Advances in Water Resources 79 (2015) 127–139

Contents lists available at ScienceDirect

Advances in Water Resources

journal homepage: www.elsevier.com/locate/advwatres

Model and parameter uncertainty in IDF relationships under climate

change
Chandra Rupa a, Ujjwal Saha b, P.P. Mujumdar a,c,⇑
a
Department of Civil Engineering, Indian Institute of Science, Bangalore, Karnataka 560 012, India
b
Department of Civil Engineering, Indian Institute of Engineering Science and Technology, Shibpur, West Bengal 711 103, India
c
Divecha Center for Climate Change, Indian Institute of Science, Bangalore, India

a r t i c l e i n f o a b s t r a c t

Article history: Quantifying distributional behavior of extreme events is crucial in hydrologic designs. Intensity Duration
Received 25 July 2014 Frequency (IDF) relationships are used extensively in engineering especially in urban hydrology, to obtain
Received in revised form 7 January 2015 return level of extreme rainfall event for a specified return period and duration. Major sources of uncer-
Accepted 22 February 2015
tainty in the IDF relationships are due to insufficient quantity and quality of data leading to parameter
Available online 12 March 2015
uncertainty due to the distribution fitted to the data and uncertainty as a result of using multiple
GCMs. It is important to study these uncertainties and propagate them to future for accurate assessment
Keywords:
of return levels for future. The objective of this study is to quantify the uncertainties arising from parame-
Bayesian method
Intensity Duration Frequency relationships
ters of the distribution fitted to data and the multiple GCM models using Bayesian approach. Posterior
Climate change distribution of parameters is obtained from Bayes rule and the parameters are transformed to obtain
Short duration rainfall return levels for a specified return period. Markov Chain Monte Carlo (MCMC) method using
Uncertainties Metropolis Hastings algorithm is used to obtain the posterior distribution of parameters. Twenty six
CMIP5 GCMs along with four RCP scenarios are considered for studying the effects of climate change
and to obtain projected IDF relationships for the case study of Bangalore city in India. GCM uncertainty
due to the use of multiple GCMs is treated using Reliability Ensemble Averaging (REA) technique along
with the parameter uncertainty. Scale invariance theory is employed for obtaining short duration return
levels from daily data. It is observed that the uncertainty in short duration rainfall return levels is high
when compared to the longer durations. Further it is observed that parameter uncertainty is large
compared to the model uncertainty.
Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction stochastic variation of the rainfall at extreme levels. Based on this

inferred model, estimates of parameters are obtained using
Quantification of precipitation extremes is important in hydro- standard statistical techniques and further inferences are drawn
logic designs for developing structural and nonstructural measures for hydrologic designs. There are two major concerns as to why
for flood protection, especially in urban areas. In order to appraise the standard approach leads to systematic underestimation
the extreme character of rainfall in urban drainage, Intensity (overestimation in a few cases) of the probability of disastrous
Duration Frequency (IDF) relationships are used. It is crucial to events. First concern is related to the non-availability of adequate
have reliable estimates of intensities of extreme rainfall for optimal quantity and quality of historical data in developing the asymp-
designing of urban infrastructure. Use of standard statistical tools totic model. The classical methods rely on sampling distribution,
for the prediction of extremes, not only in hydrology but also in which involves the behavior of a statistic over repeating sampling,
several scientific disciplines, has led to dissatisfaction. Several which requires sufficiently ’large‘ data to ascertain the correctness
examples abound in gross underestimation/overestimation of of the asymptotic model fitted to the data. The second concern is
designs based on standard procedure that subsequently cause related to parameter uncertainty, which is often overlooked or
catastrophic damage [9]. ignored most of the times. It is important to quantify these
A standard methodology for modeling rainfall extremes is uncertainties and use them in the design decisions and risk
performed by employing an asymptotic model to describe the analyses. Bayesian approach provides a coherent framework for
incorporating these uncertainties.
⇑ Corresponding author at: Department of Civil Engineering, Indian Institute of Very few attempts are made in the past to incorporate Bayesian
Science, Bangalore, Karnataka 560 012, India. approach into extreme value analysis due to computational

http://dx.doi.org/10.1016/j.advwatres.2015.02.011
0309-1708/Ó 2015 Elsevier Ltd. All rights reserved.
128 R. Chandra et al. / Advances in Water Resources 79 (2015) 127–139

difficulties in obtaining posterior/predictive distributions which Southern Quebec using regional frequency analysis, predicted a
requires integration over parameter space. Bayesian analysis has 50% decrease by 2050 for rainfall durations of two and six hours,
received due attention, however, after the introduction of and a 32% decrease for 12 and 24 h rainfall durations. Buonomo
Markov Chain Monte Carlo (MCMC) methodology [15] overcoming et al. [5] found that changes in extreme rainfall values become
the limitations due to computational aspects. MCMC allows greater as both the return period of the rainfall becomes longer
approximation of integrals by setting up a Markov chain with and the considered duration becomes shorter. Using the informa-
posterior/predictive distribution as invariant distribution of the tion provided by five general circulation models under four future
chain. Bayesian methods have the capability to address the uncer- climate scenarios of greenhouse gas emissions, Rodrıguez et al.
tainties incorporating additional sources of important information [45] reported an increase of at least 4% on the expected daily rain-
[21]. Smith and Naylor [50] examined the effect of different priors fall with a return period longer than 20 years. They observed that
using specially adapted versions of numerical quadrature to obtain increase in the expected hourly rainfall is slightly higher than the
the posterior distributions of the parameters of Weibull corresponding daily rainfall.
distribution. Their work is the first one to exemplify the com- The literature cited above is limited to addressing either the
putational feasibility in modeling the extremes, which is scientifi- uncertainties in parameters of fitted distribution for historical time
cally attractive but intractable analytically. Similar works for period or the uncertainties in GCMs and scenarios for future time
Weibull models are carried out in studying the issue of prior period. Till date, no study has been carried out applying the
elicitation [4,47]. A few investigations focused on specialized sub- Bayesian approach in addressing uncertainties in parameters,
families of Generalized Extreme Value (GEV) distributions to avoid propagation of these uncertainties into the future periods to assess
analytical and computational difficulties with many restrictions on the effect on IDF relationships that are, important in hydraulic
parameters [10]. Coles and Powell [8] reviewed the developments design of urban infrastructure. Therefore, in this paper, likely
in Bayesian methods in extreme value modeling and also discussed changes in IDF relationships and the uncertainties associated with
the modeling ideas based on MCMC techniques. Feasibility of the return levels under climate change are modeled for future time
Bayesian analysis in a situation where maximum likelihood fails period using Bayesian method. An improved Markov chain weather
– i.e., the maximum likelihood is unbounded when the parent generator is used to downscale the daily rainfall for future period
population from which the data is drawn has short tail - is also pre- using twenty six CMIP5 (Coupled Model Intercomparison Project
sented. Bayesian estimators for inference in non-regular likelihood Version 5) GCM simulations and four Representative Concentration
cases were suggested by Smith [48,49]. Coles et al. [9] reaffirmed Pathway (RCP) scenarios. Reliability Ensemble Averaging (REA) is
that classical methods do not take comprehensive account of the used for modeling uncertainties resulting from multiple GCMs
uncertainties in model and predictions and are confined to develop obtaining future daily time series for different RCP scenarios. This
an over optimistic appraisal of extremes in future that is often con- method with the assignment of weights to GCMs based on evalua-
tradicted by observed events. Huard et al. [21] obtained IDF curves tion of model is one of the methods to address GCM uncertainty
from Bayesian analysis showing the extent of uncertainties in the [18]. The ensemble averaging approach is characterized by the
curves and evaluating the risk of misinterpretation of the IDF uncertainty caused by partial ignorance, in which weights are
relationships. In their work, the method of MCMC is not employed assigned to GCMs taking into the account of model performance
in estimating the posterior distribution of parameters, but the and model convergence criteria [19]. Disaggregation from daily
numerical quadrature algorithms are used in evaluating the pos- scale to hourly and sub-hourly scales is performed using Scale
terior distribution of return levels. The integration bounds are Invariance theory. The flow chart in Fig. 1 gives the overview of
determined by approximating the pdf by bivariate normal dis- work. The following section presents the data used and over view
tribution to reduce the computational time. of the Bayesian methodology is given in Section 3. Details of
In all the above literature only the historical (or the observed) statistical downscaling using CMIP5 data, REA methodology and
time line is considered in evaluating the uncertainties for obtaining the method for obtaining the IDF relationships for sub-hourly
the IDF relationships. Nevertheless, addressing additional uncer- durations are discussed in Section 4. In Section 5, concluding
tainties due to climate change in the IDF relationships is crucial. remarks with a discussion on advantages and limitations of the
Many studies earlier have addressed GCM and scenario uncertainty approach are provided.
without considering the uncertainty in the model fitted to the data
[12,17,33]. New and Hulme [34] presented uncertainty in climate
change scenarios by developing a hierarchical model addressing Historical time Bayesian Historical probabilistic
uncertainty in future greenhouse gas emissions, climate sensitivity series analysis IDF relationships
and unpredictability in GCMs for seasonal mean temperature and
precipitation over the United Kingdom. Jones [22] used a similar Statistical downscaling
methodology for risk assessment and sensitivity analysis for irriga- with weather generator

tion demand. Allen et al. [2] applied Bayesian models to multimo- Multi model
del ensembles to characterize uncertainty, and obtained pdf of ensemble GCM
Ensemble of projected annual
precipitation data
temperature at regional scales for future climate. The uncertainties maximum precipitation series
due to GCM models are studied in developing IDF relationships for Observed hourly and sub-
future without considering the uncertainties in the parameters of Reliability Ensemble
hourly precipitation
distribution fitted to the extreme data [32,38,46,52]. The IDF Averaging (REA)
curves for the city of London using K-NN based weather generator
was developed by Prodanovic and Simonovic [42], wherein 30% Weighted projected time
Scale invariance theory:
Disaggregation from daily scale
increase in rainfall intensity is projected for a range of return series of annual maximum
to hourly and sub-hourly scale
periods and durations. Coulibaly and Shi [11] developed IDF precipitation at daily scale
Bayesian
relationships using the outputs from Canadian Global Circulation analysis
Model-2 (CGCM2) and employing Statistical Down-Scaling Model Projected probabilistic
(SDSM) downscaling methodology for Grand River and Kenora IDF relationships
Rainy River regions in Ontario. The outputs from Regional
Climate Models (RCMs) used by Mailhot et al. [25,26] for Fig. 1. Overview of the proposed framework.
 R. Chandra et al. / Advances in Water Resources 79 (2015) 127–139 129

2. Observed and model data period. The Generalized Extreme Value (GEV) distribution fits the
AMS data for all the durations considered. Cumulative
Bangalore City, the study area, is located at 12.59 °N 77.57 °E, Distribution Function (CDF) of the GEV distribution is given by
and is situated at an altitude of 920 m above mean sea level. The 8   
< exp  exp  zrl k¼0
average annual rainfall of the city is 974 mm, summer temperature n
GðzÞ ¼  zl1=k o
ranges from 18 to 38 °C and the winter temperature ranges from : exp  1 þ k r k–0; fz : 1 þ kðx  lÞ=r > 0g
12 to 25 °C. Bangalore receives rainfall from both southwest
(June to September) and northeast monsoons (October to ð1aÞ
December), with the wettest months being August, September where l is location parameter (l e R), r the scale parameter (r > 0)
and October. In the city, the summer monsoon rainfall is and k is the shape parameter (k e R).
predominantly due to convective heat transfer [44]. Convective The return level (zp) associated with the return period 1/p is
precipitation occurs for a short duration with high intensity over given by
a limited area, as the convective clouds have limited horizontal
extent. zp ¼ F 1 ðp=l; r; kÞ
( h i
Rainfall data is collected from Indian Meteorological
l  rk 1  f logð1  pÞgk for k–0 ð1bÞ
Department (IMD) Pune and Bangalore, recorded at Bangalore city zp ¼
station (IMD station No. 43295). Even though daily rainfall data is l  r log ½logð1  pÞ for k ¼ 0
available for more than hundred years (1901–2013), hourly and
The basic principle of Bayesian method is to upgrade one’s uncer-
15minute rainfall data are available for the period of 1969–2003.
tainty about parameter (h) expressed in terms of prior p(h), before
Hence, only the historical period, from 1969–2003, is considered
observing the data to post data availability, using Bayes rule. This
in the study. Precipitation flux outputs of 26 General Circulation
is achieved by modifying the prior distribution of parameters due
Models (GCMs) for historical and future periods is obtained from
to the information contained in the data, expressed through the
CMIP5 website http://pcmdi3.llnl.gov/esgcet/home.htm (accessed
likelihood function.
between the months of July and September 2012). A set of 85 sim-
The posterior distribution of parameters pðh=zÞ is obtained by
ulations are considered corresponding to 26 CMIP5 GCMs and four
multiplying the likelihood function of the GEV distribution, Lðz=hÞ
RCP scenarios. The number of GCMs considered for RCP 2.6, 4.5, 6.0
and the prior distribution of parameters (Eq. (2)). The choice of
and 8.5 scenarios are 20, 25, 16 and 24 respectively, adding to 85
prior distribution of parameters is crucial in the Bayesian analysis,
simulations. The list of GCMs and scenarios used is given in
which is discussed in the next section. The un-normalized poster-
Table 1. For calculation of perturbation factors of the weather
ior density of the parameters is given by
generator parameters, the duration 1969–1998 is considered as
the control period. Future scenarios are constructed for three time pðh=zÞ / Lðz=hÞ  pðhÞ; ð2Þ
slices: 2021-50, 2051-80 and 2071-2100. For each climate model, T
the grid point that is nearest to the Bangalore city rain gauge sta- where h is the parameter vector ðl; r; kÞ , pðhÞ is the prior dis-
tion is considered. In case of two or more possible candidate grid tribution of parameters and Lðz=hÞ is the likelihood function equal
Qm
points, the grid point with minimum Mean Absolute Percent to i¼1 f ðzi =hÞ, m is the No. of observations and f ðzi =hÞ is the
Error (MAPE) is considered. MAPE is calculated based on the conditional pdf of annual maximum rainfall (Z) given the parameter
observed and GCM simulated rainfall and return levels. The vector (h). Bayesian inference obeys the so called likelihood principle,
selected grid points for the GCM model are shown in Table 1. which states ‘‘for a given sample of data, any two probability mod-
els that have the same likelihood function yield the same inference
for h’’ [16].
3. Bayesian method Bayesian inference about the model parameters proceeds by
solving the Eq. (2) and obtaining the posterior distribution of
Given a rainfall time series, annual maximum series (AMS) or parameters. It is not possible to solve the Eq. (2) analytically lead-
partial duration series (PDS) is obtained for selected durations, ing to a search of alternative methods for its solution. We are inter-
and a suitable distribution is fitted to this AMS/PDS in the classical ested in various features of the posterior distribution, which can be
approach. The parameters of the selected distribution are obtained approximately studied if one has a large sample h1, h2, . . ., hN from
classically by various methods, namely Maximum Likelihood the posterior distribution of h. For example, for a given return per-
Estimation (MLE), method of moments, and L-moments. iod 1/p, the return level will be obtained from the sequence of h.
However, the uncertainty associated with the point estimate of The method of Markov Chain Monte Carlo (MCMC) is used for
parameters is not usually accounted for by the classical methods. drawing large sample of h, which involves setting up a Markov
Bayesian approach provides a coherent framework for incorporat- chain with pðh=zÞ as its stationary distribution and then drawing
ing the uncertainties in estimation of parameters [9]. Analysis of samples using its transition density from this Markov chain. A
extreme rainfall is often hampered by lack of sufficient data; espe- detailed explanation of MCMC method is given in Appendix A.
cially, the availability of short duration (15 min, 30 min) rainfall
data for adequate period is a great deal. When the sample size is 3.1. Prior distribution of parameters
small, effectiveness of classical methods is questionable [10]. As
the data for short duration events are sparse, optimal use of The choice of prior is crucial in Bayesian analysis and the prior
available data should be done. Bayesian methods are proven to elicitation is usually based on understanding physical rainfall
be the most desirable way of analyzing the sparse data, waiving process, knowledge of rainfall characteristics etc. within vicinity
the concerns of repeated sampling used by classical methods of study area [10]. In case of absence of such information, non-
[9,21]. Bayesian IDF relationships are worked out much the same informative distributions or arbitrary distribution with large
way as the classical IDF relationships, the main difference being variances are adopted to reflect prior ignorance [9]. Smith and
that instead of using the optimal set of parameters, the entire Naylor [50] worked with 3 parameter Weibull distribution using
probability distribution of parameters is considered. arbitrary informative priors, which are chosen to reflect a range
The annual maximum series is prepared for durations of 15 min, of potential scientific hypothesis. Singpurwalla and Song [47]
30 min, 1 h, 3 h etc. from the rainfall time series for historical worked on the 2 parameter Weibull distribution, incorporated
130 R. Chandra et al. / Advances in Water Resources 79 (2015) 127–139

Table 1
GCM model with future climate scenarios and latitude–longitude selected for the study area.

Model name Modeling center Latitude Longitude Abbreviation Future climate

scenarios
ACCESS1_0 Commonwealth Scientific and Industrial Research Organization (CSIRO) and Bureau of 12.5N 76.875 E A1 RCP 4.5, RCP 8.5
Meteorology (BOM), Australia
BCC-CSM1-1 Beijing Climate Center, China Meteorological Administration 12.55N 78.75 E BC1 RCP 2.6, RCP 4.5,
RCP 6.0, RCP 8.5
BCC-CSM1-1-M Beijing Climate Center, China Meteorological Administration 12.89N 77.6 E BC1M RCP 2.6, RCP 4.5,
RCP 6.0, RCP 8.5
BNU-ESM College of Global Change and Earth System Science, Beijing Normal University 12.55N 78.75 E BNU RCP 2.6, RCP 4.5,
RCP 8.5
CanESM2 Canadian Centre for Climate modeling and Analysis 12.55N 78.7 E CAN RCP 2.6, RCP 4.5,
RCP 8.5
CCSM4 National Center for Atmospheric Research 12.7225N 77.5 E C4 RCP 2.6, RCP 4.5,
RCP 6.0, RCP 8.5
CMCC-CMS Centro Euro-Mediterraneo per I CambiamentiClimatici 12.1N 76.8 E CMS RCP 4.5, RCP 8.5
CNRM-CM5 Centre National de RecherchesMeteorologiques/Centre Europeen de Recherche et 11.9N 77.34 E CN5 RCP 4.5, RCP 8.5
Formation Avancees en CalculScientifique
CSIRO-MK3-6-0 Commonwealth Scientific and Industrial Research Organization 12.12N 76.875 E CM60 RCP 2.6, RCP 4.5,
RCP 6.0, RCP 8.5
FGOALS-G2 LASG, Institute of Atmospheric Physics, Chinese Academy of Sciences and CESS, 12.55N 78.75 E FG2 RCP 2.6, RCP 4.5,
Tsinghua University RCP 8.5
FGOALS-S2 LASG, Institute of Atmospheric Physics, Chinese Academy of Sciences 12.4N 75.9 E FS2 RCP 2.6, RCP 4.5,
RCP 6.0, RCP 8.5
GFDL-CM3 NOAA Geophysical Fluid Dynamics Laboratory 13N 78.75 E GF3 RCP 2.6, RCP 6.0
GFDL-ESM2G NOAA Geophysical Fluid Dynamics Laboratory 13.14N 76.25 E GF2G RCP 2.6, RCP 4.5,
RCP 6.0, RCP 8.5
GFDL-ESM2M NOAA Geophysical Fluid Dynamics Laboratory 13.14N 76.25 E GF2M RCP 2.6, RCP 4.5,
RCP 6.0, RCP 8.5
GISS-E2-R NASA Goddard Institute for Space Studies 13N 76.25 E GISS RCP 4.5
HadGEM2-CC Met Office Hadley Centre (additional HadGEM2-ES realizations contributed by 12.5N 76.87 E HADC RCP 4.5, RCP 8.5
Instituto Nacional de Pesquisas Espaciais)
HadGEM2-ES Met Office Hadley Centre (additional HadGEM2-ES realizations contributed by 12.5N 76.875 E HADE RCP 2.6, RCP 4.5,
Instituto Nacional de Pesquisas Espaciais) RCP 6.0, RCP 8.5
INMCM4 Institute for Numerical Mathematics 12.75N 78 E IN4 RCP 4.5, RCP 8.5
IPSL-CM5A-LR Institut Pierre-Simon Laplace 12.31N 78.75 E IPCL RCP 2.6, RCP 4.5,
RCP 6.0, RCP 8.5
IPSL-CM5A-MR Institut Pierre-Simon Laplace 12.6N 77.5 E IPCM RCP 2.6, RCP 4.5,
RCP 6.0, RCP 8.5
MIROC5 Atmosphere and Ocean Research Institute (The University of Tokyo), National Institute 11.9N 78.75 E MI5 RCP 2.6, RCP 4.5,
for Environmental Studies, and Japan Agency for Marine-Earth Science and Technology RCP 6.0, RCP 8.5
MIROC-ESM-CHEM Japan Agency for Marine-Earth Science and Technology, Atmosphere and Ocean 12.55N 78.75 E MIEC RCP 2.6, RCP 4.5,
Research Institute (The University of Tokyo), and National Institute for Environmental RCP 6.0, RCP 8.5
Studies
MIROC-ESM Japan Agency for Marine-Earth Science and Technology, Atmosphere and Ocean 12.55N 78.75 E MIE RCP 2.6, RCP 4.5,
Research Institute (The University of Tokyo), and National Institute for Environmental RCP 6.0, RCP 8.5
Studies
MPI-ESM-LR Max Planck Institute for Meteorology 12.12N 78.75 E MPI RCP 2.6, RCP 4.5,
RCP 6.0, RCP 8.5
MRI-CGCM3 Meteorological Research Institute 12.89N. 77.62 E MRI RCP 2.6, RCP 4.5,
RCP 6.0, RCP 8.5
NorESM1-M Norwegian Climate Centre 12.31N 77.5 E NEM RCP 2.6., RCP 4.5,
RCP 6.0, RCP 8.5

moments fixed by an expert into an informative prior specified approach is more localized and cannot be used in general. Coles
by an analyst who corrects the anticipated bias and overconfidence [7] considered the parameters of the GEV distribution as indepen-
in the expert’s assessment. [13] considered a 2-parameter dent and following normal distribution with some mean and large
exponential distribution with nonconjugate prior, and also variance. There is, however, no guideline for specifying the mean
studied various 2-parameter exponential distributions with order and variance, even though there is a general principle of employing
restriction on the location parameters using Gibbs sampler. Even a prior with very high variance reflecting the absence of prior infor-
though GEV model admits no conjugate prior distribution, a few mation. Huard et al. [21] used uninformative prior distributions for
studies [3,14] obtained conjugate priors with restrictions on shape locations, scale parameters and beta distribution for shape
parameter. However for practical applications, these models are parameter. The integration bounds are obtained by approximating
not of much use. pdf by bivariate normal distribution to reduce the computational
For GEV distribution, Coles and Tawn [10] elicited prior infor- time. The posterior pdf of intensity is obtained using numerical
mation in terms of return levels (quantiles). They considered three quadrature algorithms.
return levels ðqp1 ; qp2 ; qp3 Þ for pre specified small values of In the present study, non-informative prior distributions are
p1 > p2 > p3, where 1/pi is the return period. They assumed that considered for location (l) and scale (r) parameter. The location
the priors are independent and follows gamma distribution. The parameter is considered uniform pðlÞ / 1 and the scale parameter
hyper parameters of the gamma distribution are determined as a as pðrÞ / 1=r (equivalently p(logr) / 1), following Jeffrey’s dis-
measure of location and variability in prior belief by experts. This tribution [16,21]. Based on extensive studies on scale parameter
 R. Chandra et al. / Advances in Water Resources 79 (2015) 127–139 131

[21,29,30], beta distribution (B(.)) with hyper parameters a = 6 and generator is modified to obtain both the rainfall occurrence and
b = 9 is considered. The prior distribution of parameters pðhÞ is rainfall amount. The computational procedure for downscaling
given by Eq. (3). using weather generator is summarized in the following steps.
Bðk þ 0:5=a ¼ 6; b ¼ 9Þ
pðl; r; kÞ / ð3Þ 1. For the precipitation occurrence, states of the Markov chain
r are divided into no-rain, rain of moderate intensity and rain
The choice of Beta distribution as a prior for shape parameter was of high intensity. The high intensity rainfall events are defined
initially suggested by Martins and Stedinger [29], and this choice as the events above 0.9th quantile (n > 0.9), and moderate
of prior restricts shape parameter to a plausible range intensity rainfall events are the ones below this quantile. The
(0.5 < k < 0.5) consistent with rainfall and flood flows observed transitional probabilities are estimated from the observed
worldwide [21,30]. Park [41] presented a Monte Carlo simulation data.
based systematic way of choosing the hyper-parameters of the 2. Separate distributions are fitted for moderate intensity and high
Beta prior, and conceived that the hyperparameters suggested by intensity rainfall to model the amount of precipitation. Various
Martins and Stedinger are relatively stronger and much more infor- distributions (Generalized Pareto distribution, exponential etc.)
mative than the choices made by Park [41]. are examined, and it is found that Gamma distribution is a good
fit for both moderate intensity and high intensity rainfall days.
3.2. Computational aspects The three state Gamma (for moderate intensity) + Gamma (for
high intensity) distribution has a good ability to simulate
Eq. (2) is solved using the method of MCMC by setting up a extreme precipitation events reproducing the mean and stan-
Markov chain with posterior distribution of parameters, pðh=zÞ as dard deviations.
its stationary distribution. From the posterior distribution, samples 3. The three state Gamma distribution is fitted for the GCM data
of 20,000 parameter sets ðl; r; kÞ are generated using the method for control period and for future time periods.
of MCMC and the initial 2000 samples of the chain are discarded, 4. The perturbation factors of the transition probabilities and
as a Burn-in phase. The thinning interval (i.e., how frequently to Gamma distribution parameters are calculated as the ratio
sample from the chain) is varied from 30 to 50. between future and control period GCM statistical parame-
The sets of parameters (l, r, k) obtained from the posterior dis- ters/transition probabilities.
tribution are transformed to obtain return levels using Eq. (1) lead- 5. The statistical properties (parameters of gamma distribution
ing to a sample from the corresponding distribution of 1/p year and transitional probabilities) of weather generator for future
return level. are obtained by multiplying the perturbation factors obtained
in Step 4 with the statistical properties of weather generator
4. IDF relationships for future for historical time period.
6. The projections for future daily rainfall for the three time slices,
Use of robust and accurate techniques for projections of IDF 2021-50, 2051-80 and 2071-2100, are obtained.
relationships under climate change is necessary for development
of urban infrastructure to adapt to the likely changes in flood Fig. 2 shows the plot of annual rainfall from various GCMs. It is
frequencies. Precipitation flux (converted to rainfall intensity) out- observed that the models Access1_0, CSIRO_MK3-6-0, HadGEM2–
puts of 26 General Circulation Models (GCMs) for four RCP scenar- CC and HadGEM2–EC performed very poorly, while CanESM2,
ios are considered in this study, and downscaling to station scale is CNRM-CM5, IPSL-CM5A-LR, MIROC5 and MPI-ESM-LR were able
performed with a weather generator. The IDF relationships are to match the observed values reasonably well. The underperform-
then constructed with the downscaled daily precipitation by ing GCMs (A2, HADE, HADC) and over performing GCMs (BC1, IN4)
disaggregating it to smaller time scales. The following subsections are not discarded for future analyses. Although some models may
describe the methodologies and discuss the results. perform well with respect to the historical data, the other models
which may be more accurate in simulating future projections can-
4.1. Downscaling using improved weather generator not be ruled out. The method of addressing uncertainty through
REA (explained in the subsequent section) therefore, in fact
Weather generators are the statistical models of sequences of accounts for both these features and assigns weights to models
local weather variables; they are essentially random number based both on their performance in the historical period as well
generators whose output has the same statistical properties as that as on their ability to provide ‘good’ future projections, the latter
of weather at the location. Weather generator models have been being measured with respect to the departure from the ensemble
used in numerous studies on climate change impact assessment, mean. The results from all the GCMs can be used without discard-
where the standard practice is to consider the occurrence and ing any of the underperformed/over-performed GCMs, since the
the amount of precipitation separately. Weather generators can multimodel projections vary across time and will produce increas-
be classified as parametric and non-parametric or distribution free. ingly precise estimates as the number of GCMs increase [51].
In parametric weather generators, the precipitation occurrence is Climate models are evaluated on how well they simulate the past
modeled with a Markov chain and the amount of precipitation fol- and present mean climate, its variability etc. Even though, any
lows a given parametric distribution (e.g. exponential, gamma model’s correctness come from its ability to replicate the mean cli-
etc.). A number of studies have been carried out in this area and matic conditions, climate variability and alterations from the
an extensive review of the models can be found in Wilks and observations, there is no guarantee that the model simulating pre-
Wilby [53]. Non-parametric weather generating techniques based sent climate will be correct in future. Tebaldi and Knutti [51] con-
on Kernel based multivariate probability density estimators and clude that most of models perform reasonably well with
K-NN bootstrap methods are also developed (e.g. [43,24]). The observations from twentieth century simulating realistic warming,
number of studies associated with climate change impact on the but their predictions for twenty first century diverge substantially.
IDF relationship and using it in urban drainage impact are, how- The agreement between the model and data is a necessary but not
ever, very few. In the present study, the traditional WGEN rainfall a sufficient criterion [23,54].
132 R. Chandra et al. / Advances in Water Resources 79 (2015) 127–139

3000

2500

2000
Rainfall (mm)

1500

1000

500

0
MEAN

MIEC
HADC

IPCM
CM60

GF2G

IPCL
BC1M

GF2M

GISS

HADE
BNU

CAN

MI5

NEM
MPI

MRI
CMS
BC1

CN5

FG2

GF3

MIE
IN4
OBS

FS2
A2

C4

Fig. 2. Annual rainfall during the control period (1969–1998) from 26 GCMs. Mean rainfall across all GCMs (MEAN) and the observed gauge data (OBS) are also shown.

4.2. Reliability Ensemble Average (REA) method 4. The average of the inverse of RMSE derived from Steps 1 and 3
is calculated and the new weights are proportionately used, so
The projected annual maximum rainfall time series for 85 that the sum of new weights is equal to 1 across all GCMs.
GCM simulations and for the three timelines are obtained from 5. Steps 2 to 4 are repeated till weights are converged (model con-
daily rainfall. The intermodal or GCM uncertainty resulting from vergence criterion).
the ensembles of projections generated with multiple GCMs is
addressed using Reliability Ensemble Average (REA) method. Table 2 gives the weights assigned to GCMs for four scenarios
This method, proposed by Giorgi and Mearns [19], considers and for the three time slices (2021-50, 2051-80 and 2071-2100).
two reliability criteria, one is the model performance criterion, The weights vary for three time slices substantially for a few
in which the performance of model in reproducing present-day GCMs, for example, weight of MRI-CGCM3 for 8.5 scenario is
climate is assessed, and the other is the model convergence 0.125 for the time period 2021-50 and 0.044 for 2071-2100. Even
criterion in which convergence of the simulated changes across though for most of the GCMs there is no significant change in
models is considered. Xu et al. [54] augmented the REA method weights for the three time slices, the weights are not kept constant
by considering multiple variables and statistics in the calculation across time, since model’s performance alters with time.
of model performance criterion, removing the model conver-
gence criterion. Even though this method is simple and has a
flexible framework, the original REA method, considering both 4.3. Disaggregation of daily rainfall to hourly and sub-hourly scales
model performance criterion and model convergence criterion,
is used in the present study for better accuracy. Initial weights The projections for future daily rainfall time series obtained
are assigned to GCMs on the basis of model performance by tak- needs to be disaggregated to a finer scale (15 min, 30 min etc.) in
ing the CDF deviations between observed rainfall and GCM simu- order to prepare the projected IDF relationships. Numerous deter-
lated downscaled rainfall time series for the control period ministic [39] and stochastic disaggregation [28] techniques are uti-
(1969–1998). The final weights are obtained by taking CDF lized, the cascade models based on scale invariance theory are
deviations between individual GCMs for future and weighted most widely used for rainfall disaggregation. Scaling implies that
mean CDF (derived from weighted projections of multiple the statistical properties of the process observed at various scales
GCMs) accounting for the model convergence criterion. The are governed by the same relationship [36,37]. The scaling concept
CDF deviation is calculated with Root Mean Square Error has been successfully applied to relate statistics of annual maxi-
(RMSE) at 10 points, taken at equal intervals [18]. The algorithm mum precipitation for different durations [6,31,35].
is elaborated in the following steps The relation between statistical properties of extreme rainfall
processes for various time scales is obtained using temporal down-
1. Initially, the weights are assigned following the model perfor- scaling method based on scale invariance theory in the present
mance criterion. The CDF deviations between the observed data work. The scale invariance theory is elaborated in Appendix B.
and GCM projected rainfall for control period are calculated in Scale invariance method is extended to the posterior distribution
terms of RMSE, and the inverse values of RMSE are used as of parameters from Bayesian analysis in Section 3. Log–log plots
weights proportionately, so that the sum of weights is equal of the first three rainfall Non Central Moments (NCMs) against
to one across all GCMs. These weights are considered as initial duration are prepared for a given set of parameters, and statistics
weights to GCMs. such as mean and quantiles (q = 0.025, 0.25, 0.5, 0.75 and 0.975)
2. The weighted mean CDF (FWM) for future is calculated by of the NCMs are plotted (Fig. 3). The NCMs for different quantiles
P are equidistant, conveying that the NCMs follow the same scaling.
F WM ¼ i ki  F GCMi , where FGCMi is the CDF derived from down-
scaled precipitation for ith GCM and ki is the corresponding Fig. 3 also shows the log–log plot of mean of the first three rainfall
weight (initial weight derived from Step-1). non-central moments versus rainfall duration. The log-linearity
3. The CDFs deviation for all GCMs is individually calculated in displayed in the plot specifies the scaling of the rainfall statistical
terms of RMSE from the weighted mean CDF. moments with duration for two time intervals from 15 to
30 min, and 30 min to 24 h. Therefore, two different scales are
Table 2
Weights assigned to GCMs for four RCP scenarios and three time lines using REA technique.

GCM-Model Name RCP2.6 RCP2.6 RCP2.6 GCM-Model RCP4.5 RCP4.5 RCP4.5 GCM-Model RCP6.0 RCP6.0 RCP6.0 GCM-Model RCP8.5 RCP8.5 RCP8.5
2021-2050 2051-2080 2071-2100 Name 2021-2050 2051-2080 2071-2100 Name 2021-2050 2051-2080 2071-2100 Name 2021-2050 2051-2080 2071-2100
BCC-CSM1-1 0.037 0.049 0.148 ACCESS1_0 0.024 0.032 0.033 BCC-CSM1- 0.075 0.073 0.103 ACCESS1_0 0.025 0.025 0.022
1
BCC-CSM1-1-M 0.056 0.023 0.031 BCC-CSM1- 0.042 0.033 0.077 BCC-CSM1- 0.033 0.041 0.025 BCC-CSM1- 0.074 0.090 0.088
1 1-M 1
BNU-ESM 0.042 0.038 0.030 BCC-CSM1- 0.035 0.032 0.021 CCSM4 0.062 0.155 0.114 BCC-CSM1- 0.021 0.018 0.017
1-M 1-M
CanESM2 0.046 0.056 0.098 BNU-ESM 0.019 0.038 0.046 CSIRO-MK3- 0.081 0.099 0.051 BNU-ESM 0.051 0.042 0.052
6-0
CCSM4 0.033 0.092 0.113 CanESM2 0.042 0.142 0.047 FGOALS-S2 0.029 0.047 0.091 CanESM2 0.053 0.089 0.089
CSIRO-MK3-6-0 0.059 0.035 0.053 CCSM4 0.039 0.113 0.068 GFDL-CM3 0.053 0.057 0.058 CCSM4 0.068 0.053 0.057

R. Chandra et al. / Advances in Water Resources 79 (2015) 127–139

FGOALS-G2 0.032 0.024 0.029 CMCC-CMS 0.058 0.053 0.035 GFDL- 0.056 0.060 0.051 CMCC-CMS 0.029 0.032 0.041
ESM2G
FGOALS-S2 0.021 0.021 0.019 CNRM-CM5 0.020 0.033 0.029 GFDL- 0.044 0.033 0.040 CNRM-CM5 0.051 0.133 0.068
ESM2M
GFDL-CM3 0.040 0.035 0.038 CSIRO-MK3- 0.035 0.026 0.038 HadGEM2- 0.152 0.048 0.154 CSIRO-MK3- 0.049 0.027 0.019
6-0 ES 6-0
GFDL-ESM2G 0.032 0.020 0.029 FGOALS-G2 0.056 0.037 0.042 IPSL-CM5A- 0.060 0.076 0.031 FGOALS-G2 0.036 0.036 0.034
LR
GFDL-ESM2M 0.071 0.028 0.022 FGOALS-S2 0.049 0.129 0.050 IPSL-CM5A- 0.061 0.067 0.050 FGOALS-S2 0.039 0.056 0.059
MR
HadGEM2-ES 0.015 0.047 0.100 GFDL- 0.021 0.030 0.027 MIROC5 0.074 0.057 0.054 GFDL- 0.067 0.033 0.032
ESM2G ESM2G
IPSL-CM5A-LR 0.086 0.081 0.056 GFDL- 0.079 0.021 0.028 MIROC- 0.029 0.036 0.026 GFDL- 0.024 0.063 0.076
ESM2M ESM-CHEM ESM2M
IPSL-CM5A-MR 0.049 0.038 0.034 GISS-E2-R 0.017 0.024 0.030 MIROC-ESM 0.028 0.046 0.037 HadGEM2- 0.014 0.025 0.014
CC
MIROC5 0.044 0.031 0.029 HadGEM2- 0.010 0.032 0.105 MRI-CGCM3 0.047 0.068 0.064 HadGEM2- 0.038 0.027 0.033
CC ES
MIROC-ESM- 0.025 0.070 0.021 HadGEM2- 0.071 0.009 0.023 NorESM1-M 0.118 0.037 0.052 INMCM4 0.019 0.016 0.017
CHEM ES
MIROC-ESM 0.030 0.080 0.028 INMCM4 0.027 0.013 0.019 IPSL-CM5A- 0.020 0.027 0.022
LR
MPI-ESM-LR 0.128 0.021 0.042 IPSL-CM5A- 0.099 0.012 0.019 IPSL-CM5A- 0.037 0.028 0.035
LR MR
MRI-CGCM3 0.036 0.028 0.028 IPSL-CM5A- 0.048 0.035 0.066 MIROC5 0.042 0.035 0.034
MR
NorESM1-M 0.117 0.183 0.049 MIROC5 0.034 0.024 0.031 MIROC- 0.025 0.017 0.017
ESM-CHEM
MIROC- 0.022 0.028 0.018 MIROC-ESM 0.041 0.021 0.021
ESM-CHEM
MIROC-ESM 0.033 0.018 0.022 MPI-ESM- 0.023 0.036 0.073
LR
MPI-ESM- 0.038 0.026 0.025 MRI-CGCM3 0.125 0.045 0.044
LR
MRI-CGCM3 0.033 0.039 0.071 NorESM1-M 0.027 0.024 0.035
NorESM1-M 0.047 0.023 0.030

133
134 R. Chandra et al. / Advances in Water Resources 79 (2015) 127–139

considered for the two time intervals (15 to 30 min and 30 min to Table 3
24 h), and the scaling exponents (b‘s) are estimated as slopes of Maximum rainfall non-central moments (NCMs)from classical and Bayesian analysis.

Fig. 3d. The scaling exponents and order of moments are plotted NCM Duration Classical Bayesian
for each section of the scaling regime and it is observed that scaling Mean Median SD
exponent is a linear function of order of moment, which implies
w1a 24 h 3.34 3.37 3.36 0.22
simple scaling. Table 3 gives the statistics of NCMs, and Table 4 12 h 6.18 6.25 6.22 0.42
the scaling exponents from both the classical and Bayesian meth- 6h 11.27 11.32 11.28 0.77
ods. The QQ plot between the observed rainfall quantiles and esti- 3h 20.59 20.76 20.67 1.45
mated rainfall quantiles using GEV scaling for all the durations is 1h 44.19 44.48 44.34 2.70
30 min 71.54 71.98 71.68 4.10
plotted and a good model performance with R-square value of
15 min 95.11 95.95 95.65 4.90
0.99 from Bayesian and 0.97 using classical method is obtained.
w2b 24 h 12.70 12.95 12.72 1.92
12 h 43.58 44.91 44.06 6.88
4.4. Probabilistic IDF relationships for future 6h 146.10 148.04 145.03 22.97
3h 487.01 500.28 492.01 77.31
Future daily rainfall time series is obtained from stochastic 1h 2115 2189 2156 279
30 min 5513 5690 5610 687
weather generator for the three time slices 2021-51, 2051-81, 15 min 9678 9992 9881 1096
and 2071-2100, for the 85 GCM simulations. REA method is used
w3c 24 h 55.70 57.81 54.74 17.36
for addressing the uncertainty in GCMs, and a weighted daily rain- 12 h 355.43 377.91 354.69 126.94
fall time series is obtained for each of the scenarios and time slices 6h 2223 2273 2136 709
for 24 h duration. Scale invariance method is used for disaggregat- 3h 13411 14140 13326 4128
ing the daily data to hourly and sub-hourly times, and the IDF 1h 108345 118296 113378 26089
30 min 453929 491434 474157 103221
relationships are plotted. The IDF relationships for future using
15 min 1053599 1132192 1100699 213751
classical approach are presented in Fig. 4 for RCP 8.5 scenario for
a
the three time slices (2021-50, 2051-80, 2071-2100). It is observed Unit: mm/hr.
b
Unit: mm2/hr2.
that the short duration rainfall return levels increase considerably c
Unit: mm3/hr3.
in future, emphasizing the importance of obtaining the uncertain-
ties associated with short duration return levels, which can be per-
formed by the Bayesian analysis. return levels are presented in Table 5. The uncertainty in the return
The uncertainty in the return levels for all the durations and dif- levels is considerably high for shorter durations. The range of the
ferent return periods is computed for the four RCP scenarios and REA averaged return level varies from 35.97 mm/hr for RCP 2.6 sce-
three time slices using the Bayesian method. The statistics, viz. nario to 49.7 mm/hr for RCP 8.5 scenario for near future (2021-50),
mean, standard deviation (SD), and the range, of REA averaged and for far future (2071-2100), the range varies from 40.3 mm/hr

5
q = 0.025 10
= 0.25 q = 0.025
4.5
= 0.5 9 = 0.25
= 0.75 = 0.5
4
Log of Non Central Moments - 1

= 0.975 = 0.75
8
Log of Non Central Moments - 2

= 0.975
3.5
7

3
6

2.5
5

2 4

1.5 (a) 3 (b)

1 2
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5

Log of Rainfall Duration Log of Rainfall Duration

16
q = 0.025 14
= 0.25 NCM1
14
= 0.5 NCM2
12
= 0.75 NCM3
Log of Non Central Moments - 3

12 = 0.975
10
Log of Non Central Moments

10
8

8
6

6
4

4
(c) 2 (d)
2 0
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5

Log of Rainfall Duration Log of Rainfall Duration

Fig. 3. Log–log plot of quantiles (q) of maximum rainfall non-central moments versus rainfall duration from Bayesian approach (a) NCM-1, (b) NCM-2 and (c) NCM-3 (d)
mean of the NCMs.
 R. Chandra et al. / Advances in Water Resources 79 (2015) 127–139 135

Table 4 for RCP 2.6 scenario to 75.7 mm/hr for RCP 8.5 scenario. The uncer-
Scaling exponent b from classical and Bayesian analysis. tainty in REA averaged return level for RCP 8.5 scenario is high, and
Time scale b from classical Bayesian approach increases by 53% from near future to far future, while the range
approach
b from 50th quantile b from mean increases by 12% for RCP 2.6 scenario. Fig. 5a shows the plot of
REA averaged return levels for various durations for near future,
24 h–30 min 0.7902 0.7897 0.7899
30 min–15 min 0.4108 0.4161 0.4164
RCP 8.5 scenario, 10 year return period. The box plot gives the
REA averaged return levels from Bayesian analysis showing
the uncertainty in parameters. The figure also gives the IDF

Fig. 4. IDF relationships for future time slices using classical approach for RCP 8.5 scenario.

Table 5
Mean, standard deviation (SD) and range of the REA averaged return levels for future three time slices and various durations.

Time line Statisticsa 24 h 12 h 6h 3h 1h 30 min 15 min

RCP 2.6(2021-2050) Mean 4.93 8.53 14.75 25.49 60.70 104.94 140.02
SD 0.15 0.26 0.45 0.78 1.85 3.21 4.28
Range 1.27 2.19 3.79 6.55 15.59 26.96 35.97
RCP 2.6(2051-2080) Mean 4.98 8.62 14.90 25.75 61.32 106.01 141.45
SD 0.16 0.27 0.47 0.81 1.94 3.35 4.46
Range 1.39 2.40 4.14 7.16 17.06 29.49 39.34
RCP 2.6(2071-2100) Mean 5.27 9.12 15.76 27.24 64.87 112.15 149.64
SD 0.18 0.32 0.55 0.95 2.25 3.89 5.19
Range 1.42 2.46 4.24 7.34 17.47 30.20 40.30
RCP 4.5(2021-2050) Mean 4.73 8.17 14.12 24.42 58.14 100.51 134.12
SD 0.12 0.21 0.36 0.62 1.48 2.56 3.41
Range 1.06 1.84 3.18 5.49 13.07 22.60 30.15
RCP 4.5(2051-2080) Mean 5.09 8.80 15.22 26.31 62.64 108.30 144.50
SD 0.16 0.28 0.49 0.85 2.03 3.51 4.68
Range 1.46 2.53 4.37 7.56 17.99 31.11 41.51
RCP 4.5(2071-2100) Mean 5.11 8.84 15.28 26.41 62.90 108.74 145.09
SD 0.14 0.24 0.41 0.71 1.68 2.91 3.88
Range 1.07 1.84 3.18 5.50 13.11 22.66 30.23
RCP 6.0(2021-2050) Mean 5.08 8.77 15.17 26.22 62.44 107.94 144.03
SD 0.15 0.26 0.45 0.77 1.83 3.17 4.23
Range 1.27 2.20 3.81 6.58 15.67 27.10 36.15
RCP 6.0(2051-2080) Mean 5.89 10.17 17.59 30.41 72.41 125.18 167.03
SD 0.24 0.42 0.72 1.25 2.97 5.13 6.84
Range 1.94 3.35 5.80 10.03 23.87 41.27 55.07
RCP 6.0(2071-2100) Mean 5.98 10.34 17.87 30.90 73.58 127.20 169.73
SD 0.20 0.34 0.59 1.01 2.41 4.17 5.56
Range 1.87 3.23 5.58 9.65 22.99 39.74 53.03
RCP 8.5(2021-2050) Mean 5.43 9.38 16.22 28.03 66.76 115.40 153.99
SD 0.20 0.35 0.61 1.05 2.49 4.31 5.75
Range 1.75 3.03 5.23 9.05 21.55 37.25 49.70
RCP 8.5(2051-2080) Mean 6.34 10.96 18.95 32.77 78.03 134.89 179.99
SD 0.26 0.44 0.76 1.32 3.14 5.44 7.25
Range 2.20 3.80 6.56 11.35 27.02 46.70 62.32
RCP 8.5(2071–2100) Mean 6.52 11.28 19.50 33.71 80.26 138.76 185.15
SD 0.21 0.36 0.63 1.09 2.59 4.48 5.98
Range 2.67 4.61 7.97 13.78 32.82 56.74 75.71
a
Units: mm/hr.
136 R. Chandra et al. / Advances in Water Resources 79 (2015) 127–139

relationship from classical analysis (blue line) and the IDF relation- scenario and for the three time slices (2021-51, 2051-81 and
ship for historical time period (green line). Fig. 5b shows the cumu- 2071-2100) for the 10 year return period. These results show that
lative distribution function of REA averaged return levels for the return level estimates are very sensitive for short durations and
various durations in the near future for RCP 8.5 scenario and a for higher RCP scenarios.
10 year return period. Fig. 6 shows the cumulative distribution The uncertainty in the return level is studied individually for all
function of REA averaged return level for RCP 2.6 and RCP 8.5 the 85 GCM runs and for the three time-slices (2021-51, 2051-81
and 2071-2100) using Bayesian and classical approaches. The
uncertainty in return levels for 15 min duration, near future
(2021-50) results are presented in Fig. 7 for 8.5 scenario; the REA
averaged return levels and the return levels for historical time per-
180 iod are also shown in the figure. The box plot gives the uncertainty
160 Bayesian in return levels obtained from Bayesian analysis and the return
140 level obtained from classical approach is marked as ⁄ on the box
Return level (mm/hr)

Classical
120 Historical plot. It is seen from Fig. 7 that for RCP 8.5 scenario, most of the
100 GCMs underestimated/overestimated the return levels calculated
80
using classical approach emphasizing the fact that the uncertainty
associated with the return levels is high for RCP 8.5. The mean of
60
the REA averaged return levels obtained from Bayesian analyses
40 (a) are approximately equal to that of the return level calculated from
20
classical approach, indicating that the model uncertainty can be
0 addressed using classical approach, whereas for accounting the
0.25 0.5 1 3 6 12 24
Duration (hr)

(b)

Fig. 5. (a) Variation in return levels for various durations in comparison with
Fig. 6. Cumulative distribution function (cdf) of return level for 15 min duration, for
classical approach and historical time line (b) cumulative distribution function (cdf)
RCP 2.6 and RCP 8.5 scenarios, and for the future three time slices for a 10 year
of return level for RCP 8.5 scenario and a 10 year return period for the near future
return period.
(2021–2050).

500

450

400

350
Return level (mm/hr)

300

250

200

150

100

50
Historical

NEM
BNU
BC1M

GF2M
CAN

HADC
CMS

HADE
CN5

CM60

MPI

MRI
IPCM
GF2G
REA

MIEC

MIE
A1

BC1

C4

FG2

FS2

IN4

MI5
IPCL

Fig. 7. Uncertainty in return levels for near future (2021–50) for different GCMs, REA averaged, and the uncertainty in return levels for historical time period (1969–2003) for
15 min. duration, for RCP 8.5 Scenario. Boxplot gives the uncertainty in return levels obtained from Bayesian analysis, and the return level obtained from classical approach is
shown as ⁄ on the box plot.
 R. Chandra et al. / Advances in Water Resources 79 (2015) 127–139 137

Table 6 non-informative prior distributions in case of non availability

Percentage change (between mean of return levels from Bayesian approach and of information.
classical approach) in return levels for different GCMs and different RCP scenarios.
The intermodal or GCM uncertainty resulting from the
Scenario RCP RCP RCP RCP RCP ensembles of projections generated with multiple GCMs is
GCM 2.6(2021- 4.5(2021- 6.0(2021- 8.5(2021- 8.5(2071- addressed using Reliability Ensemble Average (REA) method.
50) 50) 50) 50) 2100)
Twenty-six CMIP5 GCMs along with four RCP scenarios are con-
REA 1.4 1.2 1.4 0.8 0.8 sidered for studying the effects of climate change on extremes. It
A1 – 7.9 – 0.4 1.8
BC1 1.4 4.2 35.0 34.9 41.8
is observed that the uncertainty in short duration rainfall return
BC1M 5.7 6.9 70.7 36.0 27.2 levels is high and need to be considered in further analysis
BNU 5.2 21.2 – 4.6 57.2 which is otherwise not possible with the classical methods. It
CAN 8.6 13.1 – 20.8 3.4 is seen that uncertainty introduced due to choice of GCM is
C4 29.4 30.3 17.1 60.0 48.1
smaller than that due to parameter uncertainty. The uncertainty
CMS – 43.6 – 30.2 0.8
CN5 – 12.3 – 26.3 32.6 in return levels is large for RCP 8.5 scenario compared to the
CM60 13.3 18.1 13.6 61.1 154.5 other scenarios as most of the GCMs underestimated/over-
FG2 11.8 3.5 40.9 12.4 70.5 estimated the return levels. Quantification of uncertainty is
FS2 12.4 6.6 15.8 25.8 54.8 required to obtain improved understanding of the IDF relation-
GF3 6.2 – – – –
ships, which is further used in the hydrologic designs and
GF2G 3.1 20.8 104.2 3.6 33.7
GF2M 37.2 1.5 31.3 30.6 48.8 Bayesian methods are proven to be advantageous technically
GISS – 20.5 – – – and scientifically in assessing the uncertainties [9]. Even though
HADC – 44.8 – 39.9 56.7 a single station is considered in the present work, this can be
HADE 17.8 6.7 23.6 26.8 20.5
extended to multiple areas and combined changes in the pro-
IN4 – 32.5 – 53.1 162.1
IPCL 2.3 4.8 0.5 49.9 47.1 jected IDF relationships can be studied.
IPCM 13.5 11.7 79.4 2.9 34.6 Even though the Bayesian analysis is a more rigorous approach
MI5 10.6 0.7 12.2 29.4 147.1 in addressing the uncertainty, this approach is not usually a part of
MIEC 39.9 17.8 41.4 69.6 79.8 design and risk assessment strategies. The probabilistic IDF
MIE 13.3 13.2 79.9 6.0 57.1
relationships need to be made part of operational risk assessment
MPI 19.9 3.1 – 7.0 12.6
MRI 38.2 23.2 17.4 10.0 43.4 strategies and should be considered in design of hydrological sys-
NEM 10.3 0.4 22.1 15.1 26.9 tems and infrastructure.

Acknowledgments

parameter uncertainty, Bayesian methods are necessary. For RCP This work is carried out as a part of project ‘‘Integrated Urban
8.5 scenario, the uncertainty in return levels (for 15 min duration) Flood Management in India: Technology Driven Solutions’’ sup-
for third time slice (2071-2100), except A1, CAN and MPI, all other ported by Information Research Technology Academy (ITRA),
GCM models overestimated or underestimated the return levels Ministry of Communications and Information Technology. We sin-
from classical approach. The results for other scenarios and time cerely thank the editor and the anonymous reviewer for reviewing
slices are not presented here for conciseness. Table 6 shows per- the manuscript and offering critical comments to improve the
centage change in return levels, which is calculated as the ratio manuscript.
between deviance of mean return level from Bayesian approach
and return level from classical approach. Positive value indicates
underestimation in return levels and vice versa. It is seen that Appendix A. Method of Markov Chain Monte Carlo
the percentage deviation is huge for RCP 8.5 scenario and the per-
centage deviation is less for REA averaged return levels. MCMC is a general method to draw samples from an approxi-
mate distribution, which otherwise is not possible by standard
methods like inversion, transformation, rejection methods etc.
Since the posterior density pðhÞ (in this section, the dependence
5. Conclusions
of pð:Þ on y is suppressed for the sake of brevity), with support
Rp , is complex high dimensional joint density, drawing sample
Flooding is a very common problem in urban areas; short
using the standard methods is not possible and therefore MCMC
duration high intensity rains often cause chaos if the infras-
is used. For a target density pðhÞ, an irreducible, aperiodic, positive
tructure (such as the storm water drains, detention ponds)
recurrent Markov chain is set up with state space Rp and transition
does not cater the requirements especially in the low lying
kernel pð/; dhÞ such that p ðhÞ  pðhÞdh is the invariant dis-
areas. The data available for short duration rainfall is often less
tribution of this Markov chain satisfying
leading to large uncertainties in GEV parameters. While obtain-
Z
ing the IDF relationships, this uncertainty becomes very impor-
tant for short duration rainfalls posing an important question p ðhÞ ¼ Pð/; dhÞp ðd/Þ 8h 2 Rp
Rp
about the authenticity of standard statistical estimates.
Bayesian approach is used in this study to quantify the uncer- Then, since Limn!1 PðnÞ ð/; dhÞ ¼ p ðdhÞ, an observation from pðhÞ can
tainties in parameters of distribution fitted to the data and be simulated by simulating the chain for long time using the transi-
uncertainties due to the use of multiple GCM models. MCMC tion kernel with an initial value of /. Two algorithms namely,
method using Metropolis Hastings algorithm is used to obtain Metropolis Hastings algorithm and Gibbs sampling, allows in find-
the posterior distribution of parameters (which is obtained by ing the appropriate transition kernel pð/; dhÞ for a given p ðhÞ as
multiplying the likelihood function and the prior distribution) its invariant distribution. In the present study Metropolis Hastings
and the parameters are transformed to obtain return levels algorithm is used. The Metropolis Hastings transition kernel is given
leading to a sample. Even though Bayesian methods have an by
additional burden of fixing the prior distributions, it can be
overcome by eliciting priors based on previous studies or using Pð/; dhÞ ¼ gð/; hÞrð/; hÞdh þ að/Þg/ ðdhÞ
138 R. Chandra et al. / Advances in Water Resources 79 (2015) 127–139

Z
where gð/; hÞ P 0; gð/; /Þ ¼ 0; gð/; hÞdh ¼ 1; lðktÞ ¼ kb lðtÞ;
R p
rðktÞ ¼ kb rðtÞ;
pðhÞgðh; /Þ kðktÞ ¼ kðtÞ and
rð/; hÞ ¼ Minf ; 1g;
pð/Þgð/; hÞ Z T ðktÞ ¼ kb Z T ðtÞ
Z
að/Þ ¼ 1  gð/; hÞrð/; hÞdh and The exponent b is arrived based on the scaling properties of the
Rp
( NCMs for various durations. Step by step procedure for obtaining
1 if / 2 ðh; h þ d#Þ b is given below.
g/ ðdhÞ ¼
0 otherwise 1. Calculate the first three NCMs w1, w2 and w3 using Eq. (A2-1)
2. Plot NCMs verses duration on log–log scale and obtain the scal-
The algorithm starts with an initial value h ¼ h0 and in the nth itera- ing exponent b as
tion, if the present value of h is hn, generate a candidate value / from
the proposal density gð/; :Þ. If rðhn ; /Þ ¼ 1, then set hn+1 = / and pro- w1 ðktÞ
kb ¼ :
ceed to the next iteration. If rðhn ; /Þ < 1, generate a uniform random w1 ðtÞ
number u and set hn+1 = /, if u < rðhn ; /Þ else set hn+1 = hn and in 3. Plot NCMs verses the scaling exponent and check if the process
either case, proceed to the next iteration. Different Metropolis follows simple scaling or multiple scaling (Note that for a sim-
Hastings algorithms are built based on the proposal density. If the ple scaling process, bðdÞ ¼ bd ).
proposal density is independent of the current value in the 4. After checking the scaling behavior, the scaling exponent b is
sequence, then the algorithm is called as independence chain used for obtaining the return levels for shorter durations.
Metropolis Hastings algorithm. Many other proposal densities are
used in the Metropolis Hastings algorithm [1]. In this work,
LearnBayes package of R (http://cran.r-project.org/web/packages/ References
LearnBayes/index.html) is used, in which random walk chain is
employed which has a simple form of rð/; hÞ ¼ pðhÞ=pð/Þ with pro- [1] Albert J. Bayesian computation with R. NY: Springer; 2009, ISBN 978-0-387-
92298-0.
posal density as / ¼ h þ scale  z, where z is multivariate normal
[2] Allen MR, Stott PA, Mitchell JFB, Schnur R, Delworth TL. Quantifying the
with mean 0 and variance–covariance matrix V and scale is a posi- uncertainty in forecasts of anthropogenic climate change. Nature
tive scale parameter. This algorithm does not require knowledge of 2000;407:617–20. http://dx.doi.org/10.1038/35036559.
[3] Ashour SK, El-Adl YM. Bayesian estimation of the parameters of the extreme
normalizing constant of the posterior distribution since the target
value distribution. Egypt Stat J 1980;24:140–52.
density appears only in the computation of rð/; hÞ through the ratio [4] Berger JO, Sun D. Bayesian analysis for the Poly-Weibull distribution. J.A.S.A.
pðhÞ=pð/Þ and hence making the Metropolis Hastings algorithm a 1993;88:1412–8. http://dx.doi.org/10.1080/01621459.1993.10476426.
very attractive option for generating observations from the [5] Buonomo E, Jones R, Huntingford C, Hannaford J. On the robustness of changes
in extreme precipitation over Europe from two high resolution climate change
posterior density. simulations. Q J R Meteorol Soc 2007;133:65–81. http://dx.doi.org/10.1002/
qj.13.
[6] Burlando P, Rosso R. Scaling and multiscaling models of depth-duration-
Appendix B. Scale invariance theory frequency curves for storm precipitation. J Hydrol 1996;187:45–64. http://
dx.doi.org/10.1016/S0022-1694(96)03086-7.
The temporal downscaling method proposed here is based on [7] Coles S. An introduction to statistical modeling of extreme
values. London: Springer; 2001, ISBN 978-1-4471-3675-0.
the concept of scale invariance theory. Scaling implies that the sta- [8] Coles S, Powell EA. Bayesian methods in extreme value modelling: a review
tistical properties of the process observed at various scales are and new developments. Int Stat Rev 1996;64(1):119–36. http://www.jstor.
governed by the same relationship [36,37]. It can be shown that org/stable/1403426.
[9] Coles S, Pericchi LR, Sisson S. A fully probabilistic approach to extreme rainfall
the ensemble moments are log–log linear functions of the scale modeling. J Hydrol 2003;273:35–50. http://dx.doi.org/10.1016/S0022-
of resolution in a random cascade [40]. The slope of this scaling 1694(02)00353-0.
relationship is known as the Mandelbrot–Kahane–Peyriere (MKP) [10] Coles SG, Tawn JA. A Bayesian analysis of extreme rainfall data. Appl Stat
1996;45:463–78. http://www.jstor.org/stable/2986068.
function [27]. [11] Coulibaly P, Shi X. Identification of the effect of climate change on future
For a random variable Z, a function g(z) is said to be scaling if design standards of drainage infrastructure in Ontario, Report prepared by
g(z) is proportional to the scaled function g(kz) for all positive val- McMaster University with funding from the Ministry of Transportation of
Ontario. 82; 2005.
ues of scale factor k i.e., if g(z) is scaling then there exists a function [12] Dessai S, Lu X, Hulme M. Limited sensitivity analysis of regional climate
dist
h(k) such that g(kz) ¼ h(k) g(z), where h(k) = kb, b denotes scaling change probabilities for the 21st century. J Geophys Res 2005;110:D19108.
http://dx.doi.org/10.1029/2005JD005919.
component. This is known as ‘‘strict sense simple scaling’’ and it [13] Day DK, Lee T. Bayes computation for life testing and reliability estimation.
states ‘‘that equality in the probability distribution of the rainfall IEEE Trans Reliab 1992;41:621–6. http://dx.doi.org/10.1109/24.249600.
depth observed at two different time scales holds’’ [20,6]. This implies [14] Engelund S, Rackwitz R. On predictive distribution functions for the three
asymptotic extreme value distributions. Struct Safety 1992;11:255–8. http://
that the quantiles and raw moments of any order are also
dx.doi.org/10.1016/0167-4730(92)90018-I.
scale invariant i.e., E½Z dk  ¼ kd b E½Z d . (referred to as ‘‘wide sense [15] Gelfand A, Smith A. Sampling based approaches to calculating marginal
simple scaling’’), where d is the order of the moment. For the GEV densities. J Am Stat Assoc 1990;85:398–409. http://dx.doi.org/10.1080/
01621459.1990.10476213.
distribution, the dth order non central moment (NCM) wd can be [16] Gelman A, Carlin JB, Stern HS, Rubin DB. Bayesian data analysis: texts in
expressed as statistical science. London: Chapman & Hall/CRC Press; 2004. ISBN:
9781439840955 - CAT# K11900.
 r d  r d [17] Ghosh S, Mujumdar PP. Nonparametric methods for modeling GCM and
wd ¼ l þ ð1Þd  Cð1  dkÞ scenario uncertainty in drought assessment. Water Resour Res
k k 2007;43:W07405. http://dx.doi.org/10.1029/2006WR005351.
X
d 1 
r i  r di [18] Ghosh S, Mujumdar PP. Climate change impact assessment – uncertainty
þ d ð1Þi  l Cð1  ikÞ ðA2-1Þ modeling with imprecise probability. J Geophys Res – Atmos
i¼1
k k 2009;114:D18113. http://dx.doi.org/10.1029/2008JD011648.
[19] Giorgi F, Mearns LO. Probability of regional climate change calculated using
where Cð:Þ is the gamma function; l, r and k are the parameters of the reliability ensemble averaging (REA) method. Geophys Res Lett
the GEV distribution. 2003;30(12):1629. http://dx.doi.org/10.1029/2003GL017130.
[20] Gupta VK, Waymire E. Multiscaling properties of spatial rainfall and river flow
For simple scaling, the statistical properties of GEV distribution distributions. J Geophys Res 1990;95(D3):1999–2009. http://dx.doi.org/
for two different time scales t and kt are related as 10.1029/JD095iD03p01999.
 R. Chandra et al. / Advances in Water Resources 79 (2015) 127–139 139

[21] Huard D, Mailhot A, Duchesne S. Bayesian estimation of intensity-duration- [37] Olsson J, Berndtsson R. Temporal rainfall disaggregation based on scaling
frequency curves and of the return period associated to a given rainfall event. properties. Water Sci Technol 1998;37(11):73–9. http://dx.doi.org/10.1016/
Stochastic Environ Res Risk Assess 2010;2(D3):337–47. http://dx.doi.org/ S0273-1223(98)00318-7.
10.1007/s00477-009-0323-1. [38] Onof, Arnbjerg-Nielsen K. Quantification of anticipated future changes in high
[22] Jones RN. Analysing the risk of climate change using an irrigation demand resolution design rainfall for urban areas. Atmos Res 2009;92:350–63. http://
model. Clim Res 2000;14:89–100. http://dx.doi.org/10.3354/cr014089. dx.doi.org/10.1016/j.atmosres.2009.01.014.
[23] Knutti R, Furrer R, Tebaldi C, Cermak J, Meehl GA. Challenges in combining [39] Ormsbee LE. Rainfall disaggregation model for continuous hydrologic
projections from multiple climate models. J Clim 2010;23:2739–58. http:// modeling. J Hydraulic Eng 1989;115(4):507–25. http://dx.doi.org/10.1061/
dx.doi.org/10.1175/2009JCLI3361.1. (ASCE)0733-9429(1989)115:4(507).
[24] Lall U, Sharma A. A nearest neighbour bootstrap for time series resampling. [40] Over TM, Gupta VK. A space-time theory of mesoscale rainfall using random
Water Resour Res 1996;32(3):679–93. http://dx.doi.org/10.1029/95WR cascades. J Geophys Res 1996;101:319–31. http://dx.doi.org/10.1029/
02966. 96JD02033.
[25] Mailhot A, Duchesne S, Caya D, Talbot G. Assessment of future change in [41] Park JS. A simulation-based hyperparameter selection for quantile estimation
intensity–duration–frequency (IDF) curves for Southern Quebec using the of the generalized extreme value distribution. Math Comput Simul
Canadian Regional Climate Model (CRCM). J Hydrol 2007;347(1–2):197–210. 2005;70:227–34. http://dx.doi.org/10.1016/j.matcom.2005.09.003.
http://dx.doi.org/10.1016/j.jhydrol.2007.09.01. [42] Prodanovic P, Simonovic SP. Development of rainfall intensity duration
[26] Mailhot A, Duchesne S, Rivard G, Nantel E, Caya D, Villeneuve VP. Climate frequency curves for the City of London under the changing climate, Water
change impacts on the performance of urban drainage systems for Southern Resources Research Report no. 058, ISBN: (print) 978-0-7714-2667-4; (online)
Québec. In: Proceedings of EIC climate change technology conference. Ottawa, 978-0-7714-2668-1; 2007.
Ontario, Canada; 2006. [43] Rajagopalan B, Lall U. A k-nearest-neighbor simulator for daily precipitation
[27] Mandelbrot BB. Intermittent turbulence in self-similar cascades: divergence of and other variables. Water Resour Res 1999;35(10):3089–101. http://
high moments and dimension of the carrier. J Fluid Mech 1974;62:331–58. dx.doi.org/10.1029/1999WR900028.
http://dx.doi.org/10.1017/S0022112074000711. [44] Ramachandra TV, Mujumdar PP. Urban floods: case study of Bangalore,
[28] Margulis SA, Entekhabi D. Temporal disaggregation of satellite-derived disaster & development. J Natl Inst Disaster Manage 2009;3(2):1–98. http://
monthly precipitation estimates and the resulting propagation of error in eprints.iisc.ernet.in/id/eprint/43235.
partitioning of water at the land surface. Hydrol Earth Syst Sci [45] Rodrıguez R, Navarro Xavier, Carmen MC, Ribalaygua Jaime, Russo Beniamino,
2001;5(1):27–38. http://dx.doi.org/10.5194/hess-5-27-2001. Pouget Laurent, Redano Angel. Influence of climate change on IDF curves for
[29] Martins ES, Stedinger JR. Generalized maximum likelihood generalized the metropolitan area of Barcelona (Spain). Int J Climatol 2013;34(3):643–54.
extreme value quantile estimators for hydrologic data. Water Resour Res http://dx.doi.org/10.1002/joc.3712.
2000;28(3):737–44. http://dx.doi.org/10.1029/1999WR900330. [46] Simonovic SP, Peck A. Updated rainfall intensity duration frequency curves for the
[30] Martins ES, Stedinger JR. Historical information in a generalized maximum City of London under the changing climate, Water Resources Research Report No.
likelihood framework with partial duration and annual maximum series. 065, ISBN: (print) 978-0-7714-2819-7; (online) 978-0-7714-2820-3; 2009.
Water Resour Res 2001;37(10):2559–68. http://dx.doi.org/10.1029/ [47] Singpurwalla NZ, Song MS. Reliability analysis using Weibull lifetime data and
2000WR000009. expert opinion. IEEE Trans Reliab 1988;37:340–7. http://dx.doi.org/10.1109/
[31] Menabde M, Seed A, Pegram G. A simple scaling model for extreme rainfall. 24.3765.
Water Resour Res 1999;35(1):335–9. http://dx.doi.org/10.1029/ [48] Smith AFM. Contribution to discussion of Non-regular maximum likelihood
1998WR900012. problems by Cheng, R.C.H. and Traylor, L.J., R. Stat. Soc. 57, 30; 1995.
[32] Mirhosseini G, Srivastava Puneet, Stefanova Lydia. The impact of climate [49] Smith RL. A survey of nonregular problems. In: Proc. 47th session of the ISI.;
change on rainfall intensity–duration–frequency (IDF) curves in Alabama. Reg 1989. p. 353–72.
Environ Change 2013;13:S25–33. http://dx.doi.org/10.1007/s10113-012- [50] Smith RL, Naylor JC. A comparison of maximum likelihood and Bayesian
0375-5. estimators for the three-parameter Weibull distribution. Appl Stat
[33] Mujumdar PP, Ghosh S. Modeling GCM and scenario uncertainty using a 1987;36:358–69. http://dx.doi.org/10.2307/2347795.
possibilistic approach: application to the Mahanadi River, India. Water Resour [51] Tebaldi C, Knutti R. The use of the multi-model ensemble in probabilistic
Res 2008;44:W06407. http://dx.doi.org/10.1029/2007WR006137. climate projections. Philos Trans R Soc A 2007;2007:365. http://dx.doi.org/
[34] New M, Hulme M. Representing uncertainty in climate change scenarios: a 10.1098/rsta.2007.2076.
Monte Carlo approach. Int Assess 2000;1:203–13. http://dx.doi.org/10.1023/ [52] Tebaldi C, Mearns LO, Nychka D, Smith RL. Regional probabilities of
A:1019144202120. precipitation change: a Bayesian analysis of multimodel simulations.
[35] Nguyen VTV, Nguyen TD, Cung A. A statistical approach to downscaling of sub- Geophys Res Lett 2004;31:L24213. http://dx.doi.org/10.1029/2004GL021276.
daily extreme rainfall processes for climate-related impact studies in urban [53] Wilks DS, Wilby RL. The weather generation game: a review of stochastic
areas. Water Sci Technol: Water Supply 2007;7(2):183–92. http://dx.doi.org/ weather models. Prog Phys Geog 1999;23(3):329–57. http://dx.doi.org/
10.2166/ws.2007.053. 10.1177/030913339902300302.
[36] Olsson J. Evaluation of a scaling cascade model for temporal rainfall [54] Xu Y, Gao Xuejie, Giorgi Filippo. Upgrades to the reliability ensemble averaging
disaggregation. Hydrol. Earth Syst. Sci. 1998;2:19–30. http://dx.doi.org/ method for producing probabilistic climate-change projections. Clim Res
10.5194/hess-2-19-1998. 2010;41:61–81. http://dx.doi.org/10.3354/cr00835.