Chaos, Solitons and Fractals 19 (2004) 285–291

www.elsevier.com/locate/chaos

Fractal dimensional analysis of Indian climatic dynamics

Abstract
In this paper, we use fractal dimensional analysis to investigate the Indian climatic dynamics. We analyze time series
data of three major dynamic components of the climate––temperature, pressure and precipitation. We study how
climate variability changes from month to month and from one season to the other. We also investigate variability both
at a local level (for individual stations) and at a regional level (for groups of stations). Our studies suggest that regional
climatic models typically would not be able to predict local climate since they deal with averaged quantities. We find an
interesting effect that precipitation during the south-west monsoon is affected by temperature and pressure variability
during the preceding winter.
Ó 2003 Elsevier Ltd. All rights reserved.

1. Introduction

Fractal dimensional analysis of geophysical time series is a well established investigative tool for exploring the
dynamics. This was initiated by Mandelbrot and Wallis in their series of seminal papers [1–3] on this subject. This has
been followed up by application of the technique to various geophysical phenomena [4–6]. Fractal dimensional analysis
is particularly well suited to analyze the variability of a given time series. In this paper we concentrate on investigating
the Indian climatic dynamics through an analysis of the temperature, pressure and precipitation time series. As these
three variables form the dominant dynamic component of climate, our emphasis on them is well justified. Since the
Indian subcontinent lies at the heart of the classic monsoon region and is the area most sensitive to monsoon fluctu-
ations, fractal dimensional analysis of its climatic time series can be expected to yield some insights into how the Indian
climate variability is linked to monsoon fluctuations.
In a previous paper [7], using fractal dimensional analysis, we had introduced the concept of a climate predictability
index as a tool for studying climate dynamics and briefly explored its applications. In this paper, we undertake a de-
tailed analysis of the Indian climate dynamics using the predictability indices as the main tool. Predictability indices are
calculated for 31 recording stations spread throughout India. These indices are computed using temperature, pressure
and precipitation time series data for these stations available from the global historical climatology network (GHCN)
dataset [8]. We calculate these indices both for a season as a whole and for individual months. This enables us to study
the climate dynamics at two temporal resolutions. Next, we study individual stations as well as groups of stations. This
allows us to identify both local dynamics as well as regional dynamics.

2. Fractal dimensional analysis

We start with the assumption that climatic time series can be modeled as a stationary stochastic process. Consider a
stationary stochastic process in discrete time, fnk g, with hnk i ¼ 0 and hn2k i ¼ r2 . Here h i denotes ensemble average. If
the autocorrelation function CðnÞ ¼ hnk nkþn i scales with the lag n as

*
Corresponding author.
E-mail address: [email protected] (G. Rangarajan).

0960-0779/03/$ - see front matter Ó 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/S0960-0779(03)00042-0
286 G. Rangarajan, D.A. Sant / Chaos, Solitons and Fractals 19 (2004) 285–291

CðnÞ
nb ð2:1Þ

for large n, where 0 < b < 1, then fni g is called P a long range correlated or long memory process [9]. The reason for the
latter term is that CðnÞ decays so slowly that Nn¼1 CðnÞ diverges as N ! 1. It is well known [1,4–6,10] that most
geophysical time series and in particular the climatic time series that we wish to analyze exhibit this long memory
property.
A standard method to assess the correlation structure of fnk g is to convert the stationary process to a random walk
by using partial sums, R1 ¼ n1 ; R2 ¼ n1 þ n2 ; . . . ; Rn ¼ n1 þ n2 þ    þ nn ; . . ., where Rn is the position of the walker at
time n. The mean range of the random walk trajectory as a function of time bears specific relations with the scaling
relation Eq. (2.1). For ease of analytical evaluation we consider the mean square displacement as a measure of the range
of the random walk, which is defined as
X
n X
n1 X
n1 X
n1
hR2n i ¼ hn2i i þ 2 ðn  sÞCðsÞ ¼ nr2 þ 2n CðsÞ  2 sCðsÞ: ð2:2Þ
i¼1 s¼1 s¼1 s¼1

Let CðsÞ obey the scaling law in Eq. (2.1). The sums in the above equation are estimated as
Xn1 Xn Z n
CðsÞ
sb
sb
n1b ð2:3Þ
s¼1 s¼1 1

and
X
n1 X
n Z n
sCðsÞ
s1b
s1b  n2b : ð2:4Þ
s¼1 s¼1 1

For 0 < b < 1 this means

for large n. Conventionally, the mean square displacement is characterized by the Hurst exponent H as
hR2n i
n2H ; ð2:6Þ

where
H ¼ ð2  bÞ=2: ð2:7Þ

The random walk analysis tool that we will use in this paper is the rescaled range (R=S)
P analysis [1,10]. A brief de-
scription of this method follows. For a given dataset fni g, consider the sum Lðn; sÞ ¼ si¼1 nnþi , where Lðn; sÞ can be
regarded as the position of a random walk after s steps. Define the trend-corrected range Rðn; sÞ of the random walk as

Rðn; sÞ ¼ maxfLðn; pÞ  pLðn; sÞ=s; 1 6 p 6 sg  minfLðn; pÞ  pLðn; sÞ=s; 1 6 p 6 sg: ð2:8Þ

Let S 2 ðn; sÞ denote the variance of the dataset fnnþi gsi¼1 . If the data has long range correlation, the average rescaled
statistic QðsÞ ¼ hRðn; sÞ=Sðn; sÞin (where h in denotes the average over n) scales with s as a power law for large s:
QðsÞ
sH ; ð2:9Þ

where H is the Hurst exponent introduced earlier. This power law manifests itself as a straight line in the log–log plot of
QðsÞ vs. s.
The Hurst exponent is related to the fractal dimension D of the time series curve by the formula [11]
D ¼ 2  H: ð2:10Þ

If the fractal dimension D for the time series is 1.5, there is no correlation between amplitude changes corresponding to
two successive time intervals. Therefore, no trend in amplitude can be discerned from the time series and hence the
process is unpredictable. However, as the fractal dimension decreases to 1, the process becomes more and more pre-
dictable as it exhibits ‘‘persistence’’. That is, the future trend is more and more likely to follow an established trend [5].
As the fractal dimension increases from 1.5 to 2, the process exhibits ‘‘anti-persistence’’. That is, a decrease in the
amplitude of the process is more likely to lead to an increase in the future. Hence, the predictability again increases.
However, we will be concerned only with persistence behaviour since all geophysical time record analyzed till date [1,4–
6,10] exhibit this behaviour.
 G. Rangarajan, D.A. Sant / Chaos, Solitons and Fractals 19 (2004) 285–291 287

Table 1
Climate predictability index (PIC ) for 15 stations spread throughout India
Station SW NE
Allahabad (0.2,0.8,0.1) (0.6,0.2,0.0)
Bangalore (0.3,0.3,0.2) (0.5,0.1,0.1)
Bombay (0.5,0.7,0.5) (0.5,0.7,0.2)
Calcutta (0.9,0.7,0.2) (0.9,0.0,0.4)
Darjeeling (0.5,0.7,0.4) (0.6,0.7,0.1)
Dwaraka (0.3,0.2,0.2) (0.1,0.2,0.2)
Gauhati (0.5,0.4,0.1) (0.4,0.1,0.0)
Hyderabad (0.7,0.7,0.3) (0.8,0.5,0.2)
Kodaikanal (0.9,0.8,0.1) (0.6,0.8,0.3)
Madras (0.6,0.8,0.4) (0.1,0.7,0.0)
Nagpur (0.2,0.6,0.4) (0.2,0.2,0.0)
New Delhi (0.5,0.7,0.3) (0.3,0.7,0.4)
Shillong (0.5,0.2,0.5) (0.3,0.0,0.0)
Simla (0.8,0.2,0.2) (0.5,0.2,0.1)
Veraval (0.9,0.4,0.6) (0.4,0.4,0.1)
PIC is listed for two seasons––the south-west and north-east monsoon periods.

We obtain the fractal dimensions of the time series corresponding to temperature, pressure and precipitation for a
given location using R=S analysis to first obtain H and then Eq. (2.10). The fractal dimensions are denoted by DT , DP
and DR respectively. The R=S analysis is used merely because it has been the conventional technique used for geo-
physical time records [1,4–6,10]. Any other method would be equally adequate.
Predictability indices (denoted by PIT , PIP and PIR respectively) for temperature, pressure and precipitation are
defined as follows [7]:
PIT ¼ 2jDT  1:5j; PIP ¼ 2jDP  1:5j; PIR ¼ 2jDR  1:5j:

Here jDj denotes the absolute value of the number D. We use absolute values since predictability increases in both the
following cases––when the fractal dimension becomes less that 1.5 and when it becomes greater than 1.5. In the former
case, we have correlation (persistence) behaviour and in the latter case, anti-correlation (anti-persistence) behaviour.
However, in either case, the process becomes more predictable. Thus, use of absolute values ensures that a process with
D ¼ 1:3 has the same predictability index as a process with D ¼ 1:7.
The climate predictability index (PIC ) is defined as the collection of the above three indices:
PIC ¼ ðPIT ; PIP ; PIR Þ:

If one of these indices is close to zero, then the corresponding process approximates the usual Brownian motion and is
therefore unpredictable. If it is close to one, the process is very predictable. Note that the PIR value is not related to the
amount of precipitation, but to how precipitation changes from year to year. The rationale for introducing the climate
predictability index is as follows. In this paper, we are interested in studying the inter-relationships between the three
climatic components from the viewpoint of fractal dimensions. Hence, it is useful to have all three of them represented
in a single index. Then it is easier to see how the three sub-indices change in relation to one another as the seasons
change. Further, by introducing predictability indices instead of fractal dimensions, we focus on how predictable the
process is. This is especially useful for precipitation.
Data from 31 measuring stations spread throughout India were studied. The temperature, pressure and precipitation
time series for these stations were obtained from the GHCN dataset [8]. The dataset gives monthly mean values for the
above three climatic variables. Mean values for various seasons were obtained by averaging over the months consti-
tuting the season under study. These time series data were then studied using R=S analysis and predictability indices
calculated using the formulas given above.
Table 1 gives the predictability indices for temperature, pressure and precipitation respectively for a few represen-
tative stations. Predictability indices are listed for two seasons––south-west monsoon, north-east monsoon.

3. Data analysis

We now study how the predictability indices change with the seasons and within the seasons. We concentrate on the
following seasons––pre-south-west monsoon season (March–May), south-west monsoon season (June–September) and
288 G. Rangarajan, D.A. Sant / Chaos, Solitons and Fractals 19 (2004) 285–291

north-east monsoon season (October and November). Each of these seasons is studied separately below. Four stations
exhibit a consistent predictability pattern across the seasons. We list them first before going into a season-wise analysis.
Calcutta has a predictable temperature whereas Darjeeling, Hyderabad and Kodaikanal have a predictable pressure
throughout the year.

3.1. Pre-south-west monsoon season

First, we consider predictability indices calculated for this season as a whole. Fig. 1 gives the scatter plot of PIT vs.
PIP for this season. We see that the stations are equally distributed across all regions of the graph. Roughly, one-third of
the stations have unpredictable PIT and PIP . Even if one considers only those stations (concentrated mainly in the
north-eastern region) where it rains significantly during this season, they do not exhibit any consistent behaviour as far
as PIT and PIP are concerned. This shows that local factors play an important role and further detailed studies are
required. Predictability index for precipitation was not studied since a majority of the stations do not get significant
rainfall during this season.
Next, we consider how the predictability indices change with the months (viz. March, April and May) constituting
this season. We observe that all coastal stations (with the exception of Dwaraka) have predictable temperature during
each month. However, pressure is not predictable for all months for these stations. For central Indian stations, tem-
perature is not predictable. For north-eastern stations, pressure is not predictable.

3.2. South-west monsoon season

If one looks at the seasonal values of the predictability indices, the following regional patterns are observed. The
coastal and north-eastern stations have predictable temperature (again with the exception of Dwaraka). Stations sit-
uated in the core of the Indian monsoon (centered around Central India) have unpredictable temperature whereas
pressure is predictable. For these stations, the variability in precipitation is linked to the variability in temperature.
As for PIR , nine stations have reasonably high predictability (greater than or equal to 0.5) if one looks at the south-
west monsoon season (June–September) as a whole. However, if one looks at PIR for the individual months constituting
the season, no station has high predictability for all these months. This anomaly can be explained as follows. It is quite
possible that the total rainfall for the season follows a more consistent pattern across the years as compared to the
rainfall in individual months. The monthly rainfall could be more unpredictable because of delays in onset of monsoon
for that station etc. whereas even a delayed monsoon can give the same amount of rainfall for the entire season as
compared to a normal monsoon. The difference between the seasonal PIR and monthly PIR quantifies this for each

Fig. 1. Scatter plot of PIT vs. PIP for the pre-south-west monsoon season (March–May).
 G. Rangarajan, D.A. Sant / Chaos, Solitons and Fractals 19 (2004) 285–291 289

1.0

0.8

0.6
PIT

0.4

0.2

0.0
1 2 3 4 5 6 7 8 9 10 11 12
Months

Fig. 2. Temperature predictability graph for Nagpur showing the variation of the monthly PIT values with the months.

station. Also note that PIR is not directly related to the amount of rainfall. It only measures the predictability of the
rainfall.
Next, we looked for correlations between the south-west monsoon rainfall and the various monthly predictability
indices. Correlations were calculated using data from all stations except those which have a long rainy season
(stretching from March/April to October/November). Significant positive correlations of the south-west monsoon
rainfall (significant at the 5% level) were observed with both February PIT and March PIP . This suggests that the
temperature/pressure dynamics during the February–March period (the transition period between winter and summer)
prior to the monsoon could play a role in the south-west monsoon rainfall over the Indian subcontinent.
Next, we looked for characteristic patterns in the temperature and pressure predictability graphs (i.e. graphs of PIT
and PIP as a function of months) for a given station prior to the onset of the south-west monsoon. An example of such a
graph is given in Fig. 2 for Nagpur. No pattern was observed in the pressure predictability graphs. On the other hand,
the temperature predictability graphs for stations getting significant rainfall (greater than 10 cm monthly) during the
south-west monsoon exhibited a characteristic dip prior to the onset of the monsoon. However, there were a few ex-
ceptions. Stations at high elevations did not exhibit this pattern which is understandable since such stations typically
have a quite different climatic dynamics. The other exceptions were Veraval (situated along the West coast) and New
Delhi where other local/geographical factors probably play a role.

3.3. North-eastern monsoon season

We now study the regional distribution of PIT and PIP for the north-east monsoon season (October–November). We
first consider stations in the southern and eastern India where it rains significantly during the north-east monsoon. The
north-eastern region where it rains practically throughout the year appears to have a different dynamics dictated by
local geographical factors. However, other stations exhibit the following patterns in the predictability indices calculated
for this season. Rainfall is unpredictable for all stations. Except for three stations (considered separately below),
temperature is predictable whereas pressure is unpredictable. Therefore, it is the variability in pressure which causes
the rainfall unpredictability.
The two high elevation stations (Kodaikanal and Darjeeling) follow a different pattern. Both their temperature and
pressure are predictable whereas rainfall is unpredictable. Here, their elevation seems to play a role. The other exception
is Madras. Here temperature is unpredictable whereas pressure is predictable. This is in contrast to what is observed for
other stations. One interesting fact (which may be related to this contrary behaviour) is that Madras alone of all these
stations gets its maximum rainfall during the month of November rather than October.

4. Discussions

Most climatic studies deal with a region or a season as a whole. Even predictions for the Indian monsoon deal with
the whole subcontinent. However, for (say) a farmer at a specific location, the only thing which matters is whether
rainfall is good at her/his location in a given month. In our paper, we have therefore studied predictability of the climate
at two temporal scales––for the season as a whole and for individual months within the season. We find that even if the
seasonal rainfall is predictable at a given location, the monthly rainfall is not. For individual stations spread throughout
290 G. Rangarajan, D.A. Sant / Chaos, Solitons and Fractals 19 (2004) 285–291

the Indian subcontinent, our studies quantify the discrepancy between predictability of seasonal vs. monthly rainfall
and consequently the limitations of gross climatic models.
We have also investigated whether regional patterns exist for the predictability indices. We have found some patterns
(detailed in the previous section). But these patterns exist only for sub-regions of the Indian subcontinent and not for
the whole subcontinent. This again highlights the fact that even if monsoon rainfall is predicted to be good for the
whole subcontinent using a regional climatic model that averages over the entire region, individual regions may be hit
with drought/flood. Therefore, regional climatic models typically do not predict such local extreme events since they
deal with averaged quantities. We extend this analogy to paleoclimatic data from the continents, where data largely
represents climate of short-lived local depositional events [12,13] and does not represent the overall climate of the
region. This shows that the link between regional climate and paleoclimatic data is weak. However strong events that
influence fundamental processes do get recorded as climatic events with a time lag. The lag depends on how climate
change influences the site and dynamics of the proxies at the site.
Our observations in the previous section also support the well known fact that climate is a very complex system.
However, this complexity has been studied from a different viewpoint in this paper using predictability indices. The
complexity is manifest in the absence of any strong regional vs. local and seasonal vs. monthly correlations in the
predictability indices.
In this paper, we have also extended the analysis carried out in our previous paper [7] for the south-west monsoon.
There we had studied stations with low predictability for rainfall and analyzed causes for the same. In this paper, the
same analysis was performed for the north-east monsoon. We found that for most stations it is the unpredictability
in pressure which leads to the unpredictability in rainfall.
We also find that the dynamics of the south-west monsoon is affected by the transition from winter to summer in the
previous February–March period. This corresponds to the observation [12,14] that in years of weak glacial/low snowfall
(weak winter over the continent), the Tibetan plateau is able to warm up earlier leading to a strong monsoon. On the
other hand, a delayed or weak monsoon is due to deep snow (severe winter over the continent) [15]. In the core region of
the monsoon, pressure is consistent across the years (leading to higher PIP ) and it is the variability in temperature that
causes the variability in rainfall. However, this pattern reverses during the north-east monsoon. For this monsoon, there
is no winter to summer like transition for temperature prior to the monsoon. Further, since it has already been raining
(during the south-west monsoon) prior to this monsoon, there is a general cooling trend in temperature which makes it
predictable in the core region for this monsoon (which is now centered around south-east India). On the other hand, the
pressure is observed to become unpredictable leading to variability in the north-eastern monsoon rainfall.

Acknowledgements

We would like to thank Prof. N.V. Joshi for useful discussions. This work was supported by a research grant from
DRDO under the Engineering Mathematics Programme at IISc. GR is also associated with the Jawaharlal Nehru
Centre for Advanced Scientific Research, Bangalore as a Honorary Faculty.

References

[1] Mandelbrot BB, Wallis JR. Some long-run properties of geophysical records. Water Resour Res 1969;5:321–40.
[2] Mandelbrot BB, Wallis JR. Noah, Joseph and operational hydrology. Water Resour Res 1968;4:909–18.
[3] Mandelbrot BB, Van Ness JW. Fractional Brownian motions, fractional noises and applications. SIAM Rev 1968;10:422–37.
[4] Flugeman Jr RH, Snow RS. Fractal analysis of long-range paleoclimatic data: oxygen isotope record of Pacific core V28–239.
Pure Appl Geophys 1989;131:307–13.
[5] Hsui AT, Rust KA, Klein GD. A fractal analysis of Quarternary, Cenozoic–Mesozoic, and Late Pennsylvanian sea level changes.
J Geophys Res 1993;98B:21963–7.
[6] Turcotte DL. Fractals and chaos in geology and geophysics. New York: Cambridge University Press; 1992.
[7] Rangarajan G, Sant DA. A climate predictability index and its applications. Geophys Res Lett 1997;24:1239–42.
[8] Vose, RS et al. The global historical climatology network: long-term monthly temperature, precipitation, sea level pressure, and
station pressure data. National Oceanic and Atmospheric Administration, Asheville, 1995.
[9] Beran J. Statistics for long-memory processes. New York: Chapman & Hall; 1994.
[10] Hurst HE, Black RP, Simaika YM. Long-term storage: an experimental study. London: Constable; 1965.
[11] Voss RF. In: Pynn R, Skjeltorp A, editors. Scaling phenomena in disordered systems. New York: Plenum; 1985.
[12] Overpeck J, Anderson D, Trumbore S, Prell W. The south-west Indian monsoon over the last 18000 years. Climate Dyn
1996;12:213–25.
 G. Rangarajan, D.A. Sant / Chaos, Solitons and Fractals 19 (2004) 285–291 291

[13] Sirocko F, Sarnthein M, Erlenkeuser H, Lang H, Arnold M, Duplessy JC. Century-scale events in monsoonal climate over past
24000 years. Nature 1993;364:322–4.
[14] Van Campo E, Duplessy JC, Rossignol-Strick M. Climatic conditions deduced from a 150 Kyr oxygen-isotope-pollen record from
the Arabian sea. Nature 1982;296:56–9.
[15] Barnett TP, Dumenil L, Schlese U, Roeckner E, Latif M. The effect of Eurasian snow cover on regional and global climate
variations. J Atmos Sci 1989;46:661–8.