Environmental Earth Sciences (2019) 78:693

https://doi.org/10.1007/s12665-019-8700-7

ORIGINAL ARTICLE

Coupling fuzzy–SVR and boosting–SVR models with wavelet

decomposition for meteorological drought prediction
Kit Fai Fung1 · Yuk Feng Huang1   · Chai Hoon Koo1

Received: 4 January 2019 / Accepted: 23 November 2019 / Published online: 4 December 2019
© Springer-Verlag GmbH Germany, part of Springer Nature 2019

Abstract
Drought is a climatic occurrence of prolonged and abnormal moisture deficiency resulting from meteorological anomalies.
Despite its negative impact to agricultural activity and water resources management, drought is still a poorly comprehended
calamity, primarily due to the difficulties ascertaning its onset. Effective drought prediction is important for any development
of a sustainable natural environment. This study discusses the wavelet–boosting–support vector regression (W–BS–SVR),
multi-input wavelet–fuzzy–support vector regression (multi-input W–F–SVR) and weighted wavelet–fuzzy–support vector
regression (weighted W–F–SVR) models for meteorological drought predictions, at the downstream of the Langat River
Basin; with lead times of 1 month, 3 months, and 6 months. Drought severity is described by the Standardized Precipitation
Evapotranspiration Indices (SPEIs) with different timescales of 1 month, 3 months, and 6 months, respectively, known as
SPEI-1, SPEI-3, and SPEI-6. The observed SPEIs from 1976 to 2007 were used for model training, while the SPEIs from
2008 to 2015 were for model validation. The root-mean-square-error (RMSE), mean absolute error (MAE), coefficient of
determination (R2), and adjusted R2 were applied to assess the performance of models. In general, it was found that the
fuzzy-based hybrid model, the weighted W–F–SVR predicted well for SPEI-1, SPEI-3, and SPEI-6 cases, with lead times
of 3 and 6 months. As for the 1-month lead time predictions, the models’ performances were dominated by the temporal
variation in the SPEIs, where the weighted W–F–SVR that is capable in reducing outlier effects, performed best for high
variation SPEI-1 and SPEI-3, while the W–BS–SVR model was better for SPEI-6.

Keywords  Meteorological drought prediction · SPEI · Wavelet · Fuzzy logic · Boosting ensemble · Support vector
regression

Introduction are some rare and extreme drought events that vary signifi-
cantly in time and extent (Burke et al. 2010). These charac-
Drought is a phenomenon of prolonged and abnormal mois- teristics of droughts cause difficulties in drought mitigations.
ture deficiency caused by meteorological anomalies. The Nonetheless, hydrologic scientists are giving greater efforts
impact from drought is damaging, and accordingly, it is the to produce effective monitoring and forecasting methods to
least comprehended of all natural hazards (Pulwarty and manage droughts (Bachmair et al. 2016).
Sivakumar 2014). Compared to aridity, drought is a tem- Drought is a natural phenomenon that occurs in both
porary aberration and does not last permanently. Drought high and low precipitation regions and all climatic
events develop slowly over time and have long-lasting effects schemes (Djerbouai and Souag-Gamane 2016). Although
(Wilhite et al. 2014). This causes the determination of the Malaysia has a tropical climate and receives an average
onset and offset of droughts to be difficult. Above all, there of 2800 mm of precipitation annually, varying precipita-
tion amount and rain day occurrence have made it drought
vulnerable. For example, the variability of precipitation
* Yuk Feng Huang amount and rain days in the wet and dry seasons can be
huangyf@utar.edu.my
extreme at times, causing difficulties in sustaining water
1
Department of Civil Engineering, Lee Kong Chian Faculty storage and supply management, which generally relies
of Engineering and Science, Universiti Tunku Abdul on direct rainwater and rainwater stored in dams. In the
Rahman (UTAR) - Sungai Long Campus, Jalan Sg. Long, past, several severe droughts occurred in the basin and
Bandar Sg. Long, Cheras, 43000 Kajang, Selangor, Malaysia

13
Vol.:(0123456789)
693 
Page 2 of 18 Environmental Earth Sciences (2019) 78:693

its surrounding areas, which included the 1991 Malacca The SPEI is a meteorological drought index that com-
water crisis, the 1998 Klang Valley water crisis (El-Nino), prises the simple and multi-scalar characteristic of SPI, and
and the 2014 Selangor water crisis (Abdulah et al. 2014). the hydrologically and ecologically responding character-
Given the vulnerability of the basin to impending drought istic of the PDSI. Thus, it is able to better describe and for
events, improvements on drought monitoring and predic- monitoring different types of droughts. The SPEI had been
tion are required. tested on 11 observations from different types of climate,
Conceptually, drought is characterized with the remark- from the tropical, monsoon, Mediterranean, semi-arid, conti-
ably short fall in precipitation, which will then affect the nental, cold, and to the oceanic weather conditions (Vicente-
water demand for either ecosystem or human activities (Wil- Serrano et al. 2010). This ensured the suitability to apply
hite et al. 2014). Hence, due to the huge impact on the envi- the SPEI to a tropical country such as Malaysia, to describe
ronment and the socio-economy caused by water scarcity, the severity of the events. Furthermore, the future incre-
it is important to figure out the starting and ending dates ment of the average annual air temperature for the country
as well as severity and magnitude of droughts, also known was also reported to be 0.5–1.0 °C during period of 2030
as operational definition of droughts (Wilhite and Glantz (2020–2040), and further to 0.9–1.6 °C during the period
1985). Based on the operational definition of drought, there of 2050 (2040–2060) (MESTECC 2018). This raises sig-
are four main categories of droughts: meteorological, which nificant concern on the effects of temperature in drought
represents precipitation deficit; agricultural, which repre- formation in Malaysia, and this problem can be solved with
sents moisture deficit following declined of crops yields with using the SPEI that can incorporate both precipitation and
no relevance to surface water resources; hydrological, which temperature conditions. Similarly, many other studies also
represents insufficient flows, in terms of subsurface or sur- devoted efforts in drought monitoring and forecasting using
face water; and socioeconomic, which represents insufficient the SPEI, and proved that it is a robust index to monitor and
water supply to fulfill water demands and leads to negative analyse droughts (Annette Hernandez and Uddamri 2014;
impacts on environment, economy and society. One of the Begueria et al. 2014; Li et al. 2015; Manatsa et al. 2015;
best ways to operationally define drought is through the use Liu et al. 2016; Maca and Pech 2016; Xiao et al. 2016; Alam
of mathematical indices, also known as drought indices. et al. 2017; Byakatonda et al. 2018; Chen et al. 2018; Soh
Over the years, many drought indices have been produced et al. 2018). Therefore, the focus of this study is to predict
in the effort to evaluate the moisture deficiency correspond- droughts using the SPEI, to represent both hydrological and
ing to the time duration of the precipitation shortage. Some ecological conditions of droughts.
of the commonly used indices include the Percent of Normal, The models used for drought forecasting ranges from the
the Standardized Precipitation Index (SPI) (McKee et al. simplistic techniques to the more complex approaches. First,
1993), the Palmer Drought Severity Index (PDSI) (Palmer the stochastic models autoregressive integrated moving aver-
1965), the Crop Moisture Index (CMI) (Bordi and Sutera age (ARIMA) and Markov Chain were adopted. Stochastic
2007), the Streamflow Drought Index (SDI) (Nalbantis and models have a long history engaged in effective drought
Tsakiris 2009), and the Standardized Precipitation Evapo- forecasting (Paulo and Pereira 2007; Abebe and Foerch
transpiration Index, SPEI (Vicente-Serrano et al. 2010). 2008; Fernández-Manso et  al. 2011; Alam et  al. 2014;
These indices serve to estimate drought severity using differ- Bazrafshan et al. 2015; Mossad and Alazba 2015; Khadr
ent indicators (e.g., temperature, soil moisture, streamflow, 2016; Mahmud et al. 2016; Sun et al. 2016; Tian et al. 2016;
etc.) to represent various categories of drought. According Karthika et al. 2017; Zhang et al. 2017). However, they are
to the WMO (2012), soil moisture conditions respond to basically linear models with limited ability to consider the
precipitation anomalies over a relatively shorter timescale, nonlinearity and non-stationarity of hydrologic data (Kim
whilst groundwater, streamflow and reservoir storage better and Valdés 2003). The high dependence on one’s capability
reflect the longer term precipitation anomalies. Hence, the in terms of identification, estimation, and diagnostic checks
SPI is commonly used for drought monitoring studies (Liu for model selection is also another limitation of these sto-
et al. 2016; Stagge et al. 2015; Venkataraman et al. 2016), chastic models. As an alternative, the machine-learning
as its multi-scalar characteristic enables it to represent dif- (ML) approaches have in recent years become popular to
ferent categories of drought by considering the cumulative handle the hydrologic predictions.
effect of the precipitation deficiency. However, its univariate The MLs are well known for their high flexibility and
characteristics of only considering precipitation are said to adaptability to nonlinearity of the data. They can be used to
be unsuitable in this era of global warming, where tempera- predict droughts that have different durations, frequencies,
ture and evaporation play an increasingly important role for and intensities. Moreover, with the use of MLs, outstand-
drought monitoring (Yoon et al. 2012). Hence, the SPEI had ing performance and accuracy had been achieved in various
been introduced (Vicente-Serrano et al. 2010) in view of studies (Ganguli and Reddy 2014; Masinde 2014; Borji et al.
these problems. 2016; Rezaeianzadeh et al. 2016; Zhang et al. 2016; Ali et al.

13
Environmental Earth Sciences (2019) 78:693 Page 3 of 18  693

2017; Kousari et al. 2017; Seibert et al. 2017). Among the over the rolling hills. There is no other water source or irri-
MLs, a sub-category of the SVM that is, the Support Vector gation system available for the oil palm plantations. Accord-
Regression (SVR), is emerging in recent years. Its concept ing to the USDA (2016), oil palm yields typically decline
of the structural risk minimization that had brought about 4–12 months after a rainfall-suppressed drought event has
better generalized prediction in drought forecasting were dis- occurred. This suggests that the availability of direct pre-
covered in various studies (Belayneh and Adamowski 2013; cipitation and the climatic water demand in the study area
Chiang and Tsai 2012–2013; Jalili et al. 2014; Jalalkamali will have a direct impact on the oil palm yield production.
et al. 2015; Borji et al. 2016; Yu 2017). Hence, the SPEI with timescales of 1 month, 3 months, and
To further improve the performances, the hydrid MLs 6 months were used in this study to represent the atmos-
have also been produced for drought forecasting. To the pheric moisture conditions. As for prediction, two fuzzy
best knowledge of the authors, the very first hybrid drought (multi-input W–F–SVR and weighted W–F–SVR) and one
forecasting model in the history was produced by Mishra boosting ensemble (W–BS–SVR), SVR-based models with
et al. (2007) with the combining of the advantages of the wavelet transform were proposed in this study. To the best
ARIMA in modelling linear data and the artificial neural knowledge of the authors, the wavelet pre-processed SVR-
network (ANN) for nonlinear data. The results showed that based drought prediction models coupled with the fuzzy
the hybrid models were able to further forecast droughts or the boosting technique, using SPEI as predictor, have
with greater accuracy. Given the capability of hybrid mod- never been applied to the Langat River Basin. The fuzzy
els to produce effective forecasting, wavelet pre-processing, membership functions from Lin and Wang (2002), which
ensemble, and other techniques have also been adopted to are able to reduce outlier effects, were used to generate the
further improve the original MLs in different aspects (Ozger fuzzy membership values in the fuzzy models. These val-
et al. 2011; Shirmohammadi et al. 2013; Woli et al. 2013; ues were then used as additional inputs in the multi-input
Belayneh et al. 2014; Mehr et al. 2014; Jalalkamali et al. W–F–SVR model and as weightage of each data in the
2015; Nguyen et al. 2015; Deo et al. 2016; Memarian et al. weighted W–F–SVR model. Following this, the models were
2016; Prasad et al. 2017). then assessed and the best ones were recommended for the
As a pre-processing technique, the wavelet transforma- basin. The expected results of this study are the SPEI series
tion is an effective approach to reduce the data noise and with various timescales and lead times predicted using these
produce more accurate predictions (Adamowski and Sun three models.
2010; Ozger et al. 2012; Belayneh and Adamowski 2013;
Mehr et al. 2014; Djerbouai and Souag-Gamane 2016). As
for the boosting ensemble technique, Freund and Schapire Materials and methods
(1996) stated that it can improve the performance of a given
learning algorithm by considering the learning effects from Study area and data acquisition
the weak learners. A recent study done by Belayneh et al.
(2016) also showed that the boosting technique improved The study area comprises the Langat River Basin, located
the performance of the SVR models for the prediction of over two Peninsular Malaysia states, the Selangor and Neg-
the SPI. Furthermore, there are cases, where data points eri Sembilan states, within latitudes 2°35′40″N to 3°16′30″N
corrupted by noises are less meaningful, and it is better to and longitudes 101°17′10″E to 102°0′0″E. The basin has a
discard them, but, however, the SVR lacks this kind of abil- total area of 2400 km2 and can be divided into three dis-
ity. A method to improve the solution is through the use of tinct zones: mountainous northeast, hilly middle, and plain
fuzzy systems with the SVR, so that different input points southwest (Juahir 2008). It was reported that 60% of Lan-
can make different contributions to the learning of decision gat River Basin is used for agriculture activity (DOA 1995;
surface. These models are known as the fuzzy–SVR and they JICA 2002) with oil palm as the major crop. Oil palm planta-
have shown outstanding performances in predicting runoff tions are located downstream with an approximate 847 km2
(Wiriyarattanakul et al. 2009), dam drainage (Wiriyarattan- area. Since drought may adversely affect the oil palm planta-
akun et al. 2011), and other applications (Chaudhuri and tions activities in the area, this study mainly focuses on the
Kajal 2011; Allaoua and Laoufi 2013; Hung 2016; Edwin meteorological drought prediction on downstream/southwest
and Somasundaram 2016). area of the basin, as indicated by the pink-dashline in Fig. 1.
The aim of this study is to assess the capabilities of To study the temporal changes in climatic conditions of
hybrid ML models in predicting meteorological droughts the study area, a preliminary study on the linear trends on
at the downstream of Langat River Basin, represented by precipitation and mean temperature was carried out and
multi-scalar drought index. The downstream area of Langat is shown in Fig. 2. The temporal analysis was carried out
River Basin consists mainly of oil palm plantations, which for the timescales of 1 month, 3 months, and 6 months to
solely relies on the direct precipitation for rain-fed irrigation study the monthly, seasonal, and semi-annual trends of the

13

693 

Precipitation (P, mm) Precipitation (P, mm)

13
Precipitation (P, mm)

0
500
1000
1500
2000
0
200
400
600
800
1000
1200
0
100
200
300
400
500
Page 4 of 18

1976 1976 1976

1977 1977 1977
1978 1978 1978
1979 1979 1979
1980 1980 1980
1981 1981 1981
1982 1982 1982
1983 1983 1983
1984 1984 1984
1985 1985 1985
1986 1986 1986
1987 1987 1987
1988 1988 1988
1989 1989 1989
1990 1990 1990
1991 1991 1991
1992 1992 1992
1993 1993 1993
1994 1994 1994
1995 1995 1995
1996 1996 1996

Years
Years
Years
1997 1997 1997
1998 1998 1998
1999 1999 1999

3-month Timescale
1-month Timescale
2000 2000 2000

6-month Timescale
2001 2001 2001
2002 2002 2002
2003 2003 2003
2004 2004 2004
2005 2005 2005
2006 2006 2006
2007 2007 2007
2008 2008 2008
2009 2009 2009
2010 2010 2010
2011 2011 2011

Fig. 2  Time series of precipitation (blue data) and temperature (orange data) with their respective linear trend line
2012 2012 2012

Tmean = 0.0015x + 28.167

2013 2013 2013
Fig. 1  Map of Langat River Basin with the locations of meteorological stations and areal extent of oil palm plantations

Tmean = 0.0015x + 28.164

P = 0.2707x + 744.67
Tmean = 0.0015x + 28.156

2014 2014 2014

P = 0.1504x + 369.93
P = 0.0562x + 122.27
2015 2015 2015

26.0
27.0
28.0
29.0
30.0
31.0
25.0
26.0
27.0
28.0
29.0
30.0
31.0
24.0
26.0
28.0
30.0
32.0

Mean Temperature Mean Temperature Mean Temperature

(Tmean, oC) (Tmean, oC) (Tmean, oC)
Environmental Earth Sciences (2019) 78:693
Environmental Earth Sciences (2019) 78:693 Page 5 of 18  693

climatic condition in the study area, respectively. It can be Table 1  Categories of SPEI

seen that in all three timescales, the precipitation and mean Moisture category SPEI
temperature for the study area are showing an increasing
trend. This suggests that the increase in temperature may Extremely wet 2.00 and above
have direct contribution to the formation of drought, as the Very wet 1.50 to 1.99
atmospheric water demand such as PET will increase due to Moderately wet 1.00 to 1.49
the increase in temperature. Thus, the SPEI considers that Near normal − 0.99 to 0.99
both effects from precipitation and temperature were chosen Moderately dry − 1.00 to − 1.49
to describe drought in this study. The precipitation data were Severely dry − 1.50 to − 1.99
retrieved from the Department of Irrigation and Drainage Extremely dry − 2.00 and below
(DID) Malaysia, while the temperature data were from the
Malaysian Meteorology Department (MMD). Precipitation
station at Pejabat JPS Sg. Manggis (ID: s2815001) located Di = P − PET, (1)
at the central of basin downstream and temperature station at
Petaling Jaya (ID: 48648), both with 40 years (1976–2015) where Di is the difference between precipitation and potential
of data, were used to provide precipitation data and tempera- evapotranspiration, P is the total precipitation over a period
ture data, respectively, to generate SPEIs for the descrip- of time and PET represents the potential evapotranspiration.
tion of drought severity. The locations of the meteorologi- First, a three-parameter distribution was estimated
cal stations are shown in Fig. 1. According to the World through the log-logistic distribution f(x), as shown below:
Meteorological Organization (WMO 1994), the minimum [
𝛽 ( x − 𝛾 )𝛽−1 ( x − 𝛾 )𝛽 ]−2
requirement for the station density is one per 900 km2 for f (x) = 1+ , (2)
coastal region and one per 575 km2 for interior plains; thus, 𝛼 𝛼 𝛼
the precipitation station density seems reasonably correct in
the present study. As for the selection of station to contrib- where α represents the scale, β represents the shape, and γ
ute temperature data, station 48648 happens to be the near- represents the origin parameters, for D values in the range
est available station to the precipitation station s2815001. (γ > D<∞). Then, these parameters were determined using
Hence, the temperature data from station 48648 were used the L-moment method:
to combine with precipitation data from station s2815001 2w1 − w0
for the computation of SPEI in this study. 𝛽= (3)
6w1 − w0 − 6w2

Standardized Precipitation Evapotranspiration ( )

Index (SPEI) w0 − 2w1 𝛽
𝛼= ( ) ( ) (4)
𝛤 1 + 𝛽1 𝛤 1 − 𝛽1
The Standardized Precipitation Evapotranspiration Index
(SPEI) was introduced by Vicente-Serrano et al. (2010).
( ) ( )
The SPEI has the inherent advantages of measuring drought 1 1
severity in the multi-scale nature using simple calculations
𝛾 = w0 − 𝛼𝛤 1 + 𝛤 1− , (5)
𝛽 𝛽
and to describe drought considering both the effects of pre-
cipitation and evapotranspiration. Since the SPEI is a stand- where 𝛤 is the gamma function of β and wl (Ɩ = 0, 1, 2…)
ardized dimensionless index generated from precipitation was computed by probability weighted moments (PWMs)
and temperature data, but not on soil moisture content as through the L-moment method (Hosking and Wallis 1997):
the Palmer Drought Severity Index, it can be compared n ( )
easily across locations and timescales. Hence, the SPEI is 1∑ i − 0.35 l
wl = xi 1 − , (6)
widely accepted in drought forecasting, as it has a broader n i=1 n
range of applications than other drought indices. The SPEIs
where xi is the ordered random sample (x1 < x2 ⋯ < xn ) of
ranges from negative to positive values, indicating dry and
D and n represents the sample size. Then, the probability
wet events, and the magnitude describes the severity of the
distribution function was calculated, as shown below:
event. SPEIs can be grouped into seven categories, as shown
in Table 1. [ ( )𝛽 ]−1
In this study, the SPEI was calculated using the concept 𝛼
F(x) = 1 + . (7)
of deficit or surplus of water, which is the monthly differ- x−𝛾
ence between precipitation and potential evapotranspiration,
as shown below:

13
693 
Page 6 of 18 Environmental Earth Sciences (2019) 78:693

At last, the computed F(x) was converted into Z-stand- decreased when the timescales increased. The average
ardized normal values to obtain the SPEI values using the moving range (AMR) has been adopted in this study to
classical approximation provided by Abramowitz and Stegun quantify the variation in the time series (Montgomery and
(1965): Runger 2014). As shown in the graphical illustration in
Fig. 3, the AMR values decreased from 1.0942 (SPEI-1) to
c0 + c1 W + c2 W 2 0.6472 (SPEI-3) and 0.5622 (SPEI-6). This indicates that
SPEI = W − (8)
1 + d1 W + d2 W 2 + d3 W 3 the fluctuations decreased when the timescales increased.
Since the SPEI with shorter timescale describes drought
√ conditions by accumulating shorter period of climatic vari-
W= −2 ln (P) for P ≤ 0.5, (9)
ables, the current trend in AMR indicates that the study
where P represents the probability of exceeding a deter- area has rapid monthly changes between dry and wet
mined D value: conditions, followed by seasonal (3 months) and semi-
annual (6 months), which is in agreement with the tem-
c0 = 2.515517, c1 = 0.8082853, c2 = 0.010328;
poral changes in fluctuations shown by precipitation and
d1 = 1.432788, d2 = 0.189269, d3 = 0.001308. temperature in Fig. 2. On top of the linear trend analysis
carried out in the previous section, the non-parametric
For P > 0.5, the P was replaced by 1 − P and the sign of
Mann–Kendall and Sen’s slope trend analyses also have
the resultant the SPEI was reversed. With different time-
been used to investigate the relationship between the
scales, the fluctuation of the SPEIs varies. Based on the
temporal changes in precipitation and temperature with
standard operating procedure mentioned in Abdulah et al.
droughts. These are tabulated in Table 2. Regardless of the
(2014), the monitoring and management of droughts in
increasing trend shown in precipitation, neither increasing
Malaysia are done based on the SPI index of timescales up
nor decreasing trend is shown in all three SPEI series. This
to 6 months. Since the SPEI has the multi-scale nature that
suggests that the increase in temperature has imposed an
is similar with the SPI, the timescales of the SPEI adopted
inevitable effect on the formation of droughts in the study
in this study were also 1, 3, and 6 months and computed
area. Hence, SPEIs were used in this study along with the
using the observed data from years 1976–2015, as shown
timescales of 1 month, 3 months, and 6 months.
in Fig. 3. It can be seen that the fluctuations of the series

(a) SPEI-1
SPEI = 6E-05x - 0.0125
3.00
2.00
1.00
SPEI-1

0.00
-1.00
-2.00
-3.00
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
Year
(b) SPEI-3 SPEI = 7E-05x - 0.0183
3.00
2.00
1.00
SPEI-3

0.00
-1.00
-2.00
-3.00
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015

Year
(c) SPEI-6 SPEI = 6E-05x - 0.0155
3.00
2.00
1.00
SPEI-6

0.00
-1.00
-2.00
-3.00
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015

Year

Fig. 3  Graphical illustration of a SPEI-1 (AMR = 1.0942), b SPEI-3 (AMR = 0.6472), and c SPEI-3 (AMR = 0.5622) with their respective linear
trend line

13
Environmental Earth Sciences (2019) 78:693 Page 7 of 18  693

Table 2  Trend test results for precipitation, temperature, and SPEI under 1-month, 3-month, and 6-month timescales
Timescales 1 month 3 months 6 months
Linear slope Mann–Ken- Sen’s slope Linear slope Mann–Ken- Sen’s slope Linear slope Mann–Ken- Sen’s slope
dall dall dall

Precipitation 0.05620 0.24100 0.03900 0.15040 0.79500 0.09800 0.27070 0.86600 0.20500
(mm)
Mean tem- 0.00150 0.02800 0.00100 0.00150 0.27000 0.00100 0.00150 0.08800 0.00100
perature
(°C)
SPEI 0.00006 No trend 0.00000 0.00007 No trend 0.00000 0.00006 No trend 0.00000

Theoretical overview of model components [ ( )] ∑ ( ( )) 1 2

⃗ ,
Rreg f x⃗ = CC l∈ yi − f x��⃗i + W (10)
x ∈X
2
This section provides the theory that is useful in the
i

development of the ML models in this study. It begins where CC is a positive constant that influences a trade-off
by introducing the SVR, discrete wavelet transforma- between an approximation error and the regression vector W ⃗
tion, and concludes by boosting ensemble technique and is a design parameter. The loss function in this expression,
fuzzy membership values used to develop the hybrid mod- which is called an ∈-insensitive loss function ( l∈ ) has the
els (W–BS–SVR, multi-input W–F–SVR, and weighted advantage that it will not need all the input data for describ-
W–F–SVR). ing the regression vector W ⃗  . Further detailed descriptions
can be found in Cimen (2008).

Support vector regression (SVR)

Discrete wavelet transformation
The support vector machine (SVM) is an ML that adopts
Wavelet transforms are mathematical functions that give a
the concept of structural risk minimization in the learning
timescale representation of a given time series and its rela-
algorithm. It was introduced by Vapnik (1995) in an effort to
tionship to analyse non-stationaries. Wavelet transforms can
characterize the properties of learning machines, so that they
reveal trends in the data such as breakdown points, disconti-
can generalize well to unseen data (Kisi and Cimen 2011).
nuities local minima and maxima that other signal analysis
Compared to other MLs, the SVM has the advantage of hav-
techniques might not reveal. Wavelet analysis can also help
ing low sensitivity to the relative number of training exam-
to de-noise a particular data set (Adamowski and Sun 2010).
ples in both positive and negative classes. In addition, the
Wavelet transformation can be grouped into two categories:
concept of the structural risk minimization also prevents the
continuous wavelet transform (CWT) and discrete wavelet
SVMs from the problem of overfitting, compared to other
transform (DWT). The CWT is useful for processing dif-
MLs such as ANN, in which learning algorithm is based
ferent images and signals; however, it is not often used for
on the empirical risk minimization and seeks to minimize
predictions because of the heavy computation time required.
the training error (Chiang and Tsai 2013; Belayneh et al.
Instead, in predictions applications, the discrete wavelet
2016; Borji et al. 2016). As such, the SVM was adopted in
transform is more frequently used. The DWT is a simplified
this study instead of the ANN, one of the more popular ML
approach of the wavelet transform using an independent set
approach. The SVM can be separated into two types: the
of the wavelet scales (Cannas et al. 2006), as shown below:
support vector classification (SVC) and the support vector
( j )
regression (SVR). Since this study is primarily concerned k − m𝜏0 s0
1 ∑
with the prediction of SPEI, SVR was used. 𝜓j,m (m) = √ 𝜓 j
x(k), (11)
| j| k s0
In the regression estimation with SVR, ( ) the purpose is |s0 |
| |
{ dependency}f x⃗ betweenn a set of
to estimate a functional
sampled points {X = x��⃗1 , x��⃗2 , x��⃗3 , …
} , x��⃗i taken from R and tar- where j and m are integers that control the scale and transla-
get values Y = y1 , y2 , y3 , … , yi with yi ∈( R) (Vapnik 1995). tion, respectively, while s0 > 1 is a fixed dilation step, and
j

The main objective is to find a function f x⃗ ∈ F that mini- 𝜏0 is a translation factor that depends on the dilation step. In
mizes a risk function. For this purpose, a regularized risk the process, the DWT decomposes the time series into two
function with the smallest steepness among the functions is components: high pass (details) and low pass (approxima-
normally used and can be expressed as tion). Although DWT has simpler and faster computation,

13
693 
Page 8 of 18 Environmental Earth Sciences (2019) 78:693

it inherited the limitation of not shift invariant. Hence, the that “every proposition is either true or false”. However,
à trous algorithm or non-decimated DWT is commonly to Zadeh (1965) suggested that the true values of variables may
overcome the problem (Mallat 1998): be any real number between 0 and 1, using the concept of
+∞
fuzzy logic. Fuzzy logic violates both “excluded middle”
∑ ( ) and “contradiction” laws (Klir and Yuan 2008), which can be
Ci+1 (k) = h(l)Ci k + 2i l , (12)
l=−∞ done using the fuzzy membership function (FMF) to assign
membership value (or degree/grade of membership) between
where h is the low-pass filter and Ci+1(k) is the original time 0 and 1 to every points in the input space (universe of dis-
series. In this algorithm, the details of components Wi(k) are course). This allows the modelling tasks to be carried out
separated by subtraction from the coarser signal preceded it, with the “Tolerance towards Uncertainty and Imprecision”.
given by (Murtagh et al. 2003) Furthermore, the choice of an FMF to be used is flexible
Wi (k) = Ci−1 (k) − Ci (k), (13) and can be determined by the domain of the application as
well as its simplicity, convenience, speed, and efficiency.
where Ci(k) is the approximation of the signal and Ci-1(k) To the best knowledge of the authors, the application of the
is the coarser signal. Hence, when the decomposition level fuzzy–SVR (F–SVR) has not been explored in drought pre-
increases, smoother approximation and higher level of detail diction. However, the F–SVR models have been used for
will be generated. dam drainage prediction (Wiriyarattanakun et al. 2011) and
other applications (Chaudhuri and Kajal 2011; Allaoua and
Boosting ensemble technique Laoufi 2013; Hung 2016; Edwin and Somasundaram 2016).
Hence, the idea of fuzzy logic was adopted in this study for
The boosting ensemble technique is a method which pro- its exploration in drought prediction and its mathematical
duces a sequence of models to improve the performance of modelling ability to reduce the outlier’s effect by incorporat-
a given learning algorithm, where each succeeding model ing the concept of fuzzy membership values.
focusses on the poorly trained cases in the preceding one to
generate more accurate results (Freund and Schapire 1996). Development of hybrid drought forecasting models
This is a relatively new approach in drought prediction;
hence, it is worthy to explore in improving the effectiveness W–BS–SVR
in drought prediction. To the best knowledge of the authors,
the boosting ensemble had once previously being coupled W–BS–SVR denotes the conjunction of wavelet transfor-
with the SVR for the SPI prediction in Ethiopia (Belayneh mation, boosting ensemble, and SVR. For wavelet trans-
et al. 2016). The results of the study were promising, and so formation, Daubechies family was selected as the mother
too for other applications such as flood forecasting by Erdal wavelet (with vanishing moment, v = 2 or db2) and the
and Karakurt (2013). In boosting ensemble technique, every processes were performed in the MATLAB platform with
new learner will be fitted from the weak learners through the the observed SPEI used as the input for one-dimensional
function showed below: stationary à trous wavelet transform. The choice of db2 was
( ) decided based on the results of Maheswaran and Khosa
yn − 𝜂f xn , (14) (2012). They showed that db2 yielded better prediction
( ) efficiencies for time series with long-term features (e.g.,
where yn is the observed response, f xn is the combined monthly). Hence, it was employed as the data pre-process-
prediction from all weak learners created so far for obser- ing technique in this study to de-noise SPEIs before the
vation xn , and η is the learning rate. The idea of boosting model development. As for the optimum decomposition
ensemble is to generate better rules by combining multiple level, the process was carried out to the maximum allowable
precedent weak learners. When more weak learners are com- level (Level 5 for this study), as there is no theory concern-
bined as ensemble, the training error will drop exponentially ing the determination of the best level of decomposition for
but the computation time will also increase. Hence, the bal- time series (Maheswaran and Khosa 2012; Altunkaynak and
ance between number of weak learners, prediction accuracy, Ozger 2016).
and computation time is important. Thereafter, each decomposed SPEI time series was
boosted using the ‘fitensemble’ function in MATLAB. In
Fuzzy membership values the ‘fitensemble’ boosting process, the appropriate ensem-
ble function and the number of learning cycles are the main
In Boolean or crisp set theory, membership of an element factors affecting the model performances. Since the Least
x in a set A is defined by value of either 1 (true) or 0 (false) Squares Boosting (LS-Boost) fits for regression purposes
to each individual in the universal set X, which indicates (Friedman 1999), the “LSBoost” in MATLAB was used

13
Environmental Earth Sciences (2019) 78:693 Page 9 of 18  693

to carry out the tasks (Cordiner 2009). As for the optimal which were for estimating the radius of class and the fuzzy
number of learning cycles, it was determined by iteration membership functions, respectively:
process using ‘resume’ function, and chosen based on the
smallest MAE during the validation period. This is, because r+ = max||x+ − xi || for positive class
(17)
boosting ensemble is prone to overfitting if the number of r− = max||x− − xi || for negative class,
learning cycles is determined based on the training error,,
and hence, the MAE of validation period, which provides where r+ and r− are the radius of positive and negative
the generalization error was used. Thereafter, the boosted classes, x+ and x− are the mean of positive and negative
de-noised SPEIs were used for the SVR training to develop classes:
the W–BS–SVR model. ( )
si = 1 − ||x+ − xi ||∕ r+ + 𝜎 for positive class
The SVR models in this study were generated using the ( ) (18)
si = 1 − ||x− − xi ||∕ r− + 𝜎 for negative class,
‘fitrsvm’ function in MATLAB platform. As there is no
set rules for data separation, the SPEIs were separated into where si is the fuzzy membership values. Since the equations
years 1976–2007 (80%) for training and years 2008–2015 require the data sets to be separated into positive and nega-
(20%) for validation, which is also known as the leave-one- tive classes, the SPEIs were separated into wet (SPEI > 0)
out cross validation. This was done based on the fulfillment and dry (SPEI < or = 0) periods to carry out the tasks. As the
of 30-year period to define climate, according to the World fuzzy membership values were ready, they were imported
Meteorological Organization. Hence, 32 years (1976–2007) into ‘fitrsvm’ function for the SVR training process, as per
of data was used for training, to allow the climate pattern described in the previous section.
to be fitted into the models. In the case of nonlinear regres- For the multi-input W–F–SVR models, the gener-
sion, the choice of radial basis function (rbf) kernel was said ated fuzzy membership values were used to form multi-
to be appropriate, as it has a good ability to map nonlinear variate training points together
inputs into high dimensional space (Kecman 2001). Thus, {( with) the (de-noised )}
and targeted (SPEIs, )}from x1 , y1 , … , xi , yi to
the parameter “C” and epsilon value for rbf kernel were {( )
x1 , s1 , y1 , … , xi , si , yi  . As for the weighted W–F–SVR
estimated. For this purpose, the practical way suggested by models, the fuzzy membership values were used as weights
Cherkassky and Ma (2004) was adopted in this study, as in the SVR-training process. Both models have the same
shown below: objective to reduce the outlier’s effect, but the authors
(
| | | |
) explored to investigate which will be the better way to assign
C = max |ȳ + 3𝜎y |, |ȳ − 3𝜎y | ,
| | | | (15) degree of membership to the data points.

where ȳ and 𝜎y are the mean and standard deviation of the y

Performance measures
values of training data (targetted data):
√
ln n The following measures were adopted to assess the accuracy
𝜀 = 3𝜎 , (16) of the models in this study:
n
N |
where 𝜎 is the standard deviation of noise and n is the data ∑ |
|ŷi − yi |
The Mean Absolute Error(MAE) = (19)
size. i=1
N

�
Multi‑input W–F–SVR and weighted W–F–SVR ∑N � �2
i=1
ŷi − yi
The Root Mean Square Error (RMSE) =
Multi-input W–F–SVR and weighted W–F–SVR are the N
(20)
models developed by combining the concept of wavelet
∑N � �
transformation, fuzzy logic, and SVR. As same as the previ- � � ŷi − yi
2

ous model, wavelet transformation was used to de-noise the The Coeficient of Determination R2 = ∑i=1 � �2
N
SPEIs. Then, the fuzzy membership values for each data in i=1 yi − yi
the de-noised SPEIs were computed to represent their degree (21)
( )
of membership to the series. 1 − R2 (N − 1)
The fuzzy membership function suggested by Lin and The Adjusted R2 = 1 − , (22)
N−p−1
Wang (2002) was applied in this study to evaluate the degree
of membership of each data to reduce the outlier’s effect. where yi is the mean value taken over N, yi is the observed
This was carried out using the equations shown below, value, ŷi is the predicted value, N is the number of data

13
693 
Page 10 of 18 Environmental Earth Sciences (2019) 78:693

points, and p is the number of predictors. The MAE and Results

RMSE evaluate the similarity between the observed and
predicted values, while the R2 and adjusted R2 measure the The proposed models were used for predicting SPEI-1,
degree of correlation between the observed and predicted SPEI-3, and SPEI-6 with lead times of 1 month, 3 months,
values. and 6 months. Figures 4, 5, and 6 provide the graphical

Fig. 4  a Observed versus pre-

dicted SPEI-1 of the proposed
models for 1-month lead time.
b observed versus predicted
SPEI-3 of the proposed
models for 1-month lead time.
c Observed versus predicted
SPEI-6 of the proposed models
for 1-month lead time

13
Environmental Earth Sciences (2019) 78:693 Page 11 of 18  693

Fig. 5  a Observed versus pre-

dicted SPEI-1 of the proposed
models for 3-month lead time.
b Observed versus predicted
SPEI-3 of the proposed
models for 3-month lead time.
c Observed versus predicted
SPEI-6 of the proposed models
for 3-month lead time

illustration of the SPEI prediction results using the vari- quantify the differences, performance measures were car-
ous proposed models for lead times of 1 month, 3 months, ried out and are shown in Table 3.
and 6 months. From the figures, the difference between First of all, Table 3 shows that the optimum decomposi-
observed and predicted SPEI varies for the different tion levels of the models vary between 2 and 5. Although
models, and timescales and lead times are revealed. To the trend is not significant, it can still be observed that the

13
693 
Page 12 of 18 Environmental Earth Sciences (2019) 78:693

Fig. 6  a Observed versus pre-

dicted SPEI-1 of the proposed
models for 6-month lead time.
b Observed versus predicted
SPEI-3 of the proposed
models for 6-month lead time.
c Observed versus predicted
SPEI-6 of the proposed models
for 6-month lead time

optimum decomposition level decreased when timescales time, levels 3–4 for 3-month lead time, and levels 4–5 for
of SPEIs increased, which were levels 3–5 for SPEI-1, lev- 1-month lead time.
els 3–4 for SPEI-3, and levels 2–4 for SPEI-6. As for the For the models performance evaluation, it can be seen
changes across lead time, the trend was much significant, that the models generally improved over increasing time-
and that is, higher decomposition levels were required when scales, but deteriorated over increasing lead times. For the
prediction lead time increased: levels 2–4 for 1-month lead SPEI-1, it can be observed from Table 3 that the predictions

13
Environmental Earth Sciences (2019) 78:693 Page 13 of 18  693

for this timescale generally showed the poorest performance

Adj R2

0.798
0.854
0.861
0.719
0.715
0.690
0.703
0.418
0.650
with the highest errors and lowest correlations in this study,
which were 0.124–0.579 (MAE), 0.156–0.726 (RMSE),
0.420–0.990 (R2), and 0.418–0.988 (Adj R2). By relating to

0.854
0.861
0.720
0.716
0.801
0.692
0.703
0.420
0.650
the fluctuations in the series, this poor performance in pre-
R2
dicting SPEI-1 may due to the high variations in the series.
This is further supported by the evidence that the prediction
RMSE

0.396
0.452
0.442
0.476
0.433
0.543
0.481
0.726
0.664 of SPEI-3 and SPEI-6 has increased MAE, RMSE, R2, and
Adj R2 performance compared to those of the SPEI-1, which
6-month lead time

are 0.110–0.405, 0.138–0.481, 0.703–0.990, 0.703–0.990

0.308
0.352
0.327
0.389
0.335
0.579
0.501
0.429
0.405
MAE

and 0.090–0.352, 0.137–0.452, 0.801–0.988, 0.798–0.987,

respectively. Furthermore, the improvement of the models
WL*

accuracy over increasing timescales is observed in Figs. 4,

5 and 6, where the differences between the predicted and
4
4
4
4
4
4
4
5
4

observed series reduced over increased timescales. By com-

2
Adj R

0.912
0.879
0.900
0.854
0.864
0.838
0.822
0.798
0.800

bining these findings with the fluctuation of series over time-

scales, it is deemed that the accuracy of the models used in
this study depends on the fluctuations of the series, which
0.912
0.880
0.900
0.856
0.864
0.799
0.800
0.838
0.822

is increasing accuracy when the fluctuations in the series

2
R

decrease.
As for the performance variation over lead time, the
RMSE

0.279
0.261
0.177
0.375
0.348
0.397
0.381
0.474
0.450

models performed best in 1-month lead-time predic-

tions, in which the performance of the models ranges with
3-month lead time

0.090–0.205 (MAE), 0.137–0.249 (RMSE), 0.982–0.990

0.156
0.218
0.203
0.292
0.276
0.306
0.300
0.395
0.359
MAE

(R 2), and 0.982–0.990 (Adj R 2). This is followed by

3-month lead time that has performance measures of
WL*

0.156–0.395 (MAE), 0.177–0.474 (RMSE), 0.799–0.912

(R2), 0.798–0.912 (Adj R2); and 6-month lead time with
3
4
3
4
4
3
4
4
4

performance measures 0.308–0.579 (MAE), 0.396–0.726

2
Adj R

0.987
0.987
0.984
0.988
0.990
0.988
0.982
0.987
0.985

(RMSE), 0.420–0.861 (R2), and 0.418–0.861 (Adj R2). Other

than the performance measures, the graphical illustrations of
the models also showed significant increase in the difference
0.985
0.987
0.988
0.988
0.990
0.990
0.982
0.987
0.986

between the predicted and the observed SPEI, from Fig. 4

2
R

(1-month lead time) to Fig. 5 (3-month lead time) and from

Bold values indicate the best results of performance measures for models

Fig. 5 (3-month lead time) to Fig. 6 (6-month lead time). The

RMSE

0.147
0.152
0.137
0.153
0.138
0.156
0.158
0.249
0.162

decrease in performances showed that the learning process

became more difficult when the lead time increased.
1-month lead time

For the comparison between models, the best performing

0.108
0.107
0.119
0.102
0.090
0.126
0.110
0.205
0.124
MAE

models vary across timescales under 1-month lead time. The

Table 3  Results of performance measures for models

weighted W–F–SVR model showed the highest overall accu-

WL*

racy for the SPEI-1, with the lowest RMSE and highest R2 and
WL* = optimum wavelet decomposition level
2
2
3
3
3
3
4
3
3

Adj R2 achieved in Table 3. For MAE, it achieved the second

lowest error, which is 0.126; second to multi-input W–F–SVR
Multi-input W–F–SVR
Multi-input W–F–SVR
Multi-input W–F–SVR

Weighted W–F–SVR
Weighted W–F–SVR
Weighted W–F–SVR

model of 0.124 and this is in agreement with the high similar-

ity between the two models found in Fig. 4a. For the SPEI-
3, the weighted W–F–SVR model again showed the lowest
W–BS–SVR
W–BS–SVR
W–BS–SVR

errors (MAE = 0.306, RMSE = 0.397) and highest correlations

(R2 = 0.838, Adj R2 = 0.838) in Table 3 and highest similar-
Model

ity in Fig. 4b. As for the SPEI-6, the best performance was

achieved by the W–BS–SVR model. It has the lowest MAE
Timescales

­ 2 and Adj R
of 0.090 with the correlations of 0.987 for both R ­ 2.
SPEI-6
SPEI-3
SPEI-1

However, it was observed from Fig. 4c that it has relatively

high error for the peaks of the curves (e.g., during Aug 2009,

13
693 
Page 14 of 18 Environmental Earth Sciences (2019) 78:693

Jul 2012). Hence, the RMSE that has better measures on peak lead times increases. These may be explained using the role
error did not show the lowest value for W–BS–SVR model. of wavelet transformation in this study. Wavelet transfor-
In summary, the results showed that the fuzzy-based models mation was used as a pre-processing technique to de-noise
have similar and more accurate results for the SPEI-1 and the and improve the quality of the data. During the process,
SPEI-3; while the W–BS–SVR model has the best prediction when the decomposition level increased, it indicates that
for the SPEI-6 for the 1-month lead time predictions. more high-pass (details) components were separated from
For the case of the 3-month lead time, both Table 3 and the original time series, and a smoother approximation com-
Fig. 5 show that the weighted W–F–SVR model has the most ponent was produced. With consideration of fluctuations
accurate prediction for all the SPEI-1, SPEI-3, and SPEI- in the original SPEI that shows a decreasing trend from 1
6. In terms of performance measures, weighted W–F–SVR month to 6 months (Sect. 2.2), it is deemed that a higher
obtained the lowest errors with highest correlations for all level of decomposition is required to extract the information
the three SPEIs. As for graphical illustration, Fig. 5 also contained in the original series when it has high fluctua-
consistently showed that the weighted W–F–SVR model tions. As for the increasing trend of optimum decomposi-
has the lowest difference with the observed SPEI, especially tion level over increasing lead time, it is also believed that
during the during the period of Dec 2012–Jul 2014 in the more smoothened series is needed by the models to fit the
SPEI-6 (Fig. 5c), where the predicted SPEI from weighted data when predictions for longer lead time is needed. This
W–F–SVR model is almost the same with the observed is in agreement with the finding of Paresh (2012), which
SPEI. However, the performance of the model has decreased showed similar phenomena and explained that this is due to
compared to the accuracy in 1-month lead time, and this the higher effect of correlation of more smoothened signals
happened to the other two models as well. Among the three with flattened variability between the inputs and output.
models, the accuracy of the W–BS–SVR model decreased For the variation of model performances under different
most, e.g., MAE reduced from 0.090 (1-month lead time) timescales of the SPEI, the findings may be explained by
to 0.218 (3-month lead time) when predicting SPEI-6, the relating the differences in the models’ training algorithms
timescale that W–BS–SVR model outperformed the others with the fluctuations in the SPEI series. The main difference
in 1-month lead time. Compared to the lower decreased in in the models are that the W–BS–SVR model improve the
performance shown by the fuzzy-based models, these find- predictions by boosting and combining the precedent weak
ings suggest that the W–BS–SVR model is weak in predict- learners, while fuzzy-based models improve the predictions
ing longer lead time predictions. by reducing the outliers effect. Given the findings shown in
Similarly, the weighted W–F–SVR model also showed Fig. 3 that the SPEI-1 has the highest fluctuations in series
the lowest prediction errors and highest correlations between (Average Moving Range = 1.0942) among the three SPEIs,
the observed and predicted SPEI in the 6-month lead time followed by SPEI-3 (Average Moving Range = 0.6472), and
prediction, for all SPEI-1, SPEI-3, and SPEI-6 (Table 3 and SPEI-6 (Average Moving Range = 0.5622); fuzzy-based
Fig. 6). Among the three SPEIs, the weighted W–F–SVR models achieved accurate predictions for the SPEI-1 of all
model performed best in predicting the SPEI-6. This is in lead times can be related to their capability to reduce the
agreement with the overall performance trend over increas- effects from fluctuations using fuzzy membership values.
ing timescales, which was discussed earlier on in this sec- Since the algorithm in fuzzy-based models are to assign dif-
tion. This is followed by the multi-input W–F–SVR model ferent fuzzy membership values to the data points by refer-
and, then, the W–BS–SVR model. However, the perfor- ring to their difference from the class mean, the outliers that
mances of the models had further reduced in the case of caused high fluctuations will have lower influence in the
the 6-month lead time. By combining the findings from all training process. In other words, the fluctuations in the series
three lead times, it was found that the performances of the have been lowered in fuzzy-based models using fuzzy mem-
models decreased when the lead time increased. In addition, bership values. Fuzzy-based models also outperformed in
the weighted W–F–SVR model also found to be the model the prediction of SPEI-3; even the variations in the SPEI-3
that is least affected by the increased lead time, as shown by series is lower compared to in the SPEI-1. However, it can
the lowest deterioration of performance of the model when be observed that the difference in performance measures
lead time increased. between the fuzzy-based models and the W–BS–SVR model
is lower compared to in the SPEI-1. Thus, the advantage of
fuzzy-based models to produce accurate prediction by reduc-
Discussion ing outlier effects is getting less significant over the increas-
ing timescales of the SPEI. As for the SPEI-6, which has
Based on the findings in the previous section, it was found the lowest sensitivity to the changes in temperature and pre-
that the optimum decomposition levels decrease when time- cipitation (lowest average moving range values), the fuzzy-
scales of SPEIs increase, but increase when the prediction based models also outperformed the W–BS–SVR model,

13
Environmental Earth Sciences (2019) 78:693 Page 15 of 18  693

except for the case of a 1-month lead time that showed and 6-month lead time predictions, the fuzzy-based mod-
similar results among the three models. In summary, the els do have the advantage over the W–BS–SVR model
fuzzy-based models have advantages over the W–BS–SVR for longer lead time predictions. The reason behind the
model for the prediction of high fluctuation series in this W–BS–SVR model being not predicting well for longer
study (SPEI-1 and SPEI-3) due to its outlier effects reducing lead times could lie with its algorithm in boosting the
function. However, their advantages became less significant weak learners. As aforementioned, the correlation between
when the timescales increased (lower fluctuation), and even training and targeted values decreases when the lead time
being outperformed by the W–BS–SVR model in the case increases. Because of this, the models’ training process
of a 1-month lead time prediction of the SPEI-6. become harder and the number of poorly trained cases
As for the variation of model performances under dif- also increases. This raises the number of boosting cycles
ferent lead times, different explanations are required for needed and more preceding weak learners to be combined
each lead time due to different dominating factors. For the as a new ensemble and, thereafter, increases the complex-
1-month lead time, the models’ target period is the nearest ity of the regression. The increase in model complexity due
future to the input period of the model, compared to the to high ensembles may further burden the training process
3-month and 6-month lead time. This makes it the easiest in SVR component, in addition to the increase in train-
training among the lead times used in this study, which is ing difficulties due to the decrease in correlation between
due to the higher correlation between input and target data. training and targeted values. On the other hand, fuzzy-
Hence, the accuracy of the models was dominated by the based models come incorporated with the idea of reduc-
fluctuations in the SPEI series, which is indicated by aver- ing outlier effects by either adding the fuzzy membership
age moving range values. Given the explanations in previ- values as influencing factor in the SPEI prediction (multi-
ous discussion on the behaviour of models’ accuracy when input W–F–SVR) or giving weights (weighted W–F–SVR)
average moving range of the SPEI series increases, fuzzy- to alter the contribution of each training points. Compared
based models performed better in predicting SPEI-1, SPEI- to boosting, the complexity of each training cases remains
3, while the W–BS–SVR model is best at predicting SPEI-6 even when the training cases become more difficult. The
but with similar performances with the fuzzy models due algorithm itself does not further burden the training pro-
to the insignificant advantages among the models in this cess when lead time increases. In addition, the outliers
timescale. For the 3-month lead time, since the correlation reducing effects of the fuzzy-based models depend on the
between the input and targeted data has decreased, the accu- mean value of the positive and negative classes, and the
racy of the models started to be dominated by the difficulties fuzzy membership values do not change even if the train-
caused by the increased lead time period. The fuzzy-based ing cases become harder when lead time increases. As for
models seemed to have the ability in reducing the training the comparison between the multi-input W–F–SVR and
difficulties through reducing the outlier effects, as shown the weighted W–F–SVR models, the better performances
in the results of the fuzzy-based models outperforming the shown in the later models also tell that using fuzzy mem-
W–BS–SVR model for all SPEIs in 3-month lead time pre- bership values, as weightage for each data points is a better
diction. The general trend that optimum decomposition level approach to incorporate the imprecision tolerance concept
is increasing with the increased lead time also showed that into SVR models, compared to using them as direct influ-
reducing the variations in the SPEI series is required when encing parameters.
the lead time increases. Thus, fuzzy-based models, which In summing up, this study found that the boosting
were designed to reduce outlier effects, are superior to the ensemble technique in W–BS–SVR model is less effi-
W–BS–SVR model, and especially the weighted W–F–SVR cient than the outliers reducing effects of fuzzy member-
models, and seemed to be more efficient in reducing the ship values in the fuzzy-based models, in improving the
training difficulties and improving the prediction accuracy. prediction accuracy over increasing timescales, except
On a 6-month lead time premise, the accuracy of the mod- for the 1-month lead time prediction of SPEI-6. Further-
els in SPEI-6 predictions was dominated by the difficulties more, it was found that the fuzzy-based models are more
caused by the increased lead time period, instead of the vari- efficient in longer lead time predictions compared to the
ations in the SPEI series. The weighted W–F–SVR models W–BS–SVR model, as it was believed that the complexity
which seemed to be the most capable in improving long lead of W–BS–SVR model increases when the training cases
time prediction accuracy outshone the other two models for become harder due to the increase in lead time, which does
all three SPEIs in 6-month lead time predictions. not happen with the case of the fuzzy-based models. Thus,
Given the results showing that the multi-input compared to the W–BS–SVR model, the fuzzy-based mod-
W–F–SVR and the weighted W–F–SVR models were bet- els especially the weighted W–F–SVR model, have better
ter than the W–BS–SVR model over, both the 3-month accuracy for 3-month and 6-month lead time predictions.

13
693 
Page 16 of 18 Environmental Earth Sciences (2019) 78:693

Conclusions Adamowski J, Sun K (2010) Development of a coupled wavelet trans-

form and neural network method for flow forecasting of non-per-
ennial rivers in semi-arid watersheds. J Hydrol 390:85–91
This study was to assess the efficiency of the Alam ATMJ, Rahman MS, Sadaat AHM (2014) Computational intel-
W–BS–SVR model, multi-input W–F–SVR model and ligence techniques in earth and environmental sciences. Springer,
weighted W–F–SVR model in predicting the SPEI-1, SPEI- Dordrecht
Alam NM, Sharma Elsa Moreira GC, Jana C, Mishra PK, Sharma NK,
3, and SPEI-6 at the downstream end of the Langat River Mandal D (2017) Evaluation of drought using SPEI drought class
Basin, which has a tropical climate pattern. To produce transitions and log-linear models for different agro-ecological
models with sufficient forecasting capability, the predictions regions of India. Phys Chem Earth 100:31–43
were carried out with the respective lead times of 1-month, Ali Z, Hussain I, Faisal M, Nazir HM, Hussain T, Shad MY, Mohamd
Shoukry A, Hussain Gani S (2017) Forecasting drought using
3 months, and 6 months. Apart from improving the SVR multilayer perceptron artificial neural network model. Adv Mete-
model using the boosting ensemble technique and fuzzy orol 2017:1–9
membership values, discrete wavelet transformation was Allaoua B, Laoufi A (2013) A novel sliding mode fuzzy control based
also adopted for de-noising purposes. Based on the predicted on SVM for electric vehicles propulsion system. Energy Proc
36:120–129
SPEIs from the three models investigated, the performance Altunkaynak A, Ozger M (2016) Comparison of discrete and continu-
measures of RMSE, MAE, ­R2, and Adj ­R2 have been carried ous wavelet-multilayer perceptron methods for daily precipitation
out to evaluate the accuracies of the models. Overall results prediction. J Hydrol Eng 7(21):04016014
showed that the two fuzzy models, the weighted W–F–SVR Annette Hernandez E, Uddamri V (2014) Standardized precipitation
evaporation index (SPEI)-based drought assessment in semi-arid
and multi-input W–F–SVR outshone the W–BS–SVR, south Texas. Environ Earth Sci 71(6):2491–2501
except in the case of the 1-month lead time prediction of Bachmair S et al (2016) Drought indicators revisited: the need for a
SPEI-6. The capability of fuzzy-based models to reduce wider consideration of environment and society. Wiley Interdiscip
outlier effects has made them performed best in SPEI-1 Rev Water 3:516–536
Bazrafshan O, Salajegheh A, Bazrafshan J, Mahdavi M, Marj AF
and SPEI-3. As for the longer lead time predictions, their (2015) Hydrological drought forecasting using ARIMA models
outstanding performances are due to their nature to remain (case study: Karkheh Basin). Ecopersia 3:1099–1117
the same model complexity even with the increasing of dif- Begueria S, Vicente-Serrano SM, Reig F, Latorre B (2014) The stand-
ficulties in the training process. In general, the weighted ardized precipitation evapotranspiration index (SPEI) revisited:
parameter fitting, evapotranspiration models, tools, datasets and
W–F–SVR model produces slightly better predictions than drought monitoring. Int J Climatol 34(10):3001–3023
the multi-input W–F–SVR model. More complex ensembles Belayneh A, Adamowski J (2013) Drought forecasting using new
were created by W–BS–SVR model when the difficulties machine learning methods. J Water Level Dev 18:3–12
of the training process increased, either increasing fluctua- Belayneh A, Adamowski J, Khalil B, Ozga-Zielinski B (2014) Long-
term SPI drought forecasting in the Awash River Basin in Ethiopia
tions in SPEIs or increasing lead time prediction. Hence, using wavelet neural networks and wavelet support vector regres-
the W–BS–SVR model best only for the case with the low- sion models. J Hydrol 508:418–429
est fluctuations and shortest lead time among the conditions Belayneh A, Adamowski J, Khalil B, Quilty J (2016) Coupling machine
considered in the study. Future work should be attempted learning methods with wavelet transforms and the bootstrap and
boosting ensemble approaches for drought prediction. Atmos Res
to predict different timescales of the SPEI with different 172–173:37–47
lead times. The robustness of the models in meteorological Borji M, Malekian A, Salajegheh A, Ghadimi M (2016) Multi-time-
drought prediction can be strengthened when more diversi- scale analysis of hydrological drought forecasting using support
fied SPEIs and lead times are available for assessments. The vector regression (SVR) and artificial neural networks (ANN).
Arab J Geosci 9:725
application and the capability of these models can also be Burke E, Perry R, Brown S (2010) An extreme value analysis of
further explored in other basins with different physical and UK drought and projections of change in the future. J Hydrol
meteorological characteristics. 388(1–2):131–143
Byakatonda J, Parida BP, Moalafhi DB, Kenabatho PK (2018) Analysis
Acknowledgements  The authors would like to thank Universiti Tunku of long term drought severity characteristics and trends across
Abdul Rahman Research Fund (UTARRF) for funding this research. semiarid Botswana using two drought indices. Atmosp Res
213:492–508
Cannas B, Fanni A, Sias G, Tronci S, Zedda MK (2006) River flow
forecasting using neural networks and wavelet analysis. Geophys
Res Abstr 7:214–219
References Chaudhuri A, Kajal D (2011) Fuzzy support vector machine for bank-
ruptcy prediction. Appl Soft Comput 11(2):2472–2486
Chen S, Zhang L, Liu X, Gao M, She D (2018) The use of the SPEI
Abdulah N, Juhaimi J, Abdul Rahman K (2014) Capacity development
and TVDI to access temporal-spatial variations in drought condi-
to support National Drought Management Policy. The UN-Water
tions in the middle and lower reaches of the Yangtze River Basin,
Decade Programme on Capacity Development (UNW-DPC),
China. Adv Meteorol 2018:1–11
Hanoi
Cherkassky V, Ma Y (2004) Practical selection of SVM param-
Abebe A, Foerch G (2008) Stochastic simulation of the severity of
eters and noise estimation for SVM regression. Neural Netw
hydrological drought. Water Environ J 22:2–10
17(2004):113–126

13
Environmental Earth Sciences (2019) 78:693 Page 17 of 18  693

Chiang JL, Tsai YS (2012) Reservoir drought prediction using support Li X, He B, Quan X, Liao Z, Bai X (2015) Use of the Standardized
vector machines. Appl Mech Mater 145:455–459 Precipitation Evapotranspiration Index (SPEI) to characterize
Chiang JL, Tsai YS (2013) Reservoir drought prediction using two- the drying trend in Southwest China from 1982–2012. Remote
stage SVM. Appl Mech Mater 284–287:1473–1477 Sens 7(8):10917–10937
Cimen M (2008) Estimation of daily suspended sediments using sup- Lin CF, Wang SD (2002) Fuzzy support vector machines. IEEE
port vector machines. Hydrol Sci J 53:656–666 Trans Neural Netw Learn Syst 13(2):464–471
Cordiner A (2009) Adaboost toolbox: a matlab toolbox for adaptive Liu Z, Wang Y, Shao M, Jia X, Li X (2016) Spatiotemporal analysis
boosting. Advanced Multimedia Research Lab Oratory Informa- of multiscalar drought characteristics across the Loess Plateau
tion Communications Technology Research Institute, University of China. J Hydrol 534:281–299
of Wollongong, Wollongong Maca P, Pech P (2016) Forecasting SPEI and SPI drought indices
Deo RC, Tiwari MK, Adamowski JF, Quilty JM (2016) Forecasting using the integrated artificial neural networks. Comput Intel
effective drought index using a wavelet extreme learning machine Neurosc 2016:1–17
(W-ELM) model. Stoch Environ Res Risk Assess 31:1211–1240 Maheswaran R, Khosa R (2012) Comparative study of different
Djerbouai S, Souag-Gamane D (2016) Drought forecasting using wavelets for hydrologic forecasting. Comput Geosci 46:284–295
neural networks, wavelet neural networks, and stochastic mod- Mahmud I, Bari SH, Ur Rahman MT (2016) Monthly precipitation
els: case of the Algerois Basin in North Algeria. Water Resour forecast of Bangladesh using autoregressive integrated moving
Manag 30:2445–2464 average method. Environ Eng Res 22:162–168
DOA (1995) Land use of Selangor and Negeri Sembilan. Department Mallat SG (1998) A wavelet tour of signal processing. Academic
of Agriculture, Kuala Lumpur Press, San Diego
Edwin DJ, Somasundaram K (2016) Evolutionary fuzzy SVR modeling Manatsa D, Mushore T, Lenouo A (2015) Improved predictability
of weld residual stress. Appl Soft Comput 42:423–430 of droughts over the Southern Africa using the Standardized
Erdal HI, Karakurt O (2013) Advancing monthly streamflow prediction precipitation evapotranspiration index and ENSO. Theor Appl
accuracy of CART models using ensemble learning paradigms. J Climatol 127(1–2):259–274
Hydrol 477:119–128 Masinde M (2014) Artificial neural networks models for predicting
Freund Y, Schapire RE (1996) Experiments with a new boosting algo- effective drought index: factoring effects of precipitation vari-
rithm. In: Proceedings of the thirteenth international conference ability. Mitig Adapt Strat Gl 19(8):1139–1162
on machine learning, Bari, Italy, 3–6 July 1996. Morgan Kauf- McKee TB, Doesken NJ, Kleist J (1993) The relationship of drought
mann Publishers Inc., USA frequency and duration to timescales. In: Proceedings of the 8th
Friedman JH (1999) Greedy function approximation: a gradient boost- conference on applied climatology, 17–22 January, American
ing machine. Ann Stat 29(5):1189–1232 Meteorology Society, Boston, pp 179–184
Ganguli P, Janga Reddy M (2014) Ensemble prediction of regional Mehr AD, Kahya E, Ozger M (2014) A gene-wavelet model for long
droughts using climate inputs and the SVM-copula approach. lead time drought forecasting. J Hydrol 517:691–699
Hydrol Process 28:4989–5009 Memarian H, Bilondi MP, Rezaei M (2016) Drought prediction using
Hung JC (2016) Fuzzy support vector regression model for forecast- co-active neuro-fuzzy inference system, validation, and uncer-
ing stock market volatility. J Intell Fuzzy Syst 31(3):1987–2000 tainty analysis (case study: Birjand, Iran). Theor Appl Climatol
Jalalkamali A, Moradi M, Moradi N (2015) Application of several 125:541–554
artificial intelligence models and ARIMAX model for forecasting MESTECC (2018) Malaysia third national communication and sec-
drought using the Standardized Precipitation Index. Int J Environ ond biennial update report to the UNFCCC. Ministry of Energy,
Sci Technol 12:1201–1210 Science, Technology, Environment And Climate Change
Jalili M, Gharibshah J, Ghavami SM, Beheshtifar M, Farshi R (2014) (MESTECC), Putrajaya
Nationwide prediction of drought conditions in Iran based on Mishra A, Desai V, Singh V (2007) Drought forecasting using a
remote sensing data. IEEE Trans Comput 63:90–101 hybrid stochastic and neural network model. J Hydrol Eng
JICA (2002) The study on the sustainable groundwater resources and 12:626–638
environmental management for the Langat Basin in Malaysia. Montgomery DC, Runger GC (2014) Applied statistics and probability
Japan International Cooperation Agency (JICA), Tokyo for engineers. Wiley, Hoboken
Juahir H (2008) Water quality data analysis and modeling at Langat Mossad A, Alazba AA (2015) Drought forecasting using stochastic
River Basin. Ph.D. dissertation, Universiti Putra Malaysia models in a hyper-arid climate. Atmosphere 6:410–430
Karthika K, Krishnaveni, Thirunavukkarasu V (2017) Forecasting of Murtagh F, Starcj JL, Renaud O (2003) On neuro-wavelet modeling.
meteorological drought using ARIMA model. Indian J Agric Res Decis Support Syst 37:475–484
51:103–111 Nalbantis I, Tsakiris G (2009) Assessment of hydrological droughts
Kecman V (2001) Learning and soft computing. MIT Press, London revisited. Water Resour Manag 23:881–897
Khadr M (2016) Forecasting of meteorological drought using Hid- Nguyen LB, Li QF, Ngoc TA, Hiramatsu K (2015) Adaptive neuro-
den Markov Model (case study: the upper Blue Nile river basin, fuzzy inference system for drought forecasting in the Cai River
Ethiopia). Ain Shams Eng J 7:47–56 basin in Vietnam. J Fac Agric Kyushu Univ 60:405–415
Kim TW, Valdés JB (2003) Nonlinear model for drought forecasting Ozger M, Mishra AK, Singh VP (2011) Estimating Palmer Drought
based on a conjunction of wavelet transforms and neural networks. Severity Index using a wavelet fuzzy logic model based on mete-
J Hydrol Eng 8(6):319–328 orological variables. Int J Climatol 31:2021–2032
Kisi O, Cimen M (2011) A wavelet-support vector machine con- Ozger M, Mishra AK, Singh VP (2012) Long lead time drought fore-
junction model for monthly streamflow forecasting. J Hydrol casting using a wavelet and fuzzy logic combination model: a case
399:132–140 study in Texas. J Hydrometeorol 13:284–297
Klir GJ, Yuan B (2008) Fuzzy sets and fuzzy logic, theory and appli- Palmer WC (1965) Meteorological drought. Research Paper No. 45.
cations. Prentice Hall PTR, Upper Saddle River US Weather Bureau, Washington
Kousari MR, Hosseini ME, Ahani H, Hakimelahi H (2017) Intro- Paulo AA, Pereira LS (2007) Prediction of SPI drought class transi-
ducing an operational method to forecast long-term regional tions using Markov chains. Water Resour Manag 21:1813–1827
drought based on the application of artificial intelligence capa- Prasad R, Deo RC, Li Y, Maraseni T (2017) Input selection and per-
bilities. Theor Appl Climatol 127:361–380 formance optimization of ANN-based streamflow forecasts in

13
693 
Page 18 of 18 Environmental Earth Sciences (2019) 78:693

the drought-prone Murray Darling Basin region using IIS and Vicente-Serrano SM, Beguería S, López-Moreno JI (2010) A multisca-
MODWT algorithm. Atmos Res 197:42–63 lar drought index sensitive to global warming: the standardized
Pulwarty RS, Sivakumar MVK (2014) Information systems in a precipitation evapotranspiration index. J Clim 23(7):1696–1718
changing climate: early warnings and drought risk management. Wilhite DA, Glantz M (1985) Understanding the drought phenomenon:
Weather Clim Extrem 3:14–21 the role of definitions. Water Int 10(3):111–120
Rezaeianzadeh M, Stein A, Cox JP (2016) Drought forecasting using Wilhite DA, Sivakumar MVK, Pulwarty R (2014) Managing drought
Markov chain model and artificial neural networks. Water Resour risk in a changing climate: the role of national drought policy.
Manag 30:2245–2259 Weather Clim Extrem 3:4–13
Seibert M, Merz B, Apel H (2017) Seasonal forecasting of hydrologi- Wiriyarattanakul S, Auephanwiriyakul S, Theera-Umpon N (2009)
cal drought in the Limpopo Basin: a comparison of statistical Runoff forecasting using fuzzy support vector regression. In:
methods. Hydrol Earth Syst Sci 21:1611–1629 2008 International symposium on intelligent signal processing and
Shirmohammadi B, Moradi H, Moosavi V, Semiromi MT, Zeinali A communication systems, Thailand, 8–10 February 2009. IEEE,
(2013) Forecasting of meteorological drought using Wavelet- Bangkok, Thailand
ANFIS hybrid model for different time steps (case study: South- Wiriyarattanakun S, Ruengsiriwatanakun A, Noimanee S (2011) Drain-
eastern part of east Azerbaijan province, Iran). Nat Hazards age prediction for dam using fuzzy support vector regression. Int
69:389–402 J Comput Electr Autom Control Inf Eng 5(4):386–390
Soh YW, Koo CH, Huang YF, Fung KF (2018) Application of artifi- WMO (1994) Guide to hydrological practices. World Meteorologi-
cial intelligence models for the prediction of standardized pre- cal Organization (WMO) and Global Water Partnership (GWP),
cipitation evapotranspiration index (SPEI) at Langat River Basin, Geneva
Malaysia. Comput Electron Agr 144:164–173 WMO (2012) Standardized Precipitation Index user guide. World
Stagge JH, Kohn I, Tallaksen LM, Stahl K (2015) Modeling drought Meteorological Organization (WMO) and Global Water Partner-
impact occurrence based on meteorological drought indices in ship (GWP), Geneva
Europe. J Hydrol 530:37–50 Woli P, Jones J, Ingram K, Paz J (2013) Forecasting drought using the
Sun P, Zhang Q, Singh VP, Xiao M, Zhang X (2016) Transitional Agricultural Reference Index for drought (ARID): a case study.
variations and risk of hydro-meteorological droughts in the Tarim Weather Forecast 28:427–443
River basin, China. Stoch Environ Res Risk Assess 31:1515–1526 Xiao M, Zhang Q, Singh VP, Liu L (2016) Transitional properties of
Tian M, Wang P, Khan J (2016) Drought forecasting with Vegetation droughts and related impacts of climate indices in the Pearl River
Temperature Condition Index using ARIMA models in the Guan- basin, China. J Hydrol 534:397–406
zhong Plain. Remote Sens 8:690 Yoon J-H, Mo K, Wood EF (2012) Dynamic-model-based seasonal
USDA (2016) Malaysia: El Nino Takes a Bite Out of 2015/16 Palm Oil prediction of meteorological drought over the contiguous United
Production. International Production Estimates Division (IPAD), States. J Hydrometeorol 13:463–482
Office of Global Analysis (OGA), USA Zadeh LA (1965) Fuzzy sets. Inform Control 8:338–353
Vapnik V (1995) The nature of statistical learning theory. Springer, Zhang T, Li J, Hu R, Wang Y, Feng P (2017) Drought class transition
New York analysis through different models: a case study in North China.
Venkataraman K, Tummuri S, Medina A, Perry J (2016) 21st Century Water Sci Technol Water Supply 17:138–150
drought outlook for major climate divisions of Texas based on
CMIP5 multimodel ensemble: implications for water resource Publisher’s Note Springer Nature remains neutral with regard to
management. J Hydrol 534:300–316 jurisdictional claims in published maps and institutional affiliations.

13