J. Earth Syst. Sci.

(2020)129:42 Ó Indian Academy of Sciences

https://doi.org/10.1007/s12040-019-1306-3 (0123456789().,-volV)(0123456789(
).,-volV)

Spatio-temporal analysis and forecasting of drought

in the plains of northwestern Algeria using the
standardized precipitation index

KENZA ACHOUR1, MOHAMED MEDDI2, AYOUB ZEROUAL2,* , SENNA BOUABDELLI2,

PAMELA MACCIONI3 and TOMMASO MORAMARCO3
1
Hassiba Ben Bouali University of Chlef & GEE Laboratory, ENSH, Blida, Algeria.
2
Ecole 
Nationale Superieure d’Hydraulique, Blida, R.L GEE, Algeria.
3
Research Institute for GeoHydrological Protection (IRPI), National Research Council, Via Madonna
Alta 126, 06128 Perugia, Italy.
*Corresponding author. e-mail: [email protected]

MS received 10 March 2019; revised 21 September 2019; accepted 29 September 2019

Drought is the most frequent natural disaster in Algeria during the last century, with a severity ranging
over the territory and causing enormous damages to agriculture and economy, especially in the north-
west region of Algeria. The above issue motivated this study, which is aimed to analyse and predict
droughts using the Standardized Precipitation Index (SPI). The analysis is based on monthly rainfall data
collected during the period from 1960 to 2010 in seven plains located in the north-western Algeria. While
a drought forecast with 2 months lead-time is addressed using an artiBcial neural network (ANN) model.
Based on SPI values at different time scales (3-, 6-, 9-, and 12-months), the seven plains of north-western
Algeria are severely aAected by drought, conversely of the eastern part of the country, wherein drought
phenomena are decreased in both duration and severity. The analysis also shows that the drought
frequency changes according to the time scale. Moreover, the temporal analysis, without considering the
autocorrelation eAect on change point and monotonic trends of SPI series, depicts a negative trend with
asynchronous in change-point timing. However, this becomes less significant at 3 and 6 months’ time
scales if time series are modelled using the corrected and unbiased trend-free-pre-whitening (TFPWcu)
approach. As regards the ANN-based drought forecast in the seven plains with 2 months of lead time, the
multi-layer perceptron networks architecture with Levenberg–Marquardt calibration algorithm provides
satisfactory results with the adjusted coefBcient of determination (R2adj ) higher than 0.81 and the root-
mean-square-error (RMSE) and the mean absolute error (MAE) less than 0.41 and 0.23, respectively.
Therefore, the proposed ANN-based drought forecast model can be conveniently adopted to establish
with 2 months ahead adequate irrigation schedules in case of water stress and for optimizing agricultural
production.
Keywords. Algeria; drought forecasting; artiBcial neural networks; standardized precipitation index;
trend-free-pre-whitening.

Supplementary materials pertaining to this article are available on the Journal of Earth Science Website (http://www.ias.ac.in/
Journals/Journal˙of˙Earth˙System˙Science).
42 Page 2 of 22 J. Earth Syst. Sci. (2020)129:42

1. Introduction drought comparison among sites can be done

through a deepened spatial analysis (Paulo et al.
Drought is recognized by the Intergovernmental 2016; Dehghani et al. 2017; Guhathakurta et al.
Panel on Climate Change (SheDeld and Wood 2017), that takes the evolution of the characteris-
2008) as one of the most severe climate events that tics of drought, namely: duration, severity and
needs to be mitigated to reduce its negative intensity into account (Masud et al. 2017; Asong
impact. Overall, drought is considered as a pro- et al. 2018).
longed and abnormal state of water deBcit (Palmer In this context, SPI can be also used for drought
1965; Wilhite and Glantz 1985; Dalezios et al. forecasting as support to a rigorous water resources
1991). Nowadays, the identiBcation of drought management by the decision makers, especially to
conditions is based on drought indices allowing, on plan how to significantly address the agricultural
one hand, the determination of the threshold irrigation which is a priority worldwide and in
indicating drought at different time scales and, on particular in Africa. For the drought forecasting,
the other hand, the classiBcation of conditions statistical models have been applied in the last 50
according to their severity and the location (Mckee yrs as the ones proposed by Gabriel and Neumann
et al. 1993). The index commonly used for droughts (1962) who used the Markov chain (Lazri et al.
is the Standardized Precipitations Index (SPI), 2015), the regression method by Torranin (1973)
whose value quantiBes the deviation from the and linear stochastic models including autoregres-
average precipitation and allows comparing dry sive integrated moving average (ARIMA) and
and humid years or years showing deBcit or surplus seasonal autoregressive integrated moving average
(Mckee et al. 1993). The World Meteorological (SARIMA) applied by Mishra and Desai (2005).
Organisation (WMO) adopted this index since However, during the last two decades, artiBcial
2009, as a global tool, in order to measure meteo- intelligence (AI) models in hydrology, like the
rological drought in the ‘Lincoln declaration on artiBcial neural network (ANN), fuzzy logic (FL)
drought indices’ (Cancelliere et al. 2007). SPI is and support vector regression (SVR), have signifi-
used everywhere in the framework of meteorologi- cantly enhanced, notably for the forecasting pur-
cal drought monitoring (Byun and Kim 2010; poses (Rezaie-Balf et al. 2017; Dariane et al. 2018).
Zhang et al. 2012; Chandramouli et al. 2017) due to AI model is overall used for the management and
its simplicity, spatial invariance in the interpreta- measurement of different aspects of water resour-
tion (Guttman 1998; Vicente-Serrano et al. 2010) ces, including rainfall-runoA modelling (Shoaib
as well as for the capability to capture and repli- et al. 2016; Tarawneh and Khalayleh 2016; Zeroual
cate the observed drought events measured by et al. 2016; Tongal and Booij 2017), soil loss pre-
others indices (Byun and Kim 2010; Maccioni et al. diction in wadis (Salhi et al. 2013) and evaporation
2015; Khan et al. 2018). The 3-month SPI index forecasting (Abyaneh et al. 2011; Goyal et al.
provides an indication of short- and medium-term 2014). In particular, the adoption of this technique
moisture conditions, precipitation estimates over a for the drought forecasting based on times series of
season and this is appropriate in agricultural areas drought indices has added a new dimension to the
to highlight the nature of soil moisture (WMO planning and management of water resources
2008). The 6-month SPI provides an indication of (Mishra and Desai 2006; Belayneh et al. 2014;
precipitation trends over a season (Khan et al. Masinde 2014; Bahrami et al. 2019). Several stud-
2008), while SPI-12 should be considered for long- ies showed that AI models could give better results
term estimations to address water resources than conventional drought modelling techniques in
management (Khan et al. 2008). Overall, SPI is semi-arid regions. As far as the comparison among
assessed for a given time scale using a probability different AI approaches for hydrological and cli-
distribution function Btting the corresponding matological applications is concerned, studies have
cumulated precipitation (Mckee et al. 1993). shown that the support vector machine (SVM)
Therefore, the estimated SPI-based severity of the approach is more accurate than that of ANN and
drought represents the deBcit or surplus over a FL one (He et al. 2014; Noori et al. 2015; Ghorbani
period at a site. However, at different locations, et al. 2016; Huang et al. 2017). Specifically, for
SPI could not correspond to the same deBcits or drought forecasting using SPI, the ANN model can
surplus of precipitation, in accordance with the forecast satisfactory up to 12 months of lead-time
average precipitation recorded at a single site. This (Barua et al. 2010, 2012; Bari Abarghouei et al.
means that SPI is site-dependent and an absolute 2013; Ali et al. 2017). Conversely, other studies
J. Earth Syst. Sci. (2020)129:42 Page 3 of 22 42

showed that SVM model has enough accuracy to be of all sectors definitely. Moreover, future climate
used in long-term drought forecasting compared to scenarios by 2100 based on regional climate models
ANNs (Tarawneh and Khalayleh 2016). Likewise (RCM) showed this region more vulnerable than
hybrid models can be also adopted for drought the centre and the eastern part of Algeria (Zeroual
forecasting mainly in the case of the nonstationary et al. 2019). Therefore for the plains of the north-
hydrologic time series (Jain and Kumar 2007; western Algeria, there is a major challenge to sus-
Wang et al. 2015). wavelet transform (WT) anal- tainable water management practices (Medjerab
ysis and cuckoo search (CS) algorithm have been and Henia 2005). Enhance the understanding of the
used to improve the forecasting ability of IA spatial distribution of drought, and comparing
models (Jalalkamali et al. 2015). Examples of their severity and duration among plains is thus
hybrid models to drought forecast based on SPIs required in order to establish adequate irrigation
times series are the wavelet support vector schedules in case of water stress. An integrated
machine (WSVM) (Djerbouai and Souag-Gamane space-temporal analysis of these drought charac-
2016; Deo Ravinesh et al. 2017a, b), Wavelet-ANN teristics in seven plains namely, Mitidja, Ghriss,
(WANN) (Belayneh et al. 2014; Zhang et al. 2017; Maghnia, Sidi Bel Abbes, High, Middle and Low-
Soh et al. 2018), wavelet-adaptive neuro-fuzzy CheliA would provide valuable knowledge that is
inference system (WANFIS) (Shirmohammadi necessary for drought forecasting, as well as for the
et al. 2013; Belayneh et al. 2014; Shabri 2014), rational exploitation of these plains under the
wavelet extreme learning machine (W-ELM) (Deo future climate change. These regions are known by
Ravinesh et al. 2017a, b) and cuckoo search- their fertile farmland and during the last decades,
support vector machine (CS-SVM) (Liang et al. the water shortage created serious concerns for the
2016; Komasi et al. 2018). plains irrigation. Therefore, drought events are
Based on the above insights, this study has two here assessed by their duration and severity using
objectives. The Brst is to provide a comprehensive the SPI over various temporal scales (3, 6, 9, and
analysis of historical droughts based on SPI anal- 12 months) and then a comparison of the evolution
ysis over the northwest Algeria. The second is to of these characteristics over the seven plains is
develop always for the north-western part of done. The temporal analysis is investigated ini-
Algeria an ANN-based forecast model of drought tially without considering the eAect of autocorre-
events 2 months in advance, addressed to irrigation lation on change point and monotonic trends of
purposes. Accomplishing these two objectives may SPIs time series, thereafter, the time series are
be of paramount importance to coping with the modelled using the corrected and unbiased trend-
drought in Algeria, that during the last 50 years free-pre-whitening (TFPWcu) approach to elimi-
experienced one of the widest variations in rainfall nate the autocorrelation eAect on the variance of
regime. As regards the Brst objective, the climatic test statistics. Major spatial and temporal patterns
Cuctuations in Algeria aAected not only rainfall of change points of drought events are then iden-
patterns, but they also created persistent and tiBed using the modiBed Pettitt test proposed by
pervasive droughts conditions. Historically, the Serinaldi and Kilsby (2016). As far as the second
precipitation deBcit has been registering since 1973 objective is concerned, it attempts to coping with
aAecting differently the north of Algeria (Meddi the persistent water shortage by developing a
et al. 2010). The duration and the severity of this drought forecasting model to support decision
deBcit varies across north-western part of the makers to mitigate the water stress resulting from
country causing severe impacts on agricultural and a cycle of droughts and to ensure the irrigation in
socio-economic features for the strong dependence plains. Studies on the drought forecast using dif-
of the country’s economy on rain-fed agriculture in ferent approaches have been addressed for the
the large fertile farmland (Meddi and Meddi 2009). eastern and central part of Algeria (Djerbouai and
For the period 1970–2013, the continuous decrease Souag-Gamane 2016; Habibi et al. 2018). It is
in precipitations in areas previously considered worth noting, however, that the plains of those
temperate climate, caused very dry environment regions have not experienced the same severity of
conditions (Meddi et al. 2010; Zeroual et al. 2017) water shortage as the west of Algeria. Therefore, to
with reservoir levels dropped up to 25% of their develop a drought forecasting system for the north-
normal level and piezometric levels decreased western part of Algeria is a priority to manage in a
below 40 m for some aquifers (Demmak 2008). This rigorous way water resources in agricultural irri-
reduction has negative impact on the water supply gation. For that, considering the uncertainty
42 Page 4 of 22 J. Earth Syst. Sci. (2020)129:42

associated with water scarcity and climate change, plain area (Sßen 1998). In the supplementary
we propose an ANN-based drought forecast model material, we show in Bgure 1, the polygon sur-
using a lead time of 2 months which can be a rounding each rain gauge using the Thiessen
valuable tool for irrigation and water resources method. Furthermore, in table 1, the geographic
management, as long as the irrigation periods coordinates of rain-gauges, regions of inCuence
starting in April until June. around each rain-gauge station and their percent to
total plain area are given.
The main statistical characteristics of rainfall
2. Study zone and dataset
are calculated for each plain after determination of
the monthly areal average precipitation for each
The Algerian northwest part spreads over 250 km
plain. Their results are exhibited in table 1. SPI
from south to north and approximately 500 km
has been calculated on the seven plains for the
from west to east (cf. Bgure 1). The territory is
period 1960–2010 at different time scales, i.e., 3, 6,
characterized by mild climate, and relatively high
9, and 12 months. SPI is computed for each plain
humidity. Rainfall ranges from 400 mm in the west
by Btting an appropriate probability density func-
to 900 mm in the east.
tion to the frequency distribution of precipitation
Monthly rainfall data recorded at 20 rain-gauge
cumulated over the considered time scale (3, 6, 9,
stations in the seven plains for the period
and 12 months).
1960–2010 have been used in this study. Data are
extracted from dataset provided by the National
Meteorology ODce (ONM) and the National 3. Methods
Hydraulic Resources Agency (ANRH) (O.N.M:
http://www.meteo.dz/index.php, ANRH: http:// 3.1 Drought index
www.anrh.dz/).
For each investigated plain, the monthly areal Thom (1958) found the two-parameter gamma
average precipitation was computed after deter- probability density function can Bt well precipi-
mination of the weighting factor using the Thiessen tation time series. The gamma distribution is
polygon method. The weighting factor speciBes the deBned by its frequency or probability density
contribution of each rain gauge station to the total function as:

Figure 1. Geographical location of the seven plains included in the study.

J. Earth Syst. Sci. (2020)129:42 Page 5 of 22 42

Table 1. Statistical characteristics of annual rainfall in the studied plains.

Standard
Number of Average deviation Median Maximum Minimum
Plain rain-gauge (mm) (mm) (mm) (mm) (mm)
Mitidja 6 714 156.1 706.2 1082 449
High CheliA 2 419 89.9 432.5 652 179
Middle CheliA 3 382 85.7 379.9 601 228
Low CheliA 2 349 70.2 343.2 483 209
Ghriss 2 323 76.0 308.8 549 210
Sidi Bel Abbes 3 345 84.6 360.1 562 195
Maghnia 2 298 82.7 287.8 501 174

1 Since the gamma function is undeBned for x ¼ 0

g ðx Þ ¼ a x a1 ex=b ; for x [ 0; ð1Þ and a precipitation distribution may contain zeros,
b CðaÞ
the cumulative probability becomes:
where a ð[ 0Þ is a shape factor, b ð[ 0Þ is a scale
factor, and x [ 0 is the amount of precipitation. H ðx Þ ¼ q þ ð1  q ÞG ðx Þ; ð8Þ
CðaÞ is the gamma function which is where, q is the probability of zero precipitation.
deBned as: The cumulative probability, H(x), is then
Z 1
transformed to the standard normal random
CðaÞ ¼ ya1 ey dy: ð2Þ variable Z with mean zero and variance one,
0 which is the value of the SPI. Following the
Edwards and McKee (1997) suggested a method studies of Edwards and McKee (1997) and Lloyd-
using the approximation of Thorn (1966) for Hughes and Saunders (2002), an approximate
maximum likelihood to optimally estimate a and conversion was used here, as provided by
b as follows: Abramowitz and Stegun (1965) to convert easily
the cumulative probability to the standard normal
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi! random variable Z or SPI:
1 4A  
a^ ¼ 1þ 1þ ; ð3Þ C0 þ C1 t þ C2 t 2
4A 3 Z ¼ SPI ¼  t 
1  d1 t þ d2 t 2 þ d3 t 3 ð9Þ
x for 0\H ðx Þ  0:5;
b^ ¼ ; ð4Þ
a^
and
where  
C0 þ C1 t þ C2 t 2
P Z ¼ SPI ¼ þ t 
lnðx Þ 1  d1 t þ d2 t 2 þ d3 t 3 ð10Þ
A ¼ lnðxÞ  ; ð5Þ
n for 0:5\H ðx Þ  1:0;
n is the number of precipitation observations. The where
resulting parameters are then used to infer the vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u !ffi
cumulative probability of an observed preci- u 1
pitation event for the given month or any other t ¼ tln for 0\H ðx Þ  0:5; ð11Þ
time scale: ðH ðx ÞÞ2

Z x Z x and
1 ^
a1 x=b^
G ðx Þ ¼ g ðx Þdx ¼ x e dx: ð6Þ vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
0 b^a^Cða^Þ 0 u !
u 1
^ this equation becomes the t ¼ tln for 0:5\H ðx Þ  1:0;
Letting t ¼ x=b, ð1  H ðx ÞÞ2
incomplete gamma function:
ð12Þ
Z x
1 And c0 ¼ 2:515517; c1 ¼ 0:802853; c2 ¼ 0:010328;
G ðx Þ ¼ ^
t a1 et dt: ð7Þ
Cða^Þ 0 d1 ¼ 1:432788; d2 ¼ 0:189269; d3 ¼ 0:001308 (Mishra
42 Page 6 of 22 J. Earth Syst. Sci. (2020)129:42

and Desai 2005). According to McKee et al. (1993), where, D is the drought duration and S is the
drought classes of the SPI indicate near normal con- drought severity. Drought intensity is deBned as
ditions at 1 \ SPI \ 0, moderate drought at severity divided by duration (Dingman 1994; Shiau
1.5 \ SPI B 1, severe drought at 1.99 \ SPI 2006; Madadgar and Moradkhani 2013).
B 1.5 and extreme drought at SPI B 2.

3.2.2 Trends and shifts analysis

3.2 Spatio-temporal analysis of drought
Once the spatial evolution variability of the time
3.2.1 Limitations of SPI and drought series of SPI’s (3, 6, 9, and 12 months) in the seven
characterization plains was analyzed, we proceeded to the temporal
variability analysis according to the two following
The SPI is a standard index over a reference period. stages: in the Brst, the presence and timing of
It is obtained for a different deBcit of precipita- abrupt changes in the SPI’s time series are checked
tion, which corresponds to the difference between using the non-parametric statistical test of Pettitt
the observed and the average of the time series. (Pettitt 1979), after their autocorrelation is veriBed
However, this functionality does not allow SPIs to be (the Brst serial correlation ‘lag-1’ coefBcient). If
considered for absolute drought comparison ‘lag-1’ serial correlation coefBcient is significant at
between stations or plains if additional information 5% level, then the modiBed corrected and unbiased
is not available (Mishra 2010; Paulo et al. 2016). trend free pre-whitening (TFPWcu) procedure for
Local time series analysis leads to missing infor- Pettitt test is used. This procedure has been sug-
mation on spatial characteristics; likewise, the gested by Serinaldi and Kilsby (2016) for removing
spatial analysis does not make it possible to follow autocorrelation eAects in the time series. Other-
the dynamic evolution of the phenomenon. For the wise, the original Pettitt test is applied for
temporal analysis, the use of precipitation series of detecting change point in the time series data. In
different lengths significantly aAects the SPI values the second stage, we applied the non-parametric
even if the same probability distribution is applied original Mann–Kendall test (MK) and its above-
(Mishra 2010; Salvi and Ghosh 2016). mentioned modiBed form MMK (Hamed and
For example, Salvi and Ghosh (2016) pointed out Ramachandra Rao 1998) to assess the significance
the impact of this eAect on the frequency of meteo- of existing trends in the long-term variability of the
rological extreme dry and wet spells in the recent SPI’s time series.
past and in the future using general circulation It is worth noting that in the MK test, the data
models as well as on parameters of the gamma dis- are assumed independent and randomly ordered,
tribution using the SPI values computing from and this can lead to erroneous results if serial cor-
different precipitation record lengths over India. relation is present (Hamed and Ramachandra Rao
To avoid these problems, a spatio-temporal 1998). Even in the absence of a trend in the time
analysis based on the comparison of the evolution of series, the positive autocorrelation would increase
the characteristics of the drought, namely, duration, the chance of significant response. Various studies
gravity and intensity, is applied to the seven plains have addressed this issue in order to Bnd an
studied covering the same period from 1960 to 2010. appropriate approach that considers the signifi-
The SPI series (3, 6, 9, and 12 months) in the cance of autocorrelation in the data. Some authors
seven plains have been investigated in terms of have suggested modifying the data itself using the
spatio-temporal variability. The duration (D) of a prewhitening procedure (Storch and Navarra
drought event, as deBned by Mckee et al. (1993), is 1995), trend-free-pre-whitening (TFPW) approach
the period, during which the SPI value is contin- (Yue et al. 2002) and corrected and unbiased trend
uously negative and reaches at least value of 1 or free pre-whitening (TFPWcu) (Serinaldi and
less. The SPI values accumulation during this Kilsby 2016). Others have suggested the variance
period was used by Shiau (2006) to measure the correction of Mann–Kendal test using an empirical
drought magnitude, which is also called drought formula (Hamed and Ramachandra Rao 1998) and
severity (S): Monte Carlo simulations (Yue and Wang 2004).
X
D Pettitt test. The non-parametric Pettitt change
S¼ SPIi ; ð13Þ point test is applied for detecting change point in
i¼1 the time series (Pettitt 1979). This test has been
J. Earth Syst. Sci. (2020)129:42 Page 7 of 22 42

widely applied to detect a single change-point in and after s). The difference of the medians or
hydrological and climate series with continuous means, l^b and l^a , of the two sub-series is
data (Yue et al. 2002). computed as D ¼ l^b  l^a , and used to remove
The Pettitt test considers that a time series of the step change as follows:
a sequence of random variables Xt with t ¼
1; 2; . . . ; T; has a change point at time step s if the xt ¼ yt  D  1ft [ sg : ð17Þ
values of Xt for t ¼ 1; 2; . . . ; s have the cumulative Here 1ft [ sg is the indicator function, xt and yt are
density function (CDF) F1 ðx Þ, and the values of Xt data after step change removal and original data
for t ¼ s þ 1; s þ 2; . . . ; T have the CDF F2 ðx Þ and before step change removal, respectively, at time
F1 ðx Þ 6¼ F2 ðx Þ; with the single assumption that the t in the time series.
two CDFs are continuous.
The null hypothesis H0, which assumes there is The value of the lag-1 autocorrelation q^ of xt is
no change in the Xt time series, is tested against estimated and corrected for bias using two stage
the alternative hypothesis (H1) of change point bias corrections. At the Brst stage, it is corrected
using the non-parametric statistic deBned as: for autocorrelation of fractional Gaussian noise
  (fGn) process or Hurst–Kolmogorov process as
KT ¼ maxUt;T ; ð14Þ follows (Koutsoyiannis 2003):
where Ut;T is deBned as the sign (sgn) function  
 1 1
between the difference of each pair of value of the qK ¼ q 1  0 þ 0 ;
^ ð18Þ
T T
two sequences Xi and Xj :
where T 0 is eAective sample size for the Brst-order
X
t X
T   autoregressive [AR(1)] process and may be obtained
Ut;T ¼ sgn Xi  Xj ; ð15Þ as follows (Koutsoyiannis and Montanari 2007):
i¼1 j¼iþ1

with the sign function: sgn(y) = 1 if y [ 0, 0 if ð1  qÞ2

T0 ¼ T : ð19Þ
y = 0, 1 if y \ 0. ð1  q2 Þ  2qð1  qT Þ=T
The change-point of the series is located at KT ,
provided that the non-parametric statistic is The computation for second stage bias
significant. correction may be carried out similar to Marriott
The significance probability (p) associated with and Pope (1954), but using following generalized
the value KT is approximately given by: formulae suggested by Mudelsee (2001):
  E½q^ ’ E½q^W
6KT2  
p ’ 2 exp ; ð16Þ 2 4 2 2 2
T3 þ T2 ¼ 1  þ 2  3 q þ 2 q3 þ 2 q5 ð20Þ
n n n n n
when p is smaller than the speciBc significance
for q\0:88:
probability, for example, p = 0.05 in this study, the
null hypothesis is rejected.
2q 2 ðq  q2n1 Þ
TFPWcu adapted for the Pettitt test. To E½q^ ’ E½q^M ¼ q  þ
ðn  1Þ ðn  1Þ2 ð1  q2 Þ
redress the Pettitt test problem in the case of
presence of serial dependence in the chronological for q  0:88: ð21Þ
series, Serinaldi and Kilsby (2016) have suggested After getting the bias corrected lag-1
the corrected and unbiased TFPWcu procedure for autocorrelation coefBcient q^ , AR(1) structure is
Pettitt test. The objective of this approach is to removed to get uncorrelated series e0t of residuals as
remove the autocorrelation eAect. The determina- follows:
tion of change point by this algorithm requires the
following steps as it is mentioned in Serinaldi and e0t ¼ xt  q^ xt1 : ð22Þ
Kilsby (2016):
Finally, a pre-whitened series xt00 is obtained by
1. If the value of the KT obtained by Pettitt test combining uncorrelated series of residuals and step
using original data is significant, then, the change. The following pre-whitened series is used
position s of the possible change point is used for examining the significance of change point
to split the time series in two sub-series (before using Pettitt test.
42 Page 8 of 22 J. Earth Syst. Sci. (2020)129:42

e0t over time (tn)) and their weight values Wij. The
xt00 ¼ D  1ft [ sg þ : ð23Þ result computes the hidden layer output through
1  q^
some non-linear activation function. Beyond that,
ModiBed Mann–Kendall test (MMK). The this operation is repeated for the output neuron.
Mann–Kendall test (Mann 1945; Kendall 1975) is a These two processes are explicitly given in
commonly used non-parametric trend test. However, equation (24).
the null hypothesis corresponds to the case where the " ! #
data are independent and random. Hamed and Xm Xn

Ramachandra Rao (1998) revealed that the serial y^k ¼ uo wjk  uh xi  wij þ bjo þ bko ;
j¼1 i¼1
correlation falsiBed trends in auto-correlated time
series. The existence of positive autocorrelation in ð24Þ
the data increases the probability of detecting trend where uh and uo are the activation functions of the
when actually it does not exist, while a negative hidden neuron and the output neuron, respectively;
autocorrelation decreases the probability of detecting wij and wjk are the vector of weights connecting the
significant trend. Therefore, to eliminate the neurons between the input layer hhiii and the hidden
inCuence of the serial correlation, Hamed and layer hhjii and between the hidden layer hhjii and
Ramachandra Rao (1998) derived an empirical output layer hhkii, respectively; bjo and bko are the
formula to modify the Mann–Kendall test.
bias for the jth hidden neuron and kth output
neuron, respectively; and n and m are the number of
3.3 ArtiBcial neural networks neurons in the input layer and the hidden layer,
respectively. The most commonly used form of uð:Þ
ArtiBcial neural networks (ANNs) are non-linear in (1) is the sigmoid function, given as:
mathematical models of ‘black-box’ type. In prac-
tice, this approach is used to deBne a deterministic 1
relationship between process variables when no uðt Þ ¼ : ð25Þ
1 þ et
a priori known about the physical nature of its
generating (Govindaraju 2000). This is particularly 3.3.1 Data preparation
useful for drought forecasting, where the phe-
nomenon is a stochastic and little is understood In order to forecast the SPIi (t) 2 months ahead
about the processes that may cause it (Mishra and (i = 3, 6, 9, and 12 months), various combinations
Desai 2005). In this paper, the multilayer percep- of antecedent SPI (tn) can be used as model
tron (MLP) feed-forward network was used to input. Each of SPI series has been standardized
forecast the SPI at 2 months lead-time at different and centred, to avoid any saturation eAect that
time scales (3, 6, 9, and 12 months). This type is may be caused by the use of sigmoid function,
the simplest and most commonly used artiBcial before being divided into two sets (Mishra and
neural network design in the water resources vari- Desai 2005). Training data of about 70%
ables and hydrological domains. Moreover, most (1960–1996) are used for calibration of the vector of
studies have shown that MLP can perform better weights ðwjk ; wij Þ of the ANN with a selection of
than conventional approaches in modelling and neuron number in the hidden layer (N) and acti-
forecasting non-linear and non-stationary time vation function. The resulting ANN is then evalu-
series (Kim and Vald es 2003; Mishra and Desai ated on independent validation data (1997–2010).
2005; Belayneh and Adamowski 2012; Aher et al. Several models of ANN have been designed and
2017). Here, a MLP feed-forward network of three- examined, based on the variation of the number of
layer is used for drought SPI forecast. The MLP input and neurons in hidden layer (N) in order to
calculates a single output such as the forecasted optimize and reBne the non-linearity existing
SPI up to 2 months ahead. This one depends on the between input and output variables. For each SPI
number of previous SPIs over time (tn) received series at different time scales for the seven plains,
by Brst layer as input variables (n is the number of nine ANN designs were selected based on the dif-
time lag in month and varies between 1 and 12 ferent combination of variables of the temporal
months) and the number of nodes in each hidden series of SPI before the time (t2). The nine ANN
layer (N). designs are constructed as table 2 shows. In the
The output SPI (t) is obtained by forming a supplementary material, we illustrate in Bgure 2,
linear combination between input (previous SPIs the schematization of input and output data to
J. Earth Syst. Sci. (2020)129:42 Page 9 of 22 42

Table 2. Structures of drought forecasting models.

Design no. Input structure Output

1 SPI (t) SPI (t+2)
2 SPI (t), SPI (t1). SPI (t+2)
3 SPI (t), SPI (t1), SPI (t2). SPI (t+2)
4 SPI (t), SPI (t1), SPI (t2), SPI (t3). SPI (t+2)
5 SPI (t), SPI (t1), SPI (t2), SPI (t3), SPI (t4). SPI (t+2)
6 SPI (t), SPI (t1), SPI (t2), SPI (t3), SPI (t4), SPI (t5). SPI (t+2)
7 SPI (t), SPI (t1), SPI (t2), SPI (t3), SPI (t4), SPI (t5), SPI (t6). SPI (t+2)
8 SPI (t), SPI (t1), SPI (t2), SPI (t3), SPI (t4), SPI (t5), SPI (t6), SPI (t7) SPI (t+2)
9 SPI (t), SPI (t1), SPI (t2), SPI (t3), SPI (t4), SPI (t5), SPI (t6), SPI (t7), SPI (t8) SPI (t+2)

forecast SPI-3 2 months lead time with 3 inputs ð1  R2 Þðn  1Þ

and 3 neurons in hidden layer (3-3-1 ANN R2adj ¼ 1  ; ð28Þ
np1
architecture).
To determine the optimum input combinations sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
that should be included in the model and the 1X n  2
RMSE ¼ SPIiðobsÞ  SPIiðsimÞ ; ð29Þ
number of neurons (N) in the hidden layer, a trial N i¼1
and error procedure is used (Hirose et al. 1991;
Mishra and Desai 2006; Choi et al. 2008; Trenn 1X n  
 SPIiðobsÞ  SPIiðsimÞ ;
MAE ¼ ð30Þ
2008; Sheela and Deepa 2013; Zeroual et al. 2016) N i¼1
during the calibration and testing phases of the
ANN model. Nodes in hidden layer also allow in which SPIiðobs Þ , SPIiðsimÞ , SPIðmeanÞ represent,
taking the presence of non-stationary in the series respectively, the observed, forecasted and mean
into account, such as trends and seasonal varia- standardized precipitation index; n is the total num-
tions (Maier and Dandy 1996). The two activation ber of SPI data, p is the number of regression coefB-
functions used are as follows: sigmoid function in cients and R2 is the determination coefBcient. A good
hidden layer and a linear function in output layer. model should have a lower RMSE and MAE which
Previous studies showed that both functions allow indicates low accumulated errors (Willmott and
approaching the existing non-linear relationship Matsuura 2005; Chai and Draxler 2014; Tian et al.
between SPIs (Bari Abarghouei et al. 2011). After 2018) and R2adj close to the unity. The latter has been
construction of the nine ANN designs presented in suggested to judge goodness of model when several
table 2, the Levenberg–Marquardt (LM) back- input variables are used. Adjusted R2 penalizes the
propagation (BP) algorithm is used for the MLP adding variables that do not improve model.
feed-forward network calibration. Details on the
algorithm can be found in Zeroual et al. (2016).
4. Results

3.3.2 Model performance criteria Results are presented in two levels showing: (1) the
spatio-temporal characterization and comparison
Once the lag is adopted, data collection for
of drought episodes over the seven studied plains at
validation has been provided to the network. This
different time scales during the period 1960–2010
time, only the input vectors have been transmitted
and (2) the assessment of artiBcial neural network
to the model. Model performances are evaluated
(ANN) in drought prediction at 2 months ahead
through several indicators, as the adjusted coefB-
based on the SPI.
cient of determination (R2adj ), root-mean-square-
error (RMSE) and the mean absolute error (MAE)
between SPIs observed and SPIs forecasted by 4.1 Spatio-temporal comparison of drought
ANN model.
Pn  2 4.1.1 Spatio-temporal variability
2 i¼1 SPIi ðobsÞ  SPIi ðsimÞ
R ¼ 1  Pn  2 ; ð27Þ The SPI are computed for the 3-, 6-, 9-, and
i¼1 SPIi ðobsÞ  SPIðmeanÞ 12-month time steps using monthly precipitation
 42
Page 10 of 22
J. Earth Syst. Sci. (2020)129:42

Figure 2. Comparison of SPI in the seven plains (1960–2010) at various time scales (3, 6, 9, and 12 months).
J. Earth Syst. Sci. (2020)129:42 Page 11 of 22 42

data from the seven studied plains for the period where the horizontal lines length represents the
1960–2010. To do end, the cumulated precipitation duration in months. We distinguish a high vul-
at the different temporal scales is Btted to the nerability to drought in both Sidi Bel-Abb es and
gamma probability distribution. We note that all Maghnia plains during the decade 1980–1990.
of the time series are well Btted by a gamma dis- These two plains are located in the extreme west of
tribution at the 5% level of significance. Several Algeria, which was severely aAected by drought
recent studies have found similar results (Maccioni events during this decade.
et al. 2015). The Sidi Bel-Abbes plain experienced the most
In Bgure 2, we compare the SPI drought index in severe drought episodes in each 3-, 6- and
the seven plains (1960–2010) at various time scales 12-month time scales. They are characterized by
(3, 6, 9, and 12 months). As can be seen, the time maximum severities of 11.8, 19.6, 61.4 and maxi-
series reveal a situation mostly dominated by mum duration of 11, 22, 59 months, respectively.
drought (negative SPI) on the seven plains. The While at 9 months time scale, Maghnia plain has
temporal analysis of SPI allows ascertaining that known the most severe drought with a maximum
‘extreme drought’ character is not dominant for duration of 43 months.
the whole area and for different scales. At 6- and 12-month time scales, drought
Inspecting Bgure 2, and referring to computed episodes have decreased in both duration and
SPI for 3 months (i.e., SPI3), the values vary in the severity passing from Sidi Bel-Abb es plain, which
range [1.5, +4.6]. The smallest negative values represents the most aAected region, and shifting
are observed in the Lower CheliA plain during the to Maghnia, Low CheliA, Ghriss, Middle CheliA
period 1987 and 1988 and in the Ghriss plain for and Mitidja plains, respectively. Finally, the
2006 and 2007; while the maximal positive value of Upper CheliA plain is the slightly aAected area by
SPIs are observed in Mitidja plain during the per- such events. This plain is characterized by a col-
iod 1974 and 1975. For SPI6, values vary between lection of mountains, which plays the role
[2, +4.3]; the minimum value is observed in the of a barrier between the Mediterranean Sea and
plain of Low CheliA during the period 2005–2006, the highlands, and by high precipitation exceeding
while the maximum positive SPI is observed in the the mean values. Furthermore, we notice that the
Ghriss plain during the period 1964–1965. For drought events distribution is similar in severity
SPI9, values were included between [2.6, +3.7]; at 3 months time scale to that at 6 and 12
the lower value was recorded for the period months.
1982–1983 at Sidi Bel-Abb es plain, while the higher
positive value was noted in Ghriss plain during the 4.1.2 Trends and shifts analysis
period 1964–1965.
For SPI12, values were included between [2.9, The previous survey shows that the seven plains
+3.9]; the lowest negative values of SPI were found have been subjected to a persistent drought over
in the Sidi Bel-Abb es plain in the period (1982, the last Bve decades, although the intensity of this
1983), while the higher positive value of SPI has drought changes from one plain to the next. In
been recorded in Mitidja plain during the period what follows, we check the occurrence of breaks
(1972, 1973). Therefore, it is evident that the time and trends in the average values of SPI to explain
scale of SPI unfolds a key role in identifying the the persistence of this drought. The different time
frequency of drought, such as also found in Spain series of SPI (3, 6, 9, and 12 months) in the seven
by Vicente-Serrano et al. (2010), who showed that plains have been checked for autocorrelation before
the drought frequency changed according to the applying trend and break tests. The lag-1 serial
time scale. Based on these results, it is noted that correlation coefBcient for all SPIs series were
plains of Lower CheliA, Sidi Bel-Abb es and Ghriss computed for the study period and results of lag-1
are the most aAected by drought during the study autocorrelation coefBcients (^ q) are shown in
period regardless the SPI time scale adopted, and Bgure 4. Thus, all SPI time series are found to be
the 1983 recorded a classiBcation of extreme auto-correlated at lag-1 at 5% level (for all
drought in Sidi Bel-Abb es plain. p-values \ 0.05).
Figure 3 shows a comparison of the severity Firstly, we identify change points in the different
expressed by equation (13) of all of the drought time scale using original Pettitt’s test. In table 3,
events that occurred for each time scale for the we present the results of test statistic KT, p-value
seven plains within the study period 1960–2010, and the year where the change point was found.
42 Page 12 of 22 J. Earth Syst. Sci. (2020)129:42

Figure 3. Drought occurrence from 1960 to 2010 in the seven plains based on SPI at various time scales (3, 6, 9, and 12 months).

Figure 4. Lag-1 autocorrelation coefBcients (q)

^ for SPI series in the seven plains (1960–2010) at various time scales (3, 6, 9, and
12 months).

Without removing the eAects of autocorrelation, sub-series based on possible change point identiBed
the change point was identiBed at 5% significance from original Pettitt’s test. The difference of mean
level at the end of the 1970s and the early 1980s for (D), lag-1 autocorrelation ð^ q) after elimination of
all SPI time scales except for the Mitidja plain shift, Brst phase corrected (qk ) and second phase
where the change point occurred in the second half corrected ð^q ) lag-1 autocorrelation coefBcients are
of the 1980s decade. To remove the eAect of auto- exhibited in table 3.
correlation in the time series, we applied modiBed After removal autocorrelation, it can be seen in
Pettitt’s test after prewhitening SPI series using table 3 that the autocorrelation coefBcient consid-
TFPWcu technique. For Pettitt test using erably decreased. Then the corrected SPI series
TFPWcu as suggested by Serinaldi and Kilsby were tested for change point by using Pettitt’s test.
(2016), auto-correlated SPI time series at 3, 6, 9, Results of Pettitt’s test applied at 5% significance
and 12 months’ time scales were divided into two were changed compared with results applied on the
Table 3. Change point in mean values of SPIs series by original and modiBed Pettitt’s test using TFPWcu for the period 1960–2010 at a 5% significant level.

Original Pettitt test ModiBed Pettitt test using TFPWcu

Plain KT Year P-value D (mm) q^ qk q^ q KT Year P-value
SPI-3
Mitidja 9092 No break 0.231 0.206 0.731 0.734 0.731 0.27 4278 No break 1.236
High CheliA 12571 SPIOND 1980 0.032 0.274 0.737 0.740 0.738 0.31 4374 No break 1.210
Middle CheliA 14314 SPISON 1979 0.010 0.312 0.722 0.725 0.723 0.29 4808 No break 1.089
Low CheliA 14142 SPISON 1979 0.011 0.283 0.739 0.742 0.739 0.31 4642 No break 1.135
Ghriss 17409 SPIJAS 1976 0.001 0.470 0.719 0.722 0.719 0.24 7476 No break 0.460
Sidi Bel Abbes 19471 SPISON 1979 0.000 0.416 0.713 0.716 0.714 0.26 8978 No break 0.240
Maghnia 19115 SPIOND 1980 0.000 0.428 0.716 0.719 0.716 0.22 9068 No break 0.230
J. Earth Syst. Sci. (2020)129:42

SPI-6
Mitidja 14146 SPIOct-Mar 1986 0.009 0.297 0.8153 0.818 0.816 0.35 5732 No break 0.833
High CheliA 24605 SPINov-Apr 1980 \ 0.0001 0.388 0.8182 0.821 0.819 0.36 7229 No break 0.496
Middle CheliA 29227 SPIDec-May 1980 \ 0.0001 0.427 0.8086 0.812 0.809 0.31 7652 No break 0.420
Low CheliA 19685 SPIOct-Mar 1980 \ 0.0001 0.396 0.8266 0.830 0.827 0.38 7003 No break 0.541
Ghriss 26567 SPISep-Feb 1976 \ 0.0001 0.664 0.8117 0.815 0.812 0.28 10219 No break 0.123
Sidi Bel Abbes 29366 SPIOct-Mar 1979 \ 0.0001 0.592 0.8169 0.820 0.817 0.29 10797 No break 0.089
Maghnia 29227 SPIDec-May 1980 \ 0.0001 0.606 0.8174 0.820 0.818 0.29 10975 No break 0.080

SPI-9
Mitidja 23868 SPIDec-Aug 1986 \ 0.0001 0.45 0.83 0.84 0.84 0.21 9680 No break 0.159
High CheliA 31030 SPIDec-Aug 1986 \ 0.0001 0.57 0.83 0.84 0.83 0.20 12950 SPIOct-Jun 1986 0.021
Middle CheliA 34156 SPIDec-Aug 1979 \ 0.0001 0.68 0.83 0.83 0.83 0.17 13854 SPIAug-Apr 1984 0.011
Low CheliA 30982 SPIDec-Aug 1980 \ 0.0001 0.59 0.85 0.85 0.85 0.29 11408 No break 0.059
Ghriss 42128 SPIOct-Jun 1976 \ 0.0001 0.97 0.84 0.84 0.84 0.16 14458 SPIJan-Sep 1976 0.007
Sidi Bel Abbes 43513 SPINov-Jul 1979 \ 0.0001 0.85 0.83 0.83 0.83 0.13 16022 SPIAug-Apr 1976 0.002
Maghnia 43850 SPIFeb-Oct 1981 \ 0.0001 0.87 0.85 0.85 0.85 0.19 16872 SPIOct-Jun 1980 0.001

SPI-12
Mitidja 31534 SPIMar-Feb 1987 \ 0.0001 0.642 0.861 0.864 0.861 0.10 14727 SPIDec-Nov 1986 0.005
High CheliA 42424 SPIMar-Feb 1979 \ 0.0001 0.795 0.867 0.870 0.867 0.03 15418 SPINov-Oct 1979 0.003
Middle CheliA 45200 SPIJan-Dec 1980 \ 0.0001 0.933 0.865 0.868 0.866 0.03 17026 SPINov-Oct 1979 0.001
Low CheliA 42293 SPIJan-Dec 1980 \ 0.0001 0.853 0.897 0.901 0.898 0.01 14386 SPINov-Oct 1979 0.007
Ghriss 57928 SPINov-Oct 1976 \ 0.0001 1.306 0.896 0.899 0.896 0.11 24741 SPIApr-Feb 1976 \ 0.0001
Sidi Bel Abbes 57882 SPIDec-Nov 1979 \ 0.0001 1.111 0.885 0.888 0.885 0.05 21581 SPIAp-Feb 1979 \ 0.0001
Maghnia 57484 SPIMar-Feb 1981 \ 0.0001 1.144 0.899 0.902 0.899 0.07 22817 SPIDec-Nov 1980 \ 0.0001
Page 13 of 22 42
42 Page 14 of 22 J. Earth Syst. Sci. (2020)129:42

original SPI series where KT values are decreased observed outputs, by modifying iteratively the
while the p-values are increased. For the new SPI weight matrices W and bias vectors according to
series, the change points are observed only at the the gradient of cost function. The gradient is esti-
9- and 12-months scales at the end of the 1970s mated by a method speciBc to neurons network:
and the beginning of the 1980s. back propagation (BP) of errors by the use of
For example, for 12 months, the Ghriss plain has Levenberg–Marquardt (LM) algorithm. For each
Brstly recorded a shift at the SPIApr–Feb 1976. The testing model, the number of neurons (N) is taken
Sidi Bel-Abb es and the Low, Middle and High between 2 and 25, and thereafter the optimum was
CheliA SPI series have all identiBed a shift in the found by using the trial and error methods.
same period from November to October 1979. Results of the best ANN design in terms of
After, the SPI Maghnia series has shown the statistics performance (R2adj , RMSE and MAE) at
change point in SPIDec–Nov 1980 and Bnally Mitidja each plain and at different time scales are shown in
plain in the SPIDec–Nov 1986. table 4. Considering calibration results, it can be
We note that, after the change point, the SPI observed that the best model ANN design was
series tend significantly to dry conditions (negative found to vary from one plain to another and from
trends) as shown in Bgure 2 where the drought one-time scale to another. Similar results were
events duration have increased and became more found for the optimal number of hidden neurons
severe starting with the Sidi Bel-Abb es and (N). Over all plains and at different time scales, the
Maghnia plain in 1980. For the time series that number of hidden neurons (N) was found in
have not spotted a change point (3 and 6 months’ accordance with the laws 2n and 2n + 1. These
time scale), the tendency towards dry conditions results were advocated by Lippmann (1987), Wong
using MMK is not significant at 5% significance (1991), and Tang and Fishwick (1993) in the case
level (cf. Bgure 5). of a Multilayer Perceptron neural network where
n is the number of inputs.
4.2 Drought forecasting by artiBcial neural As shown in table 4, RMSE and MAE for the
networks (ANN) validation phase of the best ANN design range
from 0.22 to 0.41 and 0.12 to 0.23, respectively.
After the input and output variables selection The lowest value of RMSE was for the GHRISS
across each of the seven plains, we examined the plain at 12-month time scale and the highest for
nine ANN designs indicated in table 2 to Bnd the the Low CheliA and GHRISs plains at 3-month
best model structure that can capture the non- time scale, while the adjusted coefBcient of deter-
stationarity and non-linearity in the SPI series and mination (R2adj ) values obtained during validation
can be used to forecast SPI (t) 2 months ahead. To phase range from 0.81 to 0.94.
this end, the sigmoid activation function was used Initially, for SPI3 (table 4) with 5 inputs and
as the transfer function. During the calibration of a
9 neurons in hidden layer, results showed that R2adj
three-layer model, connexion coefBcients (weight)
between different layers are computed in such a of Maghnia plain is the highest compared with
way that model outputs are as close as possible to those of other plains, with an R2adj equal to 0.934,
the desired outputs. RMSE value of 0.22 and MAE equal to 0.159.
The goal of optimization during calibration The R2adj for Maghnia plain is also the highest
phase is to minimize error between forecasted and value (0.934) for SPI9 with 6 inputs and 12

Figure 5. Z scores derived from the MMK method for SPIs series in the seven plains (1960–2010) at various time scales (3, 6, 9,
and 12 months). The two red lines represent the theoretical critical n values of the MMK test at the 5% probability level.
J. Earth Syst. Sci. (2020)129:42 Page 15 of 22 42

Table 4. Statistical measures of the ANN performance for 2 month lead time in the seven
plains at different time scales.

Network
Plains Time scale architecture R2adj RMSE MAE
Mitidja SPI 3 3–08–1 0.862 0.38 0.2218
High CheliA 4–08–1 0.849 0.39 0.2233
Middle CheliA 4–09–1 0.885 0.35 0.1999
Low CheliA 3–07–1 0.850 0.41 0.2303
Ghriss 5–10–1 0.827 0.41 0.2345
Sidi Bel Abbes 3–06–1 0.878 0.33 0.1965
Maghnia 5–09–1 0.934 0.22 0.1593
Mitidja SPI 6 4–08–1 0.868 0.35 0.1737
High CheliA 5–12–1 0.896 0.29 0.1886
Middle CheliA 5–11–1 0.862 0.35 0.2044
Low CheliA 5–10–1 0.874 0.34 0.2152
Ghriss 6–13–1 0.850 0.32 0.2091
Sidi Bel Abbes 6–09–1 0.895 0.31 0.2187
Maghnia 5–10–1 0.849 0.34 0.2171
Mitidja SPI 9 4–09–1 0.818 0.37 0.1661
High CheliA 5–10–1 0.824 0.37 0.1772
Middle CheliA 5–11–1 0.892 0.30 0.169
Low CheliA 5–12–1 0.901 0.30 0.1682
Ghriss 7–13–1 0.868 0.28 0.1647
Sidi Bel Abbes 6–13–1 0.888 0.30 0.1462
Maghnia 6–12–1 0.934 0.21 0.1453
Mitidja SPI 12 4–08–1 0.848 0.26 0.1297
High CheliA 6–13–1 0.908 0.23 0.1373
Middle CheliA 7–12–1 0.926 0.23 0.1520
Low CheliA 8–15–1 0.940 0.23 0.1425
Ghriss 7–14–1 0.894 0.21 0.1404
Sidi Bel Abbes 3–06–1 0.923 0.24 0.1549
Maghnia 4–08–1 0.904 0.25 0.1711

neurons in hidden layer, and the RMSE and through the box-plot shown in Bgure 6. The
MAE values is 0.21 and 0.145, respectively (cf. agreement at 12 months time scale between the
table 4). However, for SPI6, the high CheliA with values of SPI observed and the values forecasted by
5 inputs and 12 neurons in hidden layer and the ANN clearly appears also in Bgure 7. The same
Sidi Bel-Abb es plains 6 inputs and 9 neurons in agreement was found at 3-, 6- and 9-month time-
hidden layer showed better results with R2adj scale. This was exhibited in Bgures S3, S4 and S5
values equal to 0.89 and RMSE values equal to (supplementary material).
0.29 and to 0.31, respectively. Finally, for SPI12 The foregoing clearly shows the powerful ability
of ANNs regarding 2 months early drought pre-
(table 4), the highest values of R2adj , RMSE and
diction despite the high spatial and temporal
MAE equal to 0.943, 0.23 and to 0.142, respec- variability of SPI in such area located in the
tively, have been recorded for the plain of Low transitional climate between arid (in the south)
CheliA with 8 inputs and 15 neurons in the and Mediterranean (in the north). This variability
hidden layer. was captured in modelling by the use of different
In addition, we note that, comparing ANN designs and structures of the neural network. Thus,
models performance in all time scales, the highest the mentioned models and their structure can be
value of R2adj and lowest value of RMSE and MAE used for water resources planning and management
was found for SPI12. The good agreement between and its use is also recommended for the neigh-
SPIs and observed at plains and those predicted by bouring regions characterized with very limited
ANN at different time scales can be inferred availability of water resources.
42 Page 16 of 22 J. Earth Syst. Sci. (2020)129:42

Figure 6. Box plot for observed and predicted SPI values for the seven studied plains (1997–2010) at different SPI time scales
(red colour line: median; box: Brst and third quartiles; whiskers: 99% conBdence interval; + marker: outlier).

5. Discussion climatic variability. Moreover, these observa-

tions have been conBrmed by the temporal
Drought forecasting and spatio-temporal variabil- variability of drought analysis, through SPI,
ity analysis over the seven plains using SPIs series using many statistical tests have shown, for all
and their characteristics, namely drought severity time scales and all plains, the increase in the
and duration at 3, 6, 9, and 12-month scale has number and the intensiBcation of dry conditions
shown three major significant results: as well. The results also revealed the inCuence of
serial correlation in time series on detecting
1. As regards spatio-temporal drought character-
trends and change point. Where the trends to
ization, drought events have further occurred
dry conditions when the pre-whitened SPI time
and are characterized by severity and duration
series on the different time scales become
rises for all studied plains during the end of
significant only at 9- and 12-month scales.
1970s, early 1980s and end of 1980s decade,
The same for the change points, where they
except for the Ghriss and Sidi Bel-Abb es plains
have been reproduced on the pre-whitened SPI
rising in the Brst half-decade of 2000. These
series only in 9- and 12-month scale series with a
drought spells have been conBrmed where crop
slight change, whether precede or pursue the
yields are decreased and the minimum water
primary change-point. These results are thus
level in dams has been recorded especially in the
conBrmed with existing studies (Onyutha 2016;
west of the country. During the mentioned
Piyoosh and Ghosh 2017) where serial correla-
periods of SPIs at different time scales, the
tion is recognized for having impact on trends
driest phase was located in the extreme west of
detection in auto-correlated time series, which
Algeria: plains of Maghnia, Ghriss and Sidi Bel-
may falsify the tests for detecting trends and
Abb es. They have experienced a large rainfall
change points.
deBcit and required sustained irrigation to
support agricultural production. These regions The change points have pointed out that the rise
are known by their fertility and productivity. of drought events was significant only for the sec-
Baali and Amine (2015) found same results in ond half of 1970s, early 1980s and end of 1980s
the North of Morocco (Sa€ıss plain). According decade at 9 and 12 months scales. These results are
to the 5th IPCC report (Dai et al. 2004; Stocker coherent with previous studies showing an increase
et al. 2013), the increase in droughts frequency in drought frequency, duration and severity in
and severity in the area surrounding the the southern part of the Mediterranean basin in the
Mediterranean Sea was due to the climate second half of the 20th century (Peñuelas et al.
change that involves warmer temperatures, 2001; Hoerling et al. 2012; Ramadan et al. 2013;
altered precipitation patterns and increased Tramblay et al. 2013; Spinoni et al. 2014).
J. Earth Syst. Sci. (2020)129:42 Page 17 of 22 42

Figure 7. Comparison of observed and forecasted SPI12 during the validation phase for the seven studied plains.

Likewise, Hoerling et al. (2012) have reported the Atlantic variations and Azores anticyclones,
change in cold-season precipitation of Mediter- where the Maghnia, Sidi Bel-Abb es and
ranean region during 1902–2010 period and the Ghriss plains are located. This atmospheric
intensiBcation of this trend towards drier condi- circulations produce generally a peaceful cli-
tions with raised drought frequency after 1970. matic weather and dry condition (Xoplaki
Moreover, Mariotti (2010) found that the et al. 2004). The CheliA plains and Mitidja
Mediterranean region has experienced 10 of the 12 plain are positioned amid the western and
driest winters since 1902, just in the last 20 years. eastern Mediterranean Basin that are inCu-
Similar results were obtained separately for local enced by El Niño-Southern Oscillation (ENSO)
studies in the Mediterranean region and other revealed in previous studies (Meddi et al. 2010;
semi-arid regions like Raymond et al. (2016) for Zeroual et al. 2017).
north Algeria, Maccioni et al. (2015) and Abenavoli 3. As SPIs chronological series are autoregressive
et al. (2016) in Italy, Dabanlı et al. (2017) in Tur- of order 1, we opted for prediction from 2
key, Estrela and Vargas (2012) in Spain and Bari months ahead. Moreover, the choice for predic-
Abarghouei et al. (2011) in Iran. tion by artiBcial neural networks is set on the
forecast ability of this technique in the cases of
2. As far as the spatial distribution is con-
non-stationary time-series (Allende et al. 2002).
cerned, the severity of drought events ampli-
The neural network models with different
Bes from east to west and this is more
designs give satisfying results despite the
marked at 9 months and annual scale. These
non-stationary of SPIs time-series.
eAects can be explained by the geographical
location of the plains, i.e., the extreme west Thus, the mentioned models can be used for
of the Mediterranean Basin inCuenced by water resources planning and management in order
42 Page 18 of 22 J. Earth Syst. Sci. (2020)129:42

to adjust 2 months earlier the agricultural activi- of ANN model is examined with different input
ties demand according to the spatial patterns of structures and has been validated against the
drought trend. Two months ahead are sufBcient seven studied plains. Hence, the best Levenberg–
as long as the rainy months are from November Marquardt back propagation networks design was
to March in this area and the agriculture demand found different from a plain to another and among
is in the spring season (April to June). the time scales.
Therefore, we consider that 2 months forward Based on measures of performance evaluation,
are satisfactory to develop a drought warning sys- we found that the ArtiBcial Neural Network
tem in the seven plains based on meteorological models can represent an eAective tool for pre-
information for the best management of irrigation dicting monthly SPI and an accurate drought
processes, by SPI3 evaluation and water resources, warning system with 2 months lead-time based on
by SPI6, SPI9, SPI12, respectively. Results obtained meteorological drought information. These Bnd-
for the forecast are a decision-making tool for water ings may be helpful for decision makers in order
resources management and especially for irrigation to establish adequate irrigation schedules in case
in the investigated plains. of water stress and for optimizing agricultural
production.

6. Conclusions
Acknowledgement
The vast and fertile plains of north-western
Algeria are known by their important contribution The authors wish to thank the National Agency of
to the self-sufBciency of the country in terms of Water Resources for providing material help and
food-processing production. In recent decades, the data on which reported analysis are based on.
ampliBcation of drought events and the increase of
rainfall deBcit caused drastic impacts on water
supply and agriculture in this area. Frequent
occurrence of droughts will further intensify local References
imbalance between supply and demand of water
Abenavoli M R, Leone M, Sunseri F, Bacchi M and Sorgon aA
resources. Thus, the investigation of meteorological
2016 Root phenotyping for drought tolerance in bean
drought variability and drought prediction landraces from Calabria (Italy); J. Agron. Crop Sci. 202
approaches are necessary to more sustainable 1–12, https://doi.org/10.1111/jac.12124.
water resource management in the context of Abramowitz M and Stegun I A 1965 Handbook of mathemat-
drought monitoring. ical functions with formulas, graphs, and mathematical
This work, by analysing the spatial and temporal tables (Vol. 55); Courier Corporation.
Abyaneh H Z, Nia A M, Varkeshi M B, MaroB S and Kisi O 2011
variability of meteorological drought based on the Performance evaluation of ANN and ANFIS models for
standardized precipitation index ‘SPI’ series at estimating garlic crop evapotranspiration; J. Irrig. Drain.
different time scales during the period 1960–2010, Eng., https://doi.org/10.1061/(ASCE)IR.1943-4774.0000298.
has shown a well identiBed dry trend in the whole Aher S, Shinde S, Guha S and Majumder M 2017 IdentiBcation
study zone in terms of severity and duration rising of drought in Dhalai river watershed using MCDM and
ANN models; J. Earth Syst. Sci. 126(2) 21.
from east to west especially after about 1976 and
Ali Z, Hussain I, Faisal M, Nazir H M, Hussain T, Shad M Y,
1981. As a consequence, we expected that signifi- Mohamd Shoukry A and Hussain Gani S 2017 Forecasting
cant persistence of drought events in the near drought using multilayer perceptron artiBcial neural net-
future may be enduring. work model; Adv. Meteorol., https://doi.org/10.1155/
Finally, a feed-forward three-layer ANN models 2017/5681308.
were used for prediction of drought 2 months Allende H, Moraga C and Salas R 2002 ArtiBcial neural
networks in time series forecasting: A comparative analysis;
ahead using the SPI series at different time scales Kybernetika 38(6) 685–707.
(3, 6, 9, 12 months). The best performance of Asong Z E, Wheater H S, Bonsal B, Razavi S and Kurkute S
ANN models for drought forecasting was acquired 2018 Historical drought patterns over Canada and their
based on the analysis of the adjusted coefBcient of teleconnections with large-scale climate signals; Hydrol.
determination (R2adj ), the root-mean-square-error Earth Syst. Sci. 22(6) 3105–3124.
Bahrami M, Bazrkar S and Zarei A R 2019 Modeling,
(RMSE) and the mean absolute error (MAE). prediction and trend assessment of drought in Iran using
Results indicate that ANN models provided a standardized precipitation index; J. Water Clim. Change
satisfactory forecast outcome and the best design 10(1) 181–196.
J. Earth Syst. Sci. (2020)129:42 Page 19 of 22 42
Baali A and Amine C 2015 Etude de l’impact des variations of precipitation enhancement in northern Greece; Theor.
pluviom etriques sur les Cuctuations pi ezom
etriques des Appl. Climatol., https://doi.org/10.1007/BF00867995.
nappes phr eatiques superBcielles en zone semi-aride (Cas de Dariane A B, Farhani M and Azimi S 2018 Long term
la plaine de SAI‹ SS, Nord du Maroc); Eur. Sci. J. 11(27). streamCow forecasting using a hybrid entropy model;
Bari Abarghouei H, Kousari M R and Asadi Zarch M A 2013 Water Resour. Manag. 32 1439–1451, https://doi.org/10.
Prediction of drought in dry lands through feedforward 1007/s11269-017-1878-0.
artiBcial neural network abilities; Arab. J. Geosci., https:// Dehghani M, SaghaBan B, Rivaz F and Khodadadi A 2017
doi.org/10.1007/s12517-011-0445-x. Evaluation of dynamic regression and artiBcial neural
Bari Abarghouei H, Kousari M R and Asadi Zarch M A 2011 networks models for real-time hydrological drought fore-
Prediction of drought in dry lands through feedforward casting; Arab. J. Geosci. 10(12) 266, https://doi.org/10.
artiBcial neural network abilities; Arab. J. Geosci. 6 1007/s12517-017-2990-4.
1417–1433, https://doi.org/10.1007/s12517-011-0445-x. Demmak A 2008 Drought in Algeria 1975–2000 – Impact on
Barua S, Ng A W M and Perera B J C 2012 ArtiBcial neural water resources and control strategy [La s echeresse en
network-based drought forecasting using a nonlinear aggre- Algerie des ann
ees 1975–2000 – Impact sur les ressources en
gated drought index; J. Hydrol. Eng. 17 1408–1413, eau et strategie de lutte]; 2nd MEDA Water Reg. Event,
https://doi.org/10.1061/(ASCE)HE.1943-5584.0000574. 28–30 April 2008, Marrakech, Morocco, https://doi.org/10.
Barua S, Perera B J C, Ng A W M and Tran D 2010 Drought 1016/J.IJSBE.2014.04.006.
forecasting using an aggregated drought index and artiBcial Deo Ravinesh C, Kisi O and Singh V P 2017a Drought
neural network; J. Water Clim. Chang., https://doi.org/10. forecasting in eastern Australia using multivariate adaptive
2166/wcc.2010.000. regression spline, least square support vector machine and
Belayneh A and Adamowski J 2012 Standard precipitation M5Tree model; Atmos. Res. 184 149–175, https://doi.org/
index drought forecasting using neural networks, wavelet 10.1016/j.atmosres.2016.10.004.
neural networks, and support vector regression; Appl. Deo Ravinesh C, Tiwari M K, Adamowski J F and Quilty J M
Comput. Intell. Soft Comput., https://doi.org/10.1155/ 2017b Forecasting eAective drought index using a wavelet
2012/794061. extreme learning machine (W-ELM) model; Stoch. Envi-
Belayneh A, Adamowski J, Khalil B and Ozga-Zielinski B ron. Res. Risk Assess. 31 1211–1240, https://doi.org/10.
2014 Long-term SPI drought forecasting in the Awash 1007/s00477-016-1265-z.
River Basin in Ethiopia using wavelet neural networks and Dingman S 1994 Physical Hydrology; Prentice Hall. Inc., New
wavelet support vector regression models; J. Hydrol., Jersey.
https://doi.org/10.1016/j.jhydrol.2013.10.052. Djerbouai S and Souag-Gamane D 2016 Drought forecasting
Byun H R and Kim D W 2010 Comparing the eAective using neural networks, wavelet neural networks, and
drought index and the standardized precipitation index; In: stochastic models: Case of the Algerois Basin in North
Economics of drought and drought preparedness in a climate Algeria; Water Resour. Manag., https://doi.org/10.1007/

change context, Lopez-Francos 
A.(comp.), Lopez-Francos s11269-016-1298-6.

A.(collab.). Options Mediterran 
eennes.  A. Seminaires
Ser.  Edwards D C 1997 Characteristics of 20th century drought in

Mediterran 
eens, Vol. 95, pp. 85–89. the United States at multiple time scales (No. AFIT-97-051);
Cancelliere A, Mauro G Di Bonaccorso B and Rossi G 2007 Air Force Inst of Tech Wright-Patterson Afb Oh.
Drought forecasting using the standardized precipitation Estrela T and Vargas E 2012 Drought management plans in
index. Water Resour. Manag., https://doi.org/10.1016/j. the European Union. The case of Spain; Water Resour.
jhydrol.2013.10.052. Manag., https://doi.org/10.1007/s11269-011-9971-2.
Chai T and Draxler R R 2014 Root mean square error (RMSE) Gabriel K R and Neumann J 1962 A Markov chain model
or mean absolute error (MAE)? – Arguments against for daily rainfall occurrence at Tel Aviv; Quart. J. Roy.
avoiding RMSE in the literature; Geosci. Model Dev., Meteorol. Soc. 88 90–95, https://doi.org/10.1002/qj.
https://doi.org/10.5194/gmd-7-1247-2014. 49708837511.
Chandramouli C V, Kaoukis N, Karim M and Dorworth L Ghorbani M A, Zadeh H A, Isazadeh M and Terzi O 2016 A
2017 Uses of precipitation-based climate indices in drought comparative study of artiBcial neural network (MLP, RBF)
characterization; J. Hydrol. Eng. 22 05017013, https://doi. and support vector machine models for river Cow predic-
org/10.1061/(ASCE)HE.1943-5584.0001536. tion; Environ. Earth Sci., https://doi.org/10.1007/s12665-
Choi B, Lee J H and Kim D H 2008 Solving local minima 015-5096-x.
problem with large number of hidden nodes on two-layered Govindaraju R S 2000 ArtiBcial neural networks in hydrology.
feed-forward artiBcial neural networks; Neurocomputing I: Preliminary concepts; J. Hydrol. Eng. 5 115–123, https://
71(16–18) 3640–3643. doi.org/10.1061/(ASCE)1084-0699(2000)5:2(115).
_ Mishra A K and Sßen Z 2017 Long-term spatio-
Dabanlı I, Goyal M K, Bharti B, Quilty J, Adamowski J and Pandey A 2014
temporal drought variability in Turkey; J. Hydrol. 552 Modeling of daily pan evaporation in subtropical climates using
779–792, https://doi.org/10.1016/j.jhydrol.2017.07.038. ANN, LS-SVR, Fuzzy Logic, and ANFIS; J. Expert Syst. Appl.,
Dai A, Trenberth K E and Qian T 2004 A global dataset of https://doi.org/10.1016/j.eswa.2014.02.047.
palmer drought severity index for 1870–2002: Relationship Guhathakurta P, Menon P, Inkane P M, Krishnan U and
with soil moisture and eAects of surface warming; J. Sable S T 2017 Trends and variability of meteorological
Hydrometeorol. 5 1117–1130, https://doi.org/10.1175/ drought over the districts of India using standardized
JHM-386.1. precipitation index; J. Earth Syst. Sci. 126(8) 120.
Dalezios N R, PapazaBriou Z G, Papamichail D M and Guttman N B 1998 Comparing the Palmer drought index and
Karacostas T S 1991 Drought assessment for the potential the standardize precipitation index; J. Am. Water Resour.
42 Page 20 of 22 J. Earth Syst. Sci. (2020)129:42

Assoc. 34 113–121, https://doi.org/10.1111/j.1752-1688. Liang Y, Niu D, Ye M and Hong W-C 2016 Short-term load
1998.tb05964.x. forecasting based on wavelet transform and least squares
Habibi B, Meddi M, Torfs P J, Remaoun M and Van Lanen support vector machine optimized by improved cuckoo
H A 2018 Characterisation and prediction of meteorological search; Energies, https://doi.org/10.3390/en9100827.
drought using stochastic models in the semi-arid Lippmann R P 1987 An introduction to computing with neural
CheliA–Zahrez basin (Algeria); J. Hydrol.: Regional Studies nets; IEEE ASSP Mag. 4 4–22, https://doi.org/10.1109/
16 15–31. MASSP.1987.1165576.
Hamed K H and Ramachandra Rao A 1998 A modiBed Lloyd-Hughes B and Saunders M A 2002 A drought climatol-
Mann–Kendall trend test for autocorrelated data; J. ogy for Europe; Int. J. Climatol. 22 1571–1592, https://doi.
Hydrol. 204 182–196, https://doi.org/10.1016/S0022- org/10.1002/joc.846.
1694(97)00125-X. Maccioni P, Kossida M, Brocca L and Moramarco T 2015
He Z, Wen X, Liu H and Du J 2014 A comparative study of Assessment of the drought hazard in the Tiber river basin in
artiBcial neural network, adaptive neuro fuzzy inference central Italy and a comparison of new and commonly used
system and support vector machine for forecasting river meteorological indicators; J. Hydrol. Eng. 20 05014029,
Cow in the semiarid mountain region; J. Hydrol., https:// https://doi.org/10.1061/(ASCE)HE.1943-5584.0001094.
doi.org/10.1016/j.jhydrol.2013.11.054. Madadgar S and Moradkhani H 2013 Drought analysis under
Hirose Y, Yamashita K and Hijiya S 1991 Back-propagation climate change using copula; J. Hydrol. Eng. 18 746–759,
algorithm which varies the number of hidden units; Neural https://doi.org/10.1061/(ASCE)HE.1943-5584.
Networks, https://doi.org/10.1016/0893-6080(91)90032-Z. Maier H R and Dandy G C 1996 The use of artiBcial neural
Hoerling M, Eischeid J, Perlwitz J, Quan X, Zhang T and networks for the prediction of water quality parameters;
Pegion P 2012 On the increased frequency of Mediterranean Water Resour. Res. 32 1013–1022, https://doi.org/10.
drought; J. Clim. 25 2146–2161, https://doi.org/10.1175/ 1029/96WR03529.
JCLI-D-11-00296.1. Mann H B 1945 Nonparametric tests against trend; Econo-
Huang F, Huang J, Jiang S H and Zhou C 2017 Prediction of metrica J. Econometric Soc. 13 245–259.
groundwater levels using evidence of chaos and support Mariotti A 2010 Recent changes in the mediterranean water
vector machine; J. Hydroinformatics 19(4) 586–606. cycle: A pathway toward long-term regional hydroclimatic
Jain A and Kumar A M 2007 Hybrid neural network models change? J. Clim. 23 1513–1525, https://doi.org/10.1175/
for hydrologic time series forecasting; Appl. Soft Comput. 2009JCLI3251.1.
7(2) 585–592. Marriott F H C and Pope J A 1954 Bias in the estimation of
Jalalkamali A, Moradi M and Moradi N 2015 Application of autocorrelations; Biometrika 41(3/4) 390–402.
several artiBcial intelligence models and ARIMAX model Masinde M 2014 ArtiBcial neural networks models for
for forecasting drought using the standardized precipitation predicting eAective drought index: Factoring eAects of
index; Int. J. Environ. Sci. Technol., https://doi.org/10. rainfall variability; Mitig. Adapt. Strat. Gl. 19(8)
1007/s13762-014-0717-6. 1139–1162.
Kendall M G 1975 Rank correlation measures; Charles GriDn, Masud M B, Khaliq M N and Wheater H S 2017 Projected
London, Vol. 202, 15p. changes to short- and long-duration precipitation extremes
Khan M I, Liu D, Fu Q and Faiz M A 2018 Detecting the over the Canadian Prairie Provinces; Clim. Dyn., https://
persistence of drying trends under changing climate condi- doi.org/10.1007/s00382-016-3404-0.
tions using four meteorological drought indices; Meteorol. McKee T B, Doesken N J and Kleist J 1993 The relationship of
Appl. 25(2) 184–194. drought frequency and duration to time scales; Proceedings
Khan S, Gabriel H F and Rana T 2008 Standard precipitation of the 8th Conference on Applied Climatology, Boston, MA:
index to track drought and assess impact of rainfall on Am. Meteorol. Soc. 17(22) 179–183.
watertables in irrigation areas; Irrig. Drain. Syst. 22 Meddi H and Meddi M 2009 Variabilit e des precipitations
159–177, https://doi.org/10.1007/s10795-008-9049-3. annuelles du Nord-Ouest de l’Alg erie; Sci. Chang.
Kim T-W and Vald es J B 2003 Nonlinear model for drought Planetaires/S
 echeresse
 20 57–65.
forecasting based on a conjunction of wavelet transforms Meddi M, Assani A A and Meddi H 2010 Temporal variability
and neural networks; J. Hydrol. Eng. 8 319–328, https:// of annual rainfall in the Macta and Tafna Catchments,
doi.org/10.1061/(ASCE)1084-0699(2003)8:6(319). northwestern Algeria; Water Resour. Manag. 24
Komasi M, Sharghi S and Safavi H R 2018 Wavelet and 3817–3833, https://doi.org/10.1007/s11269-010-9635-7.
cuckoo search-support vector machine conjugation for Medjerab A and Henia L 2005 R egionalisation des pluies
drought forecasting using standardized precipitation index annuelles dans l’Alg erie nord-occidentale/Regionalisation
(case study: Urmia Lake, Iran); J. Hydroinfor. 20 975–988, of annual rainfall in the north-western parts of Algeria; Rev.
https://doi.org/10.2166/hydro.2018.115. Geogr. Est. 45(2) 1–14, http://journals.openedition.org/
Koutsoyiannis D 2003 Climate change, the Hurst phe- rge/501.
nomenon, and hydrological statistics; Hydrol. Sci. J. 48 Mishra A K and Singh V P 2010 A review of drought concepts;
3–24, https://doi.org/10.1623/hysj.48.1.3.43481. J. Hydrol. 391(1–2) 202–216.
Koutsoyiannis D and Montanari A 2007 Statistical analysis of Mishra A K and Desai V R 2006 Drought forecasting using
hydroclimatic time series: Uncertainty and insights; Water feed-forward recursive neural network; Ecol. Modell. 198
Resour. Res. 43(5), https://doi.org/10.1029/2006WR005592. 127–138, https://doi.org/10.1016/j.ecolmodel.2006.04.017.
Lazri M, Ameur S, Brucker J M, Lahdir M and Sehad M 2015 Mishra A K and Desai V R 2005 Drought forecasting using
Analysis of drought areas in northern Algeria using Markov stochastic models; Stoch. Environ. Res. Risk Assess. 19
chains; J. Earth Syst. Sci. 124(1) 61–70. 326–339, https://doi.org/10.1007/s00477-005-0238-4.
J. Earth Syst. Sci. (2020)129:42 Page 21 of 22 42
Mudelsee M 2001 Note on the bias in the estimation of the SheDeld J and Wood E F 2008 Projected changes in drought
serial correlation coefBcient of AR (1) processes; Stat. occurrence under future global warming from multi-
Pap. 42(4) 517–527. model, multi-scenario, IPCC AR4 simulations; Clim.
Noori R, Deng Z, Kiaghadi A and Kachoosangi F T 2015 How Dyn. 31 79–105, https://doi.org/10.1007/s00382-007-
reliable are ANN, ANFIS and SVM techniques for predict- 0340-z.
ing longitudinal dispersion coefBcient in natural rivers? Shiau J T 2006 Fitting drought duration and severity with
J. Hydraul. Eng. 142(1) 04015039. two-dimensional copulas; Water Resour. Manag. 20
Onyutha C 2016 Statistical uncertainty in hydrometeorolog- 795–815, https://doi.org/10.1007/s11269-005-9008-9.
ical trend analyses; Adv. Meteorol., https://doi.org/10. Shirmohammadi B, Moradi H, Moosavi V, Semiromi M T and
1155/2016/8701617. Zeinali A 2013 Forecasting of meteorological drought using
Palmer W C 1965 Meteorological Drought; U.S. Weather Bur. Wavelet-ANFIS hybrid model for different time steps
Res. Pap. No. 45. (case study: Southeastern part of east Azerbaijan province,
Paulo A, Martins D and Pereira L S 2016 InCuence of precipitation Iran); Nat. Hazards, https://doi.org/10.1007/s11069-013-
changes on the SPI and related drought severity. An analysis 0716-9.
using long-term data series; Water Resour. Manag. 30 Shoaib M, Shamseldin A Y, Melville B W and Khan M M 2016
5737–5757, https://doi.org/10.1007/s11269-016-1388-5. Hybrid wavelet neuro-fuzzy approach for rainfall-runoA
Peñuelas J, Lloret F and Montoya R 2001 Severe drought modeling; J. Comput. Civil Eng., https://doi.org/10.1061/
eAects on Mediterranean woody Cora in Spain; Forest Sci. (ASCE)CP.1943-5487.0000457.
47(2) 214–218. Soh Y W, Koo C H, Huang Y F and Fung K F 2018
Pettitt A N 1979 A non-parametric approach to the change- Application of artiBcial intelligence models for the predic-
point problem; Appl. Stat. 28 126–135, https://doi.org/10. tion of standardized precipitation evapotranspiration index
2307/2346729. (SPEI) at Langat River Basin, Malaysia; Comput. Electron.
Piyoosh A K and Ghosh S K 2017 EAect of autocorrelation on Agric. 144 164–173, https://doi.org/10.1016/J.COMPAG.
temporal trends in rainfall in a valley region at the foothills 2017.12.002.
of Indian Himalayas; Stoch. Environ. Res. Risk Assess. 31 Spinoni J, Naumann G, Carrao H, Barbosa P and Vogt J 2014
2075–2096, https://doi.org/10.1007/s00477-016-1347-y. World drought frequency, duration and severity for
Ramadan H H, Beighley R E and Ramamurthy A S 2013 1951–2010; Int. J. Climatol. 34 2792–2804, https://doi.
Temperature and precipitation trends in Lebanon’s org/10.1002/joc.3875.
Largest River: The Litani Basin; J. Water Resour. Plan. Stocker T, Qin D, Plattner G and Tignor M 2013 IPCC, 2013:
Manag. 139 86–95, https://doi.org/10.1061/(ASCE)WR. Climate Change 2013: The Physical Science Basis; Contri-
1943-5452.0000238. bution of Working Group I to the Fifth Assessment Report
Raymond F, Ullmann A and Camberlin P 2016 Pr ecipita- of the Intergovernmental Panel.
tions intenses sur le Bassin M editerraneen: quelles ten- Storch H von Navarra A 1995 Analysis of Climate Variability,
dances entre 1950 et 2013? Cybergeo Eur. J. Geogr. Nat Springer-Verlag, Berlin-Heidelberg, 334p.
Paysage document 760, https://doi.org/10.4000/cybergeo. Tang Z and Fishwick P A 1993 Feedforward neural nets as
27410. models for time series forecasting; J. Comput. 5 374–385,
Rezaie-Balf M, Zahmatkesh Z and Kim S 2017 Soft computing https://doi.org/10.1287/ijoc.5.4.374.
techniques for rainfall-runoA simulation: Local non-para- Tarawneh Z and Khalayleh Y 2016 Improved estimate of
metric paradigm vs. model classiBcation methods; Water multiyear drought for water resources management studies;
Resour. Manag. 31(12) 3843–3865. J. Water Clim. Chang. 7 721–730, https://doi.org/10.2166/
Salhi C, Touaibia B and Zeroual A 2013 Les r eseaux de wcc.2016.151.
neurones et la r
egression multiple en pr ediction de l’
erosion Thom H C S 1958 A note on the Gamma Distribution; Mon.
speciBque: cas du bassin hydrographique Alg erois-Hodna- Weather Rev. 86 117–122, https://doi.org/10.1175/1520-
Soummam (Alg erie); Hydrol. Sci. J. 58 1573–1580, https:// 0493(1958)086\0117:ANOTGD[2.0.CO;2.
doi.org/10.1080/02626667.2013.824090. Thorn H C S 1966 Some methods of climatological analysis;
Salvi K and Ghosh S 2016 Projections of extreme dry and wet library.wmo.int Number 81, 53p.
spells in the 21st century India using stationary and non- Tian Y, Xu Y P and Wang G 2018 Agricultural drought
stationary standardized precipitation indices; Clim. Change prediction using climate indices based on support vector
139(3–4) 667–681. regression in Xiangjiang River basin; Sci. Total Environ.,
Sßen Z 1998 Average areal precipitation by percentage https://doi.org/10.1016/j.scitotenv.2017.12.025.
weighted polygon method; J. Hydrol. Eng. 3(1) 69–72. Tongal H and Booij M J 2017 QuantiBcation of parametric
Serinaldi F and Kilsby C G 2016 The importance of uncertainty of ANN models with GLUE method for
prewhitening in change point analysis under persistence; different streamCow dynamics; Stoch. Environ. Res. Risk
Stoch. Environ. Res. Risk Assess. 30 763–777, https://doi. Assess. 31(4) 993–1010.
org/10.1007/s00477-015-1041-5. Torranin P 1973 Unpredictability of hydrological drought;
Shabri A 2014 A hybrid wavelet analysis and adaptive neuro- In: The Second International Symposium in Hydrology,
fuzzy inference system for drought forecasting; Appl. Math. Fort Collins, Colo., Water Resources Publications,
Sci. 8 6909–6918, https://doi.org/10.12988/ams.2014.48263. pp. 595–604.
Sheela K G and Deepa S N 2013 Review on methods to Bx Tramblay Y, El Adlouni S and Servat E 2013 Trends and
number of hidden neurons in neural networks neuro-fuzzy variability in extreme precipitation indices over Maghreb
inference system for drought forecasting; Math. Probl. Eng., countries; Nat. Hazards Earth Syst. Sci. 13 3235–3248,
https://doi.org/10.1155/2013/425740. https://doi.org/10.5194/nhess-13-3235-2013.
42 Page 22 of 22 J. Earth Syst. Sci. (2020)129:42

Trenn S 2008 Multilayer perceptrons: Approximation order Yue S, Pilon P and Cavadias G 2002 Power of the Mann–
and necessary number of hidden units. IEEE Trans. Neural Kendall and Spearman’s rho tests for detecting monotonic
Networks, https://doi.org/10.1109/TNN.2007.912306. trends in hydrological series. J. Hydrol. 259, https://doi.
Vicente-Serrano S M, Begueria S and Lopez-Moreno J I 2010 A org/10.1016/S0022-1694(01)00594-7.
multiscalar drought index sensitive to global warming: The Yue S and Wang C 2004 The Mann–Kendall test modiBed by
standardized precipitation evapotranspiration index; J. Clim. eAective sample size to detect trend in serially correlated
23 1696–1718, https://doi.org/10.1175/2009JCLI2909.1. hydrological series; Water Resour. Manag. 18 201–218,
Wang L, Zeng Y and Chen T 2015 Back propagation neural https://doi.org/10.1023/B:WARM.0000043140.61082.60.
network with adaptive differential evolution algorithm for Zeroual A, Assani A and Meddi M 2017 Combined analysis of
time series forecasting; Expert Syst. Appl., https://doi.org/ temperature and rainfall variability as they relate to
10.1016/j.eswa.2014.08.018. climate indices in Northern Algeria over the 1972–2013
Wilhite D A and Glantz M H 1985 Understanding: The period; Hydrol. Res. 48 584–595.
drought phenomenon: The role of definitions; Water Int., Zeroual A, Assani A A, Meddi M and Alkama R 2019
https://doi.org/10.1080/02508068508686328. Assessment of climate change in Algeria from 1951 to 2098
Willmott C J and Matsuura K 2005 Advantages of the mean using the K€ oppen–Geiger climate classiBcation scheme;
absolute error (MAE) over the root mean square error Clim. Dyn. 52(1–2) 227–243.
(RMSE) in assessing average model performance. Clim. Zeroual A, Meddi M and Assani A A 2016 ArtiBcial neural
Res., https://doi.org/10.3354/cr030079. network rainfall-discharge model assessment under rating
WMO 2008 Guide to Hydrological Practice, Volume I, curve uncertainty and monthly discharge volume predic-
Hydrology – From Measurement to Hydrological Informa- tions; Water Resour. Manag. 30 3191–3205, https://doi.
tion; 6th edn, World Meteorological Organisation, Geneva, org/10.1007/s11269-016-1340-8.
Switzerland. Zhang Q, Li J, Singh V P and Bai Y 2012 SPI-based
Wong F S 1991 Time series forecasting using backpropagation evaluation of drought events in Xinjiang, China; Nat.
neural networks; Neurocomputing 2 147–159, https://doi. Hazards 64 481–492, https://doi.org/10.1007/s11069-012-
org/10.1016/0925-2312(91)90045-D. 0251-0.
Xoplaki E, Gonz alez-Rouco J F, Luterbacher J and Wanner H Zhang Y, Li W, Chen Q, Pu X and Xiang L 2017 Multi-models
2004 Wet season Mediterranean precipitation variability: for SPI drought forecasting in the north of Haihe River
InCuence of large-scale dynamics and trends; Clim. Dyn. 23 Basin, China; Stoch. Environ. Res. Risk Assess. 31
https://doi.org/10.1007/s00382-004-0422-0. 2471–2481, https://doi.org/10.1007/s00477-017-1437-5.

Corresponding editor: SUBIMAL GHOSH