CHAPTER 14

Statistical Analysis of Hydro-Climatic Variables

R. K. Rai, Alka Upadhyay, C. S. P. Ojha and L. M. Lye

In many instances, a time series is generally not statistically independent but

is comprised of patterns of persistence, cycles, trends or some other non-random
components. In climatology, identification of the precise nature and extent of non-
randomness in the time series of meteorological data is very important and is usually
the first step of time series analysis. Statistical evidence of non-randomness in such
time series is equated with evidence of bona fide climatic fluctuations and said to be
dependent or exhibits persistence of some sort. A climatologic time series may
consists either of wholly random variations, of wholly random changes, or both
random and non-random components. In a hydrological point of view, dependence in
the time series is associated with the hydro-climatic paths followed by moisture from
atmosphere to earth through surface and sub-surface route and back to the
atmosphere. If a time series is non-random, the time series is further investigated for
trends, persistence, periodic fluctuations or combination of these. Therefore, this
chapter presents the statistical approach applied for understanding the hydro-climatic
time series, in which three components were discussed in detail with application to
the hydro-climatic data of the Yamuna River basin of Ganga River system of India.
The components are: (a) tests of short and long term dependence; (b) trend analysis;
and (c) tests of persistence and periodicity.

14.1 Statistical Methods of Short- and Long-term Dependence

Many types of hydrologic time series exhibit significant serial correlation.

That is, the value of the random variable under consideration at one time period is
correlated with the values of the random variable at earlier time periods. When the
time series shows the persistence for a shorter period then it is referred as short-term
dependence. It can be measured through the serial correlation of the time series. The
series is said to be completely random (or not dependent) if the serial correlation is
zero. However, in practice, this situation never occurs. Therefore, for confidence in
the test of short-term dependence, it is necessary to test the time series using alternate
statistical tests.

For short term dependence, statistical tests may be divided into two sets, viz.,
non-parametric and parametric. The non-parametric test accounts for the length and

127
relative sign of the observation in the time series whereas, parametric test is based on
the values associated with the series and requires fulfilling strict assumptions of the
distribution of the data and therefore, is less robust than the non-parametric test.

14.1.1 Non-Parametric Test

The commonly used non-parametric statistical tests to investigate the short

term dependence are described as follows.

Median-Crossing Test (Fisz 1963). In this test X = 0 is considered when xi <

xmedian, and X = 1 for xi > xmedian. If the original sequence of Xs has been generated by a
purely random process, then u, the number of times zero is followed by one or one is
followed by zero, is approximately normally distributed. Initial value of u is taken as
zero (i.e., u = 0). If xi > xmedian and xi+1 < xmedian or xi < xmedian and xi+1 > xmedian then u = u
+1. The expected value of u is defined as follows.
E (u ) (n 1) / 2 (Eq. 14.1)
Var (u ) (n 1) / 4 (Eq. 14.2)
where, n is the length of sample, E(u) is the expectation of the series, and Var (u) is
the variance. The Z statistic is computed as follows.
u E (u )
Z
[Var (u )]1/ 2 (Eq. 14.3)
If Z < Z then the null hypothesis is not rejected, otherwise it is rejected and
the series is assumed to be non-random. The null hypothesis can be tested on 5% and
10% significance level. The value of Z at 5% and 10% significance level is 1.96 and
2.65, respectively.

Turning Point Test. Kendalls turning point test (Kendall and Stuart 1976)
is also based on the binary series. If xi-1 < xi > xi+1 or xi-1 > xi < xi+1 then xi = 1;
otherwise xi = 0. The total number of ones u (i.e., initial value of u = 0 and if above
stated condition is satisfied then u = u + 1), is approximately normally distributed.
The expected value of u is defined as follows.
E (u ) 2( n 2) / 3 (Eq. 14.4)
Var (u ) (16 n 29) / 90 (Eq. 14.5)
The Z statistic can be computed as follows.
u E (u )
Z (Eq. 14.6)
[Var (u )]1/ 2
If Z < Z then the null hypothesis is not rejected otherwise it is rejected and the series
is assumed to be non-random.

128
 Rank Difference Test (Meacham 1968). Values are replaced by their
relative ranks Ri with the lowest being denoted by rank 1 (R1). The statistic u is
calculated by
n
u Ri Ri 1 (Eq. 14.7)
i2

The expected value of u is estimated by using the following relationship.

E (u ) [(n 1) (n 1)] / 3 (Eq. 14.8)
The variance of u is given by
Var (u ) [(n 2) (n 1) (4 n 7)] / 90 (Eq. 14.9)
The Z statistic is estimated as follows:
u E (u )
Z (Eq.
[Var (u )]1/ 2
14.10)
If Z Z then the null hypothesis is rejected and the series is assumed to be non-
random.

Cumulative Periodogram Test (Box and Jenkins 1976). The periodogram

of a time series is defined as

2
2 2
n
n

I fj xi cos 2 if j xi sin 2 fj (Eq.
n i 1 i 1
14.11)
where f j j n is the frequency; j = 1, 2,., (n 2)/2 for n even; and j = 1, 2,
., (n 1) for n odd. The normalized cumulative periodogram is obtained from
following relation.
j
C ( f j ) I ( f i ) ns 2 (Eq.
i 1

14.12)
where S 2 is the variance of xi . For a white-noise series the plot of C f i against f i
would be scattered about a straight line joining points (0, 0) and (0.5, 1). The
approximate confidence limit lines for a truly random series are drawn at distances:

K n 2 2
1/ 2
(Eq.
14.13)
where n is even and K is 1.63 (99%), 1.36 (95%), 1.22 (90%) or 1.02 (75%).

Wald-Wolfowitz Test (Wald and Wolfowitz 1943). For a sample of size n

129
 n 1
R xi xi 1 x1 xn (Eq.
i 1
14.14)
If the elements of the sample are independent then expectation and variance are
computed as
s12 s2
E (R) (Eq.
n 1
14.15)
2
s 2 s4 s12 s2 s14 4 s12 s2 4 s1s3 s2 2 2 s4
Var ( R ) 2 (Eq.
n 1 n 1 (n 1) ( n 2)
14.16)
where,
sr x1r x2 r
xnr (Eq.
14.17)
Similarly, the Z-statistic is computed as:
R E (R)
Z (Eq.
[Var ( R )]1/ 2
14.18)
If the mean is subtracted first, s1 0 , then

s s 2 s 2
s2 2s4
2
1/ 2
s2
RN 2
, 2 4
(Eq.
n 1 n 1 n 1 (n 1) (n 2)
14.19)
If Z Z then the null hypothesis is rejected and the series is assumed to be non-
random.

Rank von Neumann ratio Test (Madansky 1988). Let r1 , r2 ,L , rn denote

the ranks associated with the xi values. The rank Von Nuemann ratio is given by
n

(r r i i 1 )2
(Eq.
i 2

n (n 1) /12
2

14.20)
Critical value of C [n 2 (n 2 1) /12] and approximate critical value of were given
by Madansky (1988). For large n, is approximately distributed as N(2, 4/n),
although Bartels recommended 20 /(5n 7) as a better approximation to the variance
of (Madansky 1988).

130
 Kendalls Rank Correlation Test. If the series is thought to have a random
component, Kendalls rank correlation test can be used to test for significance. This
measures the disarray in the data; it is particularly effective if the underlying trend is
of a linear type. This test is also referred as test and is based on the proportionate
number of subsequent observations which exceeds a particular value. The statistics u
is calculated as follows: if ( xi 1 xi ) then u u 1 , whereas initial value of u is
assigned zero (i.e. u = 0). The Z statistic is computed as:
T
Z (Eq.
[Var (T )]1/ 2
14.21)
4u
where, T 1 (Eq.
n(n 1)
14.22)
2 (2n 5)
Var (T ) (Eq.
9 n ( n 1)
14.23)
If Z Z then the null hypothesis is rejected and the series is assumed to be non-
random.

Run Test. In this test, x is replaced by 0 if xi xmedian , and x is replaced by 1

if xi xmedian . If the original sequence of X s has been generated by a purely random
process, then u, the number of times 0 is followed by 1 or 1 is followed by 0, is
approximately normally distributed. The test is worked out using the following steps.
Initially u = 0. If xi > xmedian > xi+1 or xi < xmedian < xi+1 then u = u +1. The expected
value of u is defined as follows.
E (u ) m 1 (Eq.
14.24)
where, m is the rank corresponding to the median value. The variance of u is
estimated as follows.
m (m 1)
Var (u ) (Eq.
2 m 1
14.25)
The Z statistic can be computed as follows.
u E (u )
Z (Eq.
[Var (u )]1/ 2
14.26)

131
If Z Z then the null hypothesis is not rejected otherwise it is rejected (i.e., the series
is assumed to non-random).

Spearmans Rho Test. For a sample data set of {xi , i 1, 2,......, n} , the null
hypothesis Ho of S-R test is that all the xi are independent and identically distributed.
The alternative hypothesis is that xi increases or decreases with i, then non-
randomness exists. The Z statistics of S-R test is given as follows:
n
6 {R( xi ) i}2
(Eq.
D 1 i 1

n(n 1)
2

14.27)
Var ( D ) 1/( n 1) (Eq.
14.28)

D
Z (Eq.
[Var ( D)]0.5
14.29)

where, R ( xi ) is the rank of ith observation xi in the sample of size n. The standardized
statistics Z follow the standard normal distribution Z N (0, 1) 1.96 at 5% .

14.1.2 Parametric Test

In the above section, non-parametric tests have been described, which uses the
scores, rank or sign of the data to perform the analysis. Whereas, parametric tests are
often robust which uses the time series statistics such as population mean and
standard deviation computed from the sample. In the parametric test of short-term
dependence, two tests viz. autocorrelation (Yevjevich 1976) and von Nuemann ratio
test (Madansky 1988) has been described.

Autocorrelation Test (Yevjevich 1971). Short-term dependence is usually

measured by the magnitude of the low-order autocorrelation coefficients. It is
extensively used to determine the linear dependence in the time series. The
autocorrelation function, rk , is estimated from
nk

( x x ) ( x
t t t k xt k )
rk t 1
1/ 2 (Eq.
nk nk



( x x ) ( x
t 1
t t
2

t 1
t k xt k ) 2

14.30)

132
where k is lag, xt is the observation at time t, and n is sample size. The lag-1
autocorrelation, r1 is calculated from eq. (14.30) and checked whether or not is
significantly different from the expected value as
E r1 1 n (Eq.
14.31)
and,
Var r1 n3 3n 2 4 n 2 n 2 1 (Eq.
14.32)
and,

Z r1 E r1 Var r1
0.5
(Eq.
14.33)
where, Z is the normally distributed Z-statistic at significance level. If Z Z
then the null hypothesis is not rejected otherwise it is rejected (i.e., the series is
assumed to non-random). The value of Z is 1.96 and 2.65 for 5% and 10%
significance level.

von Neumann Ratio Test (Madansky 1988). If xi , i 1, 2,..... is a time

series of hydrologic variable then the V-statistics need to be computed based on the
assumption that the series is normally distributed with N (0, 1) and is computed as
follows:
V E (V )
Z (V ) (Eq.
[Var (V )]0.5
14.34)
n

(x x i i 1 )
with, V i2
n (Eq.
(x x )
i 1
i
2

14.35)
E (V ) 2 (Eq.
14.36)
4 (n 2)
Var (V ) (Eq.
(n 2 1)
14.37)

14.1.3 Test for Long Term Dependence (Lye and Lin 1994)

133
 The long-term dependence in hydrologic time series is manifested by the fact
that the extreme events may persist for a long time. In water resources design long
sequence of hydrologic data are required (more than 100 years). In such long records
it is expected to have the extreme events. Therefore, the generation of long range
dependence effect is of vital importance to the water resources system planning.

In this chapter, long term dependence can be measured by the magnitude of

the Hurst coefficient. The Hursts coefficient k is estimated by eq. (14.38).
log( Rn / n )
k (Eq.
log(n / 2)
14.38)
where Rn range of cumulative departure from the mean, n standard deviation of
the sample, n sample length. If k = 0.5, series is independent (i.e. random in nature)
and for greater degree of dependence k > 0.5, but not exceed to unity.

Range Analysis: Computation of Rn . The range, related to the storage

capacity of the reservoir and run, the property related to the drought, are the
additional characteristics important in water resources studies that may be derived
from a time series. The range of cumulative departures from the sample mean is
related to the minimum storage capacity required to deliver the sample mean
throughout a time period equal to the length of the sample n. Range of the cumulative
departure from mean Rn can be expressed as follows.
Rn d n d n (Eq.
14.39)

In eq. (14.39), d n is the maximum value of the cumulative departure from

mean, and d n is the minimum value of cumulative departure from mean. If
zi ( z1 , z2 , z3 ,.....) is the flow series with mean of z , then
n
d n max ( zi z ) (Eq.
1i n
i 1
14.40)
n
and d min ( zi z )

n (Eq.
1i n
i 1

14.41)

If zi is the annual flow and initial storage is d n then a reservoir having capacity Rn
will be able to meet the average demand and could be full without overflow at least
once and empty at least once.

134
 Rescaled Range. The rescaled range is the range divided by the sample
standard deviation n . The rescaled range is defined by eq. (14.42):
Rn* Rn / n (Eq.
14.42)
It is proportional to N 0.5 as N for models most typically used in hydrology such
as AR and the ARMA models. Analysis made by Hurst (1951) using long records of
geophysical time series, appears to show that the rescaled range is proportional to n k
with k 0.5 (while for models such as those referred above (i.e. eq. 14.38) k 0.5 ).
This apparent discrepancy has been called the Hurst Phenomenon. i.e.,
Rn* f (n) n k (Eq.
14.43)
For random series Rn f (n) n , i.e. k 0.5 .
* 0.5

Bootstrap Method for Testing the Significance of Hursts k. To test the

significance of Hursts K, the non-parametric bootstrap approach (Efron 1979; 1981;
1982) was applied. The bootstrap samples are generated from the data of the original
samples as follows: (i) suppose that the annual flow series x1 , x2 ,..... xn are
independent observations. Each xi has the same probability of occurrence and is
equal to 1/ n , where n is the number of observations in the series; (ii) generate a
uniform random data i between 1 and n, then choose xi as one point in the bootstrap
sample. Repeat this step n times to generate a bootstrap sample of the same size n as
the original sample size; (iii) calculate the Hursts k for the bootstrap samples; (iv)
repeat steps (ii) to (iii) for a large number of times (say, 10000); (v) count the number
of times the observed k value of sample is exceeded by 10000 bootstrap k values; and
(vi) calculate the P value as follows:
Number of k kobserved
Pvalue (Eq.
Number of Bootstrap samples (i.e. 10000)
14.44)

If the Pvalue < (say 0.05) then the series has long term dependence at the specified
level (i.e. = 5%); otherwise, series does not have long term dependence. It is
important to note that the null hypothesis of this bootstrap test is that of
independence.

14.2 Statistical Methods of Trend Analysis

A steady and regular movement in a time series through which the values are
on average either increasing or decreasing is termed a trend. This type of behavior
can be local, in which case the nature of the trend is subject to change over short
intervals of time, or, on the other hand, it can be visualized as a global trend that is
135
long lasting. If a trend in a hydrologic time series appear it is, in effect, part of a low
frequency oscillatory movement induced by climatic factors or through change in
land use and catchment characteristics. There are several approaches to detecting the
trend in the time series. These approaches can be either parametric or non-parametric.
Parametric methods assume that the data are normally distributed and free from
outliers. On the other hand, non-parametric methods are free from such assumptions.
The most commonly used non-parametric tests for detecting trend in the time series is
the Mann-Kendall (MK) test (Mann 1945; Kendall 1955). It is widely used for
various climatic variables (Hirsch et al. 1982; Hirsch and Slack 1984; Lettenmaier et
al. 1994, Gan 1998; Lins & Slack 1999; Douglas et al. 2000; Burn and Elnur 2002;
Yue et al. 2002; Yue and Pilon 2004; Burn et al. 2004; Zhang et al. 2005; Aziz and
Burn 2006; Chen et al. 2007; Rai et al. 2010).

This section demonstrates the different techniques of trend analysis for

climatic variables. The statistical tests applied for the analysis are the original Mann-
Kendall, modified Mann-Kendall (MK), and Mann-Kendall with pre-whitening.

14.2.1 Original Mann-Kendall Test (Mann 1945; Kendall 1955)

The original MK test searches for a trend in a time series without specifying
whether the trend is linear or nonlinear (Khaliq et al. 2009). It works on the null
hypothesis that data are independent and randomly ordered, i.e. there is no trend or
serial correlation structure present in the observations. The Mann-Kendall test for
detecting monotonic trends in hydrologic time series is described by Yue et al. (2002).
It is based on the test statistics S, which is defined as:
n 1 n
S sgn ( x j xi ) (Eq.
i 1 j i 1

14.45)
where, x j are the sequential data values, n is the length of the data set and
1, for t 0

sgn (t ) 0, for t 0 (Eq.
1, for t 0

14.46)

The value of S indicates the direction of trend. A negative (positive) value

indicates falling (rising) trend. Mann-Kendall has documented that when n 8 , the
test statistic S is approximately normally distributed with mean and variance as
follows:
E (S ) 0 (Eq.
14.47)

136
 m
n (n 1) (2n 5) ti (ti 1)(2ti 5)
(Eq.
Var ( S ) i 1

18
14.48)
where, m is the number of tied groups and ti is the size of the i th tie group. The
standardized test statistics Z is computed as follows.
S 1
, for S 0
Var ( S )

Z MK 0 , for S 0 (Eq.
S 1
, for S 0
Var ( S )
14.49)
The standardized Mann-Kendall statistics Z follows the standard normal distribution
with zero mean and unit variance. If Z Z1 (/2), the null hypothesis about no trend
is rejected at the significance level (10% in this study).

14.2.2 Mann-Kendall Test for Autocorrelated Data

For the original Mann-Kendall test, the time series must be serially
independent. However, in many real situations the observed data are auto-correlated.
The autocorrelation in the observed data may cause misinterpretation of the trend test
results. Cox and Stuart (1955) stated that positive serial correlation among the
observations would increase the chance of a significant answer, even in the absence
of a trend. A closely related problem that has been studied is the case where
seasonality exists in the data (Hirsch et al. 1982). By dividing the observations into
separate classes according to the season and then performing the Mann-Kendall trend
test on the sum of the statistics from each season, the effect of seasonality can be
eliminated. This modification is called the seasonal Mann-Kendall test (Hirsch et al.
1982; Hirsch and Slack 1984). Although the seasonal test eliminates the effect of
seasonal dependence, it does not account for the correlation in the series within the
season (Hirsch and Slack 1984). The same problem exists when yearly time series is
considered for the analysis as it can be significantly auto correlated. Therefore, the
original Mann-Kendall test is modified so that it can account for serially dependent
data.

Modified Mann-Kendall Test. For the Modified Mann-Kendalls test, the

statistic S tends to normality for large n, with mean and variance given by:
E (S ) 0 (Eq.
14.50)
Var ( S ) n (n 1) (2n 5) /18 (Eq.
14.51)

137
where S is given by eq. (14.52):
n 1 n
S aij sgn ( x j xi ) (Eq.
i 1 j i 1

14.52)
with the same mean and variance as in eqs. (14.50) and (14.51). A modified version
of the Mann-Kendall test which is robust in the presence of autocorrelation is based
on the modified variance of S given by eq. (14.52).
n n(n 1)(2n 5) n
V * ( S ) Var ( S ) * * (Eq.
ns 18 ns
14.53)
* *
where, n / ns represents a correlation due to the autocorrelation in the data. The n / ns
is evaluated using eq. (14.54).
n 1
n 2
ns*
1 (n i )(n i 1)(n i 2) s (i)
n(n 1)(n 2) i 1
(Eq.

14.54)
In eq. (14.54), n is the actual number of the observations and s (i ) is the
autocorrelation function of the ranks of the observations. The advantage of using eqs.
(14.53) and (14.54) for the evaluation of variance of S is that there is no need of
either normalized data or their autocorrelation function. The autocorrelation of ranks
of observations s (i ) is related with the parent autocorrelation function and is given
as follows (Kendall 1955):

(i ) 2 sin s (i ) (Eq.
6
14.55)
The inverse of eq. (14.55) can therefore be used to evaluate the autocorrelation of the
ranks s (i ) that appeared in eq. (14.54) and is given by eq. (14.56).
6 (i)
s (i ) sin 1 (Eq.
2
14.56)
The significance of the trends is tested by comparing the standardized test statistics Z.
S
Z * (Eq.
[V ( S )]0.5
14.57)
A significant level of 0.1 for the autocorrelation of the ranks s (i ) was used,
which produced the best overall empirical significance level. The serial correlation of
the series can be computed as follows:

138
 nk

Cov ( xt , xt k ) ( x x ) ( x
t t t k xt k )
rk t 1
(Eq.
St St k nk nk

1/ 2

( x x ) . ( x
t 1
t t
2

t 1
t k xt k ) 2

14.58)

where, xt is the mean of the first n k values x1 , x2 ,...., xn k , and xt k is the mean of
the last n k values xk 1 , xk 2 ,...., xn . rk varies in the range of 1 rk 1 . To test the
significance of serial correlation 95 and 99% of probability level, eqs. (14.59) and
(14.60) are used respectively (Yevjevich 1971).
1 1.965(n k 1)1/ 2
rk (95%) (Eq.
nk
14.59)
For 99 % of probability level,
1 2.326 ( n k 1)1/ 2
rk (99 %) (Eq.
nk
14.60)

Mann-Kendall Test with Pre-whitening. An alternative approach to

perform the trend analysis of time series with the presence of serial correlation using
the Mann-Kendall test is to first remove the serial correlation from data and then
apply the test. Several approaches have been suggested for removing the serial
correlation from a data set prior to applying the test. The pre-whitening approach is
most common which involves computation of serial correlation and removing the
correlation if the calculated serial correlation is significant at 5% significance level
(Burn and Elnur 2002). The pre-whitening is accomplished as follows:
X t' X t 1 1 X t (Eq.
14.61)
where, X t = original time series with autocorrelation for time interval t; X t' = pre-
whitened time series; and 1 = the lag-1 autocorrelation coefficient. This pre-whitened
series is then subjected to Mann-Kendall test (i.e. eqs. 14.45 to 14.49) for detecting
the trend.

14.3 Statistical Methods of Periodicity

For analyzing the persistence and periodicity in the climatic series over the
world, considerable efforts have been made in the past by many investigators. For
example, Angell et al. (1966) investigated quasi-biennial oscillation in troposphere

139
zonal winds of the Tropics and sub-tropics; whereas it is also evident in the surface
temperature in many parts of the world (Landsberg et al. 1963).

14.3.1 Test of Persistence and Periodicity

To test the persistence and periodicity in the climatological time series,

normalized anomaly of the time series is used. A normalized series is obtained as
follows:
X sN ( j ) ( X s ( j ) X s ) / s (Eq.
14.62)

where, X sN ( j ) is the normalized anomaly of the series, X s is the observed time series
for station s during jth year, X s and s are the long-term mean and standard deviation
of annual/seasonal time series.

14.3.2 Computation of the power spectrum

Periodicity is one of the deterministic components in the time-series. Most of

the climatic, atmospheric and hydrological time-series would consist of a
combination of stochastic and deterministic components. The power spectrum is a
method of analysis that was developed to handle the problem of periodicity in
variations of natural events observed in time, such as in climatological and
hydrological time series. Power spectrum analysis, also called generalized harmonic
analysis, was derived from the principles first developed by Wiener (1930, 1949). It is
based on the premise that the time series are not necessarily composed of a finite
number of oscillations, each with a discrete wavelength, but rather that they consist of
virtually infinite number of small oscillations spanning a continuous distribution of
wavelengths. The spectrum therefore, gives the distribution of variations in a time
series over a continuous domain of all possible wavelengths.

Procedures for computing the power spectra may vary. Here, in this study, an
approach described in WMO (1966), developed by Tukey (1950) and Blackman and
Tukey (1958), was employed. A detailed description of this approach can also be
found in various textbooks: Blackman and Tukey (1958), Jenkins and Watts (1968)
and Julian (1967). It can be summarized through the following steps:

(i) First, all serial correlation coefficients of normalized climatic series are
computed for lags from L = 0 to m, where m is the maximum lag considered to be N/3
in which; N is the length of the series. The serial correlation coefficient can be
computed using eq. (14.63).

140
 n L

( x x ) ( x
t t tL xt L )
rL t 1
1/ 2 (Eq.
nL nL



( x x ) ( x
t 1
t t
2

t 1
tL xt L ) 2

14.63)

(ii) Using the values of rL , the raw spectral estimates, sk are computed using
the following set of equations:
1 1 m 1
s0 (r0 rm ) rL (Eq.
2m m L 1
14.64)
r0 2 m 1 kL 1
sk rL Cos ( ) rm (1) k ; for k 1, 2,......., m 1 (Eq.
m m L 1 m m
14.65)
1 1 m 1
sm [ r0 (1) m rm ] (1) L rL (Eq.
2m m L 1
14.66)
Smallest is the value of k longest will be the wavelength of the spectrum, i.e. shortest
wavelength is achieved at k = m.

(iii) The raw spectrum sk is then smoothened with a 3-term weighted average.
For smoothing, procedure suggested by Hanning was used (WMO 1966).
s0 ( s0 s1 ) / 2 (Eq.
14.67)
sk ( sk 1 2 sk sk 1 ) / 4 ; for k 1, 2,......., m 1 (Eq.
14.68)
sm ( sm 1 sm ) / 4 (Eq.
14.69)
The averaging procedure is performed to derive a constant estimate of the final
spectrum in terms of m 1 discrete estimates (WMO 1966).

14.3.3 Test for Statistical Significance

The procedure for evaluating the results of power spectrum analysis mention
in WMO (1966) is described below:

141
 A null hypothesis continuum is fitted to the computed spectrum. To start
with, significance of the lag-1 serial correlation coefficient r1 of the climatic series is
tested by the following equation.
(r1 )t [1 t g N 1] /( N 1) (Eq.
14.70)
where, t g 1.645 at 90 percent confidence level. The null hypothesis of the
randomness of climatic series against the serial correlation is rejected for the large
value of (r1 )t . If r1 is not significantly different from zero, then series is regarded to
be free from persistence. In this case, the appropriate null continuum is white noise.
In other words, a horizontal straight line, the value of which is everywhere equal to
the average of the values of all the m 1 raw spectral estimates (i.e., s ) in the
computed spectrum (i.e., S k s ), is taken as the most suitable theoretical approach.

On the other hand, if the computed r1 is positive and statistically significant,

serial correlation coefficients for lag-2 and lag-3 are checked to see whether they
approximate the exponential relations r2 r1 and r3 r1 (WMO 1966). If these
2 3

relations are ensured with the computed serial coefficients, the approximate null
continuum is assumed as the simple Markov red noise, whose shape depends on
unknown value of the lag-1 serial correlation coefficient for a population . Then the
null continuum can be created by following an approximate procedure. By assuming
that the sample r1 is an unbiased estimation of serial correlation coefficient , of the
Harmonic number of k between k = 0 to m are assessed:
S k s [(1 r12 ) /{1 r12 2 r1 Cos( k / m)}] (Eq.
14.71)
where, s is the average of all m 1 raw spectral estimates sk in the
computed spectrum. The resulting values of Sk can be plotted superposed on the
sample spectrum, and a smoothed curve passed through these values to reach the
required null continuum. If r1 is statistically significant but a few serial correlation
coefficients for higher lags do not show the required exponential relations with r1 ,
then doubt arises as to whether the simple Markov-type persistence is the dominant
form of non-randomness in series of climatic observations. Nevertheless, WMO
(1966) suggested that this procedure could be continued with just as before to
compute the red noise continuum for r1 .

At this stage of the power spectrum analysis a first choice of the null
continuum is made, and this selected continuum is superposed on the studied
spectrum. In this case, it would be possible to make an assessment of the spectrum for
its consistency with the chosen continuum. Then, the value of each spectral estimate
sk is compared with the local value of the null continuum. The statistic associated
with each spectral estimate is the ratio of the magnitude of the spectral estimate to the

142
local magnitude of the continuum (red noise continuum). Tukey (1950) found that the
quantity of this ratio is distributed as Chi-square divided by the degree of freedom.
The degree of freedom, , of each estimate of a computed spectrum is given as
follows.
(2 N m / 2) / m (Eq.
14.72)
In eq. (14.72) N is the length of series and m is the maximum lag. The ratio of any
sample spectral estimate sk to its local value of the red noise continuum is then
compared with critical percentage-point levels of 2 / distribution for the proper
value. This comparison produces the required statistical significance level. The 2
value can be obtained from standard statistical books. In a sample spectrum, critical
percentage-point levels of the 2 / distribution, e.g. the 0.95 confidence level, is the
same for all spectral estimates sk . The confidence limits are finally derived by
multiplying the null continuum (i.e. S k ) with the 2 / . Finally, the cycle is
computed using the relation of P 2m / L .

14.4 Application of Methodology

The application of the described methodology for dependence, trend and

periodicity is demonstrated using the hydro-climatic data of the Yamuna River basin.
The Yamuna River is a major tributary of the Ganga River system which originates
from the Shivalik range of Himalaya which after traversing of 1325 km joins with the
River Ganga at Allahabad. In 1325 km of stretch, it faces large hydro-climatic
variation, and the catchment of it also has large variation between semi arid to humid
with diversified land uses.

14.4.1 Data Used

The variables used in the analysis are categorized into two sets: direct
variables and derived variables. The variables are: rainfall (annual, monsoon and non-
monsoon rainfall), and few derived variables important for agricultural planning such
as: the onset of effective monsoon (OEM), number of rainydays (annual, monsoon
and non-monsoon) and the Aridity Index (AI).

For rainfall, one degree grid data of daily rainfall for the period 1951 to 2002
supplied by the India Meteorological Department (IMD) was used. Sixty five rain-
grids points falling under the Yamuna basin with 1 buffer were used in the analysis.
For seasonal analysis, only two seasons were considered, viz. the monsoon (July to
October) and non-monsoon (November to June).

The OEM, an important governing agro-climatic parameter for planning of

kharif crops was estimated using the Ashokraj (1979) criteria, which uses the daily

143
rainfall and potential evapotranspiration data. The potential evapotranspiration was
estimated using the Thornthwaite method (1948). A seven days rainfall spell
satisfying the following conditions are terms as OEM. These conditions are (Ashokraj
1979): (i) first day rainfall in the seven days spells should be more than evaporation
of that particular day; (ii) a day with-more than 3 mm of rainfall is considered to be a
rainy day; (iii) total rainfall during the seven days spell is more than (3 PET + 10)
mm; and (iv) at least four out of seven days are rainy days.

The aridity index (AI), an indicator of annual soil moisture deficit was
estimated using the procedure of UNEP (1993) as:
AI ( R / PET ) 100 % (Eq.14.73)
where, AI is the aridity index (%), R is the annual rainfall (mm) and PET is the annual
potential evapotranspiration (mm). AI values below 100 % shows annual moisture
deficit in average climatic conditions.

14.4.2 Analysis of Short- and Long-Term Dependence Test

As stated above, 12 statistical tests (10 non-parametric and 2 parametric tests)

for short-term dependence and a single test (i.e. Hursts coefficient with the boot strap
method) for long-term dependence were used. For the long term dependence, Hursts
K and lag-1 serial correlation for all the considered variables has been estimated. A
sample table for the OEM is given in Table 14.1, and the summary statistics is
presented in Tables 14.2.

Table 14.1. Start of onset of effective monsoon, OEM (days)a

ID Grid Point Mean Hursts r1 ID Grid Point Mean Hursts r1
Lat Long K Lat Long K
1 21.5 75.5E 180 0.7 0.049 34 25.5N 81.5E 188 0.68 0.033
N
2 22.5 74.5E 185 0.78 0.068 35 25.5N 82.5E 188 0.549 -
N 0.137
3 22.5 75.5E 178 0.662 -0.028 36 26.5N 73.5E 203 0.625 0.018
N
4 22.5 76.5E 180 0.61 0.039 37 26.5N 74.5E 198 0.735 0.156
N
5 22.5 77.5E 186 0.665 -0.069 38 26.5N 75.5E 200 0.744 0.187
N
6 22.5 78.5E 177 0.752 0.096 39 26.5N 76.5E 195 0.749 0.008
N
7 22.5 79.5E 181 0.768 0.195 40 26.5N 77.5E 196 0.767 0.09
N
8 23.5 73.5E 194 0.678 0.4 41 26.5N 78.5E 196 0.74 -
N 0.055
9 23.5 74.5E 187 0.796 0.054 42 26.5N 79.5E 193 0.787 0.112
N
10 23.5 75.5E 186 0.787 0.235 43 26.5N 80.5E 192 0.691 -
N 0.064
11 23.5 76.5E 187 0.677 -0.145 44 26.5N 81.5E 186 0.631 -
N 0.203

144
12 23.5 77.5E 182 0.632 -0.297 45 27.5N 74.5E 207 0.714 -
N 0.011
13 23.5 78.5E 182 0.632 -0.02 46 27.5N 75.5E 198 0.712 0.015
N
14 23.5 79.5E 182 0.662 0.056 47 27.5N 76.5E 196 0.687 -
N 0.056
15 23.5 80.5E 181 0.761 0.072 48 27.5N 77.5E 201 0.789 0.124
N
16 24.5 72.5E 200 0.647 0.195 49 27.5N 78.5E 196 0.812 -
N 0.053
17 24.5 73.5E 191 0.642 -0.101 50 27.5N 79.5E 193 0.73 0.037
N
18 24.5 74.5E 190 0.862 0.315 51 27.5N 80.5E 190 0.672 -
N 0.177
19 24.5 75.5E 189 0.824 0.372 52 28.5N 75.5E 203 0.704 0.057
N
20 24.5 76.5E 191 0.706 -0.099 53 28.5N 76.5E 204 0.678 -
N 0.161
21 24.5 77.5E 189 0.772 0.203 54 28.5N 77.5E 194 0.616 -
N 0.154
22 24.5 78.5E 187 0.761 0.063 55 28.5N 78.5E 193 0.832 0.277
N
23 24.5 79.5E 188 0.745 0.117 56 29.5N 76.5E 205 0.741 -
N 0.147
24 24.5 80.5E 188 0.746 0.054 57 29.5N 77.5E 190 0.65 -
N 0.084
25 24.5 81.5E 186 0.748 0.127 58 29.5N 78.5E 188 0.448 -
N 0.068
26 25.5 73.5E 195 0.621 -0.17 59 30.5N 76.5E 194 0.677 -
N 0.042
27 25.5 74.5E 196 0.75 -0.073 60 30.5N 77.5E 184 0.692 0.094
N
28 25.5 75.5E 196 0.759 0.129 61 30.5N 78.5E 173 0.687 0.13
N
29 25.5 76.5E 191 0.717 0.048 62 31.5N 76.5E 181 0.712 0.078
N
30 25.5 77.5E 189 0.694 0.041 63 31.5N 77.5E 180 0.664 0.082
N
31 25.5 78.5E 191 0.727 0.089 64 31.5N 78.5E 197 0.479 -
N 0.129
32 25.5 79.5E 189 0.697 -0.026 65 31.5N 79.5E 186 0.683 -0.05
N
33 25.5 80.5E 192 0.734 0.142
N
a
Values bold: dependence at 10% significance level; bold & italic: dependence at 5% significance
level. The lower and upper critical limit for r1 are: r1(l) = -0.248 and r1(u) = 0.208.

145
Table 14.2. Spatial summary of variables of Yamuna river basin
S. Variable Mean Value Hursts K R1
No Spatial Range SD Spatial Range SD Spatial Range SD
mean (mm) (mm) mean mean
(mm)
1 Annual 893.3 372.2- 280.3 0.652 0.436 - 0.088 0.039 (-0.329) - 0.199
rainfall 1653.5 0.887 0.591
2 Monsoon 806.8 327.0 - 259.7 0.653 0.407 - 0.097 0.028 (-0.386) - 0.220
rainfall 1568.9 0.880 0.557
3 Non- 86.5 19.5 - 100.1 0.656 0.516 0.062 0.077 (-0.207) - 0.149
monsoon 479.0 0.767 0.493
rainfall
4 Annual 88 46.0 - 25 0.678 0.476 - 0.073 0.105 (-0.118) - 0.125
rainydays 160.0 0.881 0.456
5 Monsoon 70 43 - 96 15 0.664 0.519 - 0.060 0.025 (-0.200) - 0.118
rainydays 0.821 0.416
6 Non- 17 1 - 72 14 0.695 0.534 - 0.069 0.132 (-0.194) - 0.139
monsoon 0.882 0.547
rainydays
7 OEM 190 173 - 7 0.704 0.448 - 0.074 0.031 (-0.297) - 0.138
207 0.862 0.400
8 AI 38.0 14.9 - 12.9 0.650 0.445 - 0.085 0.038 (-0.334) - 0.197
74.6 0.883 0.615

For all the tests, the tests were conducted at 5% and 10% significance levels
as these significance levels are quite appropriate for engineering practices (Lye and
Lin 1994). The results obtained for these two levels were compared and summarized
in Table 14.3 for sample variables. Different tests for independence have been
designed for different assumptions and conditions, and do not have equal power to
discriminate between the time series (Wall and Englot 1985; Lye and Lin 1994).
Sometimes various tests can give different results for the same time series. This can
be clearly seen from Table 14.3. Because of this, it is difficult to conclude whether the
time series is said to be independent or not. Under such circumstances, similar criteria
proposed by Lye and Lin (1994) were applied, i.e., if 4 out of 12 tests failed the test
of independence, the series is considered to be dependent (i.e. not-random). Based on
all the tests performed, the number of grid stations showing the significant
dependence is presented (Table 14.4), and it is clear that no time series has passed all
the tests.

Apart from this, it can be clearly seen from Table 14.1, in which only Hursts
K and r1 values are shown along with the means of the variables that there is no exact
relationship between Hursts K and r1. However, most of the time series that showed
significant r1 also showed significant long term dependence, which is expected
sometimes. To clarify this observation, a comparison of short term and long term
dependence tests was made which is presented in Table 14.5. It can be seen from the
Table 14.5 that time series of rainfall including rainydays and AI having long term
dependence can pass short term dependence tests, but it is not the case for OEM. It is
also observed from Table 14.5 that the number of time series having only long term
dependence is fewer than the number of time series having both long- and short-term
dependence for all the considered variables except for non-monsoon rainfall which

146
seems to reflect the existence of some relationship between the short term and long
term dependencies for climatic time series.

Table 14.3. Dependence as a function of test for onset of effective monsoon (OEM)
and annual aridity index (AI) series over Yamuna river basin
Number of Series (Percentage) Indicating
Dependence out of 65 series
S. No. Test
OEM AI
5% 10% 5% 10%
Short-term dependence: Non-parametric
1 Median crossing test 3 (4.62) 4 (6.15) 10 (15.38) 15 (23.08)
2 Turning Point 7 (10.77) 12 (18.46) 5 (7.69) 6 (9.23)
3 Rank difference 5 (7.69) 7 (10.77) 13 (20.00) 15 (23.08)
4 Cumulative periodogram 12 (18.46) 16 (24.62) 12 (18.46) 17 (26.15)
5 Wald-Wolfowitz 4 (6.15) 5 (7.69) 11 (16.92) 16 (24.62)
6 Rank Von Neumann 5 (7.69) 7 (10.77) 10 (15.38) 16 (24.62)
7 Kendal rank 26 (40.00) 35 (53.85) 8 (12.31) 10 (15.38)
8 Run test 5 (7.69) 11 (16.92) 11 (16.92) 16 (24.62)
9 Spearman Rho 32 (49.23) 40 (61.54) 8 (12.31) 13 (20.00)
Short-term dependence: Non-parametric
10 Autocorrelation 4 (6.15) 5 (7.69) 11 (16.92) 16 (24.62)
11 Von Neumann Ratio 5 (7.69) 7 (10.77) 12 (18.46) 16 (24.62)
Long-term dependence: Non-parametric
12 Hursts K test 17 (26.15) 27 (41.54) 7 (10.77) 12 (18.46)

Based on the above analysis of short term and long term dependence,
following remarks can be made based on the time series data of Yamuna basin:
(i) Most of the time series showing significant r1 when there is significant
long term dependence. This is not unusual as short term dependence can give
rise to some long term dependence.
(ii) Most of the time series indicating long term dependence can pass for
short term independence also except for OEM.

14.4.3 Analysis of Trend Test

All the three tests were applied to the entire range of considered variables.
The sample test results are given in Table 14.6 for OEM. Table 14.6 also includes the
values of lag-1 serial correlation along with their lower and upper limits at 95%
percent probability level. The results for all the variables are summarized in Table
14.7. The spatial variation of lag-1 serial correlation of the considered climatic series
is depicted in Figure 14.1. Figure 14.1 also comprise of spatio-temporal variation of
trend in climatic variables except the number of rainydays. The modified Mann-
Kendalls Z-statistic was used for depicting the spatio-temporal trend in the climatic
series with significance level at 10 percent. To present the, spatio-temporal variation
in the trend, the Mann-Kendalls Z-statistic value was spatially interpolated using the
Inverse Distance Weighted (IDW) technique in ArcGIS 9x. The interpolated raster
surface is based on a weighted average of the location. The value of each cell is

147
influenced mostly by nearby points and less by more distant stations. The IDW
method works on the power parameter and a search radius. The power parameter
controls the significance of calculated station values on the interpolated values. A
high power value gives more emphasis on the nearest points, and the resulting surface
will have more detail. In the present analysis, a power of six was fixed with quadrant
search radius.

Based on the analysis, it was observed that annual rainfall and monsoon
rainfall shows a declining trend whereas rising trend was observed in case of the non-
monsoon rainfall. A similar result was also observed in case of rainydays. The
number of rainydays and rainfall magnitude during the monsoon is declining. The
overall mean Z-statistic for monsoon rainfall and monsoon rainydays was 0.874
and 0.62 , respectively. However, trend statistics for annual rainfall pattern and
annual rainydays were obtained as 0.832 and 0.362 , respectively. It was
evident from Figure 14.1 that rainfall is increasing during the non-monsoon period.
The average Z - statistic for non-monsoon rainfall and rainydays were +0.209 and
+0.258, respectively. The critical limit at 95% confidence level is 1.645. OEM is
dependent on the rainfall pattern and potential evapotranspiration, and both the
variable has shown a general falling trend in the basin. Therefore, the increasing trend
in the OEM was the resulting effect of the rainfall and evapotranspiration patterns.
The overall mean Z-statistic for the OEM was +1.81. Similarly, AI is the ratio of
annual rainfall and potential evapotranspiration. Overall falling trend in the rainfall
pattern in the Yamuna basin has been observed (Figure 14.1). Trend analysis of AI
shows the falling trend in the Yamuna basin, which confirms the overall soil moisture
deficit in the Yamuna basin. The overall mean Z-statistic value for AI was -0.78. A
summary of results for the Yamuna basin is seen in Table 14.7.

Based on the analysis, the following remarks can be made:

(i) The presence of serial correlation in the time series significantly
affects the Mann-Kendalls trend analysis.
(ii) The original Mann-Kendall test overestimated the presence of
significant trend in the series than the modified Mann-Kendall and Mann-
Kendall with Pre-whitening tests. Based on the overall trend results, the
original Mann-Kendall test resulted in approximately 37% more significant
trend than the modified Mann-Kendall test.

14.4.4 Analysis of Persistence and Periodicity

Persistence Analysis. For climatic variability and changes, the definition of

persistence given by WMO (1966) is very common. According to this definition,
persistence is a tendency for successive values of the series to remember their
antecedent values, and to be influenced by them. The value of r1 has been used to
detect the possible persistence in the observed year-to-year variations of normalized
anomaly series and to examine its nature and magnitude. The approach proposed by
WMO (1966) and Matalas (1967) was widely used later in many studies related to
long-term climatic variations (e.g. Rodhe and Virji 1976; Granger 1977; Ogallo 1979;

148
Figure 14.1. Spatial pattern of lag-1 serial correlation coefficient and Mann-
Kendalls Z-Statistic

149
Figure 14.1. (Continued)

Anyadike 1993; Drosdowsky 1993; Nicholson and Palao 1993; Trkes 1998; 1999;
Trkes et al. 2002). Persistence is evident in long series of climatic observations
characterized by a positive serial correlation. Significant negative r1 are very likely
to be indicative of high-frequency oscillations, whereas significant positive r1 is
likely to be indicative of low-frequency fluctuations and persistence in climatic
series. In the study, serial correlation coefficient for all the lags L 0 to m for all
the variables, where m N / 3 . However, serial correlation coefficients up to lag 3
was assessed. Serial correlation coefficient up to lag-3 is plotted for all the climatic
variables except non-monsoon, although all the variables were analyzed.
Approximately 15 percent of grid points showed significant positive lag-1 serial
correlation coefficient in the annual rainfall series which indicated the presence of
persistence. The statistical significance of serial correlation was assessed at 90%
significance level. Little positive spatial coherence for the stations characterized by
statistically significant lag-1 serial correlation coefficient was identified. A similar
spatial variation in the time series was observed in annual and monsoon rainfall,
annual rainy days, and AI.
150
Table 14.4. Number of time series passing the test of dependence out of 65 series
No. of Test Rainfall Rainydays OEM Aridity Index (AI)
Indicating Annual Monsoon Non-monsoon Annual Monsoon Non-monsoon
Dependence 5% 10% 5% 10% 5% 10% 5% 10% 5% 10% 5% 10% 5% 10% 5% 10%
10 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
9 2 3 3 4 0 1 0 2 0 1 1 3 1 0 2 3
8 2 5 4 5 2 1 2 4 1 3 3 9 1 2 2 5
7 4 9 6 10 2 5 4 7 3 4 7 15 3 5 5 8
6 8 11 7 12 7 10 6 10 3 5 11 17 3 5 5 10
5 12 13 11 17 7 11 7 13 3 6 14 21 5 7 11 13
4 13 17 13 21 9 13 11 16 3 6 18 28 6 9 11 19
3 13 20 18 27 12 19 13 19 6 10 24 33 9 15 14 20
2 21 28 25 35 18 25 23 30 12 28 31 38 31 37 21 28
1 26 33 28 38 22 28 31 37 21 36 39 48 37 43 28 33
0 39 32 37 27 43 36 34 27 44 29 26 16 28 21 37 32
Hursts K 7 12 10 16 4 7 9 15 5 10 15 21 17 27 7 12

Table 14.5 Comparison of short-term and long-term dependence in the climatic variables
Rainfall Rainydays OEM Aridity Index (AI)
Annual Monsoon Non-monsoon Annual Monsoon Non-monsoon
5% 10% 5% 10% 5% 10% 5% 10% 5% 10% 5% 10% 5% 10% 5% 10%
Short-term 13 17 13 21 9 13 11 16 3 6 18 28 6 9 11 19
Long-term 7 12 10 16 4 7 9 15 5 10 15 21 17 27 7 12
Only short-term 8 9 7 10 9 12 5 9 1 2 10 13 2 3 7 11
Only long-term 0 0 1 1 3 3 2 4 0 1 0 2 2 1 0 0
Both short &
5 8 6 11 0 2 6 8 2 4 8 16 4 7 4 8
long-term

151
Table 14.6. Lag-1 serial correlation and Mann-Kendalls Z statistics of OEM
of the Yamuna river basin [r1(l) = -0.248, r1(u) = 0.208]
ID r1 Z-MK ID r1 Z-MK
Org Rank PW Org Rank PW
1 0.099 1.320 1.321 1.321 34 0.020 2.647 2.648 2.648
2 0.084 1.785 1.785 1.785 35 -0.118 1.358 1.359 1.359
3 0.050 1.335 1.335 1.335 36 0.042 -1.153 -1.154 -1.154
4 0.028 1.990 1.991 1.991 37 0.172 0.569 0.569 0.569
5 0.001 0.917 0.918 0.918 38 0.155 2.559 2.559 2.559
6 0.168 2.354 2.356 2.356 39 -0.036 1.991 1.993 1.993
7 0.234 2.110 2.111 1.184 40 0.144 3.002 3.003 3.003
8 0.117 3.776 0.056 3.777 41 -0.188 2.606 2.607 2.607
9 0.035 2.678 2.679 2.679 42 0.085 3.502 3.507 3.507
10 0.323 2.472 2.473 1.097 43 -0.065 2.141 2.141 2.141
11 -0.136 1.517 1.519 1.519 44 -0.194 0.924 0.924 0.924
12 -0.291 1.146 0.031 0.963 45 -0.007 0.079 0.079 0.079
13 -0.101 1.406 1.407 1.407 46 0.032 1.193 1.193 1.193
14 0.014 1.763 1.766 1.766 47 -0.069 0.371 0.372 0.372
15 0.037 2.433 2.433 2.433 48 0.003 3.270 3.271 3.271
16 0.074 1.998 1.999 1.999 49 -0.025 3.341 3.343 3.343
17 -0.097 1.240 1.241 1.241 50 0.026 1.808 1.808 1.808
18 0.307 3.658 0.020 1.965 51 -0.141 1.011 1.012 1.012
19 0.320 2.441 -0.025 1.515 52 0.046 1.125 1.131 1.131
20 -0.043 1.651 1.652 1.652 53 -0.100 0.537 0.537 0.537
21 0.019 3.792 3.793 3.793 54 -0.153 0.134 0.134 0.134
22 0.051 3.160 3.162 3.162 55 0.194 4.447 0.000 4.448
23 0.136 3.106 3.108 3.108 56 -0.047 1.047 1.096 1.096
24 0.004 2.782 2.786 2.786 57 -0.071 1.936 1.937 1.937
25 0.167 2.938 2.939 2.939 58 0.024 0.672 0.673 0.673
26 -0.157 0.537 0.538 0.538 59 0.009 1.698 1.699 1.699
27 -0.106 1.509 1.509 1.509 60 0.051 2.156 2.157 2.157
28 0.115 2.566 2.566 2.566 61 0.118 0.387 0.387 0.387
29 0.030 2.812 2.813 2.813 62 0.158 2.267 2.268 2.268
30 0.030 2.173 2.174 2.174 63 0.115 1.216 1.216 1.216
31 0.096 3.238 3.239 3.239 64 -0.056 -0.348 -0.348 -0.348
32 -0.104 3.100 3.107 3.107 65 -0.035 0.963 0.963 0.963
33 0.014 2.964 2.966 2.966

Table 14.7. Summary of trend analysis based on Modified Mann-Kendall test

S. Variable No. of Series No. of Series Indicating Significant
No Indicating Rising Falling No Overall
significant lag-1 Trend Trend Trend Trend
serial correlation
1 Annual rainfall 16 6 (8) 59 (57) - ve
2 Monsoon rainfall 18 8 (13) 57 (52) - ve
3 Non-monsoon rainfall 13 3 (2) 62 (63) + ve
4 Annual rainydays 17 4 (2) 7 (11) 54 (52) - ve
5 Monsoon rainydays 5 5 (6) 60 (59) - ve
6 Non-monsoon rainydays 22 6 (2) 1 (2) 58 (61) + ve
7 OEM 5 29 (30) 36 (35) + ve
8 Aridity Index (AI) 17 6 (8) 59 (57) - ve
(Values in the parenthesis are the results from Mann-Kendall with Pre-whitening test).

152
 Detailed tables of the serial correlation coefficients up to lag-3 are
estimated. The sample table is presented in Table 14.8. There are very few
coherent area with significant negative lag-1 serial coefficient over the basin
was observed. Looking into the figures of lag-1 and lag-2 serial correlation
pattern, it was indicated that coherent area of significant positive lag-2 serial
correlation are greater than that of the lag-1. The lag-2 serial correlation
coefficients are mostly positive but insignificant for all the variables except
for the non-monsoon rainfall, non-monsoon rainydays and OEM.

Table 14.8. Results of serial correlation and power spectrum for OEM
ID r1 r2 r3 L Conf. Cont. ID r1 r2 r3 L Conf. Cont.
Level Level
90 95 90 95
1 0.10 0.06 0.08 4 8.5 0.0 WN 36 0.04 - 0.09 WN
0.16
2 0.08 - 0.38 10 3.4 3.4 WN 37 0.17 0.08 0.16 1 34.0 0.0 WN
0.07
3 0.05 - 0.42 1 34.0 0.0 WN 2 17.0 0.0
0.03
2 17.0 0.0 38 0.16 0.11 0.08 2 17.0 0.0 WN
12 2.8 0.0 39 - 0.14 0.10 14 2.4 0.0 WN
0.04
13 2.6 2.6 40 0.14 0.02 0.33 1 34.0 0.0 WN
4 0.03 0.06 0.12 WN 9 3.8 0.0
5 0.00 - - 7 4.9 4.9 WN 41 - - 0.22 13 2.6 0.0 WN
0.18 0.03 0.19 0.02
8 4.3 4.3 14 2.4 0.0
6 0.17 0.07 0.06 1 34.0 0.0 WN 42 0.09 0.29 0.29 1 34.0 0.0 WN
7 0.23 - - 6 5.7 0.0 WN 13 2.6 0.0
0.04 0.03
9 3.8 0.0 43 - - 0.17 9 3.8 3.8 WN
0.07 0.13
8 0.12 - - WN 13 0.0 0.0
0.09 0.06
9 0.04 0.13 0.34 1 34.0 34.0 WN 44 - - - 14 2.4 2.4 WN
0.19 0.14 0.02
10 0.32 0.13 0.12 7 4.9 0.0 RN 45 - - 0.19 WN
0.01 0.01
11 - - 0.01 WN 46 0.03 - 0.10 WN
0.14 0.05 0.08
12 - 0.06 - 15 2.3 0.0 WN 47 - 0.00 - 8 4.3 4.3 WN
0.29 0.12 0.07 0.04
16 2.1 2.1 9 3.8 0.0
17 2.0 2.0 48 0.00 0.19 0.32 1 34.0 0.0 WN
13 - - - 9 3.8 0.0 WN 13 2.6 0.0
0.10 0.21 0.02
10 3.4 3.4 49 - 0.29 0.19 13 2.6 0.0 WN
0.03
14 0.01 - 0.17 8 4.3 0.0 WN 50 0.03 - 0.22 9 3.8 3.8 WN
0.19 0.10
9 3.8 3.8 10 3.4 0.0
10 3.4 0.0 51 - - 0.09 9 3.8 0.0 WN
0.14 0.11
15 0.04 0.03 0.06 WN 13 2.6 0.0
16 0.07 - - WN 52 0.05 0.19 0.17 1 34.0 34.0 WN
0.02 0.11
17 - - 0.21 10 3.4 0.0 WN 2 17.0 0.0
0.10 0.15
11 3.1 3.1 53 - - 0.20 11 3.1 0.0 WN
0.10 0.02
12 2.8 0.0 54 - - 0.13 6 5.7 0.0 WN
0.15 0.19
18 0.31 0.29 0.30 RN 11 3.1 3.1
19 0.32 0.04 - 8 4.3 4.3 12 2.8 2.8
0.01

153
20 - 0.04 0.14 11 3.1 0.0 WN 13 2.6 0.0
0.04
21 0.02 0.04 0.21 WN 55 0.19 0.30 0.07 1 34.0 34.0 WN
22 0.05 - 0.35 9 3.8 3.8 WN 56 - 0.06 0.08 WN
0.10 0.05
23 0.14 - 0.24 9 3.8 0.0 WN 57 - 0.33 0.07 2 17.0 0.0 WN
0.06 0.07
24 0.00 - 0.24 WN 15 2.3 2.3
0.13
25 0.17 - 0.18 8 4.3 4.3 WN 58 0.02 0.03 - 4 8.5 8.5 WN
0.15 0.25
9 3.8 3.8 5 6.8 6.8
26 - - - 10 3.4 0.0 WN 59 0.01 0.13 0.08 13 2.6 0.0 WN
0.16 0.05 0.08
27 - 0.03 0.15 13 2.6 2.6 WN 60 0.05 - - 8 4.3 4.3 WN
0.11 0.28 0.04
28 0.12 0.10 0.15 WN 9 3.8 3.8
29 0.03 0.12 0.14 14 2.4 2.4 WN 61 0.12 0.12 - 2 17.0 0.0 WN
0.13
30 0.03 0.09 0.10 14 2.4 0.0 WN 62 0.16 - 0.02 WN
0.03
31 0.10 0.13 0.22 WN 63 0.12 - - 5 6.8 0.0
0.19 0.08
32 - 0.26 0.14 13 2.6 2.6 WN 6 5.7 5.7 WN
0.10
33 0.01 0.20 0.21 12 2.8 0.0 WN 64 - - 0.08 8 4.3 0.0 WN
0.06 0.19
13 2.6 2.6 9 3.8 3.8
34 0.02 - 0.17 8 4.3 0.0 WN 65 - - 0.13 9 3.8 0.0 WN
0.21 0.04 0.14
9 3.8 0.0
35 - - 0.20 9 3.8 3.8 WN
0.12 0.25
10 3.4 3.4
WN: white noise, RN: Markov red noise

By analyzing the lag-3 serial correlation coefficients of the climatic

variables, it was evident that the percentage coherent area is significantly
reduced for all the variables. However, it is increase in case of non-monsoon
rainfall, non-monsoon rainydays and OEM. Based on the spatial distribution
of lag-1 serial correlation coefficient, it is indicated that a simple Markov type
persistence is experienced in the upper Himalayan, central and south-west
region of the Yamuna basin.

Periodicity. Significance of the spectral estimates was evaluated at 90

and 95 percent confidence levels of the appropriate null continuum (red or
white noise). If a time series has persistence, the spectrum over all the
wavelengths and the magnitude of the spectrum has a decreasing trend from
long to short wavelengths; and the spectrum is termed as red noise. For the
spectrum having necessary exponential relationships among r1 , r2 and r3
for simple Markov type persistence, the appropriate null hypothesis was
assumed to be a Markov red noise continuum. A series characterized by an
insignificant positive lag-1 serial correlation coefficient or a series that has a
significant positive lag-1 serial correlation coefficient but not a simple
Markov-type, and any series with negative lag-1 serial correlation coefficient
was evaluated as a white noise continuum. For all the variables, power
spectrum plots for representative stations in the Yamuna basin are prepared.
The sample plot for the OEM is presented in Figure 14.2.

154
 (10)
(19)

(27) (31)

(33)
(40)

(42)
(47)

(49) (57)

(61)
(64)
Figure 14.2. Power spectrum plot for OEM (solid line and dashed line
represents the 95% and 90% confidence limits, respectively)

155
 To evaluate the power spectrum results, generally periodicity values
computed by the equation P (= 2 m / L) for all the time series. The periodicity
along with their significance level was also computed (for examples refer
Table 14.8). Table 14.8 also comprised of the serial correlation coefficients up
to lag-3. Based on the Table 14.8, it is evident that the short-term period
fluctuation of 2.0 to 4.9 is dominant in the annual and monsoon rainfall,
whereas for non-monsoon rainfall medium and long-term period of 5.0 to
34.0 is dominant. This feature of rainfall pattern over the Yamuna basin
should be kept in mind while preparing the water resources plan. In case of
number of rainydays, the long-period is dominated over the short- periodicity.
Results of the power spectrum analysis of OEM and AI are indicated that
OEM and AI have dominating short-period fluctuations of 2.0 to 3.0 years,
which leads to high probability of meteorological and agricultural drought and
will frequently disturbs the planning of kharif crops in the Yamuna basin.

Based on the above analysis on periodicity, a good consistency

between the lag-1 serial correlation coefficient and results of the power
spectrum analysis was remarked. Positive lag-1 serial correlation coefficients
gives low frequency fluctuations, whereas negative is an indicator for the high
frequency.

14.5 Conclusions

This chapter has demonstrated the various statistical tests for

investigating the short-term and long-term dependence, trend analysis, and
periodicity in the hydrological time series.

The short term dependence is very important for small water resources
projects, which requires less period of hydrologic data. For short-term
dependence, 10 tests were used and hypothesis test statistics was performed at
5% and 10% significance level. However, it may be questioned that why so
many of tests have been applied. The answer may be that since the analysis is
based on the statistics, therefore, to ascertain the short-term dependence or
short term persistence in the series, it will be better option to establish through
various tests followed by personal judgment. In the analysis, if 4 tests out of
10 have been passed through the test, then the series is said to have
persistence.

Similar to the short term dependence, when water resources project is

planned based on the longer series (say up to 500 years), then it is required to
test the long term dependence or persistence. In this chapter, only non-
parametric test (i.e. Hurst coefficient followed by boot-strap sampling of
10000 samples) was successfully demonstrated.

156
 Once the persistence test (i.e. short- and long-term dependence) is
over, the series is subjected to the trend analysis. For trend analysis, a well-
established Mann-Kendall test has been demonstrated with their limitation. It
was appeared in the analysis, that presence of persistence plays a vital role in
the trend analysis. It may be established that either modified Mann-Kendall or
original Mann-Kendall test with removal of persistence (represented by lag-1
serial correlation) should be used to investigate the trend in hydrologic time
series.

In last, methodology for the periodicity in the hydrologic time series

has been demonstrated. Again the persistence in the time series has great
impact on the periodicity.

14.6 References

Angell, J. K., Korshover, J., and Carpenter, T. H. (1966). Note concerning

the period of the quasi-biennial oscillation. Monthly Weather Review,
94, 319323.
Anyadike, R. N. C. (1993). Seasonal and annual rainfall variations over
Nigeria. Int. J. Climatol., 13, 567580.
Ashokraj, P. C. (1979). Onset of effective monsoon and critical dry spell-A
computer based forecasting technique. Indian Agricultural Research
Institute, ICAR, New Delhi, Bulletin No. 11, India.
Aziz, O. I. A., and Burn, D. H. (2006). Trends and variability in the
hydrological regime of the Mackenzie River Basin. J. Hydrol., 319
(14), 282294.
Blackman, R. B., and Tukey, J. W. (1958). The measurement of power
spectra. Dover Publications, New York.
Box, G. E. P., and Jenkins, G. M. (1976). Time Series Analysis: Forecasting
and Control. Holden-Day, San Francisco, CA.
Burn, D. H., and Elnur, M. A. H. (2002). Detection of hydrologic trends and
variability. J. Hydrol., 255, 107122.
Burn, D. H., Cunderlink, J. M., and Pietroniro, A. (2004). Hydrological
trends and variability in the Liard river basin. Hydrol. Sci. J., 49 (1),
5367.
Chen, H., Guo, S., Xu, C. Y., and Singh, V. P. (2007). Historical temporal
trends of hydro-climatic variables and runoff response to climate
variability and their relevance in water resource management in the
Hanjiang basin. J. Hydrol., 344, 171184.
Cox, D. R., and Stuart, A. (1955). Some quick tests for trend in location and
dispersion. Biometrika, (42), 8095.
Douglas, E. M., Vogel, R. M., and Kroll, C. N. (2000). Trends in floods and
low flows in the United states: Impact of spatial correlation. J.
Hydrol., 240 (12), 90105.

157
Drosdowsky, W. (1993). An analysis of Australian seasonal rainfall
anomalies: 19501987. II: Temporal variability and teleconnection
patterns. Int. J. Climatol., 13, 111149.
Efron, B. (1979). Bootstrap methods: another look at the jackknife. Ann.
Statist., 7, 126.
Efron, B. (1981). Nonparametric standard errors and confidence intervals.
Can. J. Statist., 9, 139172.
Efron, B. (1982). The jackknife, the bootstrap, and other resampling plans.
SIAM (Soc. Ind. Appl. Math.) Monogr., 38.
Fisz, M. (1963). Probability Theory and Mathematical Statistics. Wiley, New
York.
Gan, T. Y. (1998). Hydroclimatic trends and possible climatic warming in the
Canadian Prairies. Water Resources Research, 34(11), 30093015.
Granger, O. E. (1977). Secular fluctuations of seasonal precipitation in
lowland California. Mon Weather Rev., 105, 386397.
Hirsch, R. M., and Slack, J. R. (1984). A nonparametric trend test for
seasonal data with serial dependence. Water Resour. Res., 20(6), 727
732.
Hirsch, R. M., Slack, J. R., and Smith, R. A. (1982). Techniques of trend
analysis for monthly water quality data. Water Resour. Res., 18(1),
107121.
Hurst, H. E. (1951). Long term storage capacity of reservoirs. Trans. Am.
Soc. Civ. Eng., 116, 770808.
Hurst, H. E. (1956). Methods of using long-term storage in reservoirs.
Proc. Inst. Civ. Eng., 1, 519543.
Jenkins, G. M., and Watts, D. G. (1968). Spectral analysis and its
applications. Holden-Day, San Francisco.
Julian, P. R. (1967). Variance spectrum analysis. Water Resour. Res., 3,
831845.
Kendall, M., and Stuart, A. (1976). The Advanced Theory of Statistics, Vol. 3,
Charles Griffin, London.
Kendall, M. G. (1955). Rank Correlation Methods. Charles Griffin, London.
Khaliq, M. N., Ouarda, T. B. M. J., Gachon, P., Sushma, L., and St-Hilaire, A.
(2009). Identification of hydrological trends in the presence of serial
and cross correlations: A review of selected methods and their
application to annual flow regimes of Canadian rivers. J. Hydrol.,
368, 117130.
Landsberg, H. E., Mitchell, J. M. Jr., Crutcher, H. L., and Quinlan, F. T.
(1963). Surface signs of the biennial atmospheric pulse. Monthly
Weather Review, 91 (1012), 549556.
Lettenmaier, D. P., Wood, E. F., and Wallis, J. R. (1994). Hydro-
climatological trends in the continental United States 19481988. J.
Clim., 7 (4), 586607.
Lins, H. F., and Slack, J. R. (1999). Streamflow trends in the United States.
Geophys. Res. Let., 26 (2), 227230.

158
Lye, L. M., and Lin, Y. (1994). Long-Term Dependence in Annual Peak
Flows of Canadian Rivers. Journal of Hydrology, 160, 89103.
Madansky, A. (1988). Prescriptions for Working Statisticians. Springer, New
York, 93118.
Mann, H. B. (1945). Non-parametric tests against trend. Econometrica, 13,
245259.
Matalas, N. C. (1967). Time series analysis. Water Resour. Res., 3, 817
829.
Meacham, I. (1968). Correlation in sequential datathree sample indicators.
Civ. Eng. Trans. Inst. Eng. Aust., CE10(2), 225228.
Nicholson, S. E., and Palao, I. M. (1993). A re-evaluation of rainfall vari-
ability in the Sahel. Part I. Characteristics of rainfall fluctuations. Int.
J. Climatol., 13, 371389.
Ogallo, L. (1979). Rainfall variability in Africa. Mon. Weather Rev., 107,
11331139.
Rai, R. K., Upadhyay, A., and Ojha, C. S. P. (2010). Temporal variability of
climatic parameters of Yamuna River basin: spatial analysis of
persistence, trend and periodicity. The Open Hydrol. J., 4, 184-210.
Rodhe, H., and Virji, H. (1976). Trends and periodicities in East African
rainfall data. Mon. Weather Rev., 104, 307315.
Tukey, J.W. (1950). The sampling theory of power spectrum estimates. In:
Symposium on applications of autocorrelation analysis to physical
problems. US Office of Naval Research, NAVEXOS-P-735,
Washington, DC, p. 4767.
Trkes M. (1998). Influence of geopotential heights, cyclone frequency and
southern oscillation on rainfall variations in Turkey. Int. J. Climatol.,
18, 649680.
Trkes M. (1999). Vulnerability of Turkey to desertification with respect to
precipitation and aridity conditions. Tr. J. Eng. Environ. Sci., 23,
363380.
Trkes M., Smer, U. M., and Kili, G. (2002). Persistence and periodicity
in the precipitation series of Turkey and associations with 500 hPa
geopotential heights. Climatic Research, 21, 5981.
UNEP (1993). World Atlas of Desertification. United Nations Environment
Programme (UNEP), London, 1993.
Wald, A., and Wolfowitz, J. (1943). An exact test for randomness in the non-
parametric case based on serial correlation. Ann. Math. Statist., 14,
378388.
Wall, D. J., and Englot, M. E. (1985). Correlation of annual peak flows for
Pennsylvania systems. Water Resour. Bull., 21(3), 459264.
Wiener, N. (1930). Generalized harmonic analysis. Acto Math, 55, 117258
Wiener, N. (1949). Extrapolation, interpolation, and smoothing of stationary
time series. Wiley and Technology Press, Cambridge, MA.
WMO (1966). Climatic change. WMO Technical Note 79, World
Meteorological Organization, Geneva.

159
Yevjevich, V. (1971). Stochastic Processes in Hydrology. Water Resources
Publications, Fort Collins, CO.
Yue, S., and Pilon, P. (2004). A comparison of the power of the t-test, Mann-
Kendall and bootstrap tests for trend detection. Hydrol. Sci. J., 49 (1),
2137.
Yue, S., Pilon, P., and Cavadias, G. (2002). Power of the Mann-Kendall and
Spearmans rho test for detecting monotonic trends in hydrological
series. J. Hydrol., 259, 254271.
Zhang, Q., Jiang, T., Gemmer, M., and Becker, S. (2005). Precipitation,
temperature and runoff analysis from 19952002 in the Yangtze basin,
China. Hydrol. Sci. J., 50 (1), 6580.

160