Assessment of Nonstationarity and Uncertainty in Precipitation Extremes of a River Basin Under Climate Change


In this study, an uncertainty analysis of extreme precipitation return levels was performed for the Chaliyar river basin, India, under representative concentration pathways (RCPs) 4.5 and 8.5. Weighted average projections of various climate models (for RCPs 4.5 and 8.5) using reliability ensemble averaging were used in the analysis for projecting the future extremes. To start with, the presence of nonstationarity in the observed annual maximum precipitation (AMP) series and the future ensemble averaged AMP projections were investigated. For this purpose, three generalized extreme value (GEV) models—one stationary model with constant parameters and two nonstationary models with trends in location and scale parameters—were applied to assess the goodness of fit using Akaike information criterion and likelihood ratio test. The best fit model was used in the uncertainty analysis, and the confidence bounds of extreme precipitation return levels were estimated. A nonparametric bootstrapping approach was followed in the uncertainty analysis. Results of the study suggest that a nonstationary GEV distribution with linear trend in location parameter and constant scale and shape parameters are the best fit distribution for the AMP series under the RCP scenarios, whereas the stationary GEV distribution fits the observed AMP series the best. The expected values and confidence bounds of return levels obtained from the uncertainty analysis reveal that precipitation extremes in the river basin would intensify under the projected climate change scenarios. Compared with the RCP4.5 scenario, the confidence intervals of return levels under the RCP8.5 scenario were wider, implying that uncertainty in the latter scenario is higher.


Drastic increase in the concentration of greenhouse gases in the atmosphere is considered as one of the dominant causes of climate change. This is believed to have resulted in the changes in pattern and occurrence of extreme weather and climatic events in different parts of the world [1]. An increase in temperature induced by the greenhouse gases will increase the rate of evaporation and thereby enhance the moisture content in the atmosphere [2]. This can cause more frequent and heavier rainfall events. According to the Clausius–Clapeyron relationship, the intensity of extreme precipitation events is expected to increase at the rate of 7% per 1 °C rise in temperature [3]. Analyses of the global trends in extreme rainfall events reveal increasing trends in some regions and decreasing trends in some other regions. Countries such as the USA, UK, Australia, South Africa, and India have been experiencing an increasing trend whereas a decreasing trend has been observed over Western Australia, Northern and Eastern New Zealand, parts of Central Pacific, and South East Asia [4,5,6,7,8,9].

India witnessed a number of unprecedented extreme precipitation events in the recent past which resulted in massive flooding. The floods in Kerala (2018 and 2019), Uttarakhand (2013), Chennai (2015), Bengaluru (2016), and Mumbai (2005 and 2017) are some among them [10, 11]. Based on the analysis of observed annual maximum precipitation (AMP) from 1979 to 2015, an increase in AMP is reported over many parts of India except the Gangetic Plain, northeastern India, and Jammu and Kashmir; the increase in AMP is found to be more prominent in Southern India than in Northern India. Furthermore, the frequency of precipitation extremes is projected to increase more prominently in southern and central India in the mid and end of the twenty-first century under the representative concentration pathway (RCP) 8.5 [11]. An earlier study using the special report on emissions scenarios, SRES A2 and B2, also projected an increase in the intensity and frequency of heavy precipitation events toward the end of the twenty-first century over many parts of India [12]. The present study is conducted in the Chaliyar river basin which is located in Kerala, the southernmost state of the Indian Union. This was one of the worst affected river basins in Kerala during the floods of 2019; it was badly affected during the 2018 floods as well. Frequent floods have adversely affected all important sectors of the society including socio-economy, infrastructure, transportation, agriculture, etc. [13]. Since frequent extreme precipitation events with increased magnitudes can be expected with climate change, studies focusing on the impact of climate change on these extremes would be very useful in the formulation of mitigation and adaptation strategies to cope with and reduce the negative impacts of such events on the society.

Climate extremes, being rare in occurrence, follow different laws and are harder to predict when compared with the averages [14]. Extreme value theory forms the foundation of analysis of extremes and determination of the likelihood of occurrence of extreme events. Gumbel distribution, generalized extreme value (GEV) distribution, log-Pearson Type III distribution, and generalized Pareto distribution (GPD) are the four commonly used extreme value distributions in the analysis of hydro-meteorological extremes [15,16,17,18]. The first three distributions follow a block maximum approach based on the annual maximum series, whereas the last one follows a peak over threshold approach which is based on the peak values over high thresholds [15, 19]. The distributions of classical extreme value theory are based on the assumptions of independence and stationarity of the extreme data series [20,21,22]. Later, many studies considered extending this theory to suit dependent and nonstationary extreme sequences [23, 24].

In recent years, several studies have reported that hydro-meteorological extremes violate the concept of stationarity [25, 26]. Nonstationarity in hydrologic time series can be identified if there exist some trends, periodicity, or shifts. Kundzewicz and Robson [27] suggested that short-term variability in the climate arising from very short records of data should not be confused with nonstationarity. Scientists are attributing the underlying causes of nonstationarity in hydrological extremes to climate change, low-frequency internal variability, and anthropogenic influences such as urbanization, deforestation, change in catchment characteristics, etc. [28,29,30,31,32]. Mondal and Mujumdar [33] advocated that in a climate change context, future extremes which are obtained from climate model projections may exhibit nonstationarity irrespective of the observed extremes in the past which may be stationary. Vogel et al. [34] established that the magnitude of a future flood tends to increase whereas its average recurrence interval tends to reduce if an increasing trend is observed in the flood series, i.e., a flood event associated with a high return period at present may become more frequent in future under nonstationary conditions.

The assumption of stationarity can reduce the complexity of modeling the extremes, but non stationary modeling is advised if there exists a significant trend in the extreme series [35]. In recent studies, for investigating and modeling the nonstationarity in the extreme series, nonstationary extensions of extreme value distributions are used, where the parameters are expressed either as functions of time or physical covariates [19, 32, 36, 37]. In case of a GEV distribution, trends in time are incorporated either in the location/scale parameter or both the location and scale parameters to incorporate the nonstationarity in the extreme series [25, 36]. In case of a GPD, the scale parameter is treated as a function of time [38]. The shape parameter is usually treated as a constant, independent of time, since varying it as a function of time is difficult [39]. Covariates related to physical processes such as urbanization, local temperature changes, global warming, El Niño-Southern Oscillation (ENSO) cycle, etc. are used in some studies to account for the nonstationarity in the extreme rainfall series [28, 40]. In some studies, generalized additive models for location, scale, and shape (GAMLSS) have been used in nonstationary flood frequency analysis [41]. This approach is based on semi-parametric regression, in which a parametric distribution belonging to a general distribution family such as a highly skewed/kurtotic distribution is used for the annual maximum flood series and the parameters of the distribution are modelled as linear, nonlinear, or nonparametric smoothing functions of the dependant variables [42]. Also, in some studies, nonlinear trends in nonstationary series are modeled using a conditional density network, which is a probabilistic extension of the multilayer perceptron neural network [43] and by a multi-objective genetic algorithm generated time-based covariate [44].

Quantification of uncertainty in an extreme quantile is of paramount importance in the context of assessment of risk and decision making. Uncertainty yields a range of values for the extreme quantile, which otherwise is estimated as a single value (point estimate) corresponding to a particular return period. Methods for quantification of uncertainty are well developed for a stationary case [45, 46]. Since hydrological extremes may exhibit nonstationarity in a changing climate, methods used to perform uncertainty analysis should be capable of taking nonstationary conditions into account. One of the premier works on the quantification of uncertainty of design quantile under nonstationarity is by Obeysekera and Salas [47]. They extended the delta method, profile likelihood method, and bootstrap method to nonstationary conditions for dealing with uncertainty in the parameters as well as in the return levels [47, 48]. Serinaldi and Kilsby [49] reported that the bootstrap method is the most practical one among these methods since it relies only on the available information, does not rely on a particular estimation method, and could be easily implemented for complex models. Return level uncertainty as well as parameter uncertainty can be estimated by Bayesian analysis too [25, 50]. Like parameter uncertainty, covariate uncertainty could also have some significance in the case of a nonstationary model with a number of covariates [40].

The objectives of the present study are (i) to investigate nonstationarity in the annual maximum precipitation series—observed and corresponding to the future climate (RCPs 4.5 and 8.5)—of the Chaliyar river basin by identifying an appropriate frequency distribution from stationary and nonstationary GEV models and (ii) to quantify the uncertainty in return levels as confidence intervals by performing frequency analysis in a bootstrapping framework.

Materials and Methods

Study Area

Chaliyar is a river in the south western Indian state of Kerala which is located between the Western Ghats in the east and the Arabian Sea on the west (Fig. 1). The state, most areas of which are located in the humid tropical zone, is endowed with 44 rivers, Chaliyar being the fourth longest (170 km). River Chaliyar originates from the Elambalari hills at an altitude of about 2067 m above mean sea level (MSL). The river basin spreads over an area of 2933km2, in the states of Kerala and Tamil Nadu. The basin of this Virgin River can be divided into highland, midland, lowland, and coastal plain based on topographic features. The predominant land uses are agriculture (60.04%) and forests (38.74%). Small patches of urban areas, pastures, waste lands, and rocky areas are also present. The major soil type is loam (42.74%), followed by clay (28.66%), clay loam (24.18%), and sandy loam (4.42%) [51]. The climate is humid tropical with hot summer and heavy monsoon rainfall. March and April are the hottest months, and December and January are the coolest months. The average annual precipitation in the river basin is 3012.61 mm and the average maximum and minimum temperatures are 34 °C and 24 °C, respectively [51].

Fig. 1

Study area


Rainfall data at three rain gauge stations, namely, Kottamparamba, Manjeri, and Nilambur, were collected. Daily rainfall data at the Kottamparamba station for the period 1981–2005 was obtained from the Centre for Water Resources Development and Management (CWRDM), Kunnamangalam, Kerala, whereas the daily rainfall data for Manjeri and Nilambur rain gauge stations for the period 1976–2004 were collected from the India Meteorological Department (IMD). To study the precipitation extremes due to projected climate change, daily rainfall data (interpolated to station point using inverse distance weighting method) for the RCPs 4.5 and 8.5 were obtained from eight high-resolution general circulation models of the Coordinated Regional Climate Downscaling Experiment (CORDEX) under the South Asia domain. The dynamically downscaled output of a general circulation model (GCM) is referred to as high-resolution GCM output in some studies, since the coarse resolution of the GCM output is brought down to a finer resolution by downscaling with a nested regional climate model (RCM) [50, 52]. In the present study also, the usage “high-resolution GCM” is adopted to represent the GCM which is nested with an RCM. Details of the high-resolution GCMs used in this study are presented in Table 1. CORDEX datasets are available from the climate data portal of the Centre for Climate Change Research (CCCR), Indian Institute of Tropical Meteorology (IITM), Pune, India (

Table 1 High-resolution GCMs from CORDEX-South Asia used for the study


The overall methodology used in the study is presented in Fig. 2. To start with, the AMP series were extracted from the observations and the dynamically downscaled projections of global climate models for the RCPs 4.5 and 8.5. The weighted averaged projections of the AMP series using reliability ensemble averaging (REA) were used to obtain a comprehensive realization of projections from multiple climate models. REA weights were assigned for each climate model simulation based on the ability of the model to emulate the present climate as well as the degree with which its simulations converged with the simulations of other models. Disparities of the weighted projections with the observations were compensated by a bias correction technique. The relationship between the magnitude and frequency of extremes was obtained by performing frequency analysis using extreme value distributions. In order to fit an extreme value distribution to the series of extreme precipitation or stream flow, it should satisfy the conditions of independence and stationarity [28, 63, 64]. In the present study, nonstationarity was investigated by detecting the trend in the AMP series using nonstationary extensions of the GEV distribution. For this purpose, a stationary GEV model with constant parameters and two nonstationary GEV models with time-dependent parameters were evaluated for goodness of fit on the AMP series using the Akaike information criterion (AIC). In order to validate the AIC test, a significance test called likelihood ratio test was also performed [52, 65, 66]. If the best fit model happens to be any one of the two nonstationary GEV models, the annual maximum precipitation series can be said to possess a nonstationary behavior. The best fit model can be used for the uncertainty analysis of extreme precipitation return levels. For this purpose, nonparametric bootstrapping was carried out, and the confidence intervals of return levels were obtained. Detailed discussions on these are presented in Sects. 2.3.1 to 2.3.6.

Fig. 2

Overall methodology

REA for Weighted Ensemble Projections

Projections of climate models are subjected to various sources of uncertainties. It includes uncertainties arising from anthropogenic forcing estimates, initial conditions, boundary conditions, model structure, model response, etc. As a result, projections of the future by different climate models yield different results. Using an ensemble of climate model projections is an accepted option for regional analysis [67]. In the present study, REA, a weighted averaging procedure based on reliability weights was used to obtain the ensemble averaged projection of the high-resolution GCMs listed in Table 1. Historical data of the high-resolution GCMs for the three stations were taken corresponding to the period of observed data, i.e., 1981–2005 for Kottamparamba and 1976–2004 for Manjeri and Nilambur. Future projections of the high-resolution GCMs were taken for the period 2006–2099.

Xu et al. [68] investigated the relevance of weighting schemes of REA by comparing it with a simple averaging procedure using several performance metrics and observed that REA is capable of providing finer variations of climate change signals than simple averaging for regional-scale applications. In REA, two reliability criteria, namely, model performance and model convergence criteria, are used to assign the weight to each climate model projection. Model performance criterion deals with the ability of the climate model projection to replicate the present-day climate (i.e., a measure of degree of bias) whereas model convergence criterion takes into account the convergence of the climate model projections to the projections of other models in the ensemble. In this study, the methodology proposed by Riano [69] for applying REA on the AMP series was used. Reliability (\(r_{i,s}\)) of the projection by the climate model i in a particular scenario s (e.g., RCP4.5) can be obtained as:

$$r_{i,s} = rb_{i} \times rc_{i,s}$$

where \(rb_{i}\) is the reliability factor for bias of the model i, and \(rc_{i,s}\) is the reliability factor for convergence of the model i in the scenario s.

The reliability factor for bias \(rb_{i}\) can be computed as:

$$rb_{i} = \frac{1}{t}\sum\limits_{j = 1}^{t} {(AMP_{o,j} } - AMP_{h,i,j} )^{2}$$

where \(AMP_{o,j}\) is the observed annual maximum precipitation for the jth year, \(AMP_{h,i,j}\) is the historical annual maximum precipitation for the ith model in the jth year, and t is the number of years.

The reliability factor for convergence \(rc_{i,s}\) can be computed as

$$rc_{i,s} = \frac{1}{t}\sum\limits_{j = 1}^{t} {(AMP_{s,i,j} } - \overline{AMP}_{s,j} )^{2}$$

where \(AMP_{s,i,j}\) is the annual maximum precipitation projection for a particular scenario s for the ith model in the jth year, \(\overline{AMP}_{s,j}\) is the REA-weighted averaged projection of all the models in a particular scenario s for the jth year, and t is the number of years. The calculation of \(rc_{i,s}\) requires \(\overline{AMP}_{s,j}\), and this can be computed as

$$\overline{AMP}_{s,j} = \frac{{\sum\limits_{i = 1}^{n} {(r_{i,s} \times AMP_{s,i,j} )} }}{{\sum\limits_{i = 1}^{n} {r_{i,s} } }}$$

in which \(r_{i,s}\) is unknown. Therefore, \(rc_{i,s}\) can be computed using a trial-and-error procedure, starting with an initial assumed value of \(r_{i,s}\) (say \(rb_{i}\)). After several iterations, a converged solution can be obtained for \(rc_{i,s}\). \(\overline{AMP}_{s,j}\) can also be written as

$$\overline{AMP}_{s,j} = \sum\limits_{i = 1}^{n} {(w_{i} } \times AMP_{s,i,j} )$$

where \(w_{i}\) is the weight of the \(i^{th}\) model in the ensemble represented by the term \(\frac{{r_{i,s} }}{{\sum\limits_{i = 1}^{n} {r_{i,s} } }}\) in Eq. 4.

A model will be less reliable if the values of \(rb_{i}\) and \(rc_{i,s}\) are higher, since the calculations of these terms are based on mean square errors. For easy interpretation, these factors can be converted to a score system in which the best model is represented by the highest score. This can be achieved by normalizing \(rb_{i}\) and \(rc_{i,s}\) [70] as

$$RB_{i} = \frac{{rb_{\max } - rb_{i} }}{{rb_{\max } }}$$
$$RC_{i,s} = \frac{{rc_{\max ,s} - rc_{i,s} }}{{rc_{\max ,s} }}$$

where \(RB_{i}\) and \(RC_{i,s}\) represent normalized values of \(rb_{i}\) and \(rc_{i,s}\), \(rb_{\max }\) is the maximum of the \(rb_{i}\) values of all models, and \(rc_{\max ,s}\) is the maximum of the \(rc_{i,s}\) values of all models for a scenario s.

Application of a Modified Scaling Factor Approach for Bias Correction

Climate model projections are affected by various sources of errors such as incorrect assumptions in model parameterization, incorrect initial conditions, incorrect representation of physical processes, etc. So, the direct application of these projections results in erroneous prediction of different climate-related processes [71]. Usually, bias correction methods are used to reduce the bias in climate model projections and enhance the consistency and quality of these projections. The commonly used bias correction techniques include delta change approach, linear transformation, local intensity scaling, variance scaling, power transformation, and distribution mapping [72]. In this study, the ensemble averaged AMP projections, after performing REA, were corrected for bias. For this purpose, a methodology recently developed by Mehr and Kahya [73] to correct the biases in the annual maximum series was followed. Using this method, the annual mean of the ensemble averaged AMP series from climate model projections for the historical (reference) period is matched with that of the observed AMP series by using a scaling factor. Correction is applied to the ensemble averaged AMP series under the scenario \(s\) (\(AMP_{s}\)) as

$$AMP_{c} = AMP_{s} \times \frac{{\eta (AMP_{o} )}}{{\eta (AMP_{r} )}},$$

where \(AMP_{c}\) is the bias corrected annual maximum precipitation, \(\frac{{\eta (AMP_{o} )}}{{\eta (AMP_{r} )}}\) is the scaling factor in which \(\eta\) is the long-term mean of the bracketed series, \(AMP_{o}\) is the observed AMP series, and \(AMP_{r}\) is the ensemble averaged AMP series from the high-resolution GCM output for the same reference period as that of the observed series.

Detection of Nonstationarity in the Annual Maximum Precipitation Series

Detection is the process of identifying statistical changes in the climate or climate-related systems [74]. Since extreme events are being reported more frequently with transient changes in intensity, the presence of nonstationarity has to be detected before performing frequency analysis. In order to handle the nonstationarity of hydrological extremes in frequency analysis, trends in the extreme series should be accounted [32]. In some earlier studies, the significance of trend in the extreme series was investigated either by performing a nonparametric test such as Mann-Kendall test which was not designed particularly for dealing with the extremes or by employing a linear regression which assumed the data as normally distributed [75, 76]. Clarke [75] and Katz [76] pointed out that the latter approach is not logical since it involves the assumption of normal distribution for the series of extremes, although it actually follows an extreme value distribution. In recent years, extensions of extreme value distributions that are capable of shifting over time have been developed for the detection and analysis of nonstationarity in hydrological extremes [19, 32, 76, 77]. In some studies, the trends in block maximum series have been detected by fitting a nonstationary generalized extreme value distribution with time-varying parameters on the block maximum series [78, 79]. The goodness of fit of the modified extreme value distributions was checked with methods for model selection or with tests for significance of the trend parameters [66]. Zhang et al. [80] compared the Mann-Kendall test, linear regression, and the nonstationary extreme value distribution fitting method for detecting the trend in the extreme values using Monte Carlo simulations and found that the latter always outperformed the other two methods in trend detection. Therefore, in the present study, the GEV distributions with time-dependent parameters were used to detect nonstationarity in the annual maximum precipitation series.

The cumulative distribution function (CDF) of a classical GEV distribution [19, 77] can be expressed as

$$F(x,\phi ) = \left\{ {\begin{array}{*{20}c} {\exp \left\{ { - \left[ {1 + k\left( {\tfrac{x - \mu }{\sigma }} \right)} \right]^{{ - \,\frac{1}{k}}} } \right\},\,\,\,\,\,k \ne 0} \\ {\exp \left\{ { - \exp \left[ { - \left( {\tfrac{x - \mu }{\sigma }} \right)} \right]} \right\},\,\,\,\,\,\,\,\,\,\,\,k = 0} \\ \end{array} } \right.$$

where \(\phi = \left\{ {\mu ,\,\sigma ,\,k} \right\}\) denotes a parameter set in which \(\mu\), \(\sigma\), and k are the location, scale, and shape parameters respectively. Under nonstationary conditions, the parameters of the GEV distribution become time dependent and the CDF changes to

$$F(x,\phi_{t} ) = \left\{ {\begin{array}{*{20}c} {\exp \left\{ { - \left[ {1 + k_{t} \left( {\tfrac{{x - \mu_{t} }}{{\sigma_{t} }}} \right)} \right]^{{ - \,\frac{1}{{k_{t} }}}} } \right\},\,\,\,\,\,k_{t} \ne 0} \\ {\exp \left\{ { - \exp \left[ { - \left( {\tfrac{{x - \mu_{t} }}{{\sigma_{t} }}} \right)} \right]} \right\},\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,k_{t} = 0} \\ \end{array} } \right.$$

where \(\phi_{t} = \left\{ {\mu_{t} ,\,\sigma_{t} ,\,k_{t} } \right\}\) denotes a parameter set in which \(\mu_{t}\), \(\sigma_{t}\), and \(k_{t}\) are the time-dependent location, scale, and shape parameters, respectively [76]. Usually, the location and scale parameters are treated as functions of time, and the shape parameter is considered as a constant, independent of time, since modeling the shape parameter as a function of time is complicated [39, 80]. Zhang et al. [80] pointed out that any variation of tail behavior of the distribution represented by a shape parameter can be approximately handled by variations of the scale parameter. In this study, three cases of GEV distributions were considered:

  1. 1.

    A stationary GEV distribution with location, scale, and shape parameters as constants

    (\(\mu\), \(\sigma\), k)—GEV1.

  2. 2.

    A nonstationary GEV distribution with location parameter as a linear function of time (\(\mu_{t} = \mu_{0} + \mu_{1} \,t\)), scale, and shape parameters as constants (\(\sigma_{t} = \sigma\), \(k_{t} = k\))—GEV2.

  3. 3.

    A nonstationary GEV distribution with location parameter as a linear function of time (\(\mu_{t} = \mu_{0} + \mu_{1} \,t\)), scale parameter as log-linear function of time (\(\log (\sigma_{t} ) = \sigma_{0} + \sigma_{1} \,t\)) and shape parameter as constant (\(k_{t} = k\))—GEV3.

In these models, \(\mu_{1}\) and \(\sigma_{1}\) are the slopes of the location and scale parameters which could represent the transient nature of the probability distribution with the covariate time t. The scale parameter is log-transformed to maintain it as a positive value [19, 76]. The AMP series of the observations and the RCPs 4.5 and 8.5 were fitted with these three models by the maximum likelihood method. This is one of the widely used methods of distribution fitting owing to its simplicity, ability to generate unbiased estimates, and the flexibility to be used in a nonstationary condition [32, 77].

The best model from among the three GEV models considered was selected by the model selection criteria discussed below. Using the selected model, inferences can be drawn about the characteristics of the data to which it is fitted. Also, it can be used to perform frequency analysis for estimating the return levels.

Model Selection

The main aim of model selection is to identify a model that adequately fits the data. It shall maintain a balance between the goodness of fit of the model and the number of parameters, with due consideration of the sample size [81]. Forward, backward, and stepwise selection procedures, cross validation, information criteria such as Akaike information criterion, Bayesian information criterion, and likelihood ratio test are some of the widely used statistical methods for model selection [82]. The process of model selection should be treated as crucial since the predictive ability of the model depends on the selected model [83].

In this study, a stationary GEV model and two nonstationary GEV models were fitted to the observed annual maximum series as well as to the ensemble averaged annual maximum series from future projections under the RCPs 4.5 and 8.5. To select the best model among these, Akaike information criterion was used along with the likelihood ratio test.


Akaike information criterion was developed by Akaike [84] to identify an optimal model that fits the data from a class of competing models. The model with minimum AIC value can be considered as the best predictive model. AIC reduces the possibility of overfitting of a model with the use of a penalty term. AIC can be computed as

$$AIC = - 2L + 2p$$

where L is the log-likelihood value and 2p is the penalty term with p representing the number of parameters of the model.

For easy interpretation, the AIC value of each model i can be rescaled with respect to the minimum AIC (\(AIC_{\min }\)) of the competing class. The rescaled value \(\Delta_{i}\) is given by

$$\Delta_{i} = AIC_{i} - AIC_{\min }$$

From different ranges of \(\Delta_{i}\) values, inferences can be drawn about the relative reliabilities of competing class of models. The model with \(\Delta_{i}\) value equal to 0 is regarded as the best and that with \(\Delta_{i}\) value greater than 10 is regarded as the worst. Though not the best, models that have small \(\Delta_{i}\) values (less than or equal to 2) can be considered as substantially reliable [85].

Likelihood Ratio Test

This is a hypothesis test that can be used to compare the goodness of fit of a simple model (with few parameters) and a complex model (with more parameters) if these are nested. Two models can be called nested when one model contains all the terms of the other model along with some additional terms [86]. Likelihood ratio (LR) test checks whether the use of a complex model with additional parameters could improve the goodness of fit when compared with what is obtained with a simple model. Often, the simple model is called a restricted model, and the complex model is called an unrestricted model. The null hypothesis of the test is in favor of the simple model, and the alternative hypothesis is in favor of the complex model. The null hypothesis of the test is rejected when the likelihood ratio test statistic is greater than the critical value at a certain significance level [86, 87]. The likelihood ratio test statistic (\(L_{r}\)) is given by

$$L_{r} = 2(L_{c} - L_{s} )$$

where \(L_{c}\) is the log-likelihood of the complex model, and \(L_{s}\) is the log-likelihood of the simple model.

When there are more than two models, this test can be implemented by successive pairwise evaluations of the nested models in a hierarchical manner [65, 88]. In the present study, the order of complexity (based on the number of parameters) of GEV models is GEV1 < GEV2 < GEV3. Hence, by using the LR test, two possibilities can be checked:

  1. 1.

    Whether the addition of an extra parameter (\(\mu_{1}\)) to GEV1, i.e., the GEV2 model, could give any significant improvement in the goodness of fit than that obtained using a simple model, i.e., GEV1

  2. 2.

    Whether the addition of an extra parameter (\(\sigma_{1}\)) to GEV2, i.e., the GEV3 model, could give any significant improvement in the goodness of fit than that obtained using a simpler model, i.e., GEV2

If in the first step the null hypothesis is not rejected, i.e., the simple model GEV1 is the best model for the data, it can be inferred that the use of a complex model with an additional parameter is a wrong choice, since it cannot provide any significant improvement in the goodness of fit. In this case, the second step can be avoided. If in the first step the null hypothesis is rejected and the complex model GEV2 is found to be the best model for the data, the second step can be performed to check whether significant improvement in the goodness of fit can be obtained by the addition of an extra parameter.

Estimation of Return Levels

In classical flood frequency analysis, the return level corresponding to a return period (T) is the extreme quantile for which the exceedance probability is equal to the reciprocal of the return period (i.e., 1/T) [77]. The inverse of Eq. 9 gives the expression for calculating the return level (\(x_{T}\)):

$$x_{T} = \left\{ {\begin{array}{*{20}c} {\mu - \frac{\sigma }{k}\left[ {1 - \left( { - \ln \left( {1 - \tfrac{1}{T}} \right)} \right)^{ - \,k} } \right],\,\,\,\,\,\,k \ne 0} \\ {\mu - \sigma \ln \left( { - \ln \left( {1 - \tfrac{1}{T}} \right)} \right),\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,k = 0\,} \\ \end{array} } \right.$$

In recent times, a number of methods have been developed to estimate the return levels under a nonstationary paradigm. In one of the methods, nonstationary return level for a T year return period event was derived from the effective return level concept [19, 36, 89]. The nonstationary counterpart of Eq. 14 is the expression for calculating the time-varying extreme quantile, which is usually referred to as the effective return level (\(\mathop x\nolimits_{{\mathop T\nolimits_{t} }}\)):

$$x_{{T_{t} }} = \left\{ {\begin{array}{*{20}c} {\mu_{t} - \frac{{\sigma_{t} }}{{k_{t} }}\left[ {1 - \left( { - \ln \left( {1 - \tfrac{1}{T}} \right)} \right)^{{ - \,k_{t} }} } \right],\,\,\,\,\,k_{t} \ne 0} \\ {\mu_{t} - \sigma_{t} \ln \left( { - \ln \left( {1 - \tfrac{1}{T}} \right)} \right),\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,k_{t} = 0} \\ \end{array} } \right.$$

A value representative of the effective return level, generally the one corresponding to the median value or the 95th percentile of the time-varying location parameter, is taken as the nonstationary return level [25]. Parey et al. [90] estimated the return level in a nonstationary context by redefining it as a level such that the expected value of the number of exceedances over that level in the next T years will be one. Later, Cooley [36] and Obeysekera and Salas [30, 47] came up with the extensions of expected waiting time theory for return period estimation under nonstationary conditions. In the present study, the expected waiting time approach was followed for the estimation of nonstationary return levels.

Expected Waiting Time Method

In this method, the return period is regarded as the expected time till the occurrence of the next extreme event [91]. The expressions of return period and time-varying exceedance probability proposed by Obeysekera and Salas [30, 47] were used in this study for the estimation of extreme precipitation return level. This method is based on the exceedance probabilities of a fixed threshold with time. Let \(x_{{T_{0} }}\) be a fixed threshold and Y be the random variable representing the waiting time till the first exceedance of \(x_{{T_{0} }}\) at Y = y. The probability distribution of Y is given by

$$f(y) = P(Y = y) = p_{y} \prod\limits_{t = 1}^{y - 1} {(1 - p_{t} ),\,\,\,\,y = 1,2,3,.....,y_{\max } }$$

Since return period is considered as the expectation of waiting time till the occurrence of the next extreme event, T can be obtained as

$$T = E(Y) = \sum\limits_{y = 1}^{{y_{\max } }} {y\,f(y) = \sum\limits_{y = 1}^{{y_{\max } }} {y\,p_{y} } } \prod\limits_{t = 1}^{y - 1} {(1 - p_{t} )}$$

The expected waiting time \(E(Y)\) is a function of time-varying exceedance probability \(p_{t}\) (i.e., the probability of the given threshold being exceeded each year) and is given by

$$p_{t} = 1 - F(x,\phi_{t} )$$

where \(F(x,\phi_{t} )\) is presented in Eq. 10.

Uncertainty in the Return Levels from Bootstrapping

Bootstrapping is a statistical method developed by Efron [92] in which the original sample is resampled and statistical inferences are derived. Bootstrapping is widely used in frequency analysis for the estimation of uncertainties in distribution parameters and return levels [93, 94]. There are two approaches in bootstrapping, namely, parametric and nonparametric. Parametric approach can be used when an appropriate probability distribution of a sample is known, and bootstrap samples are derived from the known distribution. This method is preferred when the sample size (n) is very small (\(n \le 60\)), since the coverage performance of the confidence interval reduces with increase in sample size [94]. The nonparametric approach is based on sampling with replacement from the original sample and is more accurate for sample sizes above 40 [94]. Compared with the parametric approach, the nonparametric approach being simple and requiring less information is widely used in frequency analysis [47, 95]. In this study, uncertainties in distribution parameters and return levels were estimated using the nonparametric approach proposed by Obeysekera and Salas [47] for use under nonstationary conditions. The steps involved in this method are as follows:

Step 1. Initially, the nonstationary GEV model was fitted to the AMP series \(x_{t}\). The return level (\(x_{{T_{0} }}\)) corresponding to return period T was obtained from Eqs. 17 and 18. Using the parameters of the fitted model and the AMP series \(x_{t}\), the standard residuals \(\overline{x}\) were calculated using the equation:

$$\overline{x} = \frac{1}{{k_{t} }}\ln \left[ {1 + k_{t} \left( {\tfrac{{x_{t} - \mu_{t} }}{{\sigma_{t} }}} \right)} \right]$$

Step 2. These residuals were resampled with replacement, and a new set of \(\overline{x}\) with a length equal to that of the original AMP series was obtained. This new set of \(\overline{x}\) was back transformed and a new maximum series of \(x_{t}\) was obtained using Eq. 19. The nonstationary GEV model was then fitted to the new \(x_{t}\) series, and the new set of parameters and return level were obtained.

The processes in step 2 were repeated a large number of times; in this study, 1000 iterations were performed.

For the stationary case, the uncertainty in distribution parameters and return levels were obtained in the same manner but with a stationary GEV distribution with constant parameters.

Results and Discussion

REA of Projections as an Average Climate Realization

The ensemble averaged AMP series for the historical and future periods were obtained by applying REA on the AMP projections of the eight high-resolution GCMs listed in Table 1. The models, whose projections for the historical period were comparable with the observations and whose future projections converged, were identified by assessing the reliability weights. Normalized values of reliability factors for bias, convergence, and the final weights of each model in the ensemble are presented in Table 2. Details regarding the number of iterations taken for the REA procedure are also presented in Table 2.

Table 2 Reliability weights of climate models

Since the factors \(RB_{i}\) and \(RC_{i,s}\), listed in Table 2, are normalized values, a higher \(RB_{i}\) value indicates lesser model bias, and a higher \(RC_{i,s}\) value indicates that the model projections exhibit a higher degree of agreement to the projections of other models in the ensemble. Therefore, a climate model with the highest weight is the best model. At Kottamparamba, the performance of ACCESS(CCAM) and CanESM(RegCM4) was found to be poor for the historical period and the RCP scenarios. IPSL(RegCM4) performed the best for the historical period and under RCP8.5, whereas the performance of CSIRO(RegCM4) was the best under RCP4.5. The best models for a particular scenario for the other stations can be identified from Table 2. The weighted averaged AMP projections obtained from REA was used in the further analyses.

Bias Correction Statistics

Even though the bias in historical projections with respect to the observed data was accounted for in the weighted averaging procedure, some amount of bias would still be left uncorrected. So, bias correction was applied on the ensemble averaged AMP series for better results. In this study, bias correction of the AMP series was performed following the methodology outlined in Sect. 2.3.2. Scaling factors of 1.07, 1.13, and 1.14 were obtained for Kottamparamba, Manjeri, and Nilambur, respectively. The ensemble averaged AMP series during the historic (raw) and future periods were multiplied by these scaling factors to obtain the bias-corrected AMP series during the respective periods. In order to assess the performance of the bias correction technique adopted, the ensemble averaged AMP series for the historic period was compared with the observed AMP series after bias correction, and the results are presented in Table 3. It can be seen that the statistics such as mean, median, 25th percentile, and 75th percentile of the bias-corrected AMP series exhibit a much better match with the observations than that obtained with the uncorrected AMP series.

Table 3 Bias correction statistics of AMP series

Appropriate Probability Distributions for the AMP Series

Nonstationarity modeling of the AMP series was implemented by introducing time trends in the parameters of the generalized extreme value distribution. The three GEV models discussed in Sect. 2.3.3 were fitted to the annual maximum precipitation series derived from the observations as well as the ensemble averaged projections for the future under RCPs 4.5 and 8.5. The model performance was evaluated using the AIC and LR tests. AIC values, \(\Delta_{i}\) values, and p values of likelihood ratio test statistics obtained from the fit of the annual maximum series of observations and ensemble averaged RCP projections with each of the three GEV models are presented in Table 4.

Table 4 Results of AIC and LR test

According to the Akaike information criterion, the model with minimum AIC value is considered to be the best fit model for the data. For the observed AMP series at the three stations, it can be inferred that GEV1, i.e., the stationary GEV distribution, fits the best since AIC is the least (\(\Delta_{i} = 0\)) for this model. For the AMP series under the RCP scenarios, the GEV2 model with the location parameter having a time-varying linear trend and constant scale and shape parameters yielded the best fit at all the three stations. The stationary model, GEV1, was found to be the worst model under both RCPs at all the stations since their \(\Delta_{i}\) values are greater than 10. Under the RCP8.5, for the Kottamparamba and Manjeri stations, the AIC values of GEV2 and GEV3 are very close (\(\Delta_{i}\) is very low). In some other cases also, \(\Delta_{i}\) values are less than 2, and these models too have substantial predictive reliability. However, in this study, the best model with \(\Delta_{i}\) equal to 0 was chosen to obtain better accuracy in the estimation of return levels.

In order to support the results of the AIC test, the likelihood ratio test was also performed. At 10% significance level, the probability associated with the LR test statistic (p value) was reviewed. As mentioned in Sect. 2.3.4, the LR test was performed in two steps. In step 1, the null hypothesis was considered in favor of GEV1 against the alternative of GEV2. If the p value is less than 10% (i.e., 0.1), the null hypothesis is rejected. It can be seen from Table 4 that for the observed AMP series, the p values at all the stations are greater than 0.1. This indicates that the null model GEV1 cannot be rejected, i.e., the stationary model is the best model for the observed series. But for the RCPs 4.5 and 8.5, the null model GEV1 is rejected since the p values are less than 0.1 at all the stations. GEV2, the nonstationary model with time-varying location parameter, was found to be the best fit model for both the RCPs in the first step.

In the second step, the suitability of a more complex model, i.e., GEV3, for the AMP series was investigated by keeping the null hypothesis in favor of GEV2 (against the alternative of GEV3). Since for the observed AMP series the stationary model GEV1 is the best fit model, there is no need for checking the suitability of a model which is more complex than GEV2. Hence, the second step was implemented only for the ensemble averaged AMP series derived for RCPs 4.5 and 8.5. The p values obtained were greater than 0.1, which means that the null model GEV2 cannot be rejected against the alternative of GEV3.

From the two-stepped LR test, the GEV2 model with time-varying location parameter was found to be the best model for the AMP series under both the RCPs, and the GEV1 model was found to be the best model for observed AMP series. The results of the LR test complemented the results of AIC test.

Uncertainty Estimates of Parameters and Return levels

A nonparametric bootstrapping algorithm was used to quantify the uncertainty associated with the return levels. Uncertainty in the distribution parameters as well as uncertainty in the return levels can be obtained from this method. To obtain the uncertainty in the observed return levels, GEV1 with constant parameters was employed; for the projected climate change scenarios, GEV2 with time-varying location parameter was used in the bootstrap method. Uncertainties in the distribution parameters obtained from 1000 runs of the bootstrapping algorithm are presented as boxplot in Figs. 3 and 4.

Fig. 3

Uncertainty in distribution parameters of the observed data at Kottamparamba, Manjeri, and Nilambur

Fig. 4

Uncertainty in distribution parameters of the future projections under RCPs 4.5 and 8.5 at Kottamparamba, Manjeri, and Nilambur

The return levels were estimated using the parameter sets obtained from bootstrapping. It yielded the uncertainty in return levels, and these are presented in the form of empirical frequency curves for the three stations Kottamparamba, Manjeri, and Nilambur (Figs. 5, 6, and 7 respectively) for return periods 10, 25, 50, and 100 years. The empirical frequency curves for RCPs 4.5 and 8.5 exhibited a shift to the right with respect to the frequency curves for the observed AMP series at all the three stations. This implies that the magnitude of extreme precipitation return levels would increase under projected climate change. The shift is more for the curve corresponding to RCP8.5 than that for the curve of RCP4.5, as the extremes projected by the former are comparatively higher.

Fig. 5

Uncertainty in extreme precipitation return levels at Kottamparamba corresponding to the return periods (a) T = 10 years, (b) T = 25 years, (c) T = 50 years, (d) T = 100 years

Fig. 6

Uncertainty in extreme precipitation return levels at Manjeri corresponding to the return periods (a) T = 10 years, (b) T = 25 years, (c) T = 50 years, (d) T = 100 years

Fig. 7

Uncertainty in extreme precipitation return levels at Nilambur corresponding to the return periods (a) T = 10 years, (b) T = 25 years, (c) T = 50 years, (d) T = 100 years

For a better understanding of the ranges of return levels, percentile-based confidence intervals for the return levels are presented in Table 5. A 90% confidence interval (\(CI_{90}\)) is obtained from the confidence limits corresponding to the 5th percentile and the 95th percentile values of the empirical cumulative distribution function. The expected values and confidence bounds of the return levels were higher under both the RCPs when compared with that for the observed AMP series. Furthermore, compared with RCP4.5, wider confidence intervals were obtained for the return levels under RCP8.5, especially for higher return periods. The finding that increase in extreme precipitation return levels and the associated uncertainty is higher under the high emission scenarios is consistent with the results of some studies in which the extremes were defined using climate indices [96, 97].

Table 5 Expected value \(E(x_{T} )\) and 90% confidence intervals (\(CI_{90}\)) of extreme precipitation return level from uncertainty analysis


Uncertainty analysis of extreme precipitation return levels was performed for the Chaliyar river basin in Kerala, India, under projected climate change. Future extremes were derived from the reliability ensemble averaged projections of eight high-resolution GCMs. A stationary GEV model and two nonstationary GEV models—one with linear trend only in the location parameter and the other with linear trend in the location parameter and log-linear trend in the scale parameter—were evaluated for goodness of fit of the observed AMP series and the ensemble averaged AMP series obtained from the high-resolution projections. The stationary GEV model was obtained as the best fit model for observed AMP series. The nonstationary GEV model with linear trend only in location parameter and constant scale and shape parameters was found to be the best model among all the three models for future ensemble averaged AMP projections. Uncertainties in the distribution parameters and the return levels were obtained by a nonparametric bootstrapping approach. The expected values of return levels and the confidence bounds were found to be higher under climate change (RCPs 4.5 and 8.5) when compared with the corresponding values estimated from the observations. Results of the study point toward the intensification of precipitation extremes under projected climate change, and this indicates that the likelihood of occurrence of floods in the river basin would be enhanced. The expected return levels of precipitation extremes and associated uncertainties could be used for the optimal planning of adaptation and mitigation strategies and for the proper management of water resources in the river basin. Hydraulic infrastructure, designed based on traditional frequency analysis, should be appropriately rehabilitated/modified/reconstructed to withstand the extremes with time varying trends under a changing climate.

Data Availability

The climate model projections used in this study are from Coordinated Regional Climate Downscaling Experiment (CORDEX). These can be accessed through the climate data portal of Centre for Climate Change Research (CCCR), Indian Institute of Tropical Meteorology (IITM), Pune, India (


  1. 1.

    Pachauri, R. K, Allen, M. R, Barros, V. R., Broome, J., Cramer, W., Christ, R., & Dubash, N. K. (2014) Climate change 2014: synthesis report. Contribution of Working Groups I, II and III to the fifth assessment report of the Intergovernmental Panel on Climate Change (pp. 151). IPCC.

  2. 2.

    Trenberth, K. E. (1998). Atmospheric moisture residence times and cycling: Implications for rainfall rates and climate change. Climatic Change, 39(4), 667–694.

    Article  Google Scholar 

  3. 3.

    Muller, C. J., O’Gorman, P. A., & Back, L. E. (2011). Intensification of precipitation extremes with warming in a cloud-resolving model. Journal of Climate, 24(11), 2784–2800.

    Article  Google Scholar 

  4. 4.

    Easterling, D. R., Meehl, G. A., Parmesan, C., Changnon, S. A., Karl, T. R., & Mearns, L. O. (2000). Climate extremes: observations, modeling, and impacts. Science, 289(5487), 2068–2074.

    CAS  Article  Google Scholar 

  5. 5.

    Frei, C., & Schär, C. (2001). Detection probability of trends in rare events: Theory and application to heavy precipitation in the Alpine region. Journal of Climate, 14(7), 1568–1584.;2.

    Article  Google Scholar 

  6. 6.

    Goswami, B. N., Venugopal, V., Sengupta, D., Madhusoodanan, M. S., & Xavier, P. K. (2006). Increasing trend of extreme rain events over India in a warming environment. Science, 314(5804), 1442–1445.

    CAS  Article  Google Scholar 

  7. 7.

    Griffiths, G. M., Salinger, M. J., & Leleu, I. (2003). Trends in extreme daily rainfall across the South Pacific and relationship to the South Pacific convergence zone. International Journal of Climatology, 23(8), 847–869.

    Article  Google Scholar 

  8. 8.

    Haylock, M., & Nicholls, N. (2000). Trends in extreme rainfall indices for an updated high quality data set for Australia, 1910–1998. International Journal of Climatology, 20(13), 1533–1615.;2-J.

    Article  Google Scholar 

  9. 9.

    Salinger, M. J., & Griffiths, G. M. (2001). Trends in New Zealand daily temperature and rainfall extremes. International Journal of Climatology, 21(12), 1437–1452.

    Article  Google Scholar 

  10. 10.

    Mishra, V., Aaadhar, S., Shah, H., Kumar, R., Pattanaik, D. R., & Tiwari, A. D. (2018). The Kerala flood of 2018: combined impact of extreme rainfall and reservoir storage. Hydrology and Earth System Sciences Discussions, 1–13, 2018.

    Article  Google Scholar 

  11. 11.

    Mukherjee, S., Aadhar, S., Stone, D., & Mishra, V. (2018). Increase in extreme precipitation events under anthropogenic warming in India. Weather and Climate Extremes, 20, 45–53.

    Article  Google Scholar 

  12. 12.

    Revadekar, J. V., Patwardhan, S. K., & Rupa Kumar, K. (2011). Characteristic features of precipitation extremes over India in the warming scenarios. Advances in Meteorology.

    Article  Google Scholar 

  13. 13.

    Riswana, K. P., & Sithara Beegam, C. R. (2019). Tale of tears: development and flood experience of Chaliyar grama panchayath, Malappuram. A Journal of Composition Theory.

  14. 14.

    Agana, N. A., Sefidmazgi, M. G., & Homaifar, A. (2015). Analysis of Nonstationary Extreme Events. MAICS, pp. 7–11.

  15. 15.

    Davison, A. C., & Smith, R. L. (1990). Models for exceedances over high thresholds. Journal of the Royal Statistical Society: Series B (Methodological)52(3), 393–425.

  16. 16.

    Gumbel, E. J. (1941). The return period of flood flows. The Annals of Mathematical Statistics12(2), 163–190.

  17. 17.

    Jenkinson, A. F. (1955). The frequency distribution of the annual maximum (or minimum) values of meteorological elements. Quarterly Journal of the Royal Meteorological Society, 81(348), 158–171.

    Article  Google Scholar 

  18. 18.

    Singh, V. (1998). Entropy-based parameter estimation in hydrology (Vol. 30). Springer Science & Business Media.

  19. 19.

    Katz, R. W., Parlange, M. B., & Naveau, P. (2002). Statistics of extremes in hydrology. Advances in Water Resources, 25(8), 1287–1304.

    Article  Google Scholar 

  20. 20.

    Bobée, B., & Rasmussen, P. F. (1995). Recent advances in flood frequency analysis. Reviews of Geophysics, 33(S2), 1111–1116.

    Article  Google Scholar 

  21. 21.

    Cunnane, C. (1989). Statistical distributions for flood frequency analysis. Operational Hydrology Report (WMO).

  22. 22.

    Volpi, E., Fiori, A., Grimaldi, S., Lombardo, F., & Koutsoyiannis, D. (2015). One hundred years of return period: Strengths and limitations. Water Resources Research, 51(10), 8570–8585.

    Article  Google Scholar 

  23. 23.

    Leadbetter, M. R., Lindgren, G., & Rootzén, H. (2012). Extremes and related properties of random sequences and processes. Springer Science & Business Media.

    Article  Google Scholar 

  24. 24.

    Watson, G. S. (1954). Extreme values in samples from-dependent stationary stochastic processes. The Annals of Mathematical Statistics25(4), 798–800.

  25. 25.

    Cheng, L., AghaKouchak, A., Gilleland, E., & Katz, R. W. (2014). Non-stationary extreme value analysis in a changing climate. Climatic Change, 127(2), 353–369.

    Article  Google Scholar 

  26. 26.

    Milly, P. C., Betancourt, J., Falkenmark, M., Hirsch, R. M., Kundzewicz, Z. W., Lettenmaier, D. P., & Stouffer, R. J. (2008). Stationarity is dead: Whither water management? Science, 319(5863), 573–574.

    CAS  Article  Google Scholar 

  27. 27.

    Kundzewicz, Z. W., & Robson, A. J. (2004). Change detection in hydrological records-a review of the methodology/revue méthodologique de la détection de changements dans les chroniques hydrologiques. Hydrological Sciences Journal, 49(1), 7–19.

    Article  Google Scholar 

  28. 28.

    Agilan, V., & Umamahesh, N. V. (2017). What are the best covariates for developing non-stationary rainfall intensity-duration-frequency relationship? Advances in Water Resources, 101, 11–22.

    Article  Google Scholar 

  29. 29.

    Prosdocimi, I., Kjeldsen, T. R., & Miller, J. D. (2015). Detection and attribution of urbanization effect on flood extremes using nonstationary flood-frequency models. Water Resources Research, 51(6), 4244–4262.

    CAS  Article  Google Scholar 

  30. 30.

    Salas, J. D., & Obeysekera, J. (2013). Revisiting the concepts of return period and risk for nonstationary hydrologic extreme events. Journal of Hydrologic Engineering, 19(3), 554–568.

    Article  Google Scholar 

  31. 31.

    Šraj, M., Viglione, A., Parajka, J., & Blöschl, G. (2016). The influence of non-stationarity in extreme hydrological events on flood frequency estimation. Journal of Hydrology and Hydromechanics, 64(4), 426–437.

    Article  Google Scholar 

  32. 32.

    Strupczewski, W. G., Singh, V. P., & Feluch, W. (2001). Non-stationary approach to at-site flood frequency modelling I. Maximum likelihood estimation. Journal of Hydrology, 248(1), 123–142.

    Article  Google Scholar 

  33. 33.

    Mondal, A., & Mujumdar, P. P. (2016). Detection of change in flood return levels under global warming. Journal of Hydrologic Engineering, 21(8), 04016021.

    Article  Google Scholar 

  34. 34.

    Vogel, R. M., Yaindl, C., & Walter, M. (2011). Nonstationarity: flood magnification and recurrence reduction factors in the United States 1. JAWRA Journal of the American Water Resources Association, 47(3), 464–474.

    Article  Google Scholar 

  35. 35.

    Bayazit, M. (2015). Nonstationarity of hydrological records and recent trends in trend analysis: a state-of-the-art review. Environmental Processes, 2(3), 527–542.

    Article  Google Scholar 

  36. 36.

    Cooley, D. (2013). Return periods and return levels under climate change. Extremes in a Changing Climate (pp. 97–114). Dordrecht: Springer.

  37. 37.

    Tramblay, Y., Neppel, L., Carreau, J., & Sanchez-Gomez, E. (2012). Extreme value modelling of daily areal rainfall over Mediterranean catchments in a changing climate. Hydrological Processes, 26(25), 3934–3944.

    Article  Google Scholar 

  38. 38.

    Yilmaz, A. G., Hossain, I., & Perera, B. J. C. (2014). Effect of climate change and variability on extreme rainfall intensity–frequency–duration relationships: a case study of Melbourne. Hydrology and Earth System Sciences, 18(10), 4065–4076.

    Article  Google Scholar 

  39. 39.

    Khaliq, M. N., Ouarda, T. B. M. J., Ondo, J. C., Gachon, P., & Bobée, B. (2006). Frequency analysis of a sequence of dependent and/or non-stationary hydro- meteorological observations: A review. Journal of Hydrology, 329(3), 534–552.

    Article  Google Scholar 

  40. 40.

    Agilan, V., & Umamahesh, N. V. (2018). Covariate and parameter uncertainty in non-stationary rainfall IDF curve. International Journal of Climatology, 38(1), 365–383.

    Article  Google Scholar 

  41. 41.

    Villarini, G., Smith, J. A., Serinaldi, F., Bales, J., Bates, P. D., & Krajewski, W. F. (2009). Flood frequency analysis for nonstationary annual peak records in an urban drainage basin. Advances in Water Resources, 32(8), 1255–1266.

    Article  Google Scholar 

  42. 42.

    Rigby, R. A., & Stasinopoulos, D. M. (2005). Generalized additive models for location, scale and shape. Journal of the Royal Statistical Society: Series C (Applied Statistics), 54(3), 507–554.

    Article  Google Scholar 

  43. 43.

    Cannon, A. J. (2010). A flexible nonlinear modelling framework for nonstationary generalized extreme value analysis in hydroclimatology. Hydrological Processes, 24(6), 673–685.

    Article  Google Scholar 

  44. 44.

    Agilan, V., & Umamahesh, N. V. (2017). Non-stationary rainfall intensity-duration-frequency relationship: a comparison between annual maximum and partial duration series. Water Resources Management, 31(6), 1825–1841.

    Article  Google Scholar 

  45. 45.

    Heo, J. H., & Salas, J. D. (1996). Estimation of quantiles and confidence intervals for the log-Gumbel distribution. Stochastic Hydrology and Hydraulics, 10(3), 187–207.

    Article  Google Scholar 

  46. 46.

    Michele, C. D., & Rosso, R. (2001). Uncertainty assessment of regionalized flood frequency estimates. Journal of Hydrologic Engineering, 6(6), 453–459.

    Article  Google Scholar 

  47. 47.

    Obeysekera, J., & Salas, J. D. (2013). Quantifying the uncertainty of design floods under nonstationary conditions. Journal of Hydrologic Engineering, 19(7), 1438–1446.

    Article  Google Scholar 

  48. 48.

    Obeysekera, J. T., & Salas, J. D. (2020). Hydrologic designs for extreme events under nonstationarity. Engineering Methods for Precipitation under a Changing Climate (pp. 63–82).

  49. 49.

    Serinaldi, F., & Kilsby, C. G. (2015). Stationarity is undead: Uncertainty dominates the distribution of extremes. Advances in Water Resources, 77, 17–36.

    Article  Google Scholar 

  50. 50.

    Das, J., & Umamahesh, N. V. (2018). Assessment of uncertainty in estimating future flood return levels under climate change. Natural Hazards, 93(1), 109–124.

    Article  Google Scholar 

  51. 51.

    Raneesh, K. Y., & Thampi, S. G. (2013). A simple semi-distributed hydrologic model to estimate groundwater recharge in a humid tropical basin. Water Resources Management, 27(5), 1517–1532.

    Article  Google Scholar 

  52. 52.

    Das, J., & Umamahesh, N. V. (2017). Uncertainty and nonstationarity in streamflow extremes under climate change scenarios over a river basin. Journal of Hydrologic Engineering, 22(10), 04017042.

    Article  Google Scholar 

  53. 53.

    Bi, D., Dix, M., Marsland, S. J., O'Farrell, S., Rashid, H., Uotila, P., Hirst, A. C., Kowalczyk, E. A., Golebiewski, M., Sullivan, A., Yan, H., Hannah, N., Franklin, C. N., Sun, Z., Vohralik, P. F., Watterson, I. G., Zhou, X., Fiedler, R., Collier, M., Ma, Y., Noonan, J. A., Stevens, L., Uhe, P., Zhu, H., Griffies, S. M., Hill, R., Harris, C., & Puri, K. (2013). The ACCESS coupled model: description, control climate and evaluation. Australian Meteorological and Oceanographic Journal, 63(1), 41–64.

  54. 54.

    Mcgregor, J. L., & Dix, M. R. (2001). The CSIRO conformal-cubic atmospheric GCM. IUTAM symposium on advances in mathematical modelling of atmosphere and ocean dynamics (pp. 197–202). Dordrecht: Springer.

  55. 55.

    Gent, P. R., Danabasoglu, G., Donner, L. J., Holland, M. M., Hunke, E. C., Jayne, S. R., et al. (2011). The community climate system model version 4. Journal of Climate, 24(19), 4973–4991.

    Article  Google Scholar 

  56. 56.

    Voldoire, A., Sanchez-Gomez, E., y Mélia, D. S., Decharme, B., Cassou, C., Sénési, S., Valcke, S., Beau, I., Alias, A., Chevallier, M. & Déqué, M. (2013). The CNRM-CM5.1 global climate model: description and basic evaluation. Climate Dynamics40(9–10), 2091–2121.

  57. 57.

    Giorgetta, M. A., Jungclaus, J., Reick, C. H., Legutke, S., Bader, J., Böttinger, M., et al. (2013). Climate and carbon cycle changes from 1850 to 2100 in MPI-ESM simulations for the Coupled Model Intercomparison Project phase 5. Journal of Advances in Modeling Earth Systems, 5(3), 572–597.

    Article  Google Scholar 

  58. 58.

    Bentsen, M., Bethke, I., Debernard, J. B., Iversen, T., Kirkevåg, A., Seland, Ø., et al. (2013). The Norwegian earth system model, NorESM1-M—Part 1: Description and basic evaluation of the physical climate. Geoscientific Model Development, 6(3), 687–720.

    Article  Google Scholar 

  59. 59.

    Arora, V. K., Scinocca, J. F., Boer, G. J., Christian, J. R., Denman, K. L., Flato, G. M., Kharin, V. V., Lee, W. G., & Merryfield, W. J. (2011). Carbon emission limits required to satisfy future representative concentration pathways of greenhouse gases. Geophysical Research Letters38(5).

  60. 60.

    Giorgi, F., Coppola, E., Solmon, F., Mariotti, L., Sylla, M. B., Bi, X., et al. (2012). RegCM4: model description and preliminary tests over multiple CORDEX domains. Climate Research, 52, 7–29.

    Article  Google Scholar 

  61. 61.

    Rotstayn, L. D., Jeffrey, S. J., Collier, M. A., Dravitzki, S. M., Hirst, A. C., Syktus, J. I., & Wong, K. K. (2012). Aerosol-and greenhouse gas-induced changes in summer rainfall and circulation in the Australasian region: a study using single-forcing climate simulations. Atmospheric Chemistry and Physics, 12(14), 6377.

    CAS  Article  Google Scholar 

  62. 62.

    Dufresne, J. L., Foujols, M. A., Denvil, S., Caubel, A., Marti, O., Aumont, O., et al. (2013). Climate change projections using the IPSL-CM5 Earth System Model: from CMIP3 to CMIP5. Climate Dynamics, 40(9–10), 2123–2165.

    Article  Google Scholar 

  63. 63.

    Jakob, D. (2013). Nonstationarity in extremes and engineering design. Extremes in a changing climate (pp. 363–417). Dordrecht: Springer.

  64. 64.

    Read, L. K., & Vogel, R. M. (2015). Reliability, return periods, and risk under nonstationarity. Water Resources Research, 51(8), 6381–6398.

    Article  Google Scholar 

  65. 65.

    Brown, C., Templin, J., & Cohen, A. (2015). Comparing the two-and three-parameter logistic models via likelihood ratio tests: A commonly misunderstood problem. Applied Psychological Measurement, 39(5), 335–348.

    Article  Google Scholar 

  66. 66.

    Mondal, A., & Mujumdar, P. P. (2017). Hydrologic extremes under climate change: Non-stationarity and uncertainty. Sustainable Water Resources Planning and Management Under Climate Change (pp. 39–60). Singapore: Springer.

  67. 67.

    Giorgi, F., & Mearns, L. O. (2003). Probability of regional climate change based on the Reliability Ensemble Averaging (REA) method. Geophysical Research Letters, 30(12).

  68. 68.

    Xu, Y., Gao, X., & Giorgi, F. (2010). Upgrades to the reliability ensemble averaging method for producing probabilistic climate change projections. Climate Research, 41(1), 61–81.

    Article  Google Scholar 

  69. 69.

    Riano, A. (2013). The Shift of Precipitation Maxima on the Annual Maximum Series using Regional Climate Model Precipitation Data. Arizona State University.

  70. 70.

    Dominguez, F., Cañon, J., & Valdes, J. (2010). IPCC-AR4 climate simulations for the Southwestern US: the importance of future ENSO projections. Climatic Change, 99(3–4), 499–514.

    Article  Google Scholar 

  71. 71.

    Ehret, U., Zehe, E., Wulfmeyer, V., Warrach-Sagi, K., & Liebert, J. (2012). HESS Opinions "Should we apply bias correction to global and regional climate model data?". Hydrology & Earth System Sciences Discussions9(4).

  72. 72.

    Switanek, M., Troch, P. A., Castro, C. L., Leuprecht, A., Chang, H. I., Mukherjee, R., & Demaria, E. M. (2017). Scaled distribution mapping: a bias correction method that preserves raw climate model projected changes. Hydrology and Earth System Sciences, 21(6), 2649–2666.

    Article  Google Scholar 

  73. 73.

    Danandeh Mehr, A., & Kahya, E. (2016). Climate change impacts on catchment-scale extreme rainfall variability: case study of Rize Province Turkey. Journal of Hydrologic Engineering, 22(3), 05016037.

    Article  Google Scholar 

  74. 74.

    Zhang, X., & Zwiers, F. W. (2013). Statistical indices for the diagnosing and detecting changes in extremes. Extremes in a Changing Climate (pp. 1–14). Dordrecht: Springer.

  75. 75.

    Clarke, R. T. (2002). Fitting and testing the significance of linear trends in Gumbel-distributed data. Hydrology and Earth System Sciences Discussions, 6(1), 17–24.

    Article  Google Scholar 

  76. 76.

    Katz, R. W. (2013). Statistical methods for nonstationary extremes. Extremes in a changing climate (pp. 15–37). Dordrecht: Springer.

  77. 77.

    Coles, S., Bawa, J., Trenner, L., & Dorazio, P. (2001). An introduction to statistical modeling of extreme values (Vol. 208). London: Springer.

    Google Scholar 

  78. 78.

    Westra, S., Alexander, L. V., & Zwiers, F. W. (2013). Global increasing trends in annual maximum daily precipitation. Journal of Climate, 26(11), 3904–3918.

    Article  Google Scholar 

  79. 79.

    Zwiers, F. W., Zhang, X., & Feng, Y. (2011). Anthropogenic influence on long return period daily temperature extremes at regional scales. Journal of Climate, 24(3), 881–892.

    Article  Google Scholar 

  80. 80.

    Zhang, X., Zwiers, F. W., & Li, G. (2004). Monte Carlo experiments on the detection of trends in extreme values. Journal of Climate, 17(10), 1945–1952.;2.

    Article  Google Scholar 

  81. 81.

    Sclove, S. L. (2011). A review of statistical model selection criteria: Application to prediction in regression, histograms, and finite mixture models. SSRN Electronic Journal.

    Article  Google Scholar 

  82. 82.

    Kadane, J. B., & Lazar, N. A. (2004). Methods and criteria for model selection. Journal of the American statistical Association, 99(465), 279–290.

    Article  Google Scholar 

  83. 83.

    Hastie, T., Tibshirani, R., & Friedman, J. (2009). Model assessment and selection. The elements of statistical learning (pp. 219–259). New York, NY: Springer.

  84. 84.

    Akaikei, H. (1973). Information theory and an extension of maximum likelihood principle. In Proc. 2nd Int. Symp. on Information Theory, pp. 267–281.

  85. 85.

    Burnham, K. P., & Anderson, D. R. (2004). Multimodel inference: understanding AIC and BIC in model selection. Sociological Methods & Research, 33(2), 261–304.

    Article  Google Scholar 

  86. 86.

    Lewis, F., Butler, A., & Gilbert, L. (2011). A unified approach to model selection using the likelihood ratio test. Methods in Ecology and Evolution, 2(2), 155–162.

    Article  Google Scholar 

  87. 87.

    Greene, W. H. (2003). Econometric Analysis. Pearson Education India.

  88. 88.

    Pol, D. (2004). Empirical problems of the hierarchical likelihood ratio test for model selection. Systematic Biology, 53(6), 949–962.

    Article  Google Scholar 

  89. 89.

    Mondal, A., & Daniel, D. (2018). Return levels under nonstationarity: The need to update infrastructure design strategies. Journal of Hydrologic Engineering, 24(1), 04018060.

    Article  Google Scholar 

  90. 90.

    Parey, S., Hoang, T. T. H., & Dacunha-Castelle, D. (2010). Different ways to compute temperature return levels in the climate change context. Environmetrics, 21(7–8), 698–718.

    Article  Google Scholar 

  91. 91.

    Gao, M. (2018). Extreme value analysis and risk communication for a changing climate. Advances in Environmental Monitoring and Assessment. IntechOpen.

  92. 92.

    Efron, B. (1979). Bootstrap? Another look at Jackknife. Annals of Statistics, 7, 1–26.

    Article  Google Scholar 

  93. 93.

    Al Mamoon, A., & Rahman, A. (2014). Uncertainty in design rainfall estimation: A review. Journal of Hydrology and Environment Research, 2(1), 65–75.

    Google Scholar 

  94. 94.

    Kyselý, J. (2008). A cautionary note on the use of nonparametric bootstrap for estimating uncertainties in extreme-value models. Journal of Applied Meteorology and Climatology, 47(12), 3236–3332.

    Article  Google Scholar 

  95. 95.

    Zucchini, W., & Adamson, P. T. (1989). Bootstrap confidence intervals for design storms from exceedance series. Hydrological Sciences Journal, 34(1), 41–48.

    Article  Google Scholar 

  96. 96.

    Ning, L., Riddle, E. E., & Bradley, R. S. (2015). Projected changes in climate extremes over the northeastern United States. Journal of Climate, 28(8), 3289–3310.

    Article  Google Scholar 

  97. 97.

    Xu, K., Xu, B., Ju, J., Wu, C., Dai, H., & Hu, B. X. (2019). Projection and uncertainty of precipitation extremes in the CMIP5 multimodel ensembles over nine major basins in China. Atmospheric Research, 226, 122–137.

    Article  Google Scholar 

Download references

Author information




The manuscript was prepared based on the discussion between Ms. Ansa Thasneem S, Dr. Chithra N R, and Dr. Santosh G Thampi. Ms. Ansa Thasneem S carried out all the numerical simulations included in the manuscript. The manuscript was prepared by Ms. Ansa Thasneem S, and this was reviewed and corrected by Dr. Santosh G Thampi and Dr. Chithra N R.

Corresponding author

Correspondence to S. Ansa Thasneem.

Ethics declarations

Conflict of Interest

The authors declare that there is no conflict of interest.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Ansa Thasneem, S., Chithra, N.R. & Thampi, S.G. Assessment of Nonstationarity and Uncertainty in Precipitation Extremes of a River Basin Under Climate Change. Environ Model Assess (2021).

Download citation


  • Climate change
  • Nonstationarity
  • Extreme precipitation
  • Uncertainty