Hydrological Sciences Journal

ISSN: 0262-6667 (Print) 2150-3435 (Online) Journal homepage: www.tandfonline.com/journals/thsj20

Investigating risk, reliability and return period

under the influence of large scale modes, and
regional hydrological variability in hydrologic
extremes

Jew Das & N. V. Umamahesh

To cite this article: Jew Das & N. V. Umamahesh (2022) Investigating risk, reliability
and return period under the influence of large scale modes, and regional hydrological
variability in hydrologic extremes, Hydrological Sciences Journal, 67:1, 65-81, DOI:
10.1080/02626667.2021.1998512

To link to this article: https://doi.org/10.1080/02626667.2021.1998512

View supplementary material

Published online: 08 Dec 2021.

Submit your article to this journal

Article views: 550

View related articles

View Crossmark data

Citing articles: 5 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=thsj20
HYDROLOGICAL SCIENCES JOURNAL
2022, VOL. 67, NO. 1, 65–81
https://doi.org/10.1080/02626667.2021.1998512

Investigating risk, reliability and return period under the influence of large scale
modes, and regional hydrological variability in hydrologic extremes
Jew Das and N. V. Umamahesh
Department of Civil Engineering, National Institute of Technology Warangal, Warangal, India

ABSTRACT ARTICLE HISTORY

Due to the present climate change scenario, it is necessary to interpret return period, risk, and reliability of Received 7 May 2021
hydrological extremes under non-stationary conditions. The present study aims to understand the crucial Accepted 27 September 2021
design parameters by introducing physically based covariates in the location parameter of the EDITOR
Generalized Extreme Value (GEV) distribution. The analysis is carried out over 30 streamflow gauging S. Archfield
stations located across the Godavari River basin, India. To compare the return period, risk and reliability
between the stationary and non-stationary conditions, the Expected Waiting Time (EWT) approach is GUEST EDITOR
used. The analysis reveals that half of the gauging stations are impacted by large-scale modes/oscillations V. Samadi
and regional hydrological variability, primarily by the Indian Summer Monsoon Index (ISMI) and pre­ KEYWORDS
cipitation. The EWT interpretation estimates that the non-stationary return period, risk, and reliability are extreme value analysis;
significantly different from under the stationary condition. Hence, a non-stationary approach can be return period; risk;
useful to water managers and policymakers in order to devise sustainable and resilient water resource non-stationarity; uncertainty
infrastructure under a climate change scenario.

1 Introduction globe (Villarini et al. 2010, Cheng and Aghakouchak 2014,

Salas and Obeysekera 2014, Bayazit 2015, Read and Vogel
The time-invariant or stationary hypothesis is the basis of tradi­ 2015, Su and Chen 2019, Das et al. 2020a, among others).
tional flood frequency analysis (FFA); it assumes the hydrologi­ In this sense, the large-scale climatic oscillations such as El
cal variables do not change over time. However, due to the Niño Southern Oscillation (ENSO), North Atlantic Oscillation
changing climate conditions and anthropogenic interventions, (NAO), Sea Surface Temperature Anomaly (SSTA), Global
the factors affecting hydrological variables no longer remain Temperature Anomaly (GTA), Southern Oscillation Index
unaltered, resulting in the presence of non-stationarity in the (SOI), and Indian Ocean Dipole (IOD) among others are
hydrological time series (Milly et al. 2008, 2015, Cheng and used as examples of climate variability to understand the
Aghakouchak 2014, Das and Umamahesh 2017). Therefore, flood mechanisms. For instance, Li and Tan (2015) used
Milly et al. (2008) stated that excluding non-stationarity would ENSO as a covariate to perform non-stationary FFA over
impose significant challenges for future water resource risk Daqinghe River basin; Das and Umamahesh (2017) and
management. Moreover, studies have shown that the assump­ Mondal and Mujumdar (2016) considered GTA as one of the
tion of stationarity can substantially underestimate hydrological covariates to examine the flood events over Wainganaga River
extreme events (Gilroy and McCuen 2012, Cheng and basin and Columbia River, respectively; and SST and IOD were
Aghakouchak 2014, Sarhadi and Soulis 2017, Feng et al. 2020, used as explanatory variables by Ray and Goel (2019) to study
Jha et al. 2021). Thus, the effects of climatic variability along the FFA under non-stationary condition over Narmada River
with the changeability due to anthropogenic activities should be basin. In addition to large-scale climate drivers, anthropogenic
taken into consideration. forcings like urbanization, reservoir operations, etc. are con­
Flood, as a frequently occurring hydrometeorological sidered as covariates for non-stationary FFA (Villarini et al.
extreme event, accounted for direct global losses of about $1 2009, Gilroy and McCuen 2012, Prosdocimi et al. 2015, Zhang
trillion and 220 000 lives from 1980 to 2013 (NatCatSERVICE et al. 2015, Ray and Goel 2019, Su and Chen 2019, among
2014). In addition, as a result of aggravating influences of others). Reservoirs with large capacities greatly control the
climate variability (e.g. global warming) and anthropogenic hydrological processes (e.g. runoff) of a catchment, which in
activity (e.g. urbanization, dam construction, irrigation), the turn imposes non-stationarity in the hydrological time series.
magnitude and occurrence rate of floods keep changing (Milly Similarly, the regional hydrological characteristics of a basin
et al. 2015, Sinha et al. 2018, Dong et al. 2019, Su and Chen also influence the flood characteristics, and climate change
2019, Chen et al. 2021, Daksiya et al. 2021). Thus, during the impact can be reflected as the spatio-temporal variability in
past few decades, in recognition of the importance of non- hydrological variables (e.g. precipitation magnitude). Thus,
stationary analysis of hydrometeorological extremes, several precipitation as an external variable was used for the non-
approaches have been developed by researchers around the stationary FFA by Su and Chen (2019).

CONTACT Jew Das [email protected] National Institute of Technology Warangal, Warangal, India
Supplemental data for this article can be accessed here.
© 2021 IAHS
66 J. DAS AND N. V. UMAMAHESH

Due to the rapidly changing climate, the risk of failure of examine the risk and return period characteristics under the
infrastructure such as roads, storm water drainage systems, prevailing non-stationary conditions over the basin in order to
dams, etc. increases along with the climatic extremes (Das et al. provide more reliable information to water managers and
2011, Jongman et al. 2014, Mallakpour et al. 2019, Nasr et al. policymakers.
2019). Therefore, under the non-stationary setting of FFA, the With this understanding, the primary objective of this
concepts of risk, reliability, and return period need to be paper is to outline and apply an integrated approach in order
revisited for the development of sustainable water resource to investigate the discrepancy in computing the risk and the
infrastructure in future. Cheng and Aghakouchak (2014) esti­ return period using the stationary and non-stationary fre­
mated that the stationarity assumption may lead to the under­ quency analysis of the high-flow series. As identified by past
estimation of extreme events by as much as 60%. Similarly, Šraj studies, it should be noted that the human interventions have
et al. (2016) advocated that to ensure the necessary future level partially contributed to induce the non-stationarity in the
of protection based on the present design criteria, the design hydrological extremes (Sinha et al. 2018, 2020). However, a
variable should be increased to encapsulate the non-stationary recent study by Hengade and Eldho (2019) revealed that cli­
behaviour of hydrological extremes. On the other hand, Du mate change has had a greater impact on the hydrology of the
et al. (2015) advocated that inappropriate estimation of statis­ Godavari Basin as compared to the land-cover changes. Thus,
tical parameters would lead to overstatement of risk. the present study includes reservoir index (RI) as an indicator
Therefore, even in the stationary case, risk can be overesti­ of human interventions and does not consider the land-cover
mated (as observed in some of our cases) unless modelled with changes. In carrying out the non-stationary analysis, we statis­
appropriate parameters. With the changing behaviour of para­ tically link GTA, IOD, SSTA, and Indian Summer Monsoon
meters, the return period of any design variable keeps chan­ Index (ISMI) as large-scale climate oscillations, soil moisture
ging from one year to another, causing difficulty in planning and precipitation as regional hydrological conditons, and RI as
and design applications. In order to handle such conditions, an anthropogenic alteration to the annual maximum series
various studies have been carried out for return period analysis (AMS) of streamflow at 30 different streamflow gauging sta­
and risk assessment under non-stationary approaches (Wigley tions. Moreover, 24 major dams are also considered to com­
2009, Parey et al. 2010, Cooley 2013, Salas and Obeysekera pute RI. Details regarding the streamflow gauging points and
2014, Jiang et al. 2019, Feng et al. 2020). dams are presented in the following section. In addition to the
To evaluate the concept of return period under non- non-stationary analysis, risk reliability and return period are
stationarity, Olsen et al. (1998) adopted the expected waiting evaluated for both stationary and non-stationary cases using
time (EWT) interpretation and presented a mathematical the EWT interpretation. The findings from the present study
representation to analyse risk and return period. Similarly, will allow for robust and effective water resource management
Wigley (2009) applied the EWT interpretation and evaluated practices under changing climatic conditions.
the changes in the risk and return period using a stochastic
simulation technique. Recently, Salas and Obeysekera (2014)
analysed the time-varying probabilities of risk and return 2 Study area and data used
period under different cases of extreme events, such as increas­ 2.1 Study area
ing, decreasing, and shifting events, using the geometric dis­
tribution.These studies vary the parameters of the underlying The geographical area of Godavari River basin is about 312
distribution with time to construct the non-stationary model. 812 km2. It is located in the Deccan Plateau lying between
However, the inclusion of more physically based or causative 16°19ʹ and 22°34ʹN latitude and between 73°24ʹ and 83°
covariates could be more effective and could produce better 04ʹE longitude. About 32% of the total area of the basin lies
modelling representation of risk and return period while con­ in the elevation zone between 500 and 750 m above mean
sidering the non-stationary approach than using time as cov­ sea level. The climate over the basin is tropical in nature,
ariate. Therefore, Zhang et al. (2015), Gao et al. (2017), and Su with the largest part of the annual precipitation occurring
and Chen (2019) among others used different climate and during the southwest monsoon season. The annual preci­
reservoir indices to analyse the non-stationary behaviour of pitation over the basin varies between 755 and 1531 mm.
flood and extreme precipitation. Though the abovementioned The authors acknowledge the report published by the
studies analysed the non-stationarity in hydrological extremes, Ministry of Water Resources, Government of India in
the comparison of risk, reliability, and return period associated 2014, which provides geographical and climatological infor­
with stationary and non-stationary approaches has not been mation over the basin. The report is freely accessible at:
explored. Therefore, the present study aims to analyse the https://indiawris.gov.in/wris/#/Basin. Figure 1(b) presents
concept of return period and risk under the influence of non- the locations of streamflow gauging stations and dams
stationarity and compare it with the stationary approach. over the Godavari River basin.
The Godavari River, the second longest river in India, is the
major source of water resources for peninsular India. The
2.2 Observational data
Godavari River contributes to the nation’s economy through
multiple strategies including irrigation, hydropower, naviga­ The daily streamflow data from 40 gauging stations were
tion, industry, etc. However, it is found that the changing obtained from the Water Resources Information System of
climate is having a significant impact on the basin hydrology India (India-WRIS) and can be downloaded from https://
(Hengade and Eldho 2019). In this sense, it is necessary to indiawris.gov.in/wris/#/RiverMonitoring. The Central
 HYDROLOGICAL SCIENCES JOURNAL 67

Figure 1. (a) Godavari River basin superimposed on a map of India. (b) Locations of streamflow gauging points and reservoirs are presented together over the basin
along with their ID numbers.

2.3 Large-scale climate covariates

Water Comission (CWC) ensures the quality of the datasets.
The starting and ending dates along with the latitude and To link the large-scale climate variations to AMS of stream­
logitude of each streamflow gauging point can be found in flow, we consider the influence of four climate-related expla­
Table S1 (in the Supplementary material). The AMS of natory variables and their possible combinations: (1) GTA, (2)
streamflow was extracted for each year at different stations. sea surface temperature (SST), (3) IOD, and (4) ISMI. Details
Similarly, the dam details are presented in Table S2. It should on the data used, as well as theoretical reasoning for their
be noted that the dam details were obtained from the report selection are presented below.
released by CWC, as mentioned in Section 2.1. The dam The availability of water resources is the consequence of the
details are used to compute the RI, which is further used as complex hydroclimatological interactions that are controlled
a covariate to represent the anthropogenic intervention. In by the atmosphere–ocean circulation. In this sense, the avail­
order to incorporate the regional hydrological variabilities, ability of water resources is greatly modulated by climate
precipitation and soil moisture data are used. The daily pre­ conditions at inter- and intra-annual scales (Ionita et al.
cipitation data are prepared by the India Meteorological 2012). It is well understood that as a result of anthropogenic
Department (IMD), Pune, and were obtained for the present activities there has been a significant increase of the mean
study at a spatial resolution of 0.25° × 0.25° from http://www. global temperature (Min et al. 2011), which is subsequently
imdpune.gov.in/Clim_Pred_LRF_New/Grided_Data_ increasing the water holding capacity of the atmosphere (i.e.
Download.html#. Similarly, the soil moisture data were according to the Clausius-Clapeyron equation). The increased
obtained from the Climate Prediction Center (CPC)soil water holding capacity leads to intense precipitation events
moisture dataset. The same dataset can be downloaded (Berg et al. 2013, Westra et al. 2013) which further increase
from https://psl.noaa.gov/data/gridded/data.cpcsoil.html. the occurrence probability of flood events. In this context,
Precipitation is the primary source of streamflow variability choosing GTA as an explanatory variable in modelling stream­
(Guimberteau et al. 2013), and past studies have established flow extreme is a reasonable selection. In the present analysis,
that the variability in the precipitation can be a result of the we use the Hadley Centre/Climatic Research Unit
impact of climate change on hydrology (Zhou et al. 2010, Temperature Version 4 (HadCRUT4)yearly GTA, computed
Prosdocimi et al. 2015). Thus, in order to incorporate the with reference to the 1961–1990 mean. For the detailed com­
confounding effects of climate variability, the total precipita­ putation of GTA, readers are advised to follow Morice et al.
tion amount seven days prior to the date of AMS of streamflow (2012). The GTA data can be downloaded freely from https://
is used as one of the covariates. It should be noted that the www.metoffice.gov.uk/hadobs/hadcrut4/.
precipitation amount is computed for each year at each gau­ Similarly, variations in SST have significant influence on the
ging station. Similarly, soil moisture storage regulates the atmospheric climate. The SST is considered one of the ENSO
streamflow intra-annual variability under different climate indices and, hence, variations in SST are likely to modulate
conditions (Ye et al. 2015). In this regard, soil moisture storage precipitation extremes globally (Alexander et al. 2009).
for the particular month of extreme streamflow is considered Therefore, reserchers also found a fairly significant association
as a covariate in the present study. between SST and flood events (Jain and Lall 2000, Yao et al.
68 J. DAS AND N. V. UMAMAHESH

2015, Cassola et al. 2016). In this context, choosing SST anom­ conditioned extreme modelling, risk and return period are
aly (SSTA; i.e. an ENSO-related index) as an explanatory vari­ investigated over various streamgauge locations. The present
able in modelling precipitation and temperature extremes is section details the extreme setting to link the AMS of stream­
a reasonable selection. In our analysis, we use the monthly flow with the large-scale climate variabilities, hydrological
mean anomaly of SST with respect to the 1981–2010 mean condition, and anthropogenic interventions. Moreover, the
over the NINO3.4 (17°E–120°W, 5°S–5°N) region. To establish EWT interpretation to compute risk and return period is
the corresponding linkage between annual streamflow extreme presented.
and SST, we averaged the SST series during the 5-month
period from November to March (Min et al. 2011, Mondal
and Mujumdar 2015). As suggested by Kenyon and Hegerl 3.1 Computation of RI
(2010), the effect of ENSO is strongest during the winter
As discussed, the hydrology over the study area is affected
season. Therefore, SSTA is averaged over the winter season
more by the meteorological variability than by alteration
(i.e. from November to March). The monthly data of SSTA can
in the land use. Here, the land-use change (primarily
be downloaded freely from https://www.cpc.ncep.noaa.gov/
changes in the built-up lands) over the study area is
data/indices/.
analysed from 2002 to 2015. The land-use data for differ­
The IOD and ISMI play a significant role in modulating the
ent years are collected from Moderate Resolution Imaging
streamflow as they control the precipitation variability (Sahu
Spectroradiometer (MODIS) Land Cover Type Product
et al. 2012, Chang et al. 2019, Das et al. 2020b). The IOD is
(MCD12Q1), which is at 500 m spatial resolution. The
characterized by the Dipole Mode Index (DMI), which is
year-wise plot of the built-up area over the study area is
defined as the difference in the sea surface temperature anom­
presented in Fig. S1 (Supplementary material). It can be
aly between the western (50°E–70°E, 10°S–10°N) and eastern
noted that there is no significant change in the built-up
(90°E–110°E, 10°S–Equator) tropical Indian Ocean (Saji et al.
area during 2002–2015. Therefore, the impact of land-use
1999). The DMI data can be obtained from http://www.jam
alteration on streamflow extreme is neglected and only the
stec.go.jp/frcgc/research/d1/iod/DATA/dmi.monthly.txt. In
influence of reservoir on streamflow is considered. The
India, the major part of precipitation occurs during the mon­
reservoir index (RI) was developed by López and Francés
soon season, which is mostly controlled by the Asian summer
(2013) in order to consider the influence of reservoir on
monsoon and particulary the Indian summer monsoon.
river floods as an external covariate in non-stationary
Therefore, ISMI can be considered a good indicator for pre­
flood frequency analysis:
cipitation variability and, hence, helpful in analysing the
streamflow extremes. The difference in the zonal winds at XM � Ai � � Ci �
RI ¼ � (1)
850 hPa over 40°E–80°E, 5°N–15°N and 70°E–90°E, 20°N– i¼1 A
T CT
30°N (Wang 2000) is characterized as ISMI. The ISMI data
can be downloaded from http://apdrc.soest.hawaii.edu/pro where M is the number of reservoirs upstream of the gauging
jects/monsoon/seasonal-monidx.html. station, and Ai and AT are the catchment area of the reservoir
and gauging station, respectively. Ci denotes the storage capa­
city of each reservoir, and CT is the mean annual runoff at the
3 Research methodology streamflow gauging station. An RI threshold value of 0.25 is
suggested by López and Francés (2013) to quantify the impact
In order to statistically model the AMS of streamflow, condi­
on river floods due to the reservoir, as the threshold value lies
tional on large-scale climate oscillations, regional hydrological
between low and high alteration.
variability, and anthropogenic interventions, we: (1) obtain the
annual series of maximum streamflow at 30 different stream­
flow gauging stations over the Godavari River basin; (2) use the
3.2 Non-stationarity modelling of streamflow
method of maximum likelihood (ML) to fit a Generalized
Extreme Value (GEV)distirbution model with parameters Recent developements in computational facilities and statisti­
that depend on different types of covariates; and (3) analyse cal techniques enable users to perform complex analysis of
the type(s) of covariate(s) that can be adopted to compute the hydroclimatological variables. In general, extreme value theory
FFA in the present scenario of climate change. Initially, the (EVT) is used to analyse the stochastic behaviour of exteme
AMS of streamflow is computed using the R package processes. From a theoretical point of view, the EVT models
“extRemes” (https://cran.r-project.org/web/packages/ the maxima and minima of environmentall processes and their
extRemes/index.html). In the present climate change scenario, return levels, through asymptotic convergence to the general­
most of the extreme weather events are controlled by ocean– ized extreme value (GEV) distribution. The GEV distribution
atmosphere interaction and hence, when modelling extreme incorporates Weibull, Gumbel, and Fréchet theoretical distri­
conditions the large-scale climate descriptors should be incor­ bution models as sub-cases (Coles 2001). In addition, the latest
porated. Likewise, the hydrological condition of human inter­ advancements in GEV modelling allow covariation of the dis­
ventions should be included while modelling the extreme tribution parameters with climate indices and descriptors, to
streamflow condition. In addition, the conditioned extreme study the influence of climate drivers on extreme events. The
modelling of the time series is even more essential to under­ cumulative distribution function (CDF) of the GEV distribu­
stand the future behaviour. Therefore, incorporating the tion is given by Equation (2) (Coles 2001, Katz et al. 2002):
 HYDROLOGICAL SCIENCES JOURNAL 69

8 � h i 1=� �
>
< exp 1 þ �ðyσ μÞ ; σ > 0; 1 þ �ðyσ μÞ > 0; ��0
Fðy; μ; σ; �Þ ¼ n h io (2)
>
: exp exp ðy μÞ ; σ > 0; � ¼ 0
σ

Table 1. Mathematical representation of the Generalized Extreme Value (GEV)

where µ, σ, and ξ are the location, scale, and shape parameters models proposed in this study.
of the GEV distribution, respectively, and y corresponds to the Group Model Location (μ) Scale (σ) Shape (�)
data series. Hence, Fðy; μ; σ; �Þ represents the CDF of the data M0 M0 μ σ �
series y having parameters µ, σ, and ξ. The tailed behaviour of M1 M11 μ0 þ μ1 PR σ �
the distribution is controlled by the shape parameter as posi­ M12 μ0 þ μ1 SM σ �
M13 μ0 þ μ1 RI σ �
tive, negative, and zero denote heavy tailed, bounded upper M2 M21 μ0 þ μ1 PR þ μ2 SM σ �
tail, and unbounded light tailed, respectively. Moreover, if ξ is M22 μ0 þ μ1 PR þ μ2 RI σ �
zero this gives a Gumbel distribution, greater than zero refers M3 M31 μ0 þ μ1 GTA σ �
M32 μ0 þ μ1 IOD σ �
to Fréchet, and less than zero indicates the Weibull distribu­ M33 μ0 þ μ1 SST σ �
tion (Cheng et al. 2014). M34 μ0 þ μ1 ISMI σ �
Here, in order to link the covariates, we linearly vary the M4 M41 μ0 þ μ1 GTA þ μ2 PR σ �
M42 μ0 þ μ1 IOD þ μ2 PR σ �
location parameter of the GEV distribution with large-scale M43 μ0 þ μ1 SST þ μ2 PR σ �
climatic covariates (GTA, IOD, SST, and ISMI), hydrological M44 μ0 þ μ1 ISMI þ μ2 PR σ �
variability (precipitation (PR), soil moisture (SM)), and anthro­
pogenic interventions (RI). In the present study, the covariates
are introduced in the location parameter, keeping the scale GEV model is chosen to attribute the significant link among
parameter constant in the GEV distribution. However, in the the covariate or combination of covariates and AMS of stream­
case of shape parameter, it is cumbersome to estimate it pre­ flow. The next section depicts the methodology for the para­
cisely in the context of the conditioned extreme modelling meter estimation of the fitted GEV models.
concept (Yilmaz and Perera 2014). Moreover, assuming the
shape parameter as a smooth function of time is quite unrealistic
(Coles 2001). Thus, the shape parameter is also kept constant. 3.3 Estimation of parameters
In order to consider all possible effects of selected covariates The parameter estimation of the GEV distribution with or
on the extreme streamflow time series, the location parameter without inclusion of the physical covariates is carried out
is conditioned linearly with covariate(s) while keeping scale using the maximum likelihood estimation (MLE) method. In
and shape parameters constant (Equation 3). The reasons for general, the likelihood method is preferred over other meth­
incorporating the covariates in the location parameter alone ods, e.g. method of moments, due to its ability to introduce
are (1) its intuitive appeal (Towler et al. 2010); (2) simplicity in various forms of parameters of the distributions which are
obtaining the esimates (Coles 2001); and (3) the possibility of conditioned with different physical covariates (Coles et al.
being a smooth function of the covariate (Mondal and 2003, Katz 2013). Moreover, the L-moments method, which
Mujumdar 2016). In order to avoid possible complexities in is more popular in hydrology than MLE, is not straightforward
the estimation with an increased number of parameters in the to apply under conditioned GEV conditions (Katz 2013).
probability model, the nonlinear association of covariates with To explain the MLE, let us define Y ¼ y1; y2;
parameters is not considered in the present study. y3; . . . . . . ::; yðn 1Þ; yðnÞ as the series of any selected extreme
P indices with n number of observations. Then, the log likelihood
μ ¼ μo þ m¼ 1 μi ci (Equation 4) is defined as
σ¼σ (3) h �y μ�i
Xn
�¼� LðθjY Þ ¼ nlogσ ð1 þ 1=�Þ log 1 þ �
i

�y σ μ�
i¼1
Xn h �y μ�i 1=�
where ci denotes the selected covariates (e.g. GTA, IOD, SST, 1þ�
i
;1 þ �
i
>0
ISMI, P, SM, and RI), μo and present the intercepts, and μi are i¼1 σ σ
the intercept and slope to the location parameter,respectively. (4)
Under the assumption of linearly conditioned dependence, the
For ξ = 0
parameters of the GEV distribution can be expressed mathe­
Xn �y μ�
matically and presented in Table 1. Table 1 shows that five LðθjY Þ ¼ nlogσ
i
groups have been developed that include 14 models. The group Xn ni ¼ 1�y σ μ�o
i
M0 presents the stationary model, whereas groups M1, M2, exp (5)
M3, and M4 depict the inclusion of covariate(s) in the location
i¼1 σ
parameter of the GEV distribution only. The performance of where L(θ) is the likelihood function of the parameter vector θ.
all 14 conditioned/non-stationary GEV models are tested Precisely, the likelihood function is the measure of how likely
against the unconditioned/stationary model, and the best the datasets are as a function of the unknown parameters of
70 J. DAS AND N. V. UMAMAHESH

GEV distribution (as GEV distribution is used in the present sample size (Coles 2001). In the present study, the non-
study), and the MLE evaluates the values of parameters that stationary model (MN) with a linear trend in the parameters
maximize the likelihood function (Katz 2013). In computation, and the stationary model (MS) with no trend in the parameters
it is suitable to work in terms of log likelihood function and are tested for the validity. It should be noted that in order to
also convenient to minimize the negative likelihood function perform hypothesis testing the stationary model is considered
rather than maximizing the likelihood function (Mondal and the null hypothesis and the non-stationary model is the alter­
Mujumdar 2016, Das and Umamahesh 2017). Moreover, the nate hypothesis. The hypothesis testing is carried out by com­
formulation of MLE can be used in the conditioned case, paring the minimized negative log likelihood (nllh) function of
where the parameters are associated with the physical covari­ MS relative to MN at a significance level of α, which is known as
ates. For instance, consider the model M11 where the location the likelihood ratio test (LR test). In LR test, twice the differ­
parameter alone is linked with the physical covariate precipita­ ence between nllh(S) and nllh(N) is computed, which has an
tion. Then the likelihood function, defined as LðμðPÞ; σ; �jYÞ, approximate Chi squared distribution for a large sample size
can be expressed as a function of four parameters, namely (Katz 2013). The LR test can be defined as
μ0; μ1; σ; and�. Therefore, optimization is carried out to obtain � �
these four parameters so that the likelihood function is max­ 2 nllhðSÞ nllhðN Þ > cα (9)
imized. In the case of conditioned GEV distribution, the like­
lihood function can be modified as where cα is the ð1 αÞ quantile of the Chi-squared distribu­
Yn tion. In the present investigation, a 5% significance level (i.e.
LðθjY Þ ¼ i¼1
f ðyi ; μðt Þ; σ ðtÞ; �ðt ÞÞ (6) 95% confidence level) is used. It is worth mentioning that the
best-conditioned fit is assessed from the p value of the Chi-
where f ðyi ; ðtÞ; σ ðt Þ; �ðtÞÞ denotes the density function of GEV squared distribution. Therefore, the null hypothesis cannot be
distribution for the parameter set μðtÞ; σ ðtÞ; �ðtÞ evaluated at rejected if the p value is greater than 0.05 in favour of the
yi. According to Coles (2001), by analogy with Equation (4) the alternate hypothesis. In the present study, the LR test is com­
conditioned log-likelihood of the GEV distribution can be puted for all 26 non-stationary combinations and compared
written as with the stationary model. The best model is chosen based on
Xn the lowest p values among all the combinations. For instance, if
LðθjY Þ ¼ logσ ðtÞ
i¼1
Xn � � �� there are multiple models with p values lower than .05 then the
yi μðtÞ
ð1 þ 1=�ðt ÞÞ log 1 þ � ð t Þ model with the lowest p value is chosen in the present study.
i¼1 σ ðt Þ
Xn � � �� 1=�ðtÞ
yi μðtÞ
1 þ � ðt Þ ;1
i¼1
� � σ ðt Þ 3.5 Computation of return level
yi μðt Þ
þ � ðt Þ >0 To quantify and assess the risk associated with stationary and
σ ðt Þ
non-stationary approaches, the return level corresponding to a
(7) particular return period (which primarily depends on the
In the present study, as discussed in Section 3.1, the scale and purpose of the application) is computed. Traditionally, the
shape parameters are not conditioned with the chosen differ­ assumption of stationarity is used to quantify the risk and
ent covariates. Hence, Equation (7) can be rearranged in terms reliability (Read and Vogel 2015). However, with sufficient
of the covariates as evidence of climate change, it is indispensable to acknowledge
Xn the transient nature of extreme events. To do otherwise would
LðθjY; CÞ ¼ logσ lead to the underestimation of estimates and improper risk
i¼1
Xn � � ��
yi μðct ; b1 management (Cheng and Aghakouchak 2014). Therefore, in
ð1 1=�Þ log 1 þ � the present study, the return level (Zp) for a given return
i¼1 σ
Xn � �
yi μðct ; b1
�� 1� period T ¼ 1=p (p, probability of occurrence) is computed as
1þ� ;1
�
i¼1
� σ � � σ^ n o
�^
yi μðct ; b1 Zp μ b �^ ¼ μ
^; σ; ^ 1 ½1 logð1 pÞ� for��0
þ� >0 �^
σ
(8)
^; σ^Þ ¼ μ
Zp ðμ ^ σ^ log½logð1 pÞ� for� ¼ 0 (10)
where C = [C1, C2, . . ., Cm]T is the vector of climate covariates,
and θ = [b1T,σ, ξ]T is the parameter vector, with b1 denoting the
In addition, the uncertainty associated with the return level is
parameter vectors of the regressions for μ.
computed using the delta method, as the inclusion of external
covariate(s) and the asymptotic normality of the maimum
likelihood estimate proved it an appropriate choice for the
3.4 Selection of the model
application (Coles 2001). In general, the delta method esti­
The parameter estimation using MLE facilitates comparison mates the confidence intervals for the maximum likelihood
among the “nested” models. To compare the significance function, where it simplifies the function to evaluate the var­
between stationary and non-stationary models, the likelihood iance. Hence, the variance can be obtained for any return level
ratio test provides an approximate procedure for a large using Equation (11).
 HYDROLOGICAL SCIENCES JOURNAL 71

� �
Var Zd T It should be noted that Equation (13) is derived for the sta­
ð1 pÞ � ÑZð1 pÞ VÑzð1 pÞ (11)
tionary case. Therefore, the return period of X (also defined as
where the variance–covariance matrix of a parameter is pre­ the expected value of X) of high flow exceeding the design
sented as V and the value of ÑZðT1 pÞ is presented as flood value for the stationary case is defined as
� � X1
@zð1 pÞ @zð1 pÞ @zð1 pÞ T ¼ Eð X Þ ¼ xf ðxÞ ¼ 1=p (14)
ÑZðT1 pÞ ¼ ; ; (12) x¼1
@μ @σ @�
However, under non-stationary conditions, the p and q values
h � � � � i no longer remain constant (Fig. 2(b)) and thus, the probable
1 2
¼ 1; � 1 yp � ; σ� 1 yp � σ� 1 yp � log yp waiting time for the first flood exceeding the design flood zq0
using the geometric probability law can be defined as
where T represents the transpose of matrix and
yp ¼ logð1 pÞ. For more details regarding the delta meth­ f ðxÞ ¼ PðX ¼ xÞ ¼ ð1 p1 Þð1 p2 Þ . . . . . . . . . ð1 px 1 Þpx
odology, readers are directed to Mondal and Mujumdar (2016) Yx 1
and Oehlert (1992). ¼ px t ¼ 1 ð1 pt Þ; x ¼ 1; 2; . . . ::; xmax
(15)

3.6 Non-stationary return period using EWT The xmax represents the time when the probability of excee­
interpretation dance becomes 1. Equation (15) is a generalization of the
geometric distribution which can be opted for the non-
Under the assumption of stationarity, suppose the design flood stationary conditions with varying paremetrs with time.
to build different water resource infrastructures is denoted by Therefore, the expected waiting time return period under non-
zq. The q is the probability of nonexceedance of the design stationary conditions can be computed using Equation (16)
flood – or, in other words, the design flood has a probability of
X1 X1 Yx 1
exceedance of p (i.e. p ¼ 1 q). Figure 2(a) schematically T ¼ Eð X Þ ¼ x ð x Þ ¼ xp x ð1 pt Þ
x¼1 x¼1 t¼1
represents the design flood, and it shows that for each year
the value of zq remains the same as the p and q values under the (16)
stationary condition. This suggests that the risk associated with To obtain the probability of exceedance ~1, the covariate value
the design flood remains the same. Therefore, failure of the should approach xmax. It should be noted that the covariate in
designed structure will occur when the flood magnitude each year is independent of the value of previous year and
exceeds the design flood in year x, x ¼ 1; 2; . . . ::; xmax , follow­ hence considered a random variable. Therefore, the covariate
ing the law of geometric probability (Mood et al. 1974, Salas series for different gauging stations is fitted to a suitable dis­
and Obeysekera 2014): tribution (selected based on the Akaike information criterion
(AIC)), and using the evaluated parameters 500 synthetic
f ðx Þ ¼ P ðX ¼ x Þ ¼ ð1 pÞx 1 p (13)
datasets are generated for each selected covariate in order to

Figure 2. (a) Schematic representation of probability of exceedance and nonexceedance under (a) stationary condition and (b) non-stationary condition.
72 J. DAS AND N. V. UMAMAHESH

obtain a long series. This approach has been used widely in However, with the increasing pace of climate change, the eva­
water resource applications and details can be obtained from luation of hydrological risk under non-stationary conditions has
Loucks and van Beek (2017). become of paramount importance. The non-stationary-based
approach enables us to provide the hydrological risk, incorpor­
ating the changes in the exceedance probability with time pt for
3.7 Computation of hydrological risk and reliability a design life of n years. Therefore, the risk of failure under the
under non-stationarity time-varying probabilities can be obtained as:
Xn Yx 1 Yn
The most probable application regarding hydrological fre­ R¼ p ð1 pt Þ ¼ 1 ð1 pt Þ (18)
quency analysis is to identify the associated risk (R) of x¼1 x t¼1 t¼1

a hydrological infrastructure for a design life of n years. In In this sense, the reliability can be defined as the probability
the present case, under the stationary condition, the R value for that no high flow exceeding the design flood will occur, and
the high flow exceeding the design flow before or at year n can can be presented as:
be obtained from the perspective of the complement, which is Yn
defined as: Rl ¼ 1 R ¼ t¼1
ð1 p t Þ (19)

For instance, for n = 3, the reliability will be

R¼1 ð1 p Þn (17)
Rl ¼ ð1 p1 Þð1 p2 Þð1 p3 Þ. Similarly, Equation (19) can
be used for any value of n.

Table 2. p values from the LR test correspond to the best model for a significance
level of 5%. 4 Results
Station p value Station p value Station p value 4.1 Covariates and their significance for AMS of
Asthi - Kanhargaon .0327 Ramakona .0143 streamflow
Bamni .0050 Keolari* - Salebardi .0418
Bhatpalli .0323 Konta .0052 Saradaput .0475 In this section, the results regarding the five different types of
Chindnar* - Mancherial .0118 Satrapur_KR_Bridge -
Degloor - Nandgaon - Somanpally .0071 models for fitting the streamflow AMS, including the p value of
Dhalegaon* .0052 Pathagudem - Sonarpal - the LR test at a 5% significance level, are analysed (Table 2).
G.R.Bridge - Pauni - Tekra - The lowest p value, which indicates the most significant asso­
Ghugus - Polavaram .0171 Tumnar .0264
Hivra .0205 Rajegaon - Wairagarh - ciation with streamflow, is extracted from all models. In addi­
Jagdalpur .0004 Rajoli .0055 Yelli - tion, the estimated model parameters for the best model are
*Represents the station with RI as one of the covariates. presented for 30 gauging stations (Table 3).

Table 3. The selected best model and associated parameters for all the gauging stations.
Estimated model parameters
Station Model ID Location Scale Shape
Asthi M0 12416:372 6436:221 0:303
Bamni M34 7146:780 þ 1744:486ISMI 3852:906 0:131
Bhatpalli M44 1049:006 þ 295:929ISMI 1:777PR 645:220 0:645
Chindnar* M0 3789:375 1922:225 0:123
Degloor M0 382:737 446:691 0:766
Dhalegaon* M34 1518:970 þ 584:505ISMI 1214:514 0:133
G.R.Bridge M0 753:658 768:755 0:765
Ghugus M0 3446:785 2237:055 0:103
Hivra M21 1636:686 þ 14:61PR 2:219SM 1387:641 0:193
Jagdalpur M41 1055:711 582:618GTA þ 3:354PR 596:742 0:426
Kanhargaon M43 305:910 280:700SST þ 1:850PR 300:785 0:548
Keolari* M0 781:968 716:696 0:557
Konta M34 4326:782 þ 599:188ISMI 2364:799 0:142
Mancherial M21 4269:204 þ 35:876PR 6:803SM 4592:281 0:236
Nandgaon M0 980:573 834:965 0:051
Pathagudem M0 9928:131 4304:776 0:351
Pauni M0 7026:929 4791:355 0:076
Polavaram M21 27148:28 þ 71:43PR 16:367SM 11609:739 0:381
Rajegaon M0 2251:508 1355:940 0:071
Rajoli M43 579:527 177:435SST þ 1:853PR 403:222 0:045
Ramakona M44 443:787 þ 120:719ISMI þ 3:053PR 517:195 0:520
Salebardi M34 667:252 þ 128:494ISMI 417:401 0:278
Saradaput M34 1950:528 þ 175:884ISMI 930:939 0:293
Satrapur_KR_Bridge M0 1371:147 1174:294 0:392
Somanpally M44 368:049 þ 146:727ISMI þ 3:7918PR 589:996 0:474
Sonarpal M0 524:264 257:379 0:254
Tekra M0 14552:095 7530:117 0:015
Tumnar M34 839:082 þ 118:946ISMI 407:085 0:433
Wairagarh M0 756:217 418:470 0:189
Yelli M0 2424:998 1890:119 0:246
*Represents the station with RI as one of the covariates.
 HYDROLOGICAL SCIENCES JOURNAL 73

From Table 2, it can be noted that the non-stationary 4.2 Return level and associated uncertainty
approach outperforms the stationary approach in 15 out of 30
The return levels and associated uncertainties (as confidence
gauging stations in the river basin at a significance level of 5%.
The outcomes show the importance of involving the climate bounds) are evaluated using Equations (10)–(12) for all gau­
change components in the design criteria and risk assessment. ging stations for 5-, 10-, 20-, 30-, 40, 50-, 60-, 75-, and 100-year
From Table 3, it is observed that the best model is obtained when return periods. However, for the sake of brevity, the return
the location parameter of the GEV distribution is conditioned levels for the above return periods for only two cases, namely
with ISMI alone or in combination with other covariates for nine Bamni and Hivra gauging stations, are presented in Figs 5 and
out of 15 gauging stations. Similarly, PR – along with some other Figure 6, respectively.
covariate – when added to the location parameter increases the It can be noted that the stationary return levels for different
likelihood as compared to the stationary analysis for nine out of return periods remain constant over the years. However, in the
15 gauging stations. The results show that the covariates ISMI case of non-stationary conditions, the return level changes along
and PR have the greatest effect on the location parameter of the with the associated covariate and does not remain constant
GEV distribution for the AMS of streamflow. The SM has throughout the years. Moreover, the uncertainty bound of the
a significant impact over three streamflow gauging points when return level increases with an increase in the return period. In
combined with P. In addition, the large-scale modes like SST and the case of a higher return period, the deviation between the
GTA improve the model performance compared to the stationary stationary and non-stationary uncertainty bound has increased
model over three out of 15 models. as compared to the lower return period. In this case, it should be
Using Equation (1), RI values over 11 different gauging noted that the stationary condition has underestimated the
stations were computed and are presented in Fig. 3 (Fig. 4) for return level and uncertainty bound, which may increase the
RI values greater (lower) than 0.25. It can be seen from the risk of failure and reduce the reliability of the water resource
figure that all stations except for three (Chindnar, Dhalegaon, infrastructures proposed using the design flood based on sta­
and Keolari) have RI values of less than 0.25. Thus, RI as an tionary conditions. Conversely, from Fig. 4, it can be noted that
external covariate is considered for those three stations only. the stationary condition has overestimated the return level and
However, from Table 3, it can be noted that the model perfor­ uncertainty bound, which may have economic consequences for
mance does not improve when considering RI as a covariate. the design of water resource infrastructures.

Figure 3. RI values obtained for streamflow gauging stations greater than 0.25.

Figure 4. RI values obtained for streamflow gauging stations of less than 0.25.
74 J. DAS AND N. V. UMAMAHESH

Figure 5. Return levels and 95% confidence interval for different return periods at Bamni gauging station under stationary and non-stationary conditions.

4.3 Comparison between stationary and non-stationary

stationary conditions is equivalent to the T2 return period
return periods
under the non-stationary approach. In other words, the return
Having obtained a suitable non-stationary model incor­ level for return period T1 under the stationarity assumption is
porating different physical covariates with the evaluated equivalent to the return period T2 for the non-stationarity
parameters (Table 3), the specific return periods (as men­ assumption. Moreover, high return periods (50, 75, and
tioned in Section 4.2) for stationary and non-stationary 100 years), commonly used in water resources planning and
approaches are compared using EWT interpretation management, are used for the comparison. With this under­
(Equations 13–16). The outcomes of this analysis are pre­ standing, T1 and T2 comparisons for different stations are as
sented in Fig. 7. follows: for Bamni station, S_50/40_NS, S_75/55_NS, S_100/
It can be observed from Fig. 7 that the stationary approach 68_NS; for Bhatpalli station, S_50/45_NS, S_75/63_NS, S_100/
overestimates the return period (i.e. underestimates the return 80_NS; for Jagdalpur station, S_50/27_NS, S_75/47_NS,
level) as compared to the non-stationary approach for 10 out S_100/69_NS; for Kanhargaon station, S_50/48_NS, S_75/
of 15 gauging stations. For comparison purposes, the notation 70_NS, S_100/90_NS; for Konta station, S_50/40_NS, S_75/
S_T1/T2_NS indicates that the T1 return period under 56_NS, S_100/71_NS; for Rajoli station, S_50/38_NS, S_75/
 HYDROLOGICAL SCIENCES JOURNAL 75

Figure 6. Return levels and 95% confidence interval for different return periods at Hivra gauging station under stationary and non-stationary conditions.

52_NS, S_100/65_NS; for Ramakona station, S_50/44_NS, 4.4 Risk assessment under stationary and non-stationary
S_75/61_NS, S_100/76_NS; for Salebardi station, S_50/ conditions
38_NS, S_75/53_NS, S_100/66_NS; for Saradaput station, The hydrological risk assocciated with the design flood under
S_50/46_NS, S_75/67_NS, S_100/86_NS; and for Tumnar sta­ stationary and non-stationary approaches is evaluated based
tion, S_50/48_NS, S_75/68_NS, S_100/87_NS. on the initial return period and design life using Equations
Similarly, over five gauging stations the stationary approach (17) and (18), respectively. Here, initial return periods such as
underestimates the return period (i.e. overestimates the return 50, 75, and 100 years are considered. The hydrological risk
level) compared to the non-stationary approach. For those results are presented in Fig. 8.
gauging stations, the T1 and T2 comparisons are as follows: The risk analysis results are quite similar to the outcomes
for Dhalegaon station, S_50/70_NS, S_75/103_NS, S_100/ from the return period analysis. It is observed from the figure
128_NS; for Hivra station, S_50/90_NS, S_75/153_NS, that the hydrological risk (as a percentage) has increased in the
S_100/203_NS; for Mancherial station, S_50/127_NS, S_75/ case of non-stationary conditions for 10 out of 15 gauging
204_NS, S_100/260_NS; for Polavaram station, S_50/60_NS, stations. The difference in the hydrological risk between sta­
S_75/82_NS, S_100/110_NS; and for Somanpally station, tionary and non-stationary conditions increases with increas­
S_50/67_NS, S_75/99_NS, S_100/127_NS. ing design life. For explanatory purposes, a design life of 100
76 J. DAS AND N. V. UMAMAHESH

Figure 7. Comparison between stationary and non-stationary return periods for different gauging stations using the EWT interpretation.

Figure 8. Stationary and non-stationary risk of failure (in %) for different design lifetimes; the dashed line (solid line) represents the non-stationary (stationary)
condition.
 HYDROLOGICAL SCIENCES JOURNAL 77

years is considered under the initial return periods of 50, 75, station to another, with no uniform spatial pattern in its
and 100 years. The hydrological risk associated with an initial influence. A similar observation was made by Mondal and
return period of 50 years and a design lifetime of 100 years is Mujumdar (2015) while modelling the non-stationarity in
86% under the stationary condition. However, under the non- extreme rainfall over India. A spatial plot of stationary and
stationary condition, the hydrological risk is evaluated as 91%, non-stationary stations over the study area is presented in
89%, 96%, 88%, 92%, 93%, 90%, 93%, 89%, and 88% for Bamni, Fig. 9. It can be seen that although some stations are close to
Bhatpalli, Jagdalpur, Kanhargaon, Konta, Rajoli, Ramakona, each other, their p values are different. For instance, stations
Salebardi, Saradaput, and Tumnar stations, respectively. with ID numbers (see Table S1) 8 (stationary, p value = .090)
Similarly, the risk related to the initial return period of and 2; stations 28 and 4 (stationary, p value = .098); stations 10
75 years (100 years) and a design life of 100 years is 74% and 26 (stationary, p value = .089); and stations 22 and 17
(63%). However, for the non-stationary condition, the hydro­ (stationary, p value = .288) are identified as group of nearby
logical risk is computed as 83% (76%), 79% (71%), 88% (79%), stations. In the present study, a 5% significance level is used,
76% (66%), 83% (75%), 85% (78%), 80% (73%), 85% (78%), such that the non-stationary model, with p values less than .05,
77% (69%), and 77% (67%) for Bamni, Bhatpalli, Jagdalpur, is selected over stationary model. It can be seen that the
Kanhargaon, Konta, Rajoli, Ramakona, Salebardi, Saradaput, selection of the non-stationary model for stations 8, 4, and
and Tumnar stations, respectively. 26 is sensitive to the selected significance level, as the difference
Conversely, for five out of 15 stations, the hydrological risk in the p value between stationary and non-stationary model is
has decreased for the non-stationary condition as compared to less. Therefore, considering a significance level of 10% (p value
the stationary condition. The computed hydrological risk for less than .1) would have resulted in selecting the non-
the 50-year (75-year, 100-year) return period and a design life stationary model for those stations. However, stations 22 and
of 100 years is 75% (60%, 49%), 74% (50%, 34%), 54% (37%, 17 are situated on different stream networks. Moreover, in the
27%), 67% (56%, 39%), and 77% (62%, 51%) for Dhalegaon, case of station 17, it is observed that a reservoir is located
Hivra, Mancherial, Polavaram, and Somanpally stations, upsteam of the station. Due to unavailability of year-by-year
respectively. The risk and reliability are complementary to data on the utilizable capacity of the reservoir, the impact of
each other (Equation 19). Therefore, the higher the risk of the reservoir on streamflow extremes is not captured.
failure, the lower the reliability of the future engineering In the present study, the selection of covariates is based on
designs, and vice versa. past evidence as discussed in Section 2.2 and 2.3 and adopted
in past studies (Mondal and Mujumdar 2015, Das et al. 2020a,
Hesarkazzazi et al. 2021, among others). In addition, correla­
5 Discussion and conclusions
tion analysis between the AMS of streamflow and selected
The recent developments in computational aspects enable us to covariates is performed using Pearson correlation analysis at
encompass such physically based covariates in extreme modelling a significance level of .05. The outcomes are presented in the
and extend the concepts of return period and risk assessment Supplementary material (Figs S2–S6) for different gauging
under non-stationary conditions. Thus, in the present study, non- stations. In addition, the p values for non-significant correla­
stationary extreme value analysis is carried out for 30 different tion are shown in the figures. The correlation analysis may able
streamflow gauging stations over Godavari River basin, and it is to capture the relation between the hydrological variability and
found that 50% of the gauging stations are influenced by large- hydrological extreme at regional scale. However, in order to
scale oscillations and regional hydrological variabilities. The influ­ identify the relation between the large-scale modes and hydro­
ence of ISMI and P is found to be more significant than that of logical extremes, methodologies like continuous wavelet and
other covariates. As stated by Su and Chen (2019), the utilizable discrete wavelet transformations (not performed in the present
capacity of a reservoir is more suitable than the total capacity for analysis) would be more useful than the correlation test. It can
analysing the impact. The year-by-year utilizable capacity of be noted that PR, ISMI, and SM show significant correlation
different reservoirs over Godavari River basin is difficult to obtain over most of the gauging stations. The Pearson correlation test
and hence was not included in the analysis. This may be the between the AMS of streamflow and RI estimates p values of .5,
reason why the influence of the dam on streamflow extremes is .06, and .48 for Chindnar, Dhalegaon, and Keolari stations,
found to be non-significant over Godavari River basin. The respectively, suggesting a weak association between AMS and
estimated parameters of the GEV distribution varying only the RI. Therefore, fitting the non-stationary models with RI may
location parameter, with constant scale and shape parameters, are not improve the model performance.
presented in Table 3. It can be seen that the shape parameter The delta method reveals that the uncertainty associated
varies from station to station, suggesting one of the three limiting with the higher return period is larger as compared to the
types as discussed in Section 3.2. In addition, the behaviour of the lower return period for both stationary and non-stationary
shape parameter in the GEV distribution is strongly affected by conditions. The variance of the return level is used to compute
the record length, and long records give a more reliable estimate the confidence interval (uncertainty range) assuming an
of the shape parameter that varies within a narrow range (Martins asymptotic normal distribution (Coles 2001). Theoretically,
and Stedinger 2000, Papalexiou and Koutsoyiannis 2013, in the case of the delta method, it is possible to show the
Ragulina and Reitan 2017). asymptotic normality of the MLE. The present finding is in
The station-wise best statistical models are chosen by the line with the work of Das and Umamahesh (2017) and Mondal
LR test for AMS of streamflow. From Table 3, it can be noted and Mujumdar (2016). It is observed that the stationary
that the most significant physical driver is different from one approach overestimates (underestimates) the return level at
78 J. DAS AND N. V. UMAMAHESH

Figure 9. Spatial map of stationary and non-stationary stations over Godavari basin.

five (10) out of 15 non-stationary models. This can be attrib­ and non-stationary conditions using the EWT interpretation,
uted to the improper parameter estimation of the fitted dis­ it was found that for a particular return level under the sta­
tribution. In particular, it can be justified for two reasons (De tionary condition, the return period has changed in the case of
Luca and Galasso 2018): in a stationary framework, the model the non-stationary condition. Therefore, the curves developed
does not correctly represent some outlier events that occurred here (Fig. 7) would be useful for making design decisions
in the past; the length of the period is not sufficient enough to under non-stationary flood regimes due to the influence of
reproduce the recent events. Hence, this can increase the risk large-scale climate oscillations and regional hydrological vari­
of failure of engineering designs and/or can significantly abilities over 15 stations of Godavari River basin. For instance,
increase the economic aspects of design criteria. It is worth to ensure an adequate level of protection for water resource
mentioning that the uncertainty bound in the return level is infrastructure over the catchment area draining at Rajoli out­
significantly influenced by the sample size. For instance, the let, for an initial design period of 50 years in the climate change
larger the size of the data sample, the lower is the uncertainty scenario, one may have to use the design level with a return
(Xu et al. 2007). In addition, the effect of sample size is more period of 75 year. In other words, the stationary return period
significant when estimating the return levels for larger return of 75 years should be considered to safeguard the event having
periods, especially when the return period is larger than the a return period of 50 years under non-stationary conditions.
size of the sample. Conversely, at Mancherial station, to ensure the stationary
It is difficult to move the return period concept from sta­ design return level corresponding to a period of 100 years,
tionary to non-stationary conditions (Cooley 2013). Moreover, one must consider a return period of 260 years under non-
it is not straightforward to answer which flood magnitude stationary conditions. Therefore, it is of paramount imporance
should be considered for flood risk management under non- to consider the consequences of climate change in the engi­
stationary conditions. However, design criteria for flood risk neering design criteria to ensure the resilience and sustainabil­
management can be formulated by using risk-based analysis ity of infrastructure in future.
for stationary and non-stationary conditions. Therefore, in the It is observed from Fig. 7 that over Dhalegaon, Hivra,
case of operational purpose, it can be defined (Serinaldi and Mancherial, Polavarm, and Somanpally stations, the high
Kilsby 2015) by extending the concept of EWT until an excee­ return period in non-stationary conditions is equivalent to
dance occurs. It is worth exploring the concept of return the low return period under stationary conditions. It is noted
period and risk under the non-stationary approach by widen­ that PR, SM and ISMI are the selected covariates that affect the
ing the geometric distribution to incorporate physically based location parameter of the GEV distribution linearly. Moreover,
covariates. When making a comparison between stationary from Table 3, it can be noted that the increase in the
 HYDROLOGICAL SCIENCES JOURNAL 79

precipitation is likely to shift the location parameter towards ● The concept of the return period was revisited under
the right side (upper tail), which will further increase the non-stationary conditions and significant differences
probability of flood events. Conversely, the decrease in the were found between stationary and non-stationary
precipitation is likely to shift the location parameter towards approaches that may affect engineering designs (mostly
the left side (lower tail), which will further decrease the prob­ hydraulic structures) in future.
ability of flood events.However, it is observed that the covari­ ● It was found that in order to assess the risk (reliability)
ate PR over Hivra, Mancherial, Polavaram, and Somanpally under the influence of physically based covariates, non-
stations has decreased over time (Fig. S7). Thus, a decrease in stationary analysis is more useful.
the precipitation is likely to shift the location parameter
towards left side (lower tail), which will reduce the probability Future work can be performed to examine the concepts of
of flood events. In the case of Dhalegaon, the dam control may return period, risk and reliability for hydrological
have an influence on controlling the flood magnitude. extremes such as precipitation, which would provide useful
However, due to the unavailability of storage volume in information regarding the design criteria for urban drainage
each year, the influence of the dam is not reflected in modelling planning. In addition, a different interpretation approach,
the location parameter. In addition, it can be noted from Fig. 3 namely expected number of exceedances (ENE), could be
that there is a significant decrease in the peak flows at used to revist the concepts of return period, risk and reliability,
Dhalegaon station over the years. With this understanding, and a comparative study could be carried out with EWT
over these five stations, the flood events are likely to decrease interpretation. Further, future covariates can be extracted
under the influence of physical covariates, and hence the from the outputs of global climate models (GCMs) to assess
equivalent return period is high in the case of non-stationary return period and risk under different climate change scenar­
conditions as compared to the stationary approach. Similarly, ios. Finally, an analysis of multivariate hydrological design
for 10 out of 15 stations the low return period under non- under the non-stationary setting would provide a more
stationary conditions is equivalent to the high return period rational design strategy under a changing climate scenario.
under stationary conditions. Interestingly, it can be noted that
the location parameter is modelled under the influence of
large-scale modes (mainly ISMI) alone or in combination Disclosure statement
with regional hydrological variability over these 10 stations. No potential conflict of interest was reported by the authors.
The non-stationary analysis can be beneficial to assess the
risk (reliability) of water resource infrastructure during its
planned project life. Thus, the analysis of risk for different ORCID
design lifetimes is carried out independently for different gau­
ging sites in Godavari River basin. It should be noted that the Jew Das http://orcid.org/0000-0002-0728-7708
N. V. Umamahesh http://orcid.org/0000-0003-0460-8956
design life of a project and the return period of design criteria
are not same. The risk is computed for 50-, 75-, and 100-year
return periods and different design lifetimes. In this sense, for Data availability statement
example (from Fig. 8), one can define that the risk (reliability)
of an engineering structure is 63% (37%) considering the 100- Data, models, and/or code generated or used during the study, including
year return period and for a 100-year design life under sta­ historical hydrometeorological data and the code for the nonstationary
analysis, are available from the corresponding author upon request.
tionary conditions at Jagdalpur gauging station. However, for
the same criteria, the risk can be computed as 79% (21%)
under non-stationary condition. As a discrepancy between References
the stationary and non-stationary conditions can be quite Alexander, L.V., Uotila, P., and Nicholls, N., 2009. Influence of sea surface
significant, it is useful to incorporate non-stationary analysis temperature variability on global temperature and precipitation
in order to make an appropriate assessment of risk (reliability) extremes. Journal of Geophysical Research, 114 (D18), D18116.
for hydraulic structure. The present analysis and application doi:10.1029/2009JD012301
may be beneficial to water managers in sustainable planning Bayazit, M., 2015. Nonstationarity of hydrological records and recent
trends in trend analysis: a state-of-the-art review. Environmental
and management of water resource infrastructure under the Processes, 2 (3), 527–542. doi:10.1007/s40710-015-0081-7
influence of global climate change and regional variability. Berg, P., Moseley, C., and Haerter, J.O., 2013. Strong increase in con­
Summarizing the results from the present analysis over vective precipitation in response to higher temperatures. Nature
the second longest river in India, the critical outcomes are as Geoscience, 6 (3), 181–185. doi:10.1038/ngeo1731
follows: Cassola, F., et al., 2016. The role of the sea on the flash floods events over
Liguria (northwestern Italy). Geophysical Research Letters, 43 (7),
3534–3542. doi:10.1002/2016GL068265
● The AMS of streamflow for 15 out of 30 gauging stations Chang, X., et al., 2019. Characterizing effects of monsoons and climate
are impacted by large-scale climatic oscillations and teleconnections on precipitation in China using wavelet coherence and
regional hydrological variabilities, with more significant global coherence. Climate Dynamics, 52 (9–10), 5213–5228.
influence of ISMI and precipitation. doi:10.1007/s00382-018-4439-1
Chen, X., et al., 2021. Urbanization and climate change impacts on future
● The return level uncertainty reveals that for higher return
flood risk in the Pearl River Delta under shared socioeconomic
periods the uncertainty bound is larger for both station­ pathways. Science of the Total Environment, 762, 143144. doi:10.1016/
ary and non-stationary conditions. j.scitotenv.2020.143144
80 J. DAS AND N. V. UMAMAHESH

Cheng, L., et al., 2014. Non-stationary extreme value analysis in a chan­ Jha, S., Das, J., and Goyal, M.K., 2021. Low frequency global-scale modes
ging climate. Climatic Change, 127 (2), 353–369. doi:10.1007/s10584- and its influence on rainfall extremes over India: nonstationary and
014-1254-5 uncertainty analysis. International Journal of Climatology, 41 (3),
Cheng, L. and Aghakouchak, A., 2014. Nonstationary precipitation 1873–1888. doi:10.1002/joc.6935
intensity-duration-frequency curves for infrastructure design in a Jiang, C., et al., 2019. Multivariate hydrologic design methods under
changing climate. Scientific Reports, 4, 1–6. doi:10.1038/srep07093 nonstationary conditions and application to engineering practice.
Coles, S., 2001. An introduction to statistical modeling of extreme values. Hydrology and Earth System Sciences, 23 (3), 1683–1704. doi:10.5194/
London: Springer Series in Statistics. doi:10.1007/978-1-4471-3675-0 hess-23-1683-2019
Coles, S., Pericchi, L.R., and Sisson, S., 2003. A fully probabilistic Jongman, B., et al., 2014. Increasing stress on disaster-risk finance due to
approach to extreme rainfall modeling. Journal of Hydrology, 273 (1– large floods. Nature Climate Change, 4 (4), 264–268. doi:10.1038/
4), 35–50. doi:10.1016/S0022-1694(02)00353-0 nclimate2124
Cooley, D., 2013. Return periods and return levels under climate change. In: Katz, R.W., 2013. Statistical methods for nonstatonary extremes. In: A.
A. AghaKouchak, et al., eds. Extremes in a changing climate: detection, AghaKouchak, et al., eds. Extremes in a changing climate: detection,
analysis and uncertainty. Dordrecht, Netherlands: Springer, 97–114. analysis and uncertainty water science and technology library. Vol. 65.
Daksiya, V., Mandapaka, P.V., and Lo, E.Y.M., 2021. Effect of climate Dordrecht: Springer Netherlands, 15–37. doi:10.1007/978-94-007-
change and urbanisation on flood protection decision-making. Journal 4479-0
of Flood Risk Management, 14 (1). doi:10.1111/jfr3.12681 Katz, R.W., Parlange, M.B., and Naveau, P., 2002. Statistics of extremes in
Das, J., Jha, S., and Goyal, M.K., 2020a. Non-stationary and copula-based hydrology. Advances in Water Resources, 25 (8–12), 1287–1304.
approach to assess the drought characteristics encompassing climate doi:10.1016/S0309-1708(02)00056-8
indices over the Himalayan states in India. Journal of Hydrology, 580, Kenyon, J. and Hegerl, G.C., 2010. Influence of modes of climate varia­
124356. doi:10.1016/j.jhydrol.2019.124356 bility on global precipitation extremes. Journal of Climate, 23 (23),
Das, J., Jha, S., and Goyal, M.K., 2020b. On the relationship of climatic 6248–6262. doi:10.1175/2010JCLI3617.1
and monsoon teleconnections with monthly precipitation over Li, J. and Tan, S., 2015. Nonstationary flood frequency analysis for annual
meteorologically homogenous regions in India: wavelet & global flood peak series, adopting climate indices and check dam index as
coherence approaches. Atmospheric Research, 238, 104889. covariates. Water Resources Management, 29 (15), 5533–5550.
doi:10.1016/j.atmosres.2020.104889 doi:10.1007/s11269-015-1133-5
Das, J. and Umamahesh, N.V., 2017. Uncertainty and nonstationarity in López, J. and Francés, F., 2013. Non-stationary flood frequency analysis in
streamflow extremes under climate change scenarios over a river basin. continental Spanish rivers, using climate and reservoir indices as
Journal of Hydrologic Engineering, 22 (10), 04017042. doi:10.1061/ external covariates. Hydrology and Earth System Sciences, 17 (8),
(ASCE)HE.1943-5584.0001571 3189–3203. doi:10.5194/hess-17-3189-2013
Das, T., et al., 2011. Potential increase in floods in California’s Sierra Loucks, D.P. and van Beek, E., 2017. An Introduction to probability,
Nevada under future climate projections. Climatic Change, 109 (S1), statistics, and uncertainty. Water resource systems planning and man­
71–94. doi:10.1007/s10584-011-0298-z agement: an introduction to methods, models, and applications. In:
De Luca, D.L. and Galasso, L., 2018. Stationary and non-stationary frame­ D.P. Loucks and E. van Beek, eds. Cham: Springer International
works for extreme rainfall time series in southern Italy. Water Publishing, 213–300. doi:10.1007/978-3-319-44234-1_6
(Switzerland), 10 (10). doi:10.3390/w10101477 Mallakpour, I., AghaKouchak, A., and Sadegh, M., 2019. Climate-induced
Dong, Q., et al., 2019. An improved nonstationary model for flood changes in the risk of hydrological failure of major dams in California.
frequency analysis and its implication for the Three Gorges Dam, Geophysical Research Letters, 46 (4), 2130–2139. doi:10.1029/
China. Hydrological Sciences Journal, 64 (7), 845–855. doi:10.1080/ 2018GL081888
02626667.2019.1596274 Martins, E.S. and Stedinger, J.R., 2000. Generalized maximum-likelihood
Du, T., et al., 2015. Return period and risk analysis of nonstationary generalized extreme-value quantile estimators for hydrologic data.
low-flow series under climate change. Journal of Hydrology, 527, Water Resources Research, 36 (3), 737–744. doi:10.1029/1999WR900330
234–250. doi:10.1016/j.jhydrol.2015.04.041 Milly, P.C.D., et al., 2008. Stationarity is dead: whither water
Feng, Y., et al., 2020. Nonstationary flood coincidence risk analysis using management? Science (80-.), 319 (5863), 573–574. doi:10.1126/
time-varying copula functions. Scientific Reports, 10 (1), 3395. science.1151915
doi:10.1038/s41598-020-60264-3 Milly, P.C.D., et al., 2015. On critiques of “Stationarity is dead: whither
Gao, L., et al., 2017. Risk of extreme precipitation under nonstationarity water management?” Water Resources Research, 51 (9), 7785–7789.
conditions during the second flood season in the Southeastern Coastal doi:10.1002/2015WR017408
Region of China. Journal of Hydrometeorology, 18 (3), 669–681. Min, S.-K., et al., 2011. Human contribution to more-intense precipita­
doi:10.1175/JHM-D-16-0119.1 tion extremes. Nature, 470 (7334), 378–381. Nature Publishing Group.
Gilroy, K.L. and McCuen, R.H., 2012. A nonstationary flood frequency doi:10.1038/nature09763
analysis method to adjust for future climate change and urbanization. Mondal, A. and Mujumdar, P.P., 2015. Modeling non-stationarity in
Journal of Hydrology, 414-415, 40–48. doi:10.1016/j.jhydrol.2011.10.009 intensity, duration and frequency of extreme rainfall over India.
Guimberteau, M., et al., 2013. Future changes in precipitation and Journal of Hydrology, 521, 217–231. doi:10.1016/j.jhydrol.2014.
impacts on extreme streamflow over Amazonian sub-basins. 11.071
Environmental Research Letters, 8 (1), 014035. doi:10.1088/1748- Mondal, A. and Mujumdar, P.P., 2016. Detection of change in flood return
9326/8/1/014035 levels under global warming. Journal of Hydrologic Engineering, 21 (8),
Hengade, N. and Eldho, T.I., 2019. Relative impact of recent climate and 04016021. doi:10.1061/(ASCE)HE.1943-5584.0001326
land cover changes in the Godavari River basin, India. Journal of Earth Mood, A.M., Graybill, F.A., and Boes, D.C., 1974. Introduction to the
System Science, 128 (4), 94. doi:10.1007/s12040-019-1135-4 theory of statistics 1974. Tokyo: McGraw-Hill Kogakusha.
Hesarkazzazi, S., et al., 2021. Stationary vs non-stationary modelling of Morice, C.P., et al., 2012. Quantifying uncertainties in global and regional
flood frequency distribution across northwest England. Hydrological temperature change using an ensemble of observational estimates: the
Sciences Journal, 66 (4), 729–744. doi:10.1080/02626667.2021.1884685 HadCRUT4 data set. Journal of Geophysical Research: Atmospheres,
Ionita, M., et al., 2012. Interannual to decadal summer drought variability 117 (D8). doi:10.1029/2011JD017187
over Europe and its relationship to global sea surface temperature. Nasr, A., et al., 2019. A review of the potential impacts of climate change
Climate Dynamics, 38 (1–2), 363–377. doi:10.1007/s00382-011-1028-y on the safety and performance of bridges. Sustainable and Resilient
Jain, S. and Lall, U., 2000. Magnitude and timing of annual maximum Infrastructure. doi:10.1080/23789689.2019.1593003
floods: trends and large-scale climatic associations for the Blacksmith NatCatSERVICE, 2014. NatCatSERVICE database. Munich: Munich RE.
Fork River, Utah. Water Resources Research, 36 (12), 3641–3651. Oehlert, G.W., 1992. A note on the delta method. American Statistician,
doi:10.1029/2000WR900183 46 (1), 27–29. doi:10.1080/00031305.1992.10475842
 HYDROLOGICAL SCIENCES JOURNAL 81

Olsen, J.R., Lambert, J.H., and Haimes, Y.Y., 1998. Risk of extreme events Su, C. and Chen, X., 2019. Assessing the effects of reservoirs on extreme
under nonstationary conditions. Risk Analysis, 18 (4), 497–510. flows using nonstationary flood frequency models with the modified
doi:10.1111/j.1539-6924.1998.tb00364.x reservoir index as a covariate. Advances in Water Resources, 124,
Papalexiou, S.M. and Koutsoyiannis, D., 2013. Battle of extreme value 29–40. doi:10.1016/j.advwatres.2018.12.004
distributions: a global survey on extreme daily rainfall. Water Towler, E., et al., 2010. Modeling hydrologic and water quality extremes in
Resources Research, 49 (1), 187–201. doi:10.1029/2012WR012557 a changing climate: a statistical approach based on extreme value
Parey, S., Hoang, T.T.H., and Dacunha-Castelle, D., 2010. Different ways theory. Water Resources Research, 46 (11), 1–11. doi:10.1029/
to compute temperature return levels in the climate change context. 2009WR008876
Environmetrics, 21 (7–8), 698–718. doi:10.1002/env.1060 Villarini, G., et al., 2009. Flood frequency analysis for nonstationary
Prosdocimi, I., Kjeldsen, T.R., and Miller, J.D., 2015. Detection and annual peak records in an urban drainage basin. Advances in
attribution of urbanization effect on flood extremes using nonstation­ Water Resources, 32 (8), 1255–1266. doi:10.1016/j.
ary flood-frequency models. Water Resources Research, 51 (6), 4244– advwatres.2009.05.003
4262. doi:10.1002/2015WR017065 Villarini, G., Smith, J.A., and Napolitano, F., 2010. Nonstationary model­
Ragulina, G. and Reitan, T., 2017. Generalized extreme value shape para­ ing of a long record of rainfall and temperature over Rome. Advances
meter and its nature for extreme precipitation using long time series in Water Resources, 33 (10), 1256–1267. doi:10.1016/j.advwatres.2010.
and the Bayesian approach. Hydrological Sciences Journal, 62 (6), 863– 03.013
879. doi:10.1080/02626667.2016.1260134 Wang, B., 2000. commentary and analysis: comments on ‘choice of south
Ray, L.K. and Goel, N.K., 2019. Flood frequency analysis of Narmada River Asian summer monsoon indices’. Bulletin of the American
Basin in India under nonstationary condition. Journal of Hydrologic Meteorological Society, 81 (4), 821–822. doi:10.1175/1520-0477(2000)
Engineering, 24 (8), 05019018. doi:10.1061/(asce)he.1943-5584.0001808 081<0821:CAA>2.3.CO;2
Read, L.K. and Vogel, R.M., 2015. Reliability, return periods, and risk Westra, S., Alexander, L.V., and Zwiers, F.W., 2013. Global increasing
under nonstationarity. Water Resources Research, 51 (8), 6381–6398. trends in annual maximum daily precipitation. Journal of Climate,
doi:10.1002/2015WR017089 26 (11), 3904–3918. doi:10.1175/JCLI-D-12-00502.1
Sahu, N., et al., 2012. IOD and ENSO impacts on the extreme Wigley, T.M.L., 2009. The effect of changing climate on the frequency of
stream-flows of Citarum river in Indonesia. Climate Dynamics, absolute extreme events. Climatic Change, 97 (1–2), 67–76.
39 (7–8), 1673–1680. doi:10.1007/s00382-011-1158-2 doi:10.1007/s10584-009-9654-7
Saji, N.H., et al., 1999. A dipole mode in the tropical Indian Ocean. Xu, Y.P., Booij, M.J., and Mynett, A.E., 2007. Propagation of discharge
Nature, 401 (6751), 360–363. doi:10.1038/43854 uncertainty in a flood damage model for the meuse river. Advances in
Salas, J.D. and Obeysekera, J., 2014. Revisiting the concepts of return Natural and Technological Hazards Research, 25, 293–310.
period and risk for nonstationary hydrologic extreme events. Journal of doi:10.1007/978-1-4020-4200-3_16
Hydrologic Engineering, 19 (3), 554–568. doi:10.1061/(ASCE)HE.1943- Yao, S., Huang, Q., and Zhao, C., 2015. Variation characteristics of rainfall
5584.0000820 in the pre-flood season of South China and its correlation with sea
Sarhadi, A. and Soulis, E.D., 2017. Time-varying extreme rainfall inten­ surface temperature of Pacific. Atmosphere (Basel), 7 (1), 5.
sity-duration-frequency curves in a changing climate. Geophysical doi:10.3390/atmos7010005
Research Letters, 44 (5), 2454–2463. doi:10.1002/2016GL072201 Ye, S., et al., 2015. Vegetation regulation on streamflow intra-annual
Serinaldi, F. and Kilsby, C.G., 2015. Stationarity is undead: uncertainty variability through adaption to climate variations. Geophysical
dominates the distribution of extremes. Advances in Water Resources, Research Letters, 42 (23). doi:10.1002/2015GL066396
77, 17–36. doi:10.1016/j.advwatres.2014.12.013 Yilmaz, A.G. and Perera, B.J.C., 2014. Extreme rainfall nonstationarity
Sinha, J., et al., 2018. Assessment of the impacts of climatic variability and investigation and intensity – frequency – duration relationship. Journal
anthropogenic stress on hydrologic resilience to warming shifts in of Hydrologic Engineering, 19 (6), 1160–1172. doi:10.1061/(ASCE)
Peninsular India. Scientific Reports, 8 (1), 13833. doi:10.1038/s41598- HE.1943-5584.0000878
018-32091-0 Zhang, Q., et al., 2015. Evaluation of flood frequency under
Sinha, J., et al., 2020. Analysing model disparity in diagnosing the climatic non-stationarity resulting from climate indices and reservoir indices
and human stresses on runoff variability over India. Journal of in the East River basin, China. Journal of Hydrology, 527, 565–575.
Hydrology, 581, 124407. doi:10.1016/j.jhydrol.2019.124407 doi:10.1016/j.jhydrol.2015.05.029
Šraj, M., et al., 2016. The influence of non-stationarity in extreme hydro­ Zhou, G., et al., 2010. Forest recovery and river discharge at the regional
logical events on flood frequency estimation. Journal of Hydrology and scale of Guangdong Province, China. Water Resources Research, 46 (9).
Hydromechanics, 64 (4), 426–437. doi:10.1515/johh-2016-0032 doi:10.1029/2009WR008829