You are on page 1of 12

INTERNATIONAL JOURNAL OF CLIMATOLOGY

Int. J. Climatol. 31: 20212032 (2011)


Published online 13 September 2010 in Wiley Online Library
(wileyonlinelibrary.com) DOI: 10.1002/joc.2215

Estimating Palmer Drought Severity Index using a wavelet


fuzzy logic model based on meteorological variables
Mehmet Ozger,a,b,c * Ashok K. Mishraa,b and Vijay P. Singha,b
a

Department of Biological and Agricultural Engineering, Texas A&M University, College Station, TX 77843-2117, USA
Department of Civil and Environmental Engineering, Texas A&M University, College Station, TX 77843-2117, USA
c Hydraulics Division, Istanbul Technical University, Maslak 34469, Istanbul, Turkey

ABSTRACT: The Palmer Drought Severity Index (PDSI) is widely used to characterize droughts. The PDSI is based
on the water balance equation over an area of concern. Calculating PDSI requires data on precipitation, temperature,
soil moisture, and the previous PDSI value. While precipitation and temperature time series data are easily available for
most locations, it is not always the case with soil moisture due to the lack of soil-moisture monitoring networks. This
study developed a wavelet fuzzy logic model (WFL) to overcome the problem. The proposed model employs commonly
available precipitation, temperature, and large-scale climate indices as predictors and PDSI as a predictand. The WFL
model is applied to ten climate divisions in Texas and its performance is compared with conventional fuzzy logic (FL)
model performance. It is shown that the WFL model outperforms the FL model. The variation of WFL model performance
along with the average wavelet spectra of precipitation time series is evaluated. Results show that the WFL model is
capable of predicting PDSI. Copyright 2010 Royal Meteorological Society
KEY WORDS

Palmer Drought Severity Index; continuous wavelet transform; fuzzy logic; average wavelet spectra

Received 27 November 2009; Revised 12 June 2010; Accepted 29 July 2010

1.

Introduction

Evaluation of droughts is important for water resources


planning and management. There are several indices
that are used to characterize drought properties. Among
them, the most used are standardized precipitation index
(SPI), deciles, Palmer drought severity index (PDSI), and
derivatives of PDSI. A drought index should represent
some basic characteristics to describe droughts satisfactorily. The three main characteristics that must be included
in the definition of a drought index are duration, magnitude, and severity (Mishra et al., 2007; Mishra and Singh,
2009). The index should also include information on the
onset and termination of a drought event. It should have
an ability to distinguish a drought from aridity. PDSI,
which is widely used in drought studies, involves all these
characteristics to define a drought.
Since its first formulation by Palmer (1965), there
have been several studies on PDSI (Szinell et al., 1998;
Heim, 2002; Ntale and Gan, 2003). Temperature and
precipitation are the most important two inputs used
in the calculation of PDSI. Guttmann (1991) examined
the sensitivity of PDSI to departures from average
temperature and precipitation conditions. It was found
that the effects of precipitation anomalies were greater
than the effects of temperature anomalies. Hu and Willson
* Correspondence to: Mehmet Ozger, Department of Biological and
Agricultural Engineering, Texas A&M University, College Station, TX
77843-2117, USA. E-mail: mehmetozger@tamu.edu
Copyright 2010 Royal Meteorological Society

(2000) investigated the temperature and precipitation


effects on the PDSI. They showed that the PDSI can be
equally affected by temperature and precipitation, when
both have similar magnitudes of anomalies. However,
estimating soil moisture from drought indices can be
a practical approach. Sims et al. (2002) studied the
possible estimation of soil moisture from PDSI and
SPI. Rao and Padmanabhan (1984) investigated the
stochastic nature of yearly and monthly PDSI, and
characterized those using stochastic models to forecast
and simulate the PDSI series. Lohani and Loganathan
(1997) used PDSI to characterize the stochastic behaviour
of droughts.
Sometimes fuzzy logic (FL) is preferred when linking
inputs to outputs in a nonlinear manner. Pesti et al. (1996)
modelled the relationship between drought characteristics
and general circulation patterns (CP) using FL. Pongracz
et al. (1999) applied fuzzy rule-based modelling for the
prediction of regional droughts using two forcing inputs,
ENSO and large scale atmospheric CPs in a typical Great
Plains state, Nebraska. These FL models are applicable
for only short-term drought forecasting.
Cutore et al. (2009) developed an artificial neural
network model to forecast Palmer Hydrological Drought
Index (PHDI) up to a 4-month lead time by considering
persistence and some climate indices. Although they
obtained high R 2 values (around 0.90) for 1 month
ahead forecasting, which is the consequence of high auto
correlation coefficient at lag-1, the R 2 values decreased
to around 0.4 for 4 months ahead forecasts. Kim and

2022

M. OZGER et al.

Valdes (2003) proposed a conjunction model, which is


the combination of discrete wavelet transform and neural
network method to forecast the PDSI values up to a 12month lead time.
PDSI is the most used index to assess the severity
of droughts. Temperature, precipitation, soil moisture,
and the previous PDSI value are required for calculation
of PDSI. Information on soil moisture includes potential evapotranspiration, recharge, loss, and runoff. Also,
incoming extraterrestrial solar radiation, relative humidity, mean monthly minimum temperature, and mean
monthly maximum temperature are used to calculate
potential evapotranspiration. Although temperature and
precipitation records are widespread, other data required
to calculate PDSI may not exist for certain locations.
Although several studies have been conducted to predict PDSI, the prediction of PDSI from simultaneous
consideration of temperature, precipitation, and some climate indices has not been pursued. Generally persistence
(lagged values) has been taken as a predictor variable to
increase the capability of prediction models. However,
past knowledge of PDSI may not exist when the areas
whose drought properties have not been investigated previously, are considered. The advantage of excluding the
previous PDSI value from predictor variables is to make
independent predictions of PDSI. In this way, one can
produce the PDSI values in the absence of soil-water
balance variables and the past knowledge of PDSI.
The objective of this study is to employ FL and
WFL models for predicting PDSI from predictor variables, which are temperature, precipitation, and climate
indices such as NINO 3.4 (is an index that represents
the sea surface temperature anomalies in eastern tropical
Pacific), PDO (Pacific decadal oscillation). The purpose
of the study and methodologies are to address specific
questions: (1) Is it possible to simulate PDSI series without using soil-moisture data, if so, with what accuracy?
(2) Can the simulated PDSI series be improved using climate indices? (3) The strength of wavelet fuzzy model
in simulating the PDSI series? (4) How can the possible effects of temperature and precipitation on droughts
be interpreted through wavelet analysis and spectral band
separation?

2.

Palmer Drought Index

The PDSI is widely used in drought evaluation studies.


The method is based on the soil-water balance equation. The climate coefficients are computed as a proportion between averages of actual versus potential values
for each of 12 months. Palmer (1965) defined climatically appropriate for existing conditions (CAFEC), which
shows the actual situation in the area of concern. The
amount of precipitation required for CAFEC can be computed from climate coefficients. Subsequently, the water
deficiency for each month is indicated by the difference,
d, between actual (P ) and CAFEC precipitation (P) as
Copyright 2010 Royal Meteorological Society

follows:
d = P P = P (PE + PR + PRO + PL) (1)




where = ET PE = R PR = RO PRO = L PL
for 12 months. The terms are actual evapotranspiration
(ET) and potential evapotranspiration (PE); recharge (R)
and potential recharge (PR); runoff (RO) and potential
runoff (PRO); net loss (L) and potential loss (PL). A
Palmer Moisture Anomaly Index (PMAI), Z, for an ith
month is then defined as follows:
Zi = Ki di

(2)

Palmer (1965) discovered that K, the weighting factor,


varied inversely with D, the mean of the absolute values
of di . An empirical relationship was suggested as follows:


12


D i Ki
(3)
Ki = 17.67Ki /
i=1

Ki  depends on the average water supply and demand,


expressed as:
Ki = 1.5 log10 [(Mi + 2.8)/Di ]

Mi = (PE + R + RO) (P + L)

(4)
(5)

where PE is the potential evapotranspiration, R is the


recharge, RO is the runoff, P is the precipitation, and L
is the loss. The PDSI is now given by
1
PDSIi = 0.897PDSIi1 + Zi
3
3.

(6)

Data

There are five distinct climate zones in Texas showing the variation from arid to sub-tropic humid zones
(Figure 1(a)). Texas is divided into ten climate divisions
by the National Climatic Data Center (Figure 1(b)). Each
climate division exhibits its own specific characteristics,
such as vegetation, temperature, humidity, rainfall, and
seasonal weather. Representative data are calculated for
each division by taking the stations which are within
the borders of that division and then averaging over all
stations. Precipitation, temperature, drought indices, and
other variables are reported using these divisions.
PDSI indicates the severity of a wet or dry spell
and is reported monthly. PDSI, which is a standardized index, is used in the assessment of meteorological
droughts. It is also considered a hydrological drought
indicator due to its relation to evapotranspiration and
soil moisture. It is capable of representing the spatial
content of droughts. While negative values stand for
dry spells, wet spells are represented by positive values. The PDSI data on a 20 latitude 30 longitude
grid were obtained from a nearest neighbour gridding
procedure of Cook et al. (1999). PDSI, precipitation
and temperature time series for each climate division
Int. J. Climatol. 31: 20212032 (2011)

2023

ESTIMATING PALMER DROUGHT SEVERITY INDEX


Climate map
36

Climate divisions

Continental

(a)

36

(b)
1

34

34
2
Sub-tropic
semi-humid

32
Arid

30

32

Sub-tropic humid

Semiarid

30

7
8

28

28

26

10

26
-106

-104

-102

-100

-98

-96

-94

-106

-104

-102

-100

-98

-96

-94

Figure 1. (a) Climate zones and (b) climate divisions for Texas.

obtained from the National Climate Data Center for


the period 19002007 can be found at NOAA website
(http://www7.ncdc.noaa.gov/CDO/CDODivisionalSelect.
jsp#).
NINO 3.4 and PDO were used as variables for
large-scale climate indices. Time-series data for NINO
3.4 region are available every month from 1856 to
2007 (http://iridl.ldeo.columbia.edu/SOURCES/.Indices/.
nino/.EXTENDED/).
The PDO Index is defined as the leading principal
component of the North Pacific monthly sea-surface
temperature variability. The monthly data covering the
period 19002007 is downloaded from the website
(http://jisao.washington.edu/pdo/PDO.latest).

4.

Methodology

4.1. Fuzzy logic


FL modelling is based on the fuzzy set theory which was
introduced by Zadeh (1965). These days many applications of the FL theory are seen in all areas of engineering.
FL can be used to relate multiple inputs with output
and has the ability to establish nonlinear relationships.
This relationship is achieved by a fuzzy inference system.
There are mainly two types of inference systems which
are Mamdani and Takagi-Sugeno (TS). While the Mamdani type inference relies on both linguistic and numerical
data, the TS inference system works only with numerical data. The TS approach has an advantage of using data
efficiently in the training procedure and makes it possible
to incorporate a suitable training algorithm, e.g. ANFIS
(Adaptive neural network fuzzy inference system).
Fuzzy rules and fuzzy sets are the main elements of
the FL modelling. On one hand, fuzzy rules provide
the connection between predictors and predictand and
on the other hand, fuzzy sets produce weights for
those rules. Fuzzy rules are in the form of IFTHEN
statements. While the part between IF and THEN is
called antecedent, the consequent part is found after
THEN. Here, the antecedent part consists of precipitation,
Copyright 2010 Royal Meteorological Society

temperature, and climate indices. The PDSI values are


taken in the consequent part. A typical fuzzy rule can be
written for this case as follows:
Ri : IF Precip is in F1 and Temp is in S1
THEN PDSI = pi1 Precip + pi2 Temp + pi0
Rk : IF Precip is in F2 and Temp is in S2
THEN PDSI = pk1 Precip + pk2 Temp + pk0
where F and S are the membership functions for
the precipitation and temperature variables that include
antecedent part parameters, and ps are the consequent
part parameters.
The fuzzy inference system consists of four steps:
(1) The predictor variables are fuzzified by assigning
membership functions to each variable. The type (Gaussian, triangular, etc.) and the number of the membership
functions are determined by the user. (2) The fuzzy rule
base is constructed based on the previous step. The rule
base consists of rules which are combinations of membership functions of predictor variables. For instance, if there
are two variables with three membership functions in the
antecedent part, there would be 3 3 = 9 rules totally in
the rule base. (3) The implication step incorporates outcomes of the antecedent part to the consequent part and
aggregate the consequent part of all rules. (4) Since the
aggregated results appear in the form of fuzzy sets, it is
required to find a one-crisp value by using defuzzification as a final step. The following equations are used to
obtain the final outcome of a fuzzy inference system:
 

x ci 2
i (x) = exp
bi

(7)

IF input 1 = n and input 2 = m THEN


output is zi = pi1 n + pi2 m + pi0
wi = 1 2

(8)
(9)

Int. J. Climatol. 31: 20212032 (2011)

2024

M. OZGER et al.
N


Final output =

wi zi

i=1
N


(10)
wi

i=1

where b and c are the antecedent part parameters; ps


are the consequent part parameters; is the membership
function; and w is the weighting of the each rule. In this
study, the ANFIS technique was employed to determine
the antecedent and consequent part parameters. Details
of this technique can be found in Jang (1993).
4.2.

Continuous wavelet transform

The continuous wavelet transform (CWT) is used to


decompose a signal into wavelets, small waves that
grow and decay over a small distance, whereas the
Fourier transform decomposes a signal into an infinite
number of sine and cosine terms losing most timelocalization information. A continuous wavelet transform of a signal produces coefficients at a given scale.
Comparison between Fourier analysis and wavelet analysis is given by Kumar and Foufoula-Georgiou (1997)
who presented only the basics regarding wavelet analysis. CWTs basis functions are scaled and shifted versions of the time-localized mother wavelet. A Morlet
wavelet is one of the many wavelet functions which
has a zero mean and is localized in both frequency and
time. Since the Morlet wavelet provides a good balance
between time and frequency localizations, it is preferred
for application and can be represented as (Torrence and
Compo, 1998; Torrence and Webster, 1999; Grinsted
et al., 2004):
() = 1/4 ei0.5

(11)

where is the dimensionless frequency, and is the


dimensionless time parameter. The wavelet is stretched
in time (t) by varying its scale (s), so that = s t. When
(a)

using wavelets for feature extraction purposes, the Morlet


wavelet (with = 6) is a good choice, since it satisfies
the admissibility condition (Farge, 1992; Torrence and
Compo, 1998).
For a given wavelet 0 (), it was assumed that Xj is
a time series of length N (Xj , i = 1, . . . , N ) with equal
time spacing t. The continuous wavelet transform of a
discrete sequence Xj is defined as a convolution of Xj
with the scaled and translated wavelet 0 ():
N


W n X (s) =

Xj

j =1

(j n)t
s

(12)

where asterisk indicates the complex conjugate. CWT


decomposes a time series into time-frequency space,
enabling the identification of both the dominant modes
of variability and how those modes vary with time.
4.3.

Wavelet fuzzy logic

Geophysical time series include different patterns, such as


periodicity, trend, noise which are the results of different
mechanisms affecting the process. Filtering such patterns
helps understand the behaviour of time series. One of
latest techniques used for filtering time series in time
and scale domains is the wavelet transform. There is a
tendency to filter the data before its use, especially in predicting problems. Several researchers (Kim and Valdes,
2003; Webster and Hoyos, 2004; Partal and Kisi, 2007;
Nourani et al., 2009) have proposed that it is better to
make predictions after decomposing both predictors and
predictand into several bands. Wavelet transform makes
it possible to separate time series into its subseries. Here,
the important question is how the significant bands can
be selected. For this purpose, Webster and Hoyos (2004)
proposed the use of average wavelet spectra obtained
from continuous wavelet transform of a variable of concern. The significant spectral bands can be selected, based
on the average wavelet spectra which show the variation of power with scales. A sample band selection for
PDSI is shown in Figure 2 along with its wavelet power
Power (%)

(b)

10

0
0

(1)
50

(2)

100

(3)

150
200

(4)

250
300
350

(5)

400

Figure 2. (a) Continuous wavelet map of PDSI series for climate division 7 and its corresponding (b) average wavelet spectra over the period
19002007. There are five significant bands detected from average spectra which are <23, 2366, 66111, 111264, and >264 months. This
figure is available in colour online at wileyonlinelibrary.com/journal/joc
Copyright 2010 Royal Meteorological Society

Int. J. Climatol. 31: 20212032 (2011)

2025

ESTIMATING PALMER DROUGHT SEVERITY INDEX

map for climate division 7. There has not been a rule


established to separate the bands so far. The important
criterion for the separation of bands is to detect the bands
that have significant power compared to others. Other
bands are separated according to their average wavelet
power, respectively. For instance, in this study, band 3
(66111 months) shows peak power and can be distinguished from others. The Morlet wavelet was employed
for the continuous wavelet transform. It can be seen from
Figure 2 that it is possible to separate the original time
series into five different significant bands. These are <23,
2366, 66111, 111264, and >264 months. Thus, at
the end, we have five different sub-series each of which
carries specific information about the process. However,
each predictor time series is separated into five different
subseries using the same spectral bands as of predictand.
Subsequently, it is required to relate each band of
predictors to the corresponding band of predictand with
a statistical scheme. Here, we used a fuzzy logic model
to establish a connection between predictors and the
predictand band. Five fuzzy models would be needed to
make predictions. Finally, all those five predicted bands
of the predictand variable are reconstructed to obtain the
final series. A schematic diagram of the overall procedure
is shown in Figure 3.

5.

Results and discussion

To predict PDSI from meteorological variables and


climate indices, FL and WFL models were applied. The
results were obtained for ten different climate divisions
in Texas. Five scenarios, each of them included different
predictand combinations (Table I), were employed to see
how combinations affect the accuracy of models.

5.1. Wavelet band separation


The selection of bands which carry significant power
is important for the model setup. The separation of
bands was made by considering the average wavelet
spectra of the PDSI series for each climate division.
We obtained different groups of spectral bands according
to the average wavelet spectra of PDSIs shown in
Figure 4. The significant bands detected from the average
wavelet spectra of PDSI are presented in Table II. The
predictors were separated into their bands according to
those intervals, identically.
The PDSI average wavelet spectra consist of several
peaks each of which represents a significant power at
the corresponding period. It is apparent from Table II
that the PDSI time series for all climate divisions can
be separated into five significant bands. While the first
band shows noisy data, the fourth and fifth bands stand
for low-frequency variation of PDSI. The higher power
is observed at around 60120-month period in climate
divisions 7 and 8, which are located in the southcentral Texas. In panhandle (climate divisions 1 and 2), a
higher power occurs for 60240 months which shows
the importance of mediumrange droughts. However,
low-frequency variation is remarkable in the arid zone
(climate division 5).
All the bands carry specific information related to the
original time series. It can be said that the bands are
rectified from the effects of processes involved in the
generation of time series and represent only one feature of
the concerned series. For instance, the higher level band
(e.g. >200 months) contains only information on longtime cycles of the concerned variable and excludes other
properties such as noisy data, trends. However, short-time
cycles (e.g. <20 months) only account for noisy data.
Predicting the more homogenous series provided by the
wavelet banding constitutes the main advantage for WFL
model. A significant improvement is expected over the
plain fuzzy logic model.
5.2. Fuzzy model versus wavelet fuzzy model

Figure 3. Flowchart of the methodology. This figure is available in


colour online at wileyonlinelibrary.com/journal/joc
Copyright 2010 Royal Meteorological Society

Fuzzy models have the ability to link predictands and


predictor in a nonlinear manner. However, in some cases
they cannot provide sufficient accuracy in the prediction.
Therefore, a different approach, which is a combination
of wavelet transforms and FL, can be preferred to reach
the desired accuracy. Here, we applied FL and WFL
methods to predict PDSI and compared them for different
scenarios in each climate division.
Prior to the application of FL and WFL models,
it is required to split data into two parts, namely,
training and testing. While the first 70 years of dataset
(19001969) was employed for training, the remaining
part (19702007) was used to validate models. Table I
presents the overall results of training and testing periods.
It is evident from the table that WFL outperforms FL. The
improvement in R 2 values provided by WFL is 100%
and more in some cases. Also, the correlation coefficient
values are over 0.70 for testing and over 0.80 for training
Int. J. Climatol. 31: 20212032 (2011)

2026

M. OZGER et al.

Table I. Results of FL and WFL modelling of PDSI for each ten climate divisions.
CDs

Inputs

FL

WFL

Train
R

CD-1

CD-2

CD-3

CD-4

CD-5

CD-6

CD-7

CD-8

CD-9

CD-10

NINO 3.4,PDO;Pcp,Temp 0.266


NINO 3.4;Pcp,Temp
0.243
PDO;Pcp,Temp
0.531
Pcp,Temp
0.237
NINO 3.4,PDO
0.036
NINO 3.4,PDO;Pcp,Temp 0.226
NINO 3.4;Pcp,Temp
0.192
PDO;Pcp,Temp
0.215
Pcp,Temp
0.204
NINO 3.4,PDO
0.080
NINO 3.4,PDO;Pcp,Temp 0.296
NINO 3.4;Pcp,Temp
0.227
PDO;Pcp,Temp
0.254
Pcp,Temp
0.220
NINO 3.4,PDO
0.064
NINO 3.4,PDO;Pcp,Temp 0.322
NINO 3.4;Pcp,Temp
0.270
PDO;Pcp,Temp
0.281
Pcp,Temp
0.269
NINO 3.4,PDO
0.047
NINO 3.4,PDO;Pcp,Temp 0.291
NINO 3.4;Pcp,Temp
0.264
PDO;Pcp,Temp
0.278
Pcp,Temp
0.258
NINO 3.4,PDO
0.065
NINO 3.4,PDO;Pcp,Temp 0.320
NINO 3.4;Pcp,Temp
0.249
PDO;Pcp,Temp
0.313
Pcp,Temp
0.244
NINO 3.4,PDO
0.116
NINO 3.4,PDO;Pcp,Temp 0.286
NINO 3.4;Pcp,Temp
0.238
PDO;Pcp,Temp
0.281
Pcp,Temp
0.229
NINO 3.4,PDO
0.087
NINO 3.4,PDO;Pcp,Temp 0.314
NINO 3.4;Pcp,Temp
0.275
PDO;Pcp,Temp
0.309
Pcp,Temp
0.258
NINO 3.4,PDO
0.064
NINO 3.4,PDO;Pcp,Temp 0.253
NINO 3.4;Pcp,Temp
0.225
PDO;Pcp,Temp
0.246
Pcp,Temp
0.201
NINO 3.4,PDO
0.002
NINO 3.4,PDO;Pcp,Temp 0.280
NINO 3.4;Pcp,Temp
0.189
PDO;Pcp,Temp
0.251
Pcp,Temp
0.184
NINO 3.4,PDO
0.081

Test
2

Correlation
coefficient

0.515
0.492
0.816
0.486
0.186
0.474
0.437
0.462
0.451
0.282
0.543
0.475
0.503
0.468
0.252
0.567
0.519
0.530
0.518
0.213
0.539
0.513
0.526
0.507
0.253
0.565
0.499
0.558
0.493
0.340
0.534
0.487
0.529
0.477
0.293
0.559
0.524
0.555
0.507
0.250
0.502
0.474
0.495
0.447
0.013
0.528
0.434
0.500
0.429
0.282

0.175
0.062
0.162
0.082
0.023
0.221
0.098
0.255
0.138
0.106
0.171
0.109
0.176
0.063
0.030
0.228
0.159
0.226
0.178
0.074
0.285
0.261
0.282
0.267
0.069
0.301
0.314
0.284
0.272
0.004
0.341
0.277
0.341
0.247
0.113
0.330
0.258
0.357
0.223
0.049
0.261
0.249
0.275
0.210
0.006
0.367
0.205
0.318
0.192
0.162

Train
2

Correlation
coefficient

0.533
0.473
0.715
0.484
0.334
0.502
0.419
0.534
0.457
0.095
0.472
0.453
0.472
0.409
0.297
0.505
0.470
0.505
0.485
0.071
0.550
0.512
0.549
0.520
0.293
0.614
0.568
0.592
0.528
0.332
0.590
0.528
0.588
0.501
0.342
0.580
0.543
0.606
0.514
0.268
0.514
0.519
0.538
0.480
0.009
0.632
0.451
0.590
0.437
0.468

0.7873
0.7537
0.7131
0.6615
0.091
0.697
0.7917
0.752
0.6452
0.115
0.674
0.657
0.648
0.669
0.145
0.698
0.682
0.656
0.648
0.111
0.645
0.619
0.606
0.588
0.129
0.664
0.642
0.644
0.631
0.170
0.698
0.680
0.678
0.668
0.228
0.652
0.649
0.639
0.647
0.135
0.676
0.658
0.662
0.656
0.140
0.544
0.524
0.527
0.506
0.086

Test
2

Correlation
coefficient

0.8882
0.8705
0.845
0.8138
0.303
0.836
0.89
0.8676
0.8067
0.342
0.822
0.811
0.805
0.819
0.383
0.836
0.826
0.810
0.805
0.336
0.815
0.798
0.790
0.778
0.387
0.815
0.801
0.803
0.794
0.412
0.836
0.825
0.824
0.817
0.477
0.810
0.808
0.802
0.807
0.372
0.824
0.814
0.816
0.812
0.377
0.756
0.742
0.745
0.731
0.337

0.4202
0.5485
0.5175
0.6035
0.050
0.334
0.4163
0.3921
0.4893
0.101
0.369
0.339
0.367
0.346
0.015
0.379
0.441
0.496
0.491
0.127
0.523
0.601
0.596
0.614
0.018
0.573
0.557
0.603
0.560
0.017
0.492
0.568
0.538
0.561
0.026
0.530
0.529
0.537
0.516
0.097
0.622
0.630
0.582
0.610
0.081
0.636
0.583
0.617
0.589
0.139

Correlation
coefficient
0.6876
0.7525
0.7198
0.7774
0.155
0.639
0.6854
0.6489
0.7149
0.334
0.715
0.707
0.728
0.681
0.258
0.703
0.726
0.766
0.764
0.088
0.740
0.790
0.779
0.796
0.105
0.760
0.750
0.782
0.755
0.358
0.727
0.763
0.752
0.765
0.365
0.764
0.761
0.775
0.765
0.192
0.797
0.806
0.786
0.800
0.298
0.812
0.778
0.799
0.784
0.421

CC, correlation coefficient.

Copyright 2010 Royal Meteorological Society

Int. J. Climatol. 31: 20212032 (2011)

2027

ESTIMATING PALMER DROUGHT SEVERITY INDEX


PREC CD-1

PDSI CD-1

25

12
10

20
Power (%)

Power (%)

8
15

10

6
4

0
0

50

100

150

200

250

300

350

400

50

100

150

200

250

300

350

400

300

350

400

300

350

400

300

350

400

300

350

400

Period (months)

Period (months)

PDSI CD-2

PREC CD-2
14

10

12
8

Power (%)

Power (%)

10
8
6

4
2
2
0

0
0

50

100

150

200

250

300

350

400

50

100

PREC CD-3

150

200

250

Period (months)

Period (months)

PDSI CD-3

12
10

Power (%)

Power (%)

8
4

6
4

0
0

50

100

150

200

250

300

350

400

50

100

Period (months)

10

8
Power (%)

Power (%)

250

PDSI CD-4

12

0
0

50

100

150

200

250

300

350

400

50

100

Period (months)

150

200

250

Period (months)
PDSI CD-5

PREC CD-5
20

10

Power (%)

15
Power (%)

200

Period (months)

PREC CD-4

150

10

0
0

50

100

150

200

250

300

350

400

Period (months)

50

100

150

200

250

Period (months)

Figure 4. Average wavelet spectra of precipitation and PDSI time series. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

in most of the cases. This significant improvement in the


model accuracy makes it possible to use these models in
practical applications.
The increase in the R 2 values by the WFL model
can be related to its setup. The main idea behind the
WFL method is based on the wavelet banding explained
above. Since WFL uses information at various spectral
Copyright 2010 Royal Meteorological Society

bands separately, it can capture and model the data


behaviour (e.g. periodicity, noise) easily compared to
the simple FL model. WFL models consist of a certain
number of FL models which is equal to the number
of separated bands from an original time series. For
instance, assume that five different spectral bands are
detected from the average wavelet spectra of PDSI.
Int. J. Climatol. 31: 20212032 (2011)

2028

M. OZGER et al.
PREC CD-6

PDSI CD-6

10

Power (%)

Power (%)

5
4

3
2

2
1

0
0

50

100

150

200

250

300

350

400

50

100

Period (months)

200

250

300

350

400

300

350

400

300

350

400

300

350

400

300

350

400

Period (months)

PREC CD-7

150

PDSI CD-7

10

Power (%)

Power (%)

5
4

3
2

2
1

0
0

50

100

150

200

250

300

350

400

50

100

Period (months)

200

250

Period (months)

PREC CD-8

5.5

150

PDSI CD-8

10

5
8

Power (%)

Power (%)

4.5

3.5
3
2.5

2
1.5

0
0

50

100

150

200

250

300

350

400

50

100

PREC CD-9

10

150

200

250

Period (months)

Period (months)

PDSI CD-9

15

Power (%)

Power (%)

10
6

5
2

0
0

50

100

150

200

250

300

350

400

50

100

Period (months)

250

PDSI CD-10

14
12

10

10
Power (%)

8
Power (%)

200

Period (months)

PREC CD-10

12

150

6
4

8
6
4

0
0

50

100

150

200

250

300

350

400

50

Period (months)

100

150

200

250

Period (months)

Figure 4. (Continued).

Subsequently, predictors are separated into the same five


bands. Thus, there are five different series all of which
contain significant spectral power. Five different fuzzy
models were established and then the results of them were
reconstructed to obtain a single series. Figure 5 shows
significant spectral bands and predicted bands of PDSI
from the identical bands of NINO 3.4, precipitation, and
temperature. Observed and predicted time series of PDSI
for climate division 7 along with the scatter diagrams
Copyright 2010 Royal Meteorological Society

of observed and predicted PDSI in the validation period


(19702007) for WFL and FL models are shown in
Figure 6.
5.3. WFL model capability in climate divisions
Considering climate divisions, the WFL model results
were evaluated throughout Texas. The WFL model performance shows variability from one division to the other,
as given in Table I, for all climate divisions in terms
Int. J. Climatol. 31: 20212032 (2011)

2029

ESTIMATING PALMER DROUGHT SEVERITY INDEX


(a)

(b)

(c)

(d)

(e)

(f)

Figure 5. (ae)Time series of the five observed and predicted wavelet bands for climate division 7 PDSIs, and (f) final reconstructed and observed
PDSI time series. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

of the R 2 values and correlation coefficients. The table


presents the results of both FL and WFL models for
training (calibration) and testing (validation) periods. In
the evaluation, the R 2 value for the testing period was
taken as an indicator of model performance. It is evident
from the table that WFL outperforms the FL model in
all cases. Precipitation and temperature time series along
with climate indices (NINO 3.4 and PDO) as predictors
were combined variously to determine the best combination. No significant effect of climate indices was seen
that increased the WFL model capabilities. It is apparent
from Table I that the capabilities of the models disappear
completely when the predictor combinations constituted
without using precipitation and temperature. Precipitation
and temperature are the driven factors in the prediction
of PDSI. To understand the model capability across climate divisions, the average wavelet spectra of PDSI along
with the predictor variables were considered. The average wavelet spectra of NINO 3.4 and PDO are depicted
in Figure 7. Different energy patterns in their spectra can
be seen from the figure. While NINO 3.4 has significant
Copyright 2010 Royal Meteorological Society

Table II. Significant bands selected from average wavelet


spectra of PDSIs for ten climate divisions.
Climate
divisions

1
2
3
4
5
6
7
8
9
10

PDSI band separation (months)


1

<33
<28
<33
<27
<33
<11
<23
<27
<16
<23

3366
2846
3355
2766
33111
1166
2366
2746
1639
2333

66111
46132
55132
66187
111132
66111
66111
46111
3978
3378

111222
132187
132222
187264
132264
111264
111264
111264
78187
78187

>222
>187
>222
>264
>264
>264
>264
>264
>187
>187

power at the 6070-month band, PDO exhibits a significant power at around 6070 and 300340-month bands.
The wavelet spectra of the temperature time series for all
Int. J. Climatol. 31: 20212032 (2011)

2030

M. OZGER et al.

(a)

(c)

(b)

Figure 6. (a) Observed and predicted time series of PDSI for climate division 7. FL and WFL models were employed to predict PDSI from
NINO 3.4, precipitation, and temperature. Scatter diagrams of observed and predicted PDSI in validation period (19712006) for (b) wavelet
fuzzy logic model (WFL), and (c) fuzzy logic (FL) model. This figure is available in colour online at wileyonlinelibrary.com/journal/joc

climate divisions show nearly the same pattern (only two


are shown, Figure 8). The significant energy is present
at 12-month (1 year) band which shows the annual cycle
of temperature variation. However, the average wavelet
spectra of precipitation changes from division to division.
The wavelet spectra of precipitation along with the corresponding PDSI time series are shown in Figure 4. It is
seen from the figure that annual cycles (significant power
at 12-month band) in precipitation are dominant for climate divisions 1, 2, 5, 6, 9, and 10. For the other climate
divisions (3, 4, 7, and 8), it is observed that low-frequency
(a)

NINO 3.4

12

bands are significant which indicate the presence of different precipitation generating mechanisms.
Since the wavelet spectra of precipitation and the
WFL model results show different patterns throughout
the climate divisions, a possible relation between these
spectra and the WFL model performance scores (R 2 ) can
be expected. Investigation of average wavelet plots along
with R 2 values reveals that the WFL model performs
better in the climate divisions where the annual cycle
of precipitation is dominant. However, the accuracy
of WFL model reduces in the places where multiple
(b)

PDO

16
14

10

12
Power (%)

Power (%)

8
6
4

10
8
6
4

2
0

2
0

50

100

150

200

250

Period (months)

300

350

400

50

100

150

200

250

300

350

400

Period (months)

Figure 7. Average wavelet spectra for (a) NINO 3.4 and (b) PDO index. This figure is available in colour online at wileyonlinelibrary.com/journal/joc
Copyright 2010 Royal Meteorological Society

Int. J. Climatol. 31: 20212032 (2011)

2031

ESTIMATING PALMER DROUGHT SEVERITY INDEX


TEMP CD-8
60

50

50

40

40

Power (%)

Power (%)

TEMP CD-5
60

30
20
10
0

30
20
10

50

100 150 200 250 300 350 400

50

Period (months)

100 150 200 250 300 350 400


Period (months)

Figure 8. Average wavelet spectra of temperature time series for (a) climate division 5 located in arid zone and (b) climate division 8
located in sub-tropic humid zone. Same pattern of variation is seen for other climate divisions. This figure is available in colour online at
wileyonlinelibrary.com/journal/joc

peaks of significant power are seen in their spectra. The


reason behind this result can be related to significant
power bands other than 12-month band which make
the prediction issue more complex. Multiple power
peaks in the wavelet spectra of precipitation show the
presence of various frequency regime combinations in the
time series. This kind of nonlinear interactions between
several physical processes leads to more disordered
PDSI series which finally makes the prediction of PDSI
difficult.

6.

Conclusions

Prediction of PDSI is achieved from precipitation, temperature, and large scale climate indices by using WFL,
which is a relatively new methodology. This method is
applied to ten climate regions in Texas to model PDSI.
The model results are compared to FL approach. The
following conclusions can be drawn from this study:
1. The WFL model predicts PDSI satisfactorily from precipitation and temperature. This enables to determine
PDSI in the absence of soil moisture information and
other parameters required for the calculation of PDSI.
Inversely, it is possible to estimate soil moisture from
the predicted PDSI values.
2. A significant improvement is obtained over the FL
model in the prediction of PDSI by using WFL which
is capable of modelling more complex systems.
3. The effect of large-scale climate indices on the prediction of PDSI is not important. While in some climate
divisions they improve the WFL model performance
slightly, in general their impact on prediction is minor.
Precipitation and temperature are the main predictors
for PDSI.
4. The evaluation of average wavelet spectra of predictor
variables reveals that only precipitation time series
exhibits different spectral patterns throughout climate
divisions. Temperature time series shows nearly the
same pattern which is a significant power at 12 months
for all climate divisions.
Copyright 2010 Royal Meteorological Society

5. The behaviour of precipitation has a significant impact


on the PDSI time series which eventually affects the
performance of the WFL model. It is found that WFL
performs better in the climate divisions where the
average wavelet spectra of precipitation show single
peak energy around 12-month cycle which indicates
the regular annual cycle. These regular precipitation
events also put the PDSI time series in order.
A future work can be the estimation of soil moisture
from PDSI. The predicted PDSI values from precipitation and temperature can be used to obtain soil moisture
for the places where a network of soil-moisture measurements does not exist.
Acknowledgements
This work was financially supported by the United States
Geological Survey (USGS, Project ID: 2009TX334G)
and Texas Water Resources Institute (TWRI) through
the project Hydrological Drought Characterization for
Texas under Climate Change, with Implications for Water
Resources Planning and Management. The authors are
thankful to the reviewers for their insightful comments
which helped improve the quality of the manuscript.
References
Cook ER, Meko DM, Stahle DW, Cleaveland MK. 1999. Drought
reconstructions for the continental United States. Journal of Climate
12: 11451162.
Cutore P, Di Mauro G, Cancelliere A. 2009. Forecasting palmer ndex
using neural networks and climatic ndexes. Journal of Hydrologic
Engineering 14(6): 588595.
Farge M. 1992. Wavelet transforms and their applications to turbulence.
Annual Review of Fluid Dynamics 24: 395457.
Grinsted A, Moore JC, Jevrejeva S. 2004. Application of the cross
wavelet transform and wavelet coherence to geophysical time series.
Nonlinear Processes in Geophysics 11: 561566.
Guttman NB. 1991. A sensitivity analysis of the Palmer Hydrologic
Drought Index. Water Resources Bulletin 27: 797807.
Heim RR. 2002. A review of twentieth-century drought indices used in
the United States. Bulletin of American Meteorological Society 83:
11491165.
Hu Q, Willson GD. 2000. Effects of temperature anomalies on
the palmer drought severity index in the central United States.
International Journal of Climatology 20: 18991911.
Int. J. Climatol. 31: 20212032 (2011)

2032

M. OZGER et al.

Jang JSR. 1993. ANFIS: adaptive network based fuzzy inference


system. IEEE Transactions on Systems, Man and Cybernetics 23(3):
665684.
Kim T, Valdes JB. 2003. Nonlinear model for drought forecasting
based on a conjunction of wavelet transforms and neural networks.
Journal of Hydrologic Engineering 8(6): 319328.
Kumar P, Foufoula-Georgiou E. 1997. Wavelet analysis for geophysical applications. Review of Geophysics 35: 385412.
Lohani VK, Loganathan GV. 1997. An early warning system for
drought management using the Palmer drought index. Journal of
the American Water Resource Association 33(6): 13751386.
Mishra AK, Singh VP. 2009. Analysis of drought SAF curves using
GCM and scenario uncertainty. Journal of Geophysical ResearchAtmosphere, American Geophysical Union 114: D06120, DOI:
10.1029/2008JD010986.
Mishra AK, Singh VP, Desai VR. 2007. Drought Characterization:
A probabilistic approach. Journal of Stochastic Environmental
Research and Risk Assessment 23: 4155.
Nourani V, Alami MT, Aminfar MH. 2009. A combined neuralwavelet model for prediction of Ligvanchai watershed precipitation.
Engineering Applications of Artificial Intelligence 22: 466472.
Ntale HK, Gan TY. 2003. Drought indices and their application to East
Africa. International Journal of Climatology 23(11): 13351357.
Palmer WC. 1965. Meteorological Drought. Research Paper No. 45,
Weather Bureau, Washington, DC, 58.

Copyright 2010 Royal Meteorological Society

Partal T, Kisi O. 2007. Wavelet and neuro-fuzzy conjunction model


for precipitation forecasting. Journal of Hydrology 342(12):
199212.
Pesti G, Shrestha B, Duckstein L, Bogardi I. 1996. A fuzzy rule-based
approach to droght assessment. Water Resources Research 32(6):
17411747.
Pongracz R, Bogardi B, Duckstein L. 1999. Application of fuzzy rulebased modeling technique to regional drought. Journal of Hydrology
224: 100114.
Rao AR, Padmanabhan G. 1984. Analysis and modeling of Palmers
drought index series. Journal of Hydrology 68: 211229.
Sims AP, Niyogi DS, Raman S. 2002. Adopting drought indices for
estimating soil moisture: a North Carolina Case Study. Geophysical
Research Letters 29(8): 1183, DOI: 10.1029/2001GL013343.
Szinell CS, Bussay A, Azentimrey T. 1998. Drought tendencies in
Hungary.
International
Journal
of
Climatology
18:
14791491.
Torrence C, Compo GP. 1998. A practical guide to wavelet analysis.
Bulletin of American Meteorological Society 79(1): 6178.
Torrence C, Webster PJ. 1999. Interdecadal changes in the ENSO
monsoon system. Journal of Climatology 12: 26792690.
Webster PJ, Hoyos CD. 2004. Prediction of monsoon rainfall and
river discharge on 15-30-day time scales. Bulletin of the American
Meteorological Society 85(11): 17451765.
Zadeh LA. 1965. Fuzzy sets. Information and Control 12(2): 94102.

Int. J. Climatol. 31: 20212032 (2011)

You might also like