-average average rainfall, the conditional probability of PMGL ≤
(5th, 25th, 50th, 75th and 95th percentiles) rainfall conditions PMGLavg is above 70% for the areas/zones covered by Alni,
were prepared for four rainfall stations as an example (Fig. Tembhurni and Kolgaon stations, which encompass about
12a–d). Obviously, for a given average rainfall, the probability 33% of the study area (4,019 km2). It indicates that PMGL
of PMGLs of lower magnitudes is smaller, whereas that of in these areas (Barshi and Madha blocks, and some parts of
PMGLs of higher magnitudes is greater. This can be explained nearby blocks) will be much lower than its average value and,
Hydrogeol J

Fig. 14 Conditional probability PMGLavg = 561.5 m MSL PMGLavg = 503.4 m MSL

of PMGL ≤ PMGLavg for the
1 1
ENSO Index Z ≤ −0.25 for La
Niña phase and Z ≤ 1.23 for El (a) Jamkhed
(b) Tembhurni
Niño phase at stations a Jamkhed, 0.8 0.8
b Tembhurni, c Solapur, and d
Bandalgi
0.6 Z ≤ -0.25 0.6

FH/Z≤z
FH/Z≤z
Z ≤ 1.23
0.4 0.4

0.2 0.2

0 0
558 560 562 564 566 498 500 502 504 506 508
PMGL (m MSL) PMGL (m MSL)

PMGLavg =467.0 m MSL PMGLavg = 452.6 m MSL

1 1
(c) Solapur (d) Bandalgi
0.8 0.8

0.6 0.6

FH/Z≤z
FH/Z≤z

0.4 0.4

0.2 0.2

0 0
460 462 464 466 468 470 472 448 450 452 454 456 458
PMGL (m MSL) PMGL (m MSL)

hence, the groundwater of these areas should be managed with Impacts of ENSO phenomenon
a high priority or an alternative water source should be uti-
lized. In addition, it is recommended to propose rainwater For evaluating the effects of ENSO phenomenon on ground-
harvesting and artificial recharge structures in these areas. water, the Frank-copula-based conditional distributions of
The conditional probabilities of PMGL ≤ PMGLavg for a giv- PMGL for different phases of ENSO were plotted, which are
en average rainfall are found in the range of 65–70% for the illustrated in Fig. 14a–d. For this study, the average of the top
areas (2,316 km2) covered by Chinchondipatil and Supa sta- 30th percentile of ENSO Index values (1.23) for 1990–2006
tions, which suggests a moderate groundwater scenario under period was considered as representative of the El Niño phase,
average rainfall conditions. The groundwater extraction from whereas the average of the bottom 30th percentile of the ENSO
these areas should be carefully monitored to protect them from Index values (−0.25) was deemed as representative of the La
falling into higher conditional probability areas. Furthermore, Niña phase. This figure reveals that with an increase in ENSO
the conditional probability values for a given average rainfall Index, the probability of PMGL for a particular interval in-
vary from 60 to 65% in the zones covered by Jamkhed, creases at a lower magnitude of PMGL, but it decreases at a
Kasegaon and Solapur stations, indicating that post- higher magnitude of PMGL. The probability of occurrence of
monsoon groundwater levels in the southern and central parts higher PMGL is greater for a negative ENSO Index (La Niña
of the study area (46% of the area; 5,674 km2) would be close phase) than for a positive ENSO Index (El Niño phase)—for
to their average values under average rainfall conditions and, example, at the Jamkhed station (Fig. 14a) for the ENSO
hence, these areas are most favorable zones for groundwater Index value of Z ≤ 1.23 (El Niño phase), the chance of occur-
extraction for domestic and irrigation needs as compared to rence of PMGL less than 563 m (above MSL) is 74%, whereas
other parts of the study area. it is about 52% in the La Niña phase (Z ≤ −0.25).
 Hydrogeol J

(a) (b)

P (PMGL ≤ PMGLavg),%
5-10

10-15

15-20

20-25

25-30

30-35

35-40

NS
Fig. 15 Probability (P) maps of PMGL ≤ PMGLavg for a El Niño and b La Niña phases. NS non-significant dependence

The conditional probability (PMGL ≤ PMGLavg) values Conclusions

during ENSO phases are determined for only Jamkhed,
Tembhurni, Solapur and Bandalgi stations (Fig. 15a,b) where In this report, Archimedean copulas were applied under lim-
dependence is statistically significant (Table 4). It is found that ited data conditions to assess the effects of the ENSO phenom-
during El Niño phase (Fig. 15a), the Tembhurni and Solapur enon and rainfall on the groundwater resource of a semi-arid
stations (covering southwestern portions of the study area) river basin of western India. With regard to the availability and
show a higher (35–40%) non-exceedence probability of continuity of hydro-climatic time-series data in the study area,
PMGL with respect to its average. This suggests that the the dataset used in this study comprised monsoon rainfall of
blocks under maximum conditional probability (PMGL ≤ nine stations, post-monsoon groundwater levels (PMGL) at
PMGLavg) will be more severely affected during El Niño years 132 sites, and the ENSO Index for the 1990–2006 period.
than the other parts of the study area. The affected blocks will Based on the salient goodness-of-fit criteria, marginal distri-
be Madha, Mohal, North-Solapur blocks and some portions of butions were selected to formulate copula-based joint distri-
Parenda and Karmala blocks encompassing an area of butions for modeling dependence between hydro-climatic var-
2,739 km2 (23% of the study area). It is also apparent from iables. Thereafter, out of the four Archimedean copula fami-
Fig. 15a that the minimum probability of non-exceedence of lies, the best-performing copula was used to derive condition-
PMGL during El Niño phase is in the range of 30–35% for al probability distributions of groundwater-level time series
Jamkhed and Bandalgi stations (central part of the study area), with respect to rainfall events and ENSO phases.
which cover 3,228 km2 (26% of the study area). On the other The analysis of the results of this study revealed that the
hand, in the La Niña phase (Fig. 15b), the non-exceedance dependence for the PMGL-Rainfall pair is positive, whereas
probability of PMGL less than or equal to its average value that for the PMGL-ENSO Index pair is negative. The PMGL
is found below 15% for Jamkhed, Bandalgi and Solapur sta- and rainfall time series are best represented respectively by the
tions covering an area of 4,283 km2 (35% of the study area); parametric Weibull and lognormal distributions, whereas the
thus, the central and southern portions of the study area will ENSO Index time series is best represented by the non-
benefit by increased PMGL during La Niña years. parametric kernel-based normal distribution. The performance
Hydrogeol J

evaluation of the Archimedean copulas family indicated that Ganguli P, Reddy MJ (2013) Analysis of ENSO-based climate variability
in modulating drought risks over western Rajasthan in India. J Earth
the Clayton copula is the best for modeling dependence be-
Sys Sci 122(1):253–269
tween PMGL and Rainfall, while the Frank copula is the best Genest C, Chebana F (2016) Copula modeling in hydrologic frequency
for the PMGL-ENSO Index pair. The spatial variation of the analysis. In: Singh VP (ed) Handbook of applied hydrology.
probability of PMGL ≤ PMGLavg for a given mean rainfall in McGraw-Hill, New York
Genest C, Favre AC (2007) Everything you always wanted to know about
the study area suggests that for managing groundwater, the
copula modeling but were afraid to ask. J Hydrol Eng ASCE 12(4):
areas having above 70% probability (33% of the study area 347–368
in the eastern and western portions) should be given higher Genest C, MacKay RJ (1986) The joy of copulas: bivariate distributions
priority. In addition, the results of the probability of PMGL ≤ with uniform marginals. Am Stat 40(4):280–283
Genest C, Nešlehová J (2012a) Copula modeling for extremes. In: El-
PMGLavg during ENSO phases indicated that 23% of the
Shaarawi AH, Piegorsch W W (eds) En cycloped ia of
study area in the southwestern portion will be severely affect- environmetrics, 2nd edn., vol 2. Wiley, Chichester, UK, pp 530–541
ed during El Niño years, but 35% of the study area in central Genest C, Nešlehová J (2012b) Copulas and copula models. In: El-
and southern portions will benefit by increased PMGL (great- Shaarawi AH, Piegorsch W W (eds) En cycloped ia of
environmetrics, 2nd edn, vol 2. Wiley, Chichester, UK, pp 530–541
er than PMGLavg) during La Niña years.
Gorelick SM, Zheng C (2015) Global change and the groundwater man-
Finally, it can be concluded that the copula-based approach agement challenge. Water Resour Res 51:3031–3051. doi:10.1002/
is very useful for understanding the impacts of environmental 2014WR016825
factors on vital groundwater resources at a basin or sub-basin Gurdak JJ, Hanson RT, Green TR (2009) Effects of climate variability on
scale. The methodology demonstrated in this study can be groundwater resources of the United States. US Geol Surv Fact
Sheet 2009-3074
replicated for the effective planning and management of water Hanson RT, Dettinger MD, Newhouse MW (2006) Relations between
resources at a basin scale under data-scarce condition, partic- climatic variability and hydrologic time series from four alluvial
ularly in the drought-prone regions of Indian subcontinent and basins across the southwestern United States. Hydrogeol J 14(7):
other parts of the world. 1122–1146
Holman IP (2006) Climate change impacts on groundwater recharge-
uncertainty, shortcomings, and the way forward? Hydrogeol J
14(5):637–647
Acknowledgements Our sincere thanks are due to the India ICSH (2017) International commission on statistical hydrology. ICHS,
Meteorology Department, Pune (Maharashtra), State Data Storage Viterbo, Italy. http://www.stahy.org/Activities/STAHYReferences/
Centre, Nashik (Maharashtra), and Groundwater Survey Development ReferencesonCopulaFunctiontopic/tabid/78/Default.aspx.
Agency, Pune, for providing hydro-meteorological data to carry out this Accessed 4 May 2017
study. We are very grateful to Prof. Christian Genest (McGill University, IRI (2017) ENSO essentials. International Research Institute for Climate
Montréal, Canada) for his technical discussions and meticulous com- and Society, New York. http://iri.columbia.edu/our-expertise/
ments on this manuscript. In addition, the help rendered by Mr. Ankit climate/enso/enso-essentials/. Accessed 15 May 2017
Shekhar and the useful discussion with Dr. Poulami Ganguli (McMaster Jones IC, Banner JL (2003) Hydrogeologic and climatic influences on
University, Hamilton, Canada) are gratefully acknowledged. spatial and interannual variation of recharge to a tropical karst island
aquifer. Water Resour Res 39:1253. doi:10.1029/2002WR001543
Kao SC, Govindaraju RS (2008) Trivariate statistical analysis of extreme
rainfall events via the Plackett family of copulas. Water Resour Res
References 44:W02415. doi:10.1029/2007WR006261
Karmakar S, Simonovic SP (2009) Bivariate flood frequency analysis,
part 2: a copula-based approach with mixed marginal distributions. J
Chary GR, Vittal KPR, Venkateswarlu B, Mishra PK, Rao GGSN,
Flood Risk Manag 2(1):32–44
Pratibha G, Rao KV, Sharma KL, Rao GR (2010) Drought hazards
and mitigation measures. In: Jha MK (ed) Natural and anthropogen- Klein B, Schumann AH, Pahlow M (2011) Copulas: new risk assessment
ic disasters: vulnerability, preparedness and mitigation. Springer, methodology for dam safety. In: Schumann AH (ed) Flood risk
The Netherlands, pp 197–236 assessment and management. Springer, The Netherlands, pp 149–
185
Chowdhary H, Singh VP (2010) Reducing uncertainty in estimates of
Luque-Espinar JA, Chica-Olmo M, Pardo-Igúzquiza E, García-Soldado
frequency distribution parameters using composite likelihood ap-
MJ (2008) Influence of climatological cycles on hydraulic heads
proach and copula-based bivariate distributions. Water Resour Res
across a Spanish aquifer. J Hydrol 354(1):33–52
46:W11516. doi:10.1029/2009WR008490
Mall RK, Gupta A, Singh R, Singh RS, Rathore LS (2006) Water re-
Deolankar SB (1980) The Deccan basalts of Maharashtra, India: their sources and climate change: an Indian perspective. Curr Sci 90(12):
potential as aquifers. Ground Water 18(5):434–437 1610–1626
DTE (2016) Down to Earth. Centre for Science and Environment, New Mishra AK, Singh VP (2010) A review of drought concepts. J Hydrol
Delhi, India, 1–15 May 2016, 35 pp 391(1):202–216
Durocher M, Chebana F, Ouarda TB (2016) On the prediction of extreme Nelsen RB (2006) An introduction to copulas. Springer, New York
flood quantiles at ungauged locations with spatial copula. J Hydrol News World India (2016) Only 1 per cent water remains in Marathwada
533:523–532 Dams. http://indianexpress.com/article/india/india-news-india/only-
Fleming SW, Quilty EJ (2006) Aquifer responses to El Niño-Southern 1-per-cent-water-remains-in-marathwada-dams-2815112/.
Oscillation, southwest British Columbia. Ground Water 44(4):595– Accessed 24 May 2016
599 NOAA (2017) MEI ranks. National Oceanic and Atmospheric
Ganguli P, Reddy MJ (2012) Risk assessment of droughts in Gujarat Administration, Maryland, United States. https://www.esrl.noaa.
using bivariate copulas. Water Resour Manag 26(11):3301–3327 gov/psd/enso/mei/table.html. Accessed 10 Dec 2014
 Hydrogeol J

PACS (2004) Drought in India: challenges and initiatives. Report of Shiau JT, Feng S, Nadarajah S (2007) Assessment of hydrological
poorest areas civil society (PACS) programme 2001–2008. PACS, droughts for the Yellow River, China using copulas. Hydrol
New Delhi, India Process 21(16):2157–2163
Perez-Valdivia C, Sauchyn D, Vanstone J (2012) Groundwater levels and Singh J (2014) Pray before you sow. Down to Earth, New Delhi. http://
teleconnection patterns in the Canadian prairies. Water Resour Res www.downtoearth.org.in/coverage/pray-before-you-sow-44955.
48:W07516. doi:10.1029/2011WR010930 Accessed 19 January 2017
Reddy MJ, Ganguli P (2012a) Bivariate flood frequency analysis of upper Sklar A (1959) Fonctions de répartition à n dimensions et leurs marges
Godavari River flows using Archimedean copulas. Water Resour [N-dimensional distribution functions and their margins]. Pub
Manag 26(14):3995–4018 l’Institut Statist l’Univ Paris 8:229–231
Reddy MJ, Ganguli P (2012b) Risk assessment of hydroclimatic variabil- Susilo GE, Yamamoto K, Imai T (2013) Modeling groundwater level
ity on groundwater levels in the Manjara basin aquifer in India using fluctuation in the tropical peatland areas under the effect of El
Archimedean copulas. J Hydrol Eng ASCE 17(12):1345–1357 Nino. Procedia Environ Sci 17:119–128
Revadekar JV, Borgaonkar HP, Kothawale DR (2012) Temperature ex- Tremblay L, Larocque M, Anctil F, Rivard C (2011) Teleconnections and
tremes over India and their relationship with El Niño-Southern interannual variability in Canadian groundwater levels. J Hydrol
Oscillation. In: Jha MK (ed) Natural and anthropogenic disasters: 410(3):178–188
vulnerability, preparedness and mitigation. Springer,
UNESCO (2009) Water in changing world. United Nations World
The Netherlands, pp 275–292
Development Report 3. UNESCO, Paris
Russo T, Lall U, Wen H, Williams M (2014) Assessment of trends in
groundwater levels across the United States. Columbia Water Center Whelan N (2004) Sampling from Archimedean copulas. Quant Finan
White Paper, Earth Institute, Columbia University, New York, 20 pp 4(3):339–352
Salvadori G, De Michele C (2004) Frequency analysis via copulas: the- Wolter K, Timlin MS (2011) El Niño/southern oscillation behaviour since
oretical aspects and applications to hydrological events. Water 1871 as diagnosed in an extended multivariate ENSO index (MEI.
Resour Res 40:W12511. doi:10.1029/2004WR003133 Ext). Int J Climatol 31(7):1074–1087
Salvadori G, De Michele C (2007) On the use of copulas in hydrology: Wong G, Lambert MF, Leonard M, Metcalfe AV (2010) Drought analysis
theory and practice. J Hydrol Eng ASCE 12(4):369–380 using trivariate copulas conditional on climatic states. J Hydrol Eng
Seeboonruang U (2014) An empirical decomposition of deep groundwa- ASCE 15(2):129–141
ter time series and possible link to climate variability. Glob Nest J Zhang L, Singh VP (2006) Bivariate flood frequency analysis using the
16(1):87–103 copula method. J Hydrol Eng 11(2):150–164