See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/374790056

Nonstationary Flood Frequency Analysis: Review of Methods and Models

Chapter · October 2023

DOI: 10.1007/978-981-99-4811-6_15

CITATIONS READS

0 176

3 authors:

Siddik Barbhuiya Meenu Ramadas

5 PUBLICATIONS 1 CITATION
Indian Institute of Technology Bhubaneswar
41 PUBLICATIONS 542 CITATIONS
SEE PROFILE
SEE PROFILE

Shanti Swarup Biswal

Indian Institute of Technology Bhubaneswar
11 PUBLICATIONS 115 CITATIONS

SEE PROFILE

All content following this page was uploaded by Siddik Barbhuiya on 19 October 2023.

The user has requested enhancement of the downloaded file.

Chapter 15
Nonstationary Flood Frequency Analysis:
Review of Methods and Models

Siddik Barbhuiya, Meenu Ramadas, and Shanti Swarup Biswal

Abstract Recent changes in the climate, land use/land cover, and field-scale water
resources allocation at the catchment scale have rendered the conventional hypoth-
esis of the stationarity of hydrologic extremes unreliable. The current understanding
of evolving patterns of hydrological variables has led to the development of
nonstationary approaches, particularly in extreme event frequency analysis. A com-
prehensive review of the different approaches for nonstationary flood frequency
analysis is presented in this chapter. The popular methods including generalized
additive models for location, scale, and shape (GAMLSS) framework; probability-
based approaches using Gumbel distribution and Log Pearson distribution III (LP 3),
Bayesian approaches, r-largest, peaks-over-threshold, time-varying moments;
among others are discussed. Additionally, the challenges associated with
nonstationary hydrological frequency analysis and future research directions in the
analysis of flood extremes are briefly addressed. It is evident that nonstationarity
needs to be incorporated in flood risk assessment framework for addressing the
likely impacts of potential future climate change in water resources management.

Keywords Flood frequency analysis · Climate change · Extreme events ·

Nonstationarity · Bayesian approach

15.1 Introduction

Among the fatal natural hazards, water resource-related extreme events such as
floods, cyclones, and droughts have been known to be the costliest and most
disastrous, across different parts of the world. Naturally occurring riverine flood
and flash flood events are characterized by overflowing of rivers into the riverbanks
as a result of heavy precipitation, inundating large areal extent comprising human
settlements and natural ecosystems (Merz et al. 2021; Das et al. 2022). In fact, the

S. Barbhuiya · M. Ramadas (✉) · S. S. Biswal

School of Infrastructure, Indian Institute of Technology, Bhubaneswar, Odisha, India
e-mail: [email protected]; [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 271
M. Pandey et al. (eds.), River, Sediment and Hydrological Extremes: Causes,
Impacts and Management, Disaster Resilience and Green Growth,
https://doi.org/10.1007/978-981-99-4811-6_15
272 S. Barbhuiya et al.

extreme events are also caused by anthropogenic influences such as urban floods,
dam break, and coastal floods. Disastrous floods can leave unprecedented impacts on
society and destruction of lives and also cause disruption to economic activities in
the region (Mudelsee et al. 2003; Ray et al. 2019; Kuang and Liao 2020; Mangukiya
and Sharma 2022; and Das et al. 2022 among others). In India, a significant portion
of land, encompassing over 40 million hectares or 12% of the geographical area, is
observed to be susceptible to flood events. Each year, floods claim the lives of more
than 1600 people and inflict damages exceeding Rs. 5600 crores (73 million USD)
(Central Water Commission 2018). Flood frequency analysis (FFA) has become an
indispensable tool in assessing the potential impacts of these flood hazards and
designing effective mitigation measures. In this regard, the term frequency or return
period that expresses the exceedance probability of the flood event is used. Through
FFA, the relationship among flood peaks, volumes, duration, and the associated
return periods can be assessed, utilizing continuous long-term data of observed flow
discharge or water levels in the river. In recent years, the phenomena of climate
change, land use/land cover change, and water resource reallocation and different
watershed-scale interventions have challenged the notion of stationarity of hydro-
logical variables that is adopted in frequency analysis and extreme event modeling
(Berghuijs et al. 2019; Milly et al. 2008; Villarini et al. 2009; Debele et al. 2017a).
Significant rise in riverine flood hazard is projected for parts of sub-Saharan Africa,
Asia, Europe, northern Russia, and specific regions in South and North America in
future periods (Merz et al. 2021). Incidentally, the trends and change patterns of
hydrological variables have been studied by numerous researchers in the form of
detection and attribution studies and encouraged the choice of nonstationary
approaches for modeling the changing risk of hydrological extremes such as floods
and droughts (Mondal and Mujumdar 2012; Serinaldi and Kilsby 2015; Singh and
Chinnasamy 2021). Therefore, these methods are increasingly being adopted for
estimation of return period and risk associated with riverine flood hazards at local-to-
regional scales (Lima et al. 2015; Mondal and Daniel 2019).
FFA methods accounting for nonstationarity have been applied to various case
studies worldwide, demonstrating their usefulness in understanding and predicting
the behavior of hydrological extremes in changing environments. The analysis
includes finding the best estimates of the time-varying parameters of probability
distributions that fit the flood variables (peak, volume, duration), using covariates
such as time, temperature, or any suitable hydroclimatic variable. The popular
methods and models adopted in nonstationary hydrologic frequency analysis include
generalized additive models for location, scale, and shape (GAMLSS) framework,
probability distribution-based models (Gumbel distribution, Log Pearson distribu-
tion III), Bayesian approaches, r-largest, peaks-over-threshold, time-varying
moments, pooled FFA, local likelihood, and quantile regression. In one of the earlier
studies by Strupczewski et al. (2001), the need to incorporate trend for accurate
analysis of flood frequency is established, wherein the temporal trends in hydrolog-
ical variables implying nonstationarity were investigated for FFA. They used annual
peak discharge series, applying both the annual maximum series and partial duration
series-based approaches for at-site frequency modeling. They had relied on the
15 Nonstationary Flood Frequency Analysis: Review of Methods and Models 273

Akaike information criterion (AIC) to identify the best nonstationary model among
different models, while the maximum likelihood method was used for model
parameter determination. The advancements in hydrological modeling over the last
few decades include use of Bayesian approaches (Cheng et al. 2014; Sharma and
Goyal 2017), as well as applications for regional flood frequency analysis, in the
context of nonstationary analysis. The application of nonstationary analysis of
extreme precipitation events in Mediterranean region is found in Tramblay et al.
(2013), where a nonstationary peaks-over-threshold model was utilized with climatic
variables as covariates. The authors used Poisson distribution and generalized Pareto
distribution for modeling the occurrence and magnitude of heavy rainfall events,
respectively, while the southern circulation patterns and monthly air temperature
were adopted as covariates. They found that the nonstationary model with climatic
covariates performed better than the classical stationary model and could simulate
future climate scenarios for understanding impacts of such changes in future. The
potential future changes in the covariates included in the model were also used to
evaluate the possible future changes in extreme precipitation events in the study area.
Das and Umamahesh (2017) had analyzed uncertainties and nonstationarity in future
streamflow projections at river basin scale under climate change scenarios: repre-
sentative concentration pathways (RCPs) 4.5 and 8.5 using the VIC-3 L model, they
and found that while stationary models were suitable for RCP4.5, nonstationary
approach was more appropriate for RCP8.5. Further, their study suggested that
nonstationary return levels were reliable for designing low-capacity hydraulic struc-
tures and highlights the role of nonstationarity in improved hydrologic modeling and
design. Similarly, a review of various return level-based metrics for hydrologic
design under nonstationary conditions can be found in Mondal and Daniel (2019).
Das and Umamahesh (2022) also studied the hydrological extremes in the Godavari
River basin, India, incorporating physically based covariates such as the Indian
Summer Monsoon Index and precipitation into the Generalized Extreme Value
distribution to incorporate nonstationarity.
Comparison of stationary and nonstationary flood frequency approaches can
provide useful insights on the drawbacks of stationary models that do not address
climatic change. Currently, we can find numerous studies that use external covariates
besides time, to improve results of FFA. In their study, Machado et al. (2015)
analyzed historical flood records of the Tagus River in Spain and found that the
estimates of extreme event magnitudes and frequencies are better modeled by
including covariates of various climate and environmental drivers. They compared
both stationary and nonstationary models including a GAMLSS model that incor-
porated both climate and catchment factors. The norming constants method (NCM)
was adopted for nonstationary FFA of flow in the Wei River in China, by Xiong et al.
(2015). In their study, the nonstationarity present in annual daily flow series and their
effect on the annual maximum flood series were modeled. The authors had consid-
ered nonstationarity using additional explanatory climatic variables, tested the NCM
on the flow data, and found that it outperformed the traditional stationary FFA
models. Šraj et al. (2016) also compared among four different models for estimating
flood quantiles at gauging sites in Slovenia. With significantly increasing trend in
274 S. Barbhuiya et al.

annual maximum discharge series at these sites, nonstationarity-based analysis of

extremes was imperative. In their study, Šraj et al. (2016) used the GEV distribution,
with parameters dependent on time and annual precipitation as covariates. The
maximum likelihood and Bayesian Monte Carlo Markov chain methods were used
for parameter estimation. Comparison of results of FFA using GAMLSS method
with different types of covariates is performed by Dègan et al. (2017), creating
nonstationary models with time and principal components obtained from empirical
orthogonal factor (EOF) analysis on climate variables (climate indices and temper-
ature). Their study found the nonstationary model using principal components as
covariates to be better at modeling change. These studies overall suggest that
nonstationarity is important in the context of FFA, and significant difference exists
between nonstationary and stationary estimates, with likely underestimation in case
of the latter approach.
There are numerous instances when nonstationary approach has reduced the
uncertainty involved in hydrologic design, specifically with advancements in Bayes-
ian model-based analysis. It is possible that stationarity assumption could lead to
unsafe designs for structures when trends are present in hydrologic variables, and
potential variations in these trends also need attention of hydrologists. Singh and
Chinnasamy (2021) performed nonstationary FFA of discharge of Periyar River in
India using generalized extreme value distribution and covariates of annual precip-
itation, urban extent, and time. When compared with stationary analysis results, they
observed that the trends in flows are present, and it is advantageous to adopt
nonstationary FFA. A study by Guo et al. (2023) has compared the performance
of nonstationary Bayesian regional flood frequency analysis (RFFA) coupled with
the linear mixed effect (LME) model with the stationary generalized least squares
(GLS) model. Indeed, the nonstationary LME-based Bayesian RFFA method
performed better than the stationary GLS-based method with respect to the deviance
information criterion (DIC). The nonstationary approach reduced uncertainty in
design flood estimation and has been recommended for nonstationary FFA of
ungauged sites.
The presented review of applications of nonstationary hydrologic frequency
analysis suggests that incorporating nonstationarity in FFA is particularly relevant
in the current scenario with possible future alterations in climatic variables and
hydrologic processes. The studies also emphasize the choice of relevant covariate
that drives the changes in these variables. There are several challenges associated
with nonstationary FFA that are evident from the review of literature. Discerning
among the natural, anthropogenic, and mixed drivers of nonstationarity is problem-
atic, as these factors may be interlinked in complex ways, and requires substantial
efforts for detecting trends and change points with limited data records. In fact, short-
term trends and multi-decadal shifts in hydrological variables may complicate
nonstationary analysis (Koutsoyiannis and Montanari 2015). Besides, the inadver-
tent use of nonstationary approach to shorter-term time series can lead to increased
uncertainty (Serinaldi and Kilsby 2015).
This chapter aims at a comprehensive review of important nonstationary FFA
approaches and their applications, challenges, and future research direction. The
15 Nonstationary Flood Frequency Analysis: Review of Methods and Models 275

popular GAMLSS framework, probability distribution-based models (Gumbel dis-

tribution, Log Pearson distribution III), Bayesian approaches, r-largest, peaks-over-
threshold, and other methods: time-varying moments, pooled FFA, local likelihood,
and quantile regression are included in the review. The objective of the review is to
provide direction to hydrologists regarding the emerging methods in nonstationary
hydrologic frequency analysis and their relative merits and demerits. With increasing
focus on nonstationarity approaches in the face of changing climate, land use, and
anthropogenic interventions, there is perennial need for developing more robust and
flexible approaches that can be adapted across spatial scales too. It is also important
to devise strategies for integrating the findings of nonstationary analysis into risk
assessment and adaptation, so that the benefits of planning and management of water
resources can be maximized, ensuring that communities are better prepared and
more resilient to future hydrologic extreme events (Chen et al. 2021; Zhou et al.
2022).

15.2 Nonstationary Frequency Analysis Approaches

15.2.1 Generalized Additive Models for Location, Scale,

and Shape (GAMLSS) Framework

15.2.1.1 Theory

The GAMLSS framework (Rigby et al. 2005) offers a flexible statistical framework
for the estimation of distribution parameters of flood variables as functions of
covariates, under assumptions of nonstationarity. The GAMLSS framework has
been widely used to model different hydrological variables, such as precipitation,
temperature, and streamflow (Archfield et al. 2016; Westra et al. 2014). In the
GAMLSS framework, the response variable y is assumed to follow a specific
distribution with parameters μ, σ, ν, τ (for location, scale, shape, and additional
shape, respectively). Each parameter is modeled as a function of covariates (Rigby
et al. 2005). The general GAMLSS model can be written as:

pk
g k ðθ k Þ = X k β k þ sik ðxik Þ, for k = 1, 2, . . . , K ð15:1Þ
i=1

where y follows a specific distribution with parameter vector θ = (μ, σ, ν, τ), gk is a

link function for the k-th (k = 1, 2, . . ., K ) parameter, θk is the k-th parameter of the
distribution (location, shape, or additional shape), Xk is known model matrix for the
k-th parameter, βk is the parameter vector for the k-th parameter, sik(xik) is the smooth
function of the covariance xik for the k-th parameter, and pk is the number of
covariates for the k-th parameter.
276 S. Barbhuiya et al.

The GAMLSS model can accommodate different choices of probability distribu-

tion and link functions. The link functions gk transform the distribution parameters to
a scale on which the linear predictors can be modeled. The choice of link function
depends on the range and interpretation of the distribution parameters. The smooth
sik(xik) functions capture nonlinear relationships between the covariates and the
distribution parameters (Wood 2006). These functions can be spline function or
local regression function. Generally, in GAMLSS framework, the distribution
parameters and smooth functions are estimated by maximizing a penalized likeli-
hood function. This maximization process ensures the goodness of fit of the model
(Rigby et al. 2005).

15.2.1.2 Scope

Debele et al. (2017a, 2017b) provide a detailed process of use of GAMLSS in flood
frequency studies. The challenges of nonstationarity-based approach are discussed in
Debele et al. (2017a), and they also suggest the GAMLSS framework as the most
popular for nonstationary statistical analysis. The relative advantages and weak-
nesses are also presented in their work. A comparison of three methods—maximum
likelihood, weighted least squares-two stage (WLS/TS), and GAMLSS for compu-
tation of design flood quantiles under nonstationarity was performed for Polish and
Norwegian catchments by Debele et al. (2017b). Their study recommended using a
multi-model approach to minimize errors associated with model formulation across
different length datasets. While GAMLSS performed best in overall estimation of
design flood quantiles with longer datasets, the WLS/TS provides better accuracy for
shorter time series analysis. In another study, Chen et al. (2021) did flood frequency
analysis for several gauging stations in the United Kingdom and examined the
choice of covariates. The nine covariates chosen included rainfall variability-related
and atmospheric circulation pattern-based variables, to model the inherent
nonstationarity in flood records. Their study found that the simplest choice—a
time-varying nonstationary flood model may not always be the most appropriate,
and physically based covariates can offer better nonstationary models. Even the use
of multiple covariates is recommended to improve the analysis, to simulate the
effects of climate change. A framework for assessing the uncertainty of
nonstationary FFA with possible application in hydrologic design, water supply,
and reservoir regulation was proposed by Zhou et al. (2022), for the Hanjiang River
of China. They combined GAMLSS, copula model, and Bayesian uncertainty
processor (BUP) techniques for design flood estimation with information on uncer-
tainty of the estimates. In this study, precipitation and reservoir index were
covariates. Their framework addresses the modeling of uncertainty in
nonstationary FFA.
15 Nonstationary Flood Frequency Analysis: Review of Methods and Models 277

15.2.2 FFA Using Probability Distributions

GEV distribution is widely used for modeling extreme events such as floods,
droughts, and other hydroclimatic extremes. The GEV family includes three types
of extreme value distributions, namely, the Gumbel, Frechet, and Weibull distribu-
tions. The Gumbel distribution has been widely used to model extreme event
characteristics in a nonstationary context. Log Pearson distribution III (LP 3) is
another popular choice of distribution for flood frequency analysis, which considers
the skewness, kurtosis, and other moments of the variable data (Stedinger
et al. 1993).

15.2.2.1 Generalized Extreme Value (GEV) Distribution

The cumulative distribution function (CDF) of the GEV distribution is given by

Coles (2001):

x-μ - 1ξ
F ðxÞ = exp - 1 þ ξ  ð15:2Þ
σ

where x is the random variable of interest, μ is the location parameter, σ is the scale
parameter, and ξ is the shape parameter. In a nonstationary context, any of the
parameters can be modeled as functions of time or other covariates. For example,
suppose we have a covariate, t, representing time. We can model the location and
scale parameters as functions of time in a linear fashion:

μðt Þ = μ0 þ μ1 t
σ ðt Þ = σ 0 þ σ 1 t ð15:3Þ
ξðt Þ = ξ 0 þ ξ 1 t

where μ0, μ1, σ 0, σ 1, ξ0, and ξ1 are the regression coefficients to be estimated.

15.2.2.2 Gumbel Distribution

The Gumbel distribution probability density function (PDF) f(x) and cumulative
distribution function (CDF) F(x) are as follows:

1 x-μ x-μ
f ðxÞ = exp⁡½ - ð Þ - exp⁡ð - Þ; FðxÞ
σ σ σ
x-μ
= exp⁡½ - exp⁡ð - Þ ð15:4Þ
σ
278 S. Barbhuiya et al.

where x is the random variable, μ is the location parameter, and σ is the scale
parameter. Incorporating the time-varying parameters μt, σ t previously discussed
into these PDF and CDF expressions results in:

1 x - μðt Þ x - μðt Þ
f ðx, t Þ = exp - - exp - ; F ðx, t Þ
σ ðt Þ σ ðt Þ σ ðt Þ
x - μ ðt Þ
= exp - exp - ð15:5Þ
σ ðt Þ

This approach allows to account for nonstationarity in the data, provided the
correct estimates of the parameters are obtained through the analysis.

15.2.2.3 Log Pearson Distribution III (LP 3)

The LP 3 distribution is specified by its parameters: location (μ), scale (σ), and shape
(ξ). In this approach, the goal is to fit a Pearson type III distribution to the base-10
logarithms of the annual flood maxima. The parameters estimated are mean μ,
standard deviation σ, and skew coefficient ξ of the log-transformed data (Griffis
and Stedinger 2007). By setting parameters μ, σ, and ξ as functions of time or other
covariates as previously discussed, it is possible to capture the changes in flood
magnitudes and frequencies over time (Vogel and Wilson 1996). By incorporating
time-varying parameters, the LP 3 distribution can better represent the nonstationary
behavior of hydrological variables, providing more accurate flood frequency esti-
mates in the context of changing conditions.

15.2.2.4 Scope

An important step of FFA is the choice of best distribution to represent the charac-
teristics of the extreme events. There are numerous probability distributions that can
be adopted and depend on the data and results of goodness of fit tests. In the study by
Gruss et al. (2022) that performed annual maxima FFA for ten rivers in the Czech
Republic and Poland, they compared three-parameter distributions including the
log-normal, Weibull, generalized extreme value (GEV), and Pearson type III distri-
butions. Their methodology was flexible as it allowed for the choice of best fit
distribution for FFA. Nonstationary FFA models were found to be superior over
stationary models for flood estimation, and among the distributions, Weibull and
log-Normal distributions were found to be the most suitable for lower and upper
quantiles, respectively.
15 Nonstationary Flood Frequency Analysis: Review of Methods and Models 279

15.2.3 Bayesian Models

15.2.3.1 Theory

Bayesian model framework has been employed to perform robust nonstationary FFA
by integrating prior knowledge, uncertainty, and model updating based on new data
(Khaliq et al. 2006; Salas et al. 2018). The analysis is used to estimate the posterior
distribution of the parameters of the probability distributions and the time-varying
functions. The framework is best suitable to model uncertainties related to
hydrological FFA.
The posterior distribution of the parameters can be expressed as (Salas et al.
2018):

pðθjyÞ / pðyjθÞ  pðθÞ ð15:6Þ

where p(θ| y) is the posterior distribution, p(y| θ) is the likelihood function that
represents the probability of observing the data y given the parameters θ, p(θ) is
the prior distribution that encodes our knowledge about the parameters before
observing the data, and y denotes the observed data.

15.2.3.2 Scope

Bayesian approaches have been successfully utilized to model distributions in flood

frequency models involving GAMLSS, Gumbel, and LP 3 distributions (Khaliq
et al. 2006). This allows for the incorporation of expert knowledge and historical
information into the modeling process, which can lead to more accurate and reliable
predictions of extreme events. In their FFA study, Lima and Lall (2010) demon-
strated that the Hierarchical Bayesian models can provide accurate estimates of
monthly/annual flood discharge probability distribution parameters for ungauged
sites in Brazil. They performed nonstationary model development using
reconstructed natural inflow series from over 40 gauging points representing a
wider range of catchment areas. In fact, the developed regional-scale approach was
used for augmenting records of flow at sites with missing data as well as to estimate
flow at ungauged sites. Similarly, studies by Ouarda and El (2011) detail the scope of
nonstationary frequency analysis models in hydrology, with a specific focus on the
Bayesian approaches. They had used the nonstationary generalized maximum like-
lihood estimation method and the reversible jump Monte Carlo Markov Chain
(MCMC) model for hydrologic frequency analysis.
280 S. Barbhuiya et al.

15.2.4 Other Methods

Few other popular models used in nonstationary FFA are discussed in this section.
The r-largest and peaks-over-threshold (POT) approaches have been useful for
modeling nonstationary extreme events, using annual maxima as well as POT data
(Douglas et al. 2000; Mudelsee et al. 2003).
The r-largest approach is a method for modeling extreme events by fitting a
distribution to the r largest order statistics within a specified period (Douglas et al.
2000). This technique focuses on modeling the most extreme floods and provides a
means of estimating the return levels associated with rare discharge levels. The
peaks-over-threshold approach, on the other hand, models exceedances over a
predefined threshold u using the generalized Pareto distribution (GPD) (Mudelsee
et al. 2003). The GPD is described by its f(x) and F(x) as follows:

1 x-u - ð1þξÞ - 1 x-u - ð1þξÞ - 1

f ðxÞ =  1þξ ; F ðxÞ = 1 - 1 þ ξ ð15:7Þ
σ σ σ

where the symbols as defined previously. Both the r-largest and POT approaches can
be used in conjunction with Bayesian approach for analysis of floods with due
consideration for uncertainty assessment.
Time-varying moments, pooled flood frequency analysis, local likelihood, and
quantile regression also offer flexibility in modeling extreme events, under climate
change, land use change, and human interventions, but are less popular methods for
nonstationary flood frequency analysis (Hejazi and Markus 2009; Villarini et al.
2009; Vogel et al. 2011; Wilson et al. 2010).
Time-varying moments provide a method for estimating the moments of the
underlying distribution as functions of time. Strupczewski et al. (2001) used the
nonstationary approach by incorporating temporal trend in the first two moments of
the distributions: mean and variance. The time-varying mean μt and variance σ 2t were
modeled using smooth functions such as splines or linear regression.
Pooled flood frequency analysis is a method that combines data from multiple
sites to improve the estimation of flood quantiles (Vogel et al. 2011). The pooling
group is determined by a similarity criterion, such as geographical proximity or
hydrological similarity. Flood quantiles are then estimated using a weighted average
of the site-specific quantiles.
Local likelihood is a nonparametric approach that estimates the distribution
parameters using a weighted likelihood function (Khaliq et al. 2006).

Lð θ Þ = ½f ðyi j θðxi ÞÞwðxi ,xÞ ð15:8Þ

i

where L(θ) is the likelihood function, f(yi| θ(xi)) is the probability density function,
and w(xi, x) are the weights. The weights depend on the distance between the xi
observation and the point of interest x, typically following a kernel function.
15 Nonstationary Flood Frequency Analysis: Review of Methods and Models 281

Quantile regression models the quantiles of the response variable as a function of

covariates, allowing for the estimation of flood quantiles under nonstationary con-
ditions (Khaliq et al. 2006). The quantile regression model is specified as follows:

Qy ðτjX Þ = X βðτÞ ð15:9Þ

where Qy(τ| X) is the conditional quantile function, τ is the quantile level, X is the
matrix of covariates, and β(τ) are the quantile-specific coefficients.

15.3 Case Studies

15.3.1 Review of Literature

An exhaustive list of case studies that have contributed to the understanding and
development of nonstationary flood frequency analysis is presented in Table 15.1.
The compiled information showcases the evolution of the modeling approaches and
the relative popularity of models and highlights the growing importance of
nonstationary approaches in flood frequency analysis.

15.3.2 Application: Case Study of Barmanghat Subbasin,

India

We present findings of case study of nonstationary and stationary FFA of discharge

data of Barmanghat subbasin of the Narmada River Basin in this section. The basin
area is predominantly agricultural land. In this case, we performed FFA on the
annual maximum streamflow data using a temporal nonstationary GEV model.
The location and scale parameters of the GEV distributions are modeled as linear
functions of the time covariate, while the shape parameter remains constant. Normal
priors are utilized for parameter estimation in this case study. This analysis is
implemented using the nonstationary extreme value analysis (NEVA) software
package developed by Cheng et al. (2014). This analysis provides posterior proba-
bilities with uncertainty interval during estimation of return levels and utilizes
Bayesian approach for parameter estimation of models.
Results presented in Fig. 15.1a show the flood discharge versus return level
curves obtained from the FFA for time covariate value of 45 years from the initial
observation (as an example). The Akaike Information Criterion (AIC) and Bayesian
Information Criterion (BIC) performance metrics used to evaluate model efficiency
both suggest that nonstationarity approach is best for streamflow analysis of
Barmanghat. With an AIC value of 499.43, the nonstationary model is significantly
better than the stationary model with AIC of 505.76. Similarly, the BIC value for the
282 S. Barbhuiya et al.

Table 15.1 Overview of popular nonstationary flood frequency analysis studies

Author(s) Objective Methodology Findings
Strupczewski Investigated time trends in Maximum likelihood Trend incorporation nec-
et al. (2001) hydrological method essary for accurate
nonstationarity modeling
Lima and Lall Developed hierarchical Hierarchical Bayesian Accurate estimation of
(2010) Bayesian models for models parameters; useful for
regional and at-site trends record augmentation
Ouarda and Discussed nonstationary Nonstationary GMLE Bayesian model efficient
El (2011) frequency analysis method and reversible for hydrological quantiles
models jump MCMC
Tramblay Analyzed nonstationary Poisson and general- Nonstationary model
et al. (2013) model for heavy rainfall ized Pareto with climatic covariates
events distribution superior to stationary
model
Xiong et al. Developed NCM for NCM NCM outperformed tra-
(2015) nonstationary FFA ditional FFA models
Machado Evaluated stationary and Stationary and Nonstationary modeling
et al. (2015) nonstationary flood fre- nonstationary models improved rare flood
quency approaches probability estimates
Šraj et al. Compared models for Stationary and Nonstationary model
(2016) flood quantile estimation nonstationary GEV with annual precipitation
models had best performance
Debele et al. Analyzed seasonal flow Maximum likelihood, GAMLSS exhibited bet-
(2017b) maxima two-stage, and ter performance
GAMLSS methods
Das and Analyzed uncertainties in VIC-3L model projec- Stationary for RCP4.5;
Umamahesh streamflow projections tions under RCPs 4.5 nonstationary for RCP8.5
(2017) and 8.5
Debele et al. Discussed challenges in GAMLSS package GAMLSS useful, but
(2017a) nonstationarity models issues remain in
and practical requirements applicability
Dègan et al. Analyzed nonstationary Stationary and Nonstationary models
(2017) flood frequency nonstationary models crucial for long periods
Singh and Investigated Stationary and Nonstationary FFA
Chinnasamy nonstationarity of Periyar nonstationary models methods recommended
(2021) River
Chen et al. Revisited the stationary Nine candidate Rainfall variability domi-
(2021) assumption in FFA covariates in nant driver for flooding
nonstationary models
Gruss et al. Compared distributions Weibull, Log-Normal, Nonstationary models
(2022) for stationary and non GEV, and Pearson superior; Weibull and
stationary data Type III-based models Log-Normal most
suitable
Zhou et al. Developed approach for GAMLSS, Copula, Reliable probabilistic
(2022) nonstationary FFA and BUP techniques interval estimations for
uncertainty design floods
(continued)
15 Nonstationary Flood Frequency Analysis: Review of Methods and Models 283

Table 15.1 (continued)

Author(s) Objective Methodology Findings
Das and Examined hydrological GEV distribution with Nonstationary approach
Umamahesh extremes under physically based provided valuable
(2022) nonstationary conditions covariates insights
Guo et al. Investigated performance LME and GLS models Nonstationary
(2023) of nonstationary Bayesian LME-based Bayesian
RFFA RFFA method performed
better

Fig. 15.1 Results of nonstationary and stationary FFA performed for Barmanghat Subbasin: (a)
return level curves constructed based on a nonstationary model, (b) return level curves based on a
stationary model, and (c) varying effective return levels as a function of time

nonstationary model (507.2) is also better than the stationary model with BIC of
509.64. The comparison of return levels between the presented nonstationary anal-
ysis (in Fig. 15.1a) and the stationary model (Fig. 15.1b) demonstrates the differ-
ences between results of both approaches. Fig. 15.1c further shows the changes in
284 S. Barbhuiya et al.

effective return level as a function of time chosen as the covariate in the present case
study.

15.4 Concluding Remarks

Nonstationary flood frequency analysis methods have become increasingly impor-

tant in understanding and predicting hydrological extremes in changing environ-
ments. Various studies have been conducted worldwide, employing different
approaches and techniques to address nonstationarity in flood frequency modeling.
In summary, nonstationary flood frequency analysis methods have demonstrated
their usefulness in various case studies, offering improved modeling accuracy and
reducing uncertainty in the face of changing environmental conditions. These
methods provide valuable insights for flood management, infrastructure design,
and flood mitigation projects.
The vast amount of literature reviewed in this domain emphasizes that
nonstationary flood frequency analysis is important for addressing the impacts of
climate change, land use changes, and human activities in water resources manage-
ment and has capability to provide reliable estimates of flood risk probabilities,
design flood levels, and convey the associated uncertainties to users, compared to
stationary approaches. Additionally, for integrating nonstationary flood frequency
analysis into policy making, following parameters are required to be met:
(i) understanding of local hydrological regimes and nonstationary factors;
(ii) creation of public awareness on the importance of nonstationary flood frequency
analysis; (iii) identification of appropriate nonstationary methods tailored to the
study area; (iv) conversations with stakeholders, such as policymakers, engineers,
and researchers for participatory resilience development; and (v) regular update and
review of flood adaptation and mitigation policies and strategies to ensure their
effectiveness and relevance.

15.4.1 Challenges

The comprehensive review of flood studies and existing approaches in nonstationary

FFA has helped us to summarize the different challenges that are faced by hydrol-
ogists (Das et al. 2022). The major challenges are as follows:
• Data availability and quality: Nonstationary flood frequency analysis often
requires long-term and high-quality data to capture the underlying trends and
changes in hydrological extremes. However, acquiring such data is often chal-
lenging due to data gaps, inconsistencies, and limited spatial coverage.
• Selection of appropriate covariates: Identifying relevant and significant covariates
that can explain the nonstationarity in flood frequency is a difficult task.
15 Nonstationary Flood Frequency Analysis: Review of Methods and Models 285

Researchers must carefully select the covariates based on their physical relevance
and statistical significance in relation to the regional flood data.
• Model selection and validation: Numerous nonstationary models have been
proposed in the literature, each with their own advantages and limitations.
Choosing the appropriate model for a specific case and validating its performance
are essential steps in nonstationary flood frequency analysis.
• Uncertainty quantification: Nonstationary flood frequency analysis has numerous
sources of uncertainty, including those due to parameter estimation, model
structure, and input data. Quantifying and accounting for these uncertainties is
crucial for providing reliable and robust estimates of design floods.
• Computational complexity: Nonstationary flood frequency analysis often
involves complex mathematical models and advanced statistical techniques that
can be computationally demanding. Development of efficient algorithms and
software packages is necessary for practical applications.

15.4.2 Future Direction

With the exhaustive review of applications and scope of nonstationary FFA carried
out in this chapter, we have been able to assess the future scope of the research in this
domain. Notably, the following important thrust areas are to be addressed in the
analysis to utilize the maximum benefits of this exercise in hydrologic design and
risk adaptation:
• Integration of climate change projections: Incorporating climate change projec-
tions into nonstationary flood frequency analysis can help assess the potential
impacts of climate change on future flood risk and inform adaptation strategies.
• Development of regional nonstationary models: Developing regional
nonstationary flood frequency models can help overcome data limitations and
improve the estimation of flood quantiles for ungauged sites, which is particularly
relevant for regions with limited data availability.
• Improvement of model performance and uncertainty quantification: Further
research is needed to develop and evaluate new nonstationary models, as well
as to refine existing ones and to improve their performance and uncertainty
quantification in flood frequency analysis.
• Cross-disciplinary collaboration: Collaboration between hydrologists, meteorol-
ogists, and climate scientists can help better understand the drivers of
nonstationarity in flood frequency and develop more accurate and robust models.
• Application of machine learning and artificial intelligence: The use of machine
learning and artificial intelligence techniques can help identify patterns and
relationships in large, complex datasets and may provide novel insights into
nonstationary flood frequency analysis.
Ultimately, advancing the science of nonstationary FFA is essential for improv-
ing the design and management of water resources infrastructure and better adapting
286 S. Barbhuiya et al.

to the impacts of natural and anthropogenic climate change on hydrological systems.

By addressing the aforementioned challenges and research gaps and addressing the
future scope for research in this domain, nonstationary FFA can be improved,
standardized, and merged with flood risk management protocols to provide timely,
accurate, and reliable estimates of flood risk and build the resilience of flood
mitigation infrastructure.

References

Archfield SA, Hirsch RM, Viglione A, Blöschl G (2016) Fragmented patterns of flood change
across the United States. Geophys Res Lett. https://doi.org/10.1002/2016GL070590
Berghuijs WR, Harrigan S, Molnar P, Slater LJ, Kirchner JW (2019) The relative importance of
different flood-generating mechanisms across Europe. Water Resour Res. 55(6):4582–4593.
https://doi.org/10.1029/2019WR024841
Central Water Commission (2018) Annual report 2018–19. Central Water Commission, Ministry of
Water Resources, Government of India, New Delhi
Chen M, Papadikis K, Jun C (2021) An investigation on the non-stationarity of flood frequency
across the UK. J Hydrol 597. https://doi.org/10.1016/j.jhydrol.2021.126309
Cheng L, AghaKouchak A, Gilleland E, Katz RW (2014) Non-stationary extreme value analysis in
a changing climate. Clim Chang. https://doi.org/10.1007/s10584-014-1254-5
Coles S (2001) An introduction to statistical modeling of extreme values. Springer, London. https://
doi.org/10.1007/978-1-4471-3675-0
Das J, Umamahesh NV (2017) Uncertainty and nonstationarity in streamflow extremes under
climate change scenarios over a River Basin. J Hydrol Eng 22(10):1–13. https://doi.org/10.
1061/(asce)he.1943-5584.0001571
Das J, Umamahesh NV (2022) Investigating risk, reliability and return period under the influence of
large scale modes, and regional hydrological variability in hydrologic extremes. Hydrol Sci J
67(1):65–81. https://doi.org/10.1080/02626667.2021.1998512
Das J, Manikanta V, Nikhil Teja K, Umamahesh NV (2022) Two decades of ensemble flood
forecasting: a state-of-the-art on past developments, present applications and future opportuni-
ties. Hydrol Sci J 67(3):477–493. https://doi.org/10.1080/02626667.2021.2023157
Debele SE, Bogdanowicz E, Strupczewski WG (2017a) Around and about an application of the
GAMLSS package to non-stationary flood frequency analysis. Acta Geophys 65(4):885–892.
https://doi.org/10.1007/s11600-017-0072-3
Debele SE, Multimodel GÁ, Monte Á (2017b) A comparison of three approaches to non-stationary
flood frequency analysis. Acta Geophys 65(4):863–883. https://doi.org/10.1007/s11600-017-
0071-4
Dègan A, Adéchinaalamou E, N’Tcha M’Po Y, Afouda A (2017) Non-stationary flood frequency
analysis using additive terms for location, scale and shape parameters in the Ouémé River basin
(Benin, West Africa). Int J Curr Eng Technol 7(2):556–570
Douglas EM, Vogel RM, Kroll CN (2000) Trends in floods and low flows in the United States:
Impact of spatial correlation. J Hydrol 240(1–2):90–105. https://doi.org/10.1016/S0022-1694
(00)00336-X
Griffis VW, Stedinger JR (2007) Log-Pearson type 3 distribution and its application in flood
frequency analysis. II: Parameter estimation methods. J Hydrol Eng 12(5):492–500. https://
doi.org/10.1061/(ASCE)1084-0699(2007)12:5(492)
Gruss Ł, Wiatkowski M, Tomczyk P, Pollert J, Pollert J (2022) Comparison of three-parameter
distributions in controlled catchments for a stationary and non-stationary data series. Water
(Switzerland) 14(3). https://doi.org/10.3390/w14030293
15 Nonstationary Flood Frequency Analysis: Review of Methods and Models 287

Guo S, Xiong L, Chen J, Guo S, Xia J, Zeng L, Xu CY (2023) Nonstationary regional flood
frequency analysis based on the Bayesian Method. Water Resour Manag 37(2):659–681. https://
doi.org/10.1007/s11269-022-03394-9
Hejazi MI, Markus M (2009) Impacts of urbanization and climate variability on floods in North-
eastern Illinois. J Hydrol Eng 14(6):606–616. https://doi.org/10.1061/(asce)he.1943-5584.
0000020
Khaliq MN, Ouarda TBMJ, Ondo J-C, Gachon P, Bobée B (2006) Frequency analysis of a sequence
of dependent and/or non-stationary hydro-meteorological observations: a review. J Hydrol
329(3–4):534–552. https://doi.org/10.1016/j.jhydrol.2006.03.004
Koutsoyiannis D, Montanari A (2015) Meurtre par imprudence de concepts scientifiques: le cas de
la stationnarité. Hydrol Sci J 60(7–8):1174–1183. https://doi.org/10.1080/02626667.2014.
959959
Kuang D, Liao KH (2020) Learning from floods: linking flood experience and flood resilience. J
Environ Manag 271:111025. https://doi.org/10.1016/j.jenvman.2020.111025
Lima CHR, Lall U (2010) Spatial scaling in a changing climate: a hierarchical Bayesian model for
non-stationary multi-site annual maximum and monthly streamflow. J Hydrol 383(3–4):
307–318. https://doi.org/10.1016/j.jhydrol.2009.12.045
Lima CHR, Lall U, Troy TJ, Devineni N (2015) A climate informed model for nonstationary flood
risk prediction: application to Negro River at Manaus, Amazonia. J Hydrol 522:594–602.
https://doi.org/10.1016/j.jhydrol.2015.01.009
Machado MJ, Botero BA, Benito G (2015) Flood frequency analysis of historical flood data under
stationary and non-stationary modelling. Hydrol Earth Syst Sci. 2561–2576. https://doi.org/10.
5194/hess-19-2561-2015
Mangukiya NK, Sharma A (2022) Flood risk mapping for the lower Narmada basin in India: a
machine learning and IoT-based framework. Nat Hazards 113(2):1285–1304. https://doi.org/10.
1007/s11069-022-05347-2
Merz B, Blöschl G, Vorogushyn S, Dottori F, Aerts JC, Bates P et al (2021) Causes, impacts and
patterns of disastrous river floods. Nature Reviews Earth & Environment 2(9):592–609. https://
doi.org/10.1038/s43017-021-00195-3
Milly PCD, Betancourt J, Falkenmark M, Hirsch RM, Kundzewicz ZW, Lettenmaier DP, Stouffer
RJ (2008) Climate change: stationarity is dead: whither water management? Science 319(5863):
573–574. https://doi.org/10.1126/science.1151915
Mondal A, Daniel D (2019) Return levels under nonstationarity: the need to update infrastructure
design strategies. J Hydrol Eng 24(1):1–11. https://doi.org/10.1061/(asce)he.1943-5584.
0001738
Mondal A, Mujumdar PP (2012) On the basin-scale detection and attribution of human-induced
climate change in monsoon precipitation and streamflow. Water Resour Res. 48(10):
2011WR011468. https://doi.org/10.1029/2011WR011468
Mudelsee M, Börngen M, Tetzlaff G, Grünewald U (2003) No upward trends in the occurrence of
extreme floods in central Europe. Nature 425(6954):166–169. https://doi.org/10.1038/
nature01928
Ouarda J, El S (2011) Bayesian nonstationary frequency analysis of hydrological variables
1. JAWRA. 47(3). https://doi.org/10.1111/j.1752-1688.2011.00544.x
Ray K, Pandey P, Pandey C, Dimri AP, Kishore K (2019) On the recent floods in India. Curr Sci
117(2):204–218. https://www.jstor.org/stable/27138236
Rigby RA, Stasinopoulos DM, Lane PW (2005) Generalized additive models for location, scale and
shape. J R Stat Soc Ser C Appl Stat 54(3):507–554. https://doi.org/10.1111/j.1467-9876.2005.
00510.x
Salas JD, Obeysekera J, Vogel RM (2018) Techniques for assessing water infrastructure for
nonstationary extreme events: a review. Hydrol Sci J 63(3):325–352. https://doi.org/10.1080/
02626667.2018.1426858
Serinaldi F, Kilsby CG (2015) Stationarity is undead: uncertainty dominates the distribution of
extremes. Adv Water Resour 77:17–36. https://doi.org/10.1016/j.advwatres.2014.12.013
 288 S. Barbhuiya et al.

Sharma A, Goyal MK (2017) A comparison of three soft computing techniques, Bayesian regres-
sion, support vector regression, and wavelet regression, for monthly rainfall forecast. Int J Intell
Syst 26(4):641–655. https://doi.org/10.1515/jisys-2016-0065
Singh N, Chinnasamy P (2021) Non-stationary flood frequency analysis and attribution of
streamflow series: a case study of Periyar River, India. Hydrol Sci J 66(13):1866–1881.
https://doi.org/10.1080/02626667.2021.1968406
Šraj M, Viglione A, Parajka J, Blöschl G (2016) The influence of non-stationarity in extreme
hydrological events on flood frequency estimation. J Hydrol Hydromech. 426–437. https://doi.
org/10.1515/johh-2016-0032
Stedinger JR, Vogel RM, Foufoula-Georgiou E (1993) Frequency Analysis of Extreme Events. In:
Maidment DR (ed) Handbook of Hydrology. McGraw-Hill, New York, NY, USA, pp 18.1–
18.66
Strupczewski WG, Singh VP, Feluch W (2001) Non-stationary approach to at-site flood frequency
modelling I. Maximum likelihood estimation. J Hydrol 248(1–4):123–142. https://doi.org/10.
1016/S0022-1694(01)00397-3
Tramblay Y, Neppel L, Carreau J, Najib K (2013) Non-stationary frequency analysis of heavy
rainfall events in southern France. Hydrol Sci J 58(2):280–294. https://doi.org/10.1080/
02626667.2012.754988
Villarini G, Serinaldi F, Smith JA, Krajewski WF (2009) On the stationarity of annual flood peaks
in the continental United States during the 20th century. Water Resour Res 45(8):1–17. https://
doi.org/10.1029/2008WR007645
Vogel RM, Wilson I (1996) Probability distribution of annual maximum, mean, and minimum
streamflows in the United States. J Hydrol Eng 1(April):69–76
Vogel RM, Yaindl C, Walter M (2011) Nonstationarity: flood magnification and recurrence
reduction factors in the United States. JAWRA J Am Water Resour Assoc 47(3):464–474.
https://doi.org/10.1111/j.1752-1688.2011.00541.x
Westra S, Thyer M, Leonard M, Kavetski D, Lambert M (2014) A strategy for diagnosing and
interpreting hydrological model nonstationarity. Water Resour Res 50(6):5090–5113. https://
doi.org/10.1002/2013WR014719
Wilson D, Hisdal H, Lawrence D (2010) Has streamflow changed in the Nordic countries? - Recent
trends and comparisons to hydrological projections. J Hydrol 394(3–4, 334):–346. https://doi.
org/10.1016/j.jhydrol.2010.09.010
Wood SN (2006) On confidence intervals for generalized additive models based on penalized
regression splines. Aust N Z J Stat 48(4):445–464. https://doi.org/10.1111/j.1467-842x.2006.
00450.x
Xiong L, Du T, Xu C, Guo S (2015) Non-stationary annual maximum flood frequency analysis
using the norming constants method to consider non-stationarity in the annual daily flow series.
Springer, Cham. https://doi.org/10.1007/s11269-015-1019-6
Zhou Y, Guo S, Xu CY, Xiong L, Chen H, Ngongondo C, Li L (2022) Probabilistic interval
estimation of design floods under non-stationary conditions by an integrated approach. Hydrol
Res 53(2):259–278. https://doi.org/10.2166/nh.2021.007

View publication stats