Abstract: In recent years, several drought monitoring indexes have found application to describe and compare droughts among different
time periods and regions as well as to forecast the evolution of ongoing droughts, in order to select appropriate mitigation measures and
policies for water resources management under shortage risk conditions. However, limited efforts have been made to investigate the
possibilities of using information conveyed by large-scale climatic indexes to improve the forecasting ability of drought indexes, provided
they exert some influence on the climatic variability in a region. The aim of the paper is to develop models for forecasting Palmer
Hydrological Drought Index series in Sicily 共Italy兲 based on artificial neural networks, and to extend such models in order to include
information from large-scale climatic indexes. First, the influence of North Atlantic Oscillation 共NAO兲 and European Blocking 共EB兲
indexes on Palmer index series, computed on areal monthly precipitation from 1955 until 1999 in Sicily, has been investigated by means
of a correlation analysis. Results indicate that NAO and EB series are significantly correlated with Palmer index series for winter and
autumn months, with special reference to the last decades. Then, forecasting models based on neural networks have been developed, using
different approaches. The comparison between the prediction for winter and autumn months obtained by either including or not including
the NAO and EB indexes within the forecasting model indicates some improvements in terms of R2 when NAO and especially EB are
considered. A different model behavior has been observed for the spring and summer predictions for which no significant improvements
in terms of model predictive capability, due to the introduction of the climatic indexes as input variables, have been observed.
DOI: 10.1061/共ASCE兲HE.1943-5584.0000028
CE Database subject headings: Forecasting; Climatology; Water resources; Droughts; Water supply; Neural networks.
forecasting, provided they exert some influence on the climatic −2.0⬍ PHDI⬍ −1.5 Mild drought
variability in a region. −3.0⬍ PHDI⬍ −2.0 Moderate drought
The aim of the paper is to develop models for forecasting −4.0⬍ PHDI⬍ −3.0 Severe drought
PHDI series in Sicily 共Italy兲 based on artificial neural networks PHDI⬍ −4.0 Extreme drought
共ANNs兲, and extending such models in order to include informa-
tion deriving from large scale climatic indexes. In particular, an
evaluation of whether considering NAO and EB indexes informa-
tion leads to improvements in the forecasting PHDI values is A drought classification proposed by Karl 共1986兲 for the
carried out by means of Monte Carlo analysis. United States and reported in Table 1, indicates extreme drought
The PHDI was selected due to its capability to assess water for PHDI艋 −4.0 with an approximate cumulative frequency
deficiencies at large geographical scale allowing a reasonable less than 4%, severe drought for −3.9艋 PHDI艋 −3.0 共frequency
comparison of drought conditions both in space and time, 5–10%兲, and mild to moderate drought for −2.9艋 PHDI艋 −1.5
whereas the adoption of ANNs is justified by their ability in mod- 共frequency 11–27%兲. The index can also express wet conditions;
eling and forecasting nonlinear and nonstationary time series in i.e., 0 is normal, +2 is moderate rainfall, etc.
hydrology, due to their intrinsic nonlinear property and flexibility The advantage of the Palmer index is that it is standardized to
for modeling 共ASCE 2000a,b兲. local climate, so it can be applied to different countries to dem-
onstrate relative dry or wet conditions.
Although the Palmer index is mainly used in United States
Palmer Drought Index 共Lohani and Loganathan 1997; Kim and Valdes 2003兲, it has
been applied also to other areas such as Australia and South
The original PDSI was developed as a descriptor of meteorologi- Africa 共Smith et al. 1993兲, Europe 共Briffa et al. 1994兲, and Medi-
cal drought and its computation is based on the comparison be- terranean areas 共Cancelliere et al. 1996兲 confirming that, despite
tween the actual precipitation over an area with the amount of its limitations, PHDI can be considered an appropriate index for
precipitation required to maintain the average supply of that area drought monitoring.
共climatically appropriate for existing conditions兲 共Palmer 1965兲.
The method is based on the estimation of a soil water balance,
usually on a monthly basis, over a calibration period using his- North Atlantic Oscillation and European Blocking
torical records of hydrological variables to derive climatic coeffi- Effects on Precipitation and Droughts in Europe
cients needed for the computation of the monthly values of the
index. The NAO and EB have long been recognized as two of the pat-
As the evaluation of the original PDSI relies on a backtracking terns of atmospheric variability in the Northern Hemisphere.
procedure to determine the beginning and the end of the drought, However, they have become the subject of a wider interest only
it cannot be used for real time monitoring. Therefore, the slightly recently.
modified version of the index, the PHDI, is generally used for Historically, NAO has been defined as an index that measures
operational purposes and has been applied here. the difference of normalized surface pressure between two areas
Computation of PHDI is based on the application of the recur- over the North Atlantic Ocean, Ponta Delgada in the Azores and
sive formula Stykkisholmur in Iceland, but recently, researchers have shown
that, during the winter season, stations located in Iberia can be
Hi used with some advantages over Ponta Delgada 共Hurrel 1995;
PHi = 0.897PHi−1 + 共1兲 Jones et al. 1997兲.
3
Several studies have established links between the NAO
where i = 12共y − 1兲 + t; y = 1 , . . . , n indicates the year; t = 1 , . . . , 12 phases and precipitation in western Europe and the Mediterranean
the month; and basin 共Hurrell 1995; Qian et al. 2000兲 showing a significant de-
pendence between these two variables. Such control exerted by
Hi = 共Pi − P̂i兲Kt 共2兲 NAO on the precipitation regime is related to corresponding
changes in the associated activity of North-Atlantic storm tracks
Pi being the precipitation on interval i, P̂i the climatologically that affect the western European border 共Goodess and Jones
appropriate precipitation for existing conditions at the same inter- 2002兲. Over the past decade NAO has remained in a positive
val, and Kt a monthly standardization factor. phase during winter, contributing significantly to the recent
PHDI values are centered around zero and range approxi- wintertime warmth across Europe and to cold conditions in the
mately between −5.0 and 5.0 with negative values indicating northwestern Atlantic 共Hurrell et al. 2001兲. During the positive
drought of different severities. phases of NAO, the North Atlantic westerlies, which provide
attention because they could cause climatic anomalies by block- Palmer series have been computed from areal averaged
ing the western currents for several weeks, so that the Atlantic monthly precipitation series from 1926 until 2003, obtained by
perturbations are forced to abandon their usual trajectories. applying the Thiessen polygons method on 40 precipitation
Therefore, together with the NAO index, the EB index consti- stations in Sicily. The selected stations are included in the
tutes a second atmospheric circulation pattern for the European drought monitoring bulletin published on the website of the
area 共Rogers 1990; Tibaldi and Molteni 1990; Pavan et al. Sicilian Regional Agency for Waste and Water—Water Obser-
2002a,b兲. As NAO and EB are strongly correlated, respectively, vatory, formerly the Hydrographic Service 关Rossi and Cancelliere
with the first and third principal components of the Atlantic vari- 共2002兲, http://www.uirsicilia.it兴 in which drought indexes
ability, they are uncorrelated. 共SPI and Palmer兲 are currently computed for real time drought
The seasonal temperature and precipitation anomalies over the monitoring.
European area are deeply connected to these two large-scale Regarding the NAO series, following Jones et al. 共1997兲, it
patterns and high correlation among these patterns and precipita- was decided to adopt the Gibraltar-Iceland NAO index developed
tion for the Mediterranean and the Alpine areas was proved by by the Climatic Research Unit of the University of East Anglia,
Quadrelli et al. 共2001a,b兲. U.K., for the same period 共1926–2003兲 共http://gcmd.nasa.gov/
records兲. Finally, the EB series has been computed as the third
principal component of the Z500 from the reanalysis data 共NCEP兲
ANN Forecasting Model for the European area from 90°W to 60°E and from 20°N to
90°N. The available EB data set covers the observed period
ANNs have been extensively used in the past two decades for the 1955–1999 共Cacciamani et al. 2007兲.
development of a large number of forecasting models in many The correlation between the PHDI index in a given month and
scientific areas. ANNs are able to mimic a large class of nonlinear the climatic index 共either NAO or EB兲, averaged over the previ-
functions, and therefore they are ideal candidates to develop em- ous h 共h = 1 , . . . , 4兲, months has been investigated, in agreement
pirical 共regression type兲 models. An ANN model is based on a with the procedure already applied by Vicente-Serrano 共2005兲 and
connectionist approach to computation involving several transfor- Lopez-Moreno et al. 共2007兲.
mation elements 共neurons兲, interconnected and distributed over Let PHDIyj+k be the value of the Palmer index at month t + k
different layers. 共t = 1 , . . . , 12; k = 0 , . . . , 3兲 and year y, and Z共h兲
y,t the value of the
In general, ANNs consist of an input layer, through which climatic index 共either NAO or EB兲 at month t for a given time
information reaches the network, one or more hidden layers, into averaging scale h 共h = 1 , . . . , 4兲.
which information is routed, and a layer of output, through which The Pearson correlation coefficient between the two variables,
the results are provided 共ASCE 2000a,b兲. for a given time averaging scale h, has been computed as
The advantages of ANNs over other regressive models are
that 共1兲 their application does not require a prior knowledge of
the process because ANNs have black-box properties; 共2兲 ANNs 兺ny=1共PHDIy,t+k − PHDI兲共Zy,t − Z̄兲
are inherently nonlinear as neurons activate a nonlinear filter rt,k = 共3兲
called an activation function; and 共3兲 the generalizing capabilities 关兺ny=1共PHDIy,t+k − PHDI兲2兺ny=1共Zy,t − Z̄兲2兴0.5
of ANNs make them attractive for a large class of problems in
hydrology. in which PHDI and Z̄ represent the mean values for the consid-
In the present case, ANNs have been adopted to model the ered variables.
relationship between future values of PHDI index and present In order to compute Kendall’s , let us consider the vector of
values of both PHDI and large-scale climatic indexes. Thus, in- variables 共PHDIy,t+k , Zy,t兲. Two distinct pairs of observations
dependent variables, i.e., an input vector consisting of monthly 共PHDI j,t+k , Z j,t兲 and 共PHDIm,t+k , Zm,t兲 are said to be concordant if
PHDI and climatic indexes data are used to generate the depen- 共PHDI j,t+k − PHDIm,t+k兲 共Z j,t − Zm,t兲 ⬎ 0, and discordant if
dent variable values, i.e., the output vector of monthly Palmer 共PHDI j,t+k − PHDIm,t+k兲 共Z j,t − Zm,t兲 ⬍ 0. Kendall’s has been then
index predictions. computed as = 共c − d兲 / 共c + d兲, where c denotes the number of
Single hidden layer feed-forward networks have been used concordant pairs, and d is the number of discordant pairs.
in the present work. Based on experiments performed on this case Correlation analysis has been carried out for each month t.
study, the logistic activation function for the neurons of input As an example, in Tables 2 and 3, correlation values between
and hidden layers and the linear function for the neurons of output PHDI and the considered climatic series are shown for the month
layer were chosen. Moreover, following Swingler 共1996兲 and of March considering the three periods of simultaneous observa-
Berry and Linoff 共1997兲, the maximum number of neurons in tion of different length 共i.e., 1955–1999, 1970–1999, and 1980–
the hidden layer has been set equal to double the input layer 1999兲.
dimension in order to reduce calibration errors and to avoid over- It is evident how the correlation significantly increases as the
Table 3. Pearson and Kendall’s 共in Brackets兲 Correlation Coefficient between EB Series 共Averaged over 1 – 4 months兲 and Palmer Series Simultaneous
and Shifted 1, 2, or 3 Months Ahead
Palmer
EB averaging scale h
Period Month 共month兲 Same month 1 month ahead 2 months ahead 3 months ahead
1955–1999 March 1 0.40 共0.26兲 0.39 共0.25兲 0.39 共0.28兲 0.40 共0.28兲
2 0.25 共0.14兲 0.21 共0.12兲 0.19 共0.13兲 0.19 共0.14兲
3 0.37 共0.24兲 0.35 共0.25兲 0.37 共0.27兲 0.38 共0.29兲
4 0.44 (0.30) 0.43 (0.34) 0.44 (0.36) 0.45 (0.37)
1970–1999 March 1 0.44 (0.30) 0.41 共0.27兲 0.45 (0.30) 0.45 (0.30)
2 0.26 共0.16兲 0.24 共0.17兲 0.28 共0.18兲 0.32 共0.23兲
3 0.42 共0.28兲 0.42 (0.30) 0.49 (0.32) 0.52 (0.36)
4 0.51 (0.37) 0.52 (0.41) 0.60 (0.43) 0.62 (0.47)
1980–1999 March 1 0.59 (0.38) 0.60 (0.36) 0.63 (0.40) 0.65 (0.40)
2 0.42 (0.31) 0.43 (0.35) 0.42 (0.37) 0.45 (0.37)
3 0.62 (0.37) 0.64 (0.42) 0.68 (0.43) 0.72 (0.47)
4 0.63 (0.44) 0.65 (0.47) 0.71 (0.51) 0.74 (0.53)
Note: Bold type= significant values for a significance level ␣ = 5%.
Table 4. Pearson and Kendall’s 共in Brackets兲 Correlation Coefficient between EB Series 共Averaged over 1 – 4 Months兲 and Palmer Series Simultaneous
and Shifted 1, 2, or 3 Months Ahead
Palmer
EB averaging scale h
Period Month 共month兲 Same month 1 month ahead 2 months ahead 3 months ahead
1955–1999 June 1 −0.23 共−0.06兲 −0.17 共−0.02兲 −0.22 共−0.09兲 −0.21 共−0.16兲
2 −0.08 共−0.05兲 −0.06 共−0.04兲 −0.13 共−0.13兲 −0.16 共−0.19兲
3 −0.04 共0.02兲 0.02 共0.07兲 −0.06 共−0.01兲 −0.03 共−0.04兲
4 −0.16 共−0.02兲 −0.14 共−0.01兲 −0.19 共−0.06兲 −0.13 共−0.07兲
1970–1999 June 1 −0.29 共−0.04兲 −0.23 共−0.05兲 −0.28 共−0.11兲 −0.24 共−0.15兲
2 −0.09 共−0.01兲 −0.11 共−0.07兲 −0.17 共−0.17兲 −0.14 共−0.18兲
3 −0.13 共0.04兲 −0.10 共0.01兲 −0.13 共−0.07兲 −0.05 共−0.04兲
4 −0.27 共0.01兲 −0.30 共−0.05兲 −0.29 共−0.07兲 −0.17 共−0.05兲
1980-1999 June 1 −0.19 共0.00兲 −0.13 共−0.02兲 −0.20 共−0.01兲 −0.25 共−0.12兲
2 −0.05 共0.08兲 −0.10 共−0.07兲 −0.20 共−0.13兲 −0.24 共−0.18兲
3 −0.06 共0.09兲 −0.04 共−0.01兲 −0.11 共−0.08兲 −0.06 共−0.08兲
4 −0.11 共0.11兲 −0.12 共0.00兲 −0.18 共−0.04兲 −0.11 共−0.02兲
Fig. 1. R2 between observed and predicted PHDI by using the Fig. 2. R2 between observed and predicted PHDI by using the
two proposed approaches for the months of January–March 共time two proposed approaches for the months of July–September 共time
horizon= 1 , 2 , 3 , 4 months ahead兲 horizon= 1 , 2 , 3 , 4 months ahead兲
Fig. 3. R2 between observed and predicted PHDI by using the Fig. 4. R2 between observed and predicted PHDI by using the
ANN model for the months of January–March 共time horizon ANN model for the months of July–September 共time horizon
= 1 , 2 , 3 , 4 months ahead, Direct approach兲 = 1 , 2 , 3 , 4 months ahead, Direct approach兲
Thus, an extension of the Direct model is proposed here in when climatic indexes are included in the model is significant.
order to include information deriving from large scale climatic Similar results 共here not shown for brevity兲 have been obtained
indexes and to assess whether considering the information of for the months of October–December.
NAO and EB indexes leads to improvements in forecasting PHDI Nonetheless, different results have been observed for the
values. A similar extension of the Iterative model has not been months of July–September as shown in Fig. 4.
implemented as it would require the use of other models to fore- For this period of the year, considering the climatic indexes as
cast future NAO and EB values. predictor variables does not improve significantly the predictive
Next, prediction of the PHDI共t + k兲 was carried out by select- capabilities of the model with respect to the case when only cur-
ing as predictor variables PHDI共t兲 and NAO共t兲 or PHDI共t兲 and rent value of Palmer index is taken into account.
EB共t兲. On the basis of the preliminary correlation analysis, both Indeed, although some small differences in R2 can be ob-
NAO and EB have been averaged over the previous 4 months. served, the R2 values obtained through the models having EB or
Further, in order to assess whether increases in R2 obtained NAO as predictor variable generally lie within the confidence
when NAO and/or EB are considered as exogenous variables are intervals. This was somehow expected, due to the limited corre-
due to the higher number of model parameters or to an effective lation observed between Palmer index and climatic indexes in
dependence between the Palmer values in Sicily and the climatic such months.
indexes, R2 confidence intervals have been computed through a Similar results, indicating no influence of climatic indexes
Monte Carlo procedure. More specifically, 100 NAO series and on forecasting PHDI, have been obtained for the months of April–
100 EB series uncorrelated with the observed Palmer series have June.
been generated and the corresponding 100 ANN predictions and Finally, Fig. 5 illustrates the observed and forecasted PHDI
the associated R2 values have been computed for the proposed time series at Lag 3 related to the months of January–March. The
models. Confidence intervals at 10% significance level have then plot confirms from a qualitative point of view the better predictive
been estimated by considering the lower and upper 5% sample capability obtained through the application of the model based on
quantiles of the R2 series. Thus, if R2 computed on observed the EB index with respect to the model based only on the PHDI
series lie above the upper confidence limits, the better predictive index or the one based on NAO index.
capabilities of the models are presumably due to an effective in- Fig. 6 illustrates the observed and forecasted PHDI time series
fluence of the NAO and EB indexes and not to the higher number at Lag 2 related to the months of July–September 共validation
of model parameters. period兲. From Fig. 6 the very similar predictive capability of all
The computed R2 corresponding to the ANNs for the months investigated models is confirmed.
of January–March considering different input variables are pre-
sented in Fig. 3. Also in Fig. 3, the values of R2 obtained through
application of ANNs having only PHDI共t兲 as the predictor vari- Conclusions
able, following the Direct approach, are also shown.
For each model, observed R2 values are reported as bars, Advances in understanding the influence of large-scale climatic
whereas the intervals between the lower and upper 5% quantiles patterns on the climate in a region can potentially contribute to a
of generated R2 values are indicated as black lines. better forecast of drought conditions. In the present paper an at-
From Fig. 3 the following considerations can be made. First, tempt is made to verify whether the use of the North Atalantic
including climatic indexes as predictor variables leads to an im- Oscillation and European Blocking indexes could improve fore-
provement of the predictive capability for the months of January– casting of Palmer PHDI index, with reference to the Sicily region.
March. This is particularly evident for the case when the EB First, the influence of NAO and EB indexes on Palmer index
index is taken into account and for increasing forecasting time series, computed on areal monthly precipitation from 1955 until
horizon. 1999 in Sicily, has been investigated by means of a correla-
Further, the R2 values obtained through the models with EB or tion analysis. Results indicate that NAO series are significantly
NAO as predictor variables always lie above the upper confidence anticorrelated with Palmer index series for winter months and
limit thus indicating that the improvement in prediction ability especially in the last decades. The EB series are significantly
Fig. 5. Observed and predicted PHDI varying the ANNs models for the months of January–March 共validation phase, Direct approach兲
correlated with Palmer index series both for winter and autumn casting model for the evaluation of the PHDI future values.
months and especially in the last decades. The comparison between the prediction for winter and autumn
The prediction of the PHDI index for up to 4 months ahead months 共for which an high correlation between Palmer index and
was, at first, carried out by adopting as predictor variable the climatic indexes has been shown兲 obtained by either including or
observed index in the previous months. not including the NAO and EB indexes within the model, indi-
Two forecasting schemes have been adopted to forecast at cates some improvements in terms of R2 when the NAO and
different time horizons. In the first scheme, named the Direct especially EB are considered.
approach, different ANNs have been developed for each forecast- A different model behavior has been observed for the spring
ing horizon. In the second scheme, named the Iterative approach, and summer predictions, for which no significant improvements
ANNs able to predict 1 month ahead have been considered, and in terms of model predictive capability due to the introduction of
forecasted values are then used as input to forecast at lag ⬎1. the climatic indexes as input variables have been observed.
The comparison between the two forecasting schemes has The overall conclusion of the paper is that NAO and EB in-
shown some improvements in terms of predictive capability dexes exert some influence on drought evolution in Sicily. Thus,
when the Iterative approach for the winter and autumn months information provided by such indexes can be usefully adopted to
is applied. No significant differences in predictive quality for improve the forecast of drought events in terms of PHDI values
the other months have been obtained varying the forecasting both in autumn and winter period.
approaches. Further research is needed to extend the Iterative model in
Therefore, the information provided by NAO and EB indexes order to include information deriving from large scale climatic
has been included, following the Direct approach, within a fore- indexes.
Fig. 6. Observed and predicted PHDI varying the ANNs models for the months of July–September months 共validation phase, Direct
approach兲