Science of the Total Environment 548–549 (2016) 110–121

Contents lists available at ScienceDirect

Science of the Total Environment

journal homepage: www.elsevier.com/locate/scitotenv

Airborne castanea pollen forecasting model for ecological and

allergological implementation
G. Astray a,b, M. Fernández-González c, F.J. Rodríguez-Rajo c, D. López b, J.C. Mejuto a
a
Physical Chemistry Department, Faculty of Science, University of Vigo, 32004 Ourense, Spain
b
Department of Geological Sciences, College of Arts and Sciences, Ohio University, 45701 Athens, USA
c
Department of Plant Biology and Soil Sciences, Faculty of Sciences, University of Vigo, 32004 Ourense, Spain

H I G H L I G H T S G R A P H I C A L A B S T R A C T

• It was detected an increasing trend of

total annual Castanea pollen concentra-
tions.
• Castanea pollen relationships with pre-
cipitation and temperature were inves-
tigated.
• A strong relationship between pollen
and precipitation, and temperature
were found.
• The implemented models are based on
artificial neural networks (ANN).
• The results show that ANNs are a valid
tool to predict Castanea pollen.

a r t i c l e i n f o a b s t r a c t

Article history: Castanea sativa Miller belongs to the natural vegetation of many European deciduous forests prompting impacts
Received 24 September 2015 in the forestry, ecology, allergological and chestnut food industry fields. The study of the Castanea flowering rep-
Received in revised form 5 December 2015 resents an important tool for evaluating the ecological conservation of North-Western Spain woodland and the
Accepted 7 January 2016
possible changes in the chestnut distribution due to recent climatic change. The Castanea pollen production
Available online 20 January 2016
and dispersal capacity may cause hypersensitivity reactions in the sensitive human population due to the rela-
Editor: D. Barcelo tionship between patients with chestnut pollen allergy and a potential cross reactivity risk with other pollens
or plant foods. In addition to Castanea pollen's importance as a pollinosis agent, its study is also essential in
Keywords: North-Western Spain due to the economic impact of the industry around the chestnut tree cultivation and its
Castanea pollen beekeeping interest. The aim of this research is to develop an Artificial Neural Networks for predict the Castanea
Artificial Neural Networks pollen concentration in the atmosphere of the North-West Spain area by means a 20 years data set.
Modelling It was detected an increasing trend of the total annual Castanea pollen concentrations in the atmosphere during
Time series analysis the study period. The Artificial Neural Networks (ANNs) implemented in this study show a great ability to predict
Castanea pollen concentration one, two and three days ahead. The model to predict the Castanea pollen concen-
tration one day ahead shows a high linear correlation coefficient of 0.784 (individual ANN) and 0.738 (multiple
ANN). The results obtained improved those obtained by the classical methodology used to predict the airborne
pollen concentrations such as time series analysis or other models based on the correlation of pollen levels
with meteorological variables.
© 2016 Elsevier B.V. All rights reserved.

http://dx.doi.org/10.1016/j.scitotenv.2016.01.035
0048-9697/© 2016 Elsevier B.V. All rights reserved.
 G. Astray et al. / Science of the Total Environment 548–549 (2016) 110–121
1. Introduction
Castanea sativa Miller belongs to the natural vegetation of many
European deciduous forests. During the summer months, the tree
flowers (Rodríguez-Rajo et al., 2005) rising the atmospheric pollen
levels as consequence of its mainly anemophilous behaviour (Jäger
and Litschaner, 1999). The Castanea pollen production is very high, in
the order of 1000 to 1500 grains per anther, which results in 10–12 bil-
lion grains per tree released during the pollen season (Jato et al., 2001;
Rudow and Conedera, 2001). Due to its small size the chestnut pollen is
easily transportable by the wind and clearly recognized as an example
of medium and long-range pollen transport (Frei, 1997; García-Mozo
et al., 2006; Rizzi Longo and Pizzulin Sauli, 2010). Its presence has
been verified in areas without any nearby chestnut populations, exceed-
ing a distance of 100 km and 700 m of altitude a.s.l. (Frei, 1997; Peeters
and Zoller, 1988). In North-Western Spain, the Castanea pollen abun-
dance and high dispersal capacity (Jato et al., 2001; Rodríguez-Rajo
et al., 2005) may cause hypersensitivity reactions in the sensitive
human population due to its moderate allergenic capacity (Jäger and
Litschaner, 1999; Cosmes Martín et al., 2005). Type I allergy to chestnut
pollen represents a major cause of pollinosis in sub-Mediterranean
areas (Hirschwehr et al., 1993) due to its major allergen Cas a 1 sensiti-
zation (Kos et al., 1993). The scientific interest in this taxon augmented
during recent years because available clinical data in the NW part of Ibe-
rian Peninsula shows an increase in skin reactivity from the 2–8% sensi-
tisation to Castanea pollen among pollinosis sufferers during the 1990's
(Arenas et al., 1996; Belmonte et al., 1998; Ferreiro et al., 2002) to 17% of
sensitivity in the 2000's (Aira et al., 2000). Moreover, recent findings
noted the relationship between patients with chestnut pollen allergy
and a potential cross reactivity risk with other pollens or plant foods
(Sánchez-Monge et al., 2006). The oral allergy syndrome is a distinctive
type of allergy to food resulting from direct contact between food and
the oral mucosa, normally affecting patients who are allergic to pollens
such as chestnut (Antico, 1996). In addition to Castanea pollen's impor-
tance as a pollinosis agent, its study is also essential in North-Western
Spain due to the economic impact of the industry around the chestnut
tree cultivation. Its fruit is collected, transformed and exported to differ-
ent countries for human consumption. The chestnut tree is also of bee-
keeping interest because it characterizes the most of the important
honeys of the area (Seijo et al., 1997). Finally, the study of the Castanea
flowering represents an important tool for evaluating the ecological
conservation of North-Western Spain woodland and the possible
changes in the chestnut distribution due to hotter spring–summer tem-
peratures in recent years (Jato et al., 2013).
During recent years forecasting airborne pollen concentrations is a
hot topic (Aznarte et al., 2007) because the pollen allergy is a very com-
mon disease in European and world population. Consequently one of
the Aerobiology targets is to obtain predictive models for the airborne
pollen concentrations (Sánchez-Mesa et al., 2002). Nowadays, there
are advanced techniques such as Artificial Neural Networks (ANN here-
after) or the Support Vector Machines (SVM) that have also been used
for forecasting air quality parameters (Csépe et al., 2014). ANNs are a
computational method formed by individual cells that perform compu-
tational calculations similar to the way the human brain works, learning
from training data (Sánchez-Mesa et al., 2002). Multi Layer Perceptron
(MLP) Neural Network is a particular case of ANN, the most successful
implementation, and it is applied in environmental science to develop
predictive models (Csépe et al., 2014). ANNs are used in very different
fields such as; i) Engineering, for example to study the solar energy po-
tential in Turkey (Sözen et al., 2005), or to predict the fracture parame-
ters of concrete (Ince, 2004), ii) in Computer Science for estimating the
effort required for developing an information system (Heiat, 2002), iii)
in Hydrology for predict the river flow forecast in reservoir manage-
ment (Baratti et al., 2003), or used in hybrid systems model for river
flow forecasting (Araujo et al., 2011; Toro et al., 2013), iv) in Medicine
for gene expression data analysis (Tan and Pan, 2005) v) in daily life
111
for bus arrival time prediction at bus stop with multiple routes (Yu
et al., 2011), vi) to assure the authenticity of food in an important
wine Spain's region (Gómez-Meire et al., 2014). In recent years, neural
network has been used successfully to forecast different bioaerosol par-
ticles such as grass pollen in the Southern part of the Iberian Peninsula
(Sánchez-Mesa et al., 2002), Betula pollen in different parts of Europe
such us in Spain (Castellano-Méndez et al., 2005) or in Poland (Puc,
2012), or even Alternaria and Pleospora spore airborne concentrations
in Central Italy (Tomassetti et al., 2013), etc.… In general, all ANN
models present better results than other kinds of models like linear re-
gressions, though the ANNs complexity is greater.
An artificial neural network can model complex and non-linear pro-
cesses through different layers (input layer, intermediate layers and
output layer) trained by back propagation algorithm (Csépe et al.,
2014), which is the most used algorithm to relate input variables to out-
put variables (Aznarte et al., 2007). The learning process of neural net-
works is based on the relationship change between the different
neurons in neural network. The parameter that defines this relationship
or importance value is called weight. The weight together with the bias
associated to each neuron, changes throughout the training process to
adjust the outputs of the neural network to the value of the training
cases, allowing the neural network learning during the operations in
the training phase (Venkatasubramanian et al., 2003). A number of
data points is reserved for a validation of the results obtained in the
training phase. In this way we can see if the training phase produced a
predictive model that is reliable.
The aim of this research is to develop an Artificial Neural Networks
for predict the airborne Castanea pollen concentration during one, two
and three days ahead in the North-West Spain area.
2. Material and methods
2.1. Study area
The study was carried out in Ourense, situated in a depression at
139 m. a.s.l., in Northwestern Spain (42°20′N and 7°52′W). The climate
of Ourense is oceanic, with a strong Mediterranean influence. During
the last 30 years, the region presents a mean annual temperature of
14.2 °C, a minimum average temperature of 8.2 °C, and a maximum av-
erage temperature of 20.2 °C (Jato et al., 2007). Annual rainfall is
794 mm, with very irregular distribution over the year; average sum-
mer rainfall is only 21.6 mm (Martínez Cortizas and Pérez-Alberti,
1999). In biogeographic terms, the study area belongs to the Euro-
Siberian Region, Atlantic European Province and Cantabrian-Atlantic
Subprovince (Rivas-Martínez et al., 2002). The characteristic vegetation
is formed by the Holco mollis–Quercetum pyrenaica association with the
presence of Q. pyrenaica Willd., Q. suber L. and Q. robur L., within the tree
stratum, and C. sativa as an accompanying species. Castanea cultivation
is widespread in this region, forming dense forests especially in areas
between 600 and 900 m above sea level.
2.2. Aerobiological survey
A Hirst-type LANZONI VPPS 2000 volumetric 7-day recording sam-
pler (Hirst, 1952) was used to collect the airborne pollen from 1993 to
2012. The sampler was situated on the roof of the Sciences Faculty (ap-
proximately 20 m above ground level). The sampler is calibrated to han-
dle a flow of 10 l of air per minute, thus matching the human breathing
rate. Pollen grains impact on a cylindrical drum covered by Melinex film
coated with a 2% silicon solution as trapping surface. The drum was
changed weekly, and the Melinex film coated was cut into seven pieces
and placed in separate glass slides (Galán et al., 2007). Daily values were
expressed as number of pollen grains per cubic metre of air. Pollen grain
identification was performed using a microscope equipped with a 40×/
0.95 lens. Pollen counts were conducted using the model proposed by
the R.E.A., consisting in four continuous longitudinal traverses along
112 G. Astray et al. / Science of the Total Environment 548–549 (2016) 110–121
the 24-h slide (Galán et al., 2007). The following Atmospheric Pollen
Season (APS) characteristics were analysed: first and last dates, length,
maximum pollen count and date recorded, and total pollen. The APS
was calculated as 95% of the annual total pollen, the first days, produc-
ing up to 2.5% of total production, were removed from the calculations
as well as the final days (from 97.5 to 100.0% of total production)
(Andersen, 1991).
2.3. Meteorological survey
The daily meteorological data used (maximum, mean, and mini-
mum temperatures; relative humidity; rainfall, solar radiation and
wind speed and direction) were supplied by automatic station of the
Spanish Meteorological Agency (AEMET) from a weather station located
in roof of Ourense Science Faculty (42° 50'N, 7°51'W) at 10 m of the vol-
umetric sampler.
2.4. Artificial neural networks
2.4.1. Neural network model training
The first step to develop a neural network model is to choose the
training and validation cases for the different neural networks that we
will be deployed. In this study, we used data collected between 1993
and 2012. The bulk of training cases are formed by data collected be-
tween 1993 and 2010. Then, to validate the different neural networks
models implemented, we have reserved the data collected in years
2011 and 2012.
Once we have chosen the different cases comprising each stage of
development of the neural network, we proceed to the development
of the various models based on trial and error technique. During the de-
velopment of these models, the neural network is able to learn from the
training cases and extrapolate this knowledge to previously reserved
cases (untrained cases). These cases, validation cases, will allow us to
see the true power of prediction of artificial neural networks.
Artificial Neural networks base their predictive power in the large
number of interconnections between each of the different neurons
that comprise the neural systems (input neurons, intermediate neurons
or output neurons). Each of these neurons is connected to the neurons
of the preceding layer and subsequent layer throughout an algorithm
that enables network learning based on the training cases. All training
and validation processes of the different neural networks have been im-
plemented in a HP ProLiant Server with an Intel Xeon processor with
RAM memory of 8 GB.
The training of the neural network starts with the introduction of the
information in the first layer of the neural network (input layer, formed
by input neurons). The information, constituted by the training cases is
entered as a vector (Eq. (1)) and it is propagated to the first intermedi-
ate layer by the propagation function (Eq. (2)) (Fig. 1).
x ¼ ðx1 ; x2 ; x3 ; …; xn Þ
X
M A time series is a sequence of data points measured at regular time
Sn ¼ wmn xm þ bn
m¼1
The propagation function (implemented in each intermediate and
output neuron) collects the information from the input case (training
values) and converts all this information into a single signal (single re-
sponse). The input values from input neurons (M) is processed with
the weight (wmn, value of the importance) of the connection linking
the intermediate neuron (n) with the previous neuron (m) from
which the information comes, and with the bias value (bn), associated
to the intermediate neuron n. The single data response is treated by
the activation function (implemented in each intermediate and output
neuron) to provide an output neuron signal (yn) (Gopalsamy, 2004).
This process is repeated for all neurons in intermediate (n) and output
layer (o) to generate a final predicted value in the output layer (yo)
(Fig. 1). Different activation functions are used in research but the
most usually are the sigmoidal function (Banakar and Azeem, 2008),
that we used in this study and in another study conducted by our re-
search group (Gonzalez-Temes et al., 2013), (Eq. (3)).
1
yo ¼ ð3Þ
1 þ e−So
We develop a predictive model based in artificial neural network to
calculate the Castanea pollen concentration in one ([Castanea + 1]), two
([Castanea + 2]), and three days ahead ([Castanea + 3]) from thirteen
variables; i) Julian day (Day), ii) precipitation (P), iii) humidity (H),
iv) maximum temperature (Tmax), v) minimum temperature (Tmin),
vi) average temperature (Tav), vii) solar radiation (S), viii) hours with-
out winds or calm (C), ix–xii) four wind directions; Northeast (W1),
Southeast (W2), Southwest (W3) and Northwest (W4), and finally,
xiii) pollen concentration at day zero ([Castanea]).
All developed models will be named, with a feature terminology, to
facilitate their identification. We normally used the following terminol-
ogy (Astray et al., 2013; Astray et al., 2014): Nin − (N1 − N2 − N3) −
Nout, Nin and Nout are the input and output layer, and N1, N2 and N3
are the neurons in the first, second and third intermediate layer (Fig. 1).
2.4.2. Identification of the best neural network
As we develop predictive models we perform the calculations of the
fits for each neural network model to guide the training process at all
time. To determine the power of adjustment various parameters are cal-
culated for training and validation phase for each implemented neuro-
nal network model. We calculated the Root Mean Square Error
(RMSE) (Equation 4) and the Average Percentage Deviations (APD)
(Eq. (5)) that they are most commonly used parameters in our depart-
ment in Ourense.
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
uX N  2
u
u ½Castaneapred −½Castaneareal
t i¼1
RMSE ¼ ð4Þ
N
N ½Castanea
X !
 pred −½Castaneareal 
 ½Castanea   100
i¼1 real
APD ¼ ð5Þ
N
The aim of this study is to implement prediction model based on Ar-
tificial Neural Networks, to predict the output variable. Choosing the
best neural network should be based on the lowest RMSE in validation
phase because these cases, as stated above, can help us to understand
how the network behaves with unknown cases.
ð1Þ 2.5. Time series analysis
ð2Þ intervals, for example, in our case daily. With the time series analysis
we tried to find an auto-correlation within the variable (using just a sin-
gle time series — univariate) or to find a cross-correlation between dif-
ferent measured variables (using several series simultaneously —
multivariate analysis) (Sahu et al., 2009). Auto-correlation function con-
sidered the relationship between time series data (is imperative that
not miss any data point in this series) and provides information about
the influence of events in the time series data (Sahu et al., 2009) using
the auto-correlation value ranging between − 1 and 1 over the time
lag. The auto-correlation slope has different behaviour depending on
how the variable influence behaves over time and if the variable is cyclic
or not, when the changes in the variable takes a long time to dissipate,
the slope of the autocorrelogram is low, while when the event has a
short influence, the slope decreases rapidly (Lee and Lee, 2000; Sahu
 G. Astray et al. / Science of the Total Environment 548–549 (2016) 110–121 113

Fig. 1. Diagram of a neural network topology 13-5-3 and operation procedure of an intermediate artificial neuron.

et al., 2009). In an autocorrelogram, the time that the correlation coeffi- correlation function shows relationship between two variables. The
cient takes to reduce to zero is the time that the effect of a change in the cross-correlation value ranging between − 1 and 1 over the time lag
variable last in the system and it is called the regulation time. The cross- that determines the transfer velocity of the studied system (influence

Fig. 2. A — Auto-correlation functions for Castanea pollen (continuous line), precipitation (dot line) and maximum temperature (dash line) and B — Magnified portion of the auto-
correlation functions from zero to 120 days.
114 G. Astray et al. / Science of the Total Environment 548–549 (2016) 110–121

Fig. 3. A — Cross-correlations functions for precipitation and Castanea pollen (line) and cross-correlation function for maximum temperature and Castanea pollen (dash-dot line), and B —
Magnified portion of the cross-correlation functions.

of events) (Sahu et al., 2009). In this study, auto-correlation and cross-
correlation functions were applied to study Castanea pollen concentra-
tion in Ourense using the time series data for Julian day, Castanea pollen
concentration, precipitation and maximum temperature. For auto-
correlation and cross-correlation analysis we used the computer pro-
gramme PAST (Hammer et al., 2001).
3. Results and discussion
Castanea pollen is found in the North-Western Spain atmosphere
during three months in the summer, and represents between 5% and
7% of total annual pollen. The pollination period is lengthy, with an av-
erage of 66 days, and a trend to increase during the last years. The chest-
nut Atmospheric Pollen Season (APS) begins around the first days of
June and ends at the middle of August. The highest concentrations are
normally attained in early July, with a maximum peak attained on July
8th 2011. There were also fluctuations in the total pollen values during
the years under study, with a minimum of 450 total pollen registered
during the APS in 1998, and a maximum of 3105 total pollen during
the 2012 pollen season. During the last decade of the past century a de-
cline in the Castanea plants population has been observed as conse-
quence to blight caused by Phytophtora cambivora and Phytophtora
cinnamomi. The recent economical increasing interest about its cultiva-
tion as well as the grafting stocks of resistant species to the fungus used,
C. crenata Siebold & Zucc. (Silva and Rigueiro, 1992), prompted an in-
creasing trend of the total annual Castanea pollen concentrations in
the atmosphere. Therefore, this wide range of variations in the Castanea
pollen concentrations detected during the last years associated to the
medium and long-range pollen transport ability of the chestnut pollen
make difficult the development of accurate prediction models
(Rodríguez-Rajo et al., 2005).
It should be noted that in previous studies (Iglesias-Otero, 2013), the
inclusion of Castanea pollen from previous days as input variables pollen
concentrations was studied, however the results don't improve the
models presented in this paper. This may be due to introducing more
input variables the system, including more noise in the system, and
due this the artificial neural is not able to correctly generalize.
3.1. Time series analysis of pollen concentration
Classical correlation models use uni-dimensional nonlinear equa-
tions and the variables with the greatest prediction capacity
(Rodríguez-Rajo et al., 2010). An auto-correlation analysis and cross-
correlation analysis were applied over a period 2006–2008. Due to the
large amount of data unavailable, it has been possible to study only
the period 2006–2008 and only for variables, Julian day, Castanea pollen,
precipitation and maximum temperature. Thus, the studied period was
of three years. In Fig. 2 and Fig. 3 we can see the auto-correlation and
cross-correlation analysis of these three variables. The variables
analysed in this section are those that should have more influence on
the prediction of Castanea pollen one, two and three days ahead.
These variables were chosen considering the results provided by other

Fig. 4. A — Cross-correlations functions for Castanea pollen vs. Castanea one day ahead (continuous line), Castanea vs. Castanea two days ahead (dot line) and cross-correlation function for
Castanea vs. Castanea three days ahead (dash line), and B — Magnified portion of the cross-correlation functions.
 G. Astray et al. / Science of the Total Environment 548–549 (2016) 110–121 115

studies reported in literature (Sánchez-Mesa et al., 2002; Csépe et al.,
2014), and taking into account that in order to perform time series anal-
ysis is necessary to have all data for the selected variable.
In Fig. 2A we can see the auto-correlation function for Castanea pol-
len, precipitation and maximum temperature. We can see in Fig. 2A that
the auto-correction for Castanea pollen decreases rapidly to 0.828 after
one day and reach zero value in approximately 35 days. We can see in the
same auto-correlation function that the distance between two Castanea
maximum peaks register a correlation value of 0.674. The position of these
peaks and their different size, suggests that the Castanea pollen is close to
the seasonality of 365 days and that the variability in the Castanea pollen con-
centration in these two years is significant (0.828 vs. 0.674). The differences in
time for the appearance of the peak auto-correlation are probably due to the
effect of all the other variables, such as precipitation and wind velocity. In
terms of precipitation auto-correlation function, as excepted, decreases
much faster than the Castanea pollen auto-correlation function going from
0.240 to zero in just 14 days. In this sense we can see as the precipitation
auto-correlation function does not show a marked seasonality as in the case
of pollen auto-correlation. Finally, we can see how the temperature auto-
correlation function decreases slowly, obviously the temperature is influ-
enced by the seasons, and it shows a perfect seasonal regularity (lag between
peaks 365 days). In Fig. 2B we can see a magnified section for the first days of
the auto-correlation functions and we can appreciate more clearly the differ-
ent slopes of each auto-correlation functions.

Fig. 5. Predicted values of individual ANN versus real values of Castanea pollen during the validation years 2011 and 2012 to one day ahead (A and B), two days ahead (C and D) and three
days ahead (E and F).
116 G. Astray et al. / Science of the Total Environment 548–549 (2016) 110–121

To identify the relationship between Castanea pollen and the other
two variables; precipitation and maximum temperature, different
cross-correlation functions were calculated (Fig. 3). The cross-
correlation function in Fig. 3 shows the cross-correlation coefficient vs.
lag time between two variables; i) Precipitation-Castanea pollen and
ii) Maximum temperature-Castanea pollen.
The cross-correlograms show that the correlation coefficient be-
tween precipitation-Castanea pollen has a maximum cross-correlation
values around 0.105 (p = 2.12 ∙ 10− 3) in the 240th Julian day, and a
minimum − 0.094 at lag 1 with p = 1.83 · 10− 3. Precipitation does
not show a great influence on Castanea pollen concentration but
shows a weak inverse correlation at zero lags suggesting that precipita-
tion reduces slightly the concentration of pollen in the air. Finally, cross-
correlation function for maximum temperature and Castanea pollen
show a positive cross-correlation at zero lag (0.347, p = 2.58 · 10−32)
and at 313 to 353 lags, and a high negative cross-correlation value in
the region of Julian day between 169 (−0.326 and p b 0.001) and 201
(−0.311 and p b 0.001). These values suggest that high pollen concen-
tration are related to high temperatures and the lag for the second max-
imum and the minimum show the seasonal yearly effect. The highest
concentration of pollen days occur when the temperature increases pro-
moting pollen explosions.

Fig. 6. Predicted values of multiple ANN versus real values of Castanea pollen during the validation years 2011 and 2012 to one day ahead (A and B), two days ahead (C and D) and three
days ahead (E and F).
 G. Astray et al. / Science of the Total Environment 548–549 (2016) 110–121 117

Table 1 As mentioned above, we have concentration data and climatological

Characteristics of the main pollen season (MPS) of Castanea (1993–2012): starting Julian measurements from 1993 to 2012. In theory we have a matrix
day (S), ending Julian day (E), length in days (L), average value (Av), maximum value
(Max), day of the maximum (DMax), total concentration in the main pollen season (Ts)
7.305 × 14 with a total of 102270 data, however, there are some
and annual concentration (T). empty registers due to lack of measurement of pollen concentration
or just a bug in the data meteorological station that, in a specific mo-
S E L (Av) (Max) DMax Ts T
ment did not record any of the variables. Initially, we developed neural
1993 167 223 57 24 78 189 1345 1392 networks with data available every day but this approach made the
1994 163 224 62 16 77 194 975 982
training phase difficult for different neural networks implemented as
1995 152 213 62 13 71 176 831 851
1996 162 232 71 12 67 195 886 890 the training was very prolonged in computer time. The neural network
1997 148 211 64 10 86 185 667 672 models did not converge and/or their fits were no satisfactory. Seeing all
1998 162 214 53 8 45 193 450 450 this problems another strategy were tried. We proceeded to use only
1999 166 207 42 25 181 191 1038 1042 data from pollen seasons, adding a safety margin, because the remain-
2000 164 220 57 13 66 177 635 637
2001 160 214 55 12 112 181 673 679
der of the year the pollen concentration is practically zero and allergic
2002 165 217 53 29 193 196 1558 1573 periods are not reported. In this study we used data collected between
2003 154 195 42 13 86 174 529 543 Julian day 120th and 270th.
2004 158 207 50 15 69 187 657 662 Our first option was to implement neural network models to predict
2005 162 209 48 33 318 188 1584 1636
individual outputs separately, that it is, pollen concentration at one, two
2006 150 243 94 19 138 192 1784 1792
2007 150 237 88 16 105 192 1336 1357 and three days head. However, we tried to develop other models to cal-
2008 161 245 85 27 212 197 2335 2349 culate simultaneously the different pollen concentrations (called multi-
2009 152 236 85 17 106 189 1418 1442 ple neural networks) similar to those conducted by Iglesias-Otero
2010 163 228 66 35 206 187 2291 2312 (Iglesias-Otero, 2013). In this article, we will implement individual
2011 140 242 103 24 362 180 2467 2495
models and joint models, and then compare the results to see which
2012 170 246 77 40 209 197 3105 3256
gives less errors.
Average 158 223 66 20 139 188 1328 1351 For this study we have implemented over one thousand two hun-
Maximum 170 246 103 40 362 197 3105 3256
dred neural network models with different input variables, different ar-
Minimum 140 195 42 8 45 174 450 450
chitectures and training cycles, with the aim to identify the best neural
network model.
To determine the most acceptable prediction window Castanea pol-
len we studied the cross-correlation functions for Castanea variables. In
the Fig. 4, the cross-correlation functions for this three pair of variables, 3.2.1. Training phase of ANNs
Castanea vs. Castanea one, two, and three days ahead. The cross- When all different neural networks were implemented we
correlation value decrease from 0.828 (p b 0.001) for one day ahead to proceeded to calculate the fit for each one. Table 2 shows the best imple-
0.657 (p b 0.001) for three days ahead. These results show that we mented neural network for individual and multiple neural network for
can predict a Castanea Pollen Concentration with an accuracy of 65.7% one, two, and three days ahead for training phase (1993–2010). We
three days ahead if we know the pollen concentration for day one. The can see, in Table 2, the linear fit coefficient (R2T) and the Root Mean
accuracy four days ahead fall to 60.9%. We assume that a prediction win- Square Error (RMSET). The individual neural networks present a good
dow with an accuracy of 65.0% is the minimum permissible. R2 coefficient, between 0.851 for one day ahead and 0.851 for three
days ahead. In terms of RMSE the values vary between 8.026 pollen/
m3 for one day ahead to 11.877 pollen/m3 for three days ahead. These
3.2. Artificial Neural Networks neural network present different topologies, and different training cy-
cles, with the exception of the two and three days ahead that's present
Develop an artificial neural networks for predicted pollen concentra- the same training cycles. In the cases of multiple neural networks, we
tion requires a great implementation of networks, using trial and error can see that the neural network to predict the pollen concentration to
method, to obtain a group of Neural Network with good fits to deter- one, two and three days ahead is the same, namely, a neural network
mine with the best accuracy the pollen concentration one, two, and with thirteen neurons in the input layer, twenty-eight neurons in the in-
three day ahead. termediate layer and three neurons in the output layer.

Table 2
Fits for training phase of best neural networks for each output (best fits in grey shadow). Topology is to the internal structure of neural network (Architecture), Cycles are the number of
training cycles for this neural network, R2T is the square correlation coefficient and RMSET is the root mean square error for training phase (pollen/m3). The subscripts correspond to the
prediction window one, two and three days ahead.

Individual neural networks

Topology Cycles R2T + 1 R2T + 2 R2T + 3 RMSET + 1 RMSET + 2 RMSET + 3
13–24–1 6,400 0.851 8.026
13–25–1 25,600 0.859 9.640
13–28–1 25,600 0.851 11.877

Multiple neural networks

2
Topology Cycles R T+1 R2T + 2 R2T + 3 RMSET + 1 RMSET + 2 RMSET + 3
13–28–3 128,000 0.812 0.798 0.807 9.169 9.536 9.316
13–28–3 128,000 0.812 0.798 0.807 9.169 9.536 9.316
13–28–3 128,000 0.812 0.798 0.807 9.169 9.536 9.316
118 G. Astray et al. / Science of the Total Environment 548–549 (2016) 110–121

Table 3
Fits for validation phase (average for 2011 and 2012) of best neural networks for each output (best fits in grey shadow). Topology is to the internal structure of neural network (architec-
ture), Cycles are the number of training cycles for this neural network, R2V is the square correlation and RMSEV is the root mean square error for validation phase (pollen/m3). The sub-
scripts correspond to the prediction window one, two and three days ahead.

Individual neural networks

Topology Cycles R2V + 1 R2V + 2 R2V + 3 RMSEV + 1 RMSEV + 2 RMSEV + 3
13–28–1 400 0.784 22.596
13–7–1 25,600 0.584 32.393
13–19–1 200 0.489 34.796

Multiple neural networks

Topology Cycles R2V + 1 R2V + 2 R2V + 3 RMSEV + 1 RMSEV + 2 RMSEV + 3
13–26–3 1,000 0.738 0.453 0.367 26.154 33.472 34.773
13–4–3 128,000 0.659 0.590 0.516 30.810 33.383 34.381
13–3–3 3,500 0.621 0.561 0.570 26.591 28.948 30.270

3.2.2. Validation phase of ANNs
The results obtained for the training phase, can be said to be good,
both for a day, two or three days ahead, however, the main purpose of
this study is to develop a model that is able to determine the concentra-
tion of Castanea pollen for cases not previously analysed, that is, for the
validation cases (2011 and 2012). We proceeded in the same manner
and we identified the best neural network for validation cases
(Table 3). In this Table, we can see the fits for validation phase have,
as expected, worse fits than the best neural networks implemented in
the training phase (Table 2).
We can see that the networks offer better results for the validation
phase differ clearly the best result available for the training phase. As ex-
pected, the adjustments for the different days differ. The best results are
obtained for ANN+1 obtaining good results in terms of R2, 0.784 for in-
dividual ANN, and 0.738 for multiple ANN. As we increase the prediction
window, the adjustments down to reach a coefficient of determination
of 0.489 and 0.570 for individual and multiple ANN+3, respectively.
In general, the individual networks show best fits than multiple net-
works. A clear case is the ANN one day ahead, where the individual net-
works presents better fits in terms of R2 (0.784 vs. 0.738). In cases for
ANNs two days ahead, we can confirm that they have similar power pre-
diction in terms of determination coefficient (0.584 vs. 0.590), in terms
of root mean square error the individual ANN+2 presents better fits
than multiple ANN+ 2 (32.393 vs. 33.383). Nevertheless, there is a
significant change in Castanea pollen prediction three days ahead. In
this case, individual ANN+3 presents a R2 of 0.489, that's significantly
less than the R2 for multiple ANN+3, 0.570, and they present a large dif-
ference in terms of RMSE (34.796 vs. 30.270).
If we consider separately the behaviour for two-year validation
(Table 4), we can see the adjustments for the different implemented
models differ substantially depending of validation year. Castanea
pollen concentration depends from the year because the different
years present different start and final dates (atmospheric pollen
season), influenced by meteorological variables often with great
variability, this explains the large difference between the fittings
of the years 2011 and 2012. The fittings have the highest predictive
power for one day ahead, on average to one day have a correlation
coefficient of 0.784 (0.873 in 2011 and 0.714 in 2012) with an aver-
age RMSEV of 22.596 pollen/m3 (17.387 in 2011 and 26.553 in 2012)
and, as can be seen in Fig. 5A–B, the fits for validation year 2011 are
good compared to the validation year 2012. Fig. 5C, D, E and F
present the fits for ANN+ 2 and ANN+ 3 and present a underperform
fit due to the increase of the prediction window. Note that the
maximum values are underestimated when we expand the time
window.
If we consider the behaviour of multiple ANN we can see the ad-
justments for the different implemented models differ substantially
compared with the individual ANN (Fig. 5 and Fig. 6). This behaviour

Table 4
Fits for validation phase (2011 and 2012) of best neural networks for each output (best fits in grey shadow). Topology is to the internal structure of neural network (architecture), Cycles
are the number of training cycles for this neural network, R2V is the square correlation and RMSEV is the root mean square error for validation phase (pollen/m3). The subscripts correspond
to the prediction window one, two and three days ahead.

Individual neural networks

2011 2012
Topology Cycles R2V + 1 R2V + 2 R 2
V+3 R2V + 1 R2V + 2 R2V + 3
13–28–1 400 0.873 0.714
13–7–1 25,600 0.603 0.571
13–19–1 200 0.406 0.570

Multiple neural networks

2011 2012
Topology Cycles R2V + 1 R2V + 2 R2V + 3 R2V + 1 R2V + 2 R2V + 3
13–26–3 1,000 0.760 0.434 0.370 0.719 0.513 0.441
13–4–3 128,000 0.573 0.507 0.425 0.739 0.666 0.599
13–3–3 3,500 0.530 0.474 0.518 0.731 0.660 0.632
 G. Astray et al. / Science of the Total Environment 548–549 (2016) 110–121 119

Table 5
Importance (%) for input variable i) Julian day (Day), ii) precipitation (P), iii) humidity (H), iv) maximum temperature (Tmax), v) minimum temperature (Tmin), vi) average temperature
(Tav), vii) solar radiation (S), viii) hours without winds or calm (C), ix-xii) four wind directions; Northeast (W1), Southeast (W2), Southwest (W3) and Northwest (W4), xiii) Castanea pol-
len concentration of actually day ([Castanea]) for the best ANN implemented to predict Castanea pollen concentration one ([Castanea +1]), two ([Castanea +2]), and three days ahead
([Castanea +3]).

Variable Individual networks Multiple networks

[Castanea+1] [Castanea +2] [Castanea +3] [Castanea +1] [Castanea +2] [Castanea +3]

Day 14.669 14.177 25.640 12.431 26.894 30.181

P 2.660 6.452 1.601 6.059 5.895 5.120
H 4.790 8.408 2.547 6.431 7.819 5.048
Tmax 5.120 5.966 3.111 5.574 3.071 7.528
Tmin 5.024 3.945 11.627 7.156 4.229 3.311
Tav 3.731 4.092 4.789 3.477 4.534 9.649
S 11.338 8.182 3.799 5.495 3.652 0.848
C 8.419 9.122 2.686 7.656 7.919 7.498
W1 9.853 9.424 3.108 9.729 5.359 10.712
W2 3.726 8.108 0.393 3.798 5.210 6.213
W3 3.140 7.831 2.313 4.860 5.980 3.819
W4 10.576 4.433 9.702 6.822 5.541 2.581
[Castanea] 16.954 9.860 28.682 20.512 13.897 7.491

may be due to that the better prediction networks are, generally, and
based on our experience, ANNs with a single output. For predictions
one day ahead the results are worse than in the case of individual
networks (Table 4), in the case of two day prediction, we can say
that the adjustments are similar, in average terms, nevertheless for
prediction three days ahead multiple network presents better results
than the individual network (0.489 vs. 0.570, in terms of R2 , and
34.796 vs. 30.270 in terms of RMSE, Table 3). The possible reason to
this is the less influence for variables Julian day, and pollen concen-
tration. The improved for adjustments three days ahead is clearly
visible in the Fig. 6E–F.
It is visible that predicted values for the Castanea peaks are
underestimated. ANN are able to model behaviour between ranges of
values used in training phase, but for unknown conditions the ANN
model presents a poor extrapolation (Kordon, 2010). The prediction is
underestimated due to the nature of the data used to develop the
model. The pollen concentration data for validation phase are higher
than for training years (Table 1). In fact, inside the twenty highest values
of data collected, twelve of them are validation data. For this reason,
when the model try to predict pollen concentration for anomalous
years (with peaks higher than training peaks), the model is not able to
predict with accuracy these peaks.
As discussed above, we have been used thirteen input variables for
all neural networks developed because they can tell us that the most im-
portant variables to predict each Castanea pollen concentrations. To
achieve this, we must consider the weights between input layer to in-
termediate layer, and from them, to determine the importance (in
term of percentage) for each variable (Table 5).
The importance of variables vary between individual ANN, Fig. 7
to the left, and multiple ANN, Fig. 7 to the right, besides this, the im-
portance change with the time window prediction for Castanea pol-
len concentration. This variation of the importance on input
variables can be clearly seen in when we represent it in a scatter
plot, in Figures below we can see that the most important variables
to predict the Castanea pollen concentration, one, two and three
days ahead, are Julian day and Castanea pollen. The rest of input var-
iables have an importance between 12 and 0.39% (Table 5), besides
this, we can see that the Julian day is more important to predict
Castanea pollen concentration in multiple ANN than in individual
ANN (Table 5 and Fig. 7).
Sánchez-Mesa et al. (2002) developed an ANN model to predict
daily grass pollen concentration one day ahead, in the southern part of
the Iberian Peninsula, with a certain accuracy degree using previous
meteorological parameters and pollen variables. They can observe that
the best results were obtained with artificial neural networks instead
of other models such as linear regressions. In this ANN model, we can
observe good accuracy around 90% (Sánchez-Mesa et al., 2002). The
most important variable for the model developed is the pollen values
of previous days (Sánchez-Mesa et al., 2002). The same behaviour is ob-
served in our model to predict the Castanea pollen concentration one

Fig. 7. Variation of importance for each input variable for Castanea pollen prediction for one, two, and three days ahead by individual ANN (to the left) and multiple ANN (to the right).
120 G. Astray et al. / Science of the Total Environment 548–549 (2016) 110–121
day ahead. When the time windows decrease, the importance value de-
creased, generally. Iglesias-Otero et al. (2015) develop three different
neural network models to predict Plantago pollen in northwestern
part of Spain, with one, two and three days ahead. The ANN models pre-
sented good correlation coefficients for one day ahead (0.669) and de-
creased gradually for two (0.573), and three days ahead (0.464). This
demonstrated that pollen prediction suffers a decrease in prediction ac-
curacy when the prediction window is increased (Iglesias-Otero et al.,
2015). The most important variables to predict pollen concentration,
one day ahead, were Julian day and pollen concentration, while to pre-
dict pollen concentration two days ahead were Jualian day and rainfall
(Iglesias-Otero et al., 2015). The same behaviour is observed in our
model to predict the Castanea pollen concentration one day ahead. In
our ANN model to predict concentration two days ahead we can observe
a decrease in the importance of pollen concentration, together with an
increased for precipitation. Discriminant analysis, and neural networks
models, to forecast the severity of the Poaceae pollen season in Cordoba
(Spain), were used by Sánchez Mesa et al., (2005). Once again, the neu-
ral model implemented by the authors improves the adjustment obtain-
ed for validation phase using discriminant analysis to predict Poaceae
pollen (50% percentage of accuracy versus 16.70% obtained by discrim-
inant analysis) (Sánchez Mesa et al., 2005). Grinn-Gofroń and Strzelczak
studied the non-linearity and non-normality of Alternaria pollen spores
and the relationships with meteorological data in Szczecin (Poland)
(Grinn-Gofron and Strzelczak, 2008). In this work daily mean, mini-
mum and maximum temperature, and dew point temperature were
the most strongly associated variables with the concentration of
Alternaria spores in Spearman's rank correlation, meanwhile for the de-
veloped ANN, only the dew point has big influence (Grinn-Gofron and
Strzelczak, 2008). Puc developed an ANN model to predict airborne
Betula pollen concentrations from meteorological conditions in Szczecin
(Poland), and it shows a good correlation coefficient, R, upper 0.85 (Puc,
2012). Sensitivity analysis indicated that the most important variables
are maximum temperature and humidity (Puc, 2012). It seems clear
that the measurement location, and the type of particle sampled, pres-
ent a strongly influence in the importance of the variables used for
prediction.
4. Conclusions
The results obtained in this work are important for the forestry, ecol-
ogy, allergological and chestnut food industry. It was detected an in-
creasing trend of the total annual Castanea pollen concentrations in
the atmosphere during the study period as consequence of the recent
economical increasing interest about its cultivation and the grafting
stocks of resistant species to fungus.
In this study we analysed the combination of time series analysis and
artificial neural networks to provide useful information about the rela-
tionship between Castanea pollen concentration with Julian day, precip-
itation and maximum temperature. The time series reported that the
most influential variable is the pollen concentrations (0.82 to 0.66).
Also, it allows us to determine the influence of precipitation and maxi-
mum temperature in the daily evolution of pollen. The model to predict
the Castanea pollen concentration one day ahead shows a high linear
correlation coefficient of 0.784 (individual ANN) and 0.738 (multiple
ANN). The results obtained improved those obtained by the classical
methodology used to predict the airborne pollen concentrations such
as time series analysis or other models based on the correlation of pollen
levels with meteorological variables.
Taking into account the different results obtained for the different
models implemented we can say that artificial neural networks have
proved to be a valid tool to predict Castanea pollen concentration for
one day ahead. Also they are a valid tool to predict Castanea pollen
two, and three days ahead, but in these cases the adjustments are
slightly worse. The wide variations in the Castanea pollen concentra-
tions as well as its medium and long-range pollen transport ability
make it difficult to apply prediction models. However, the three dif-
ferent models obtained can be used, is recommended to optimize the
model with the addition of new experimental data of subsequent
years.
Acknowledgments
Astray G. thanks Xunta de Galicia, Consellería de Cultura, Educación
e Ordenación Universitaria, for the Postdoctoral grant (Plan I2C),
P.P.0000 421S 140.08.
References
Aira, M.J., Ferreiro, M., Iglesias, I., Jato, V., Marcos, C., Varela, S., et al., 2000. Aeropalinología
de cuatro ciudades de Galicia y su incidencia sobre la sintomatología alérgica
estacional. XIII Simposio de Palinología, A.P.L.E.. Cartagena, España.
Andersen, T.B., 1991. A model to predict the beginning of the pollen season. Grana 30,
269–275.
Antico, A., 1996. Oral allergy syndrome induced by chestnut (Castanea sativa). Ann. Aller-
gy Asthma Immunol. 76, 37–40.
Araujo, P., Astray, G., Ferrerio-Lage, J.A., Mejuto, J.C., Rodriguez-Suarez, J.A., Soto, B., 2011.
Multilayer perceptron neural network for flow prediction. J. Environ. Monit. 13,
35–41.
Arenas, L., González, C., Tabares, J.M., Iglesias, I., Méndez, J., Jato, V., 1996. Sensibilización
cutánea a pólenes en pacientes afectos de rinoconjuntivitis-asma en la población de
Ourense en el año 1994–95. Proceedings 1st European Symposium on Aerobiology.
Santiago de Compostela, España.
Astray, G., González-Temes, M., López, D.L., Mejuto, J.C., 2014. Density, viscosity and re-
fractive index prediction of binary and ternary mixtures systems of ionic liquid.
Mediterr. J. Chem. 3, 972–986.
Astray, G., Gálvez, J.F., Mejuto, J.C., Moldes, O.A., Montoya, I., 2013. Esters flash point pre-
diction using artificial neural networks. J. Comput. Chem. 34, 355–359.
Aznarte, J.L.M., Benítez Sánchez, J.M., Lugilde, D.N., de Linares Fernández, C., de la Guardia,
C.D., Sánchez, F.A., 2007. Forecasting airborne pollen concentration time series with
neural and neuro-fuzzy models. Expert Syst. Appl. 32, 1218–1225.
Banakar, A., Azeem, M.F., 2008. Artificial wavelet neural network and its application in
neuro-fuzzy models. Applied Soft Computing Journal 8, 1463–1485.
Baratti, R., Cannas, B., Fanni, A., Pintus, M., Sechi, G.M., Toreno, N., 2003. River flow forecast
for reservoir management through neural networks. Neurocomputing 55, 421–437.
Belmonte, J., Roure, J.M., March, X., 1998. Aerobiology of Vigo, North-Western Spain: at-
mospheric pollen spectrum and annual dynamics of the most important taxa, and
their clinical importance for allergy. Aerobiologia 14, 155–163.
Castellano-Méndez, M., Aira, M.J., Iglesias, I., Jato, V., González-Manteiga, W., 2005. Artifi-
cial neural networks as a useful tool to predict the risk level of Betula pollen in the air.
Int. J. Biometeorol. 49, 310–316.
Cosmes Martín, P.M., Moreno Ancillo, A., Domínguez Noche, C., Gutiérrez Vivas, A.,
Belmonte Soler, J., Roure Nolla, J.M., 2005. Sensitization to Castanea sativa pollen
and pollinosis in northern Extremadura (Spain). Allergol. Immunopathol. 33,
145–150.
Csépe, Z., Makra, L., Voukantsis, D., Matyasovszky, I., Tusnády, G., Karatzas, K., et al., 2014.
Predicting daily ragweed pollen concentrations using Computational Intelligence
techniques over two heavily polluted areas in Europe. Sci. Total Environ. 476–477,
542–552.
Ferreiro, M., Dopazo, A., Aira, M.J., 2002. Incidence of pollinosis in the city of A Coruña:
correlation with aerobiological data. J. Investig. Allergol. Clin. Immunol. 12, 124–129.
Frei, T., 1997. Pollen distribution at high elevation in Switzerland: evidence for medium
range transport. Grana 36, 34–38.
Galán, C., Cariñanos, P., Alcázar, P., Domínguez, E., 2007. Spanish Aerobiology Network:
Management and Quality Manual: University of Córdoba.
García-Mozo, H., Pérez-Badia, R., Fernández-González, F., Galán, C., 2006. Airborne pollen
sampling in Toledo, central Spain. Aerobiologia 22, 55–66.
Gómez-Meire, S., Campos, C., Falqué, E., Díaz, F., Fdez-Riverola, F., 2014. Assuring the au-
thenticity of northwest Spain white wine varieties using machine learning tech-
niques. Food Res. Int. 60, 230–240.
Gonzalez-Temes, M., Astray, G., Morales, J., Mejuto, J.C., Astray, G., 2013. Prediction of the
penetration of drugs by artificial neural networks. Iberian Conference on Information
Systems and Technologies, CISTI.
Gopalsamy, K., 2004. Stability of artificial neural networks with impulses. Appl. Math.
Comput. 154, 783–813.
Grinn-Gofron, A., Strzelczak, A., 2008. Artificial neural network models of relationships
between Alternaria spores and meteorological factors in Szczecin (Poland). Int.
J. Biometeorol. 52, 859–868.
Hammer, Ø., Harper, D.A.T., Ryan, P.D., 2001. Past: Paleontological statistics software
package for education and data analysis. Palaeontol Electron 4, XIX–XXX.
Heiat, A., 2002. Comparison of artificial neural network and regression models for esti-
mating software development effort. Inf. Softw. Technol. 44, 911–922.
Hirschwehr, R., Jager, S., Horak, F., Ferreira, F., Valenta, R., Ebner, C., et al., 1993. Allergenss
from birch pollen and pollen of the European chestnut share common epitopes. Clin.
Exp. Allergy 23, 755–761.
Hirst, J.M., 1952. An automatic volumetric spore trap. Annals of Applied Biology Volume
39 (2), 257–265.
 G. Astray et al. / Science of the Total Environment 548–549 (2016) 110–121
Iglesias-Otero, M.A., 2013. Bioinformática aplicada á aerobioloxía. Thesis, Universidad de
Vigo, Ph. D.
Iglesias-Otero, M.A., Fernández-González, M., Rodríguez-Caride, D., Astray, G., Mejuto, J.C.,
Rodríguez-Rajo, F.J., 2015. A model to forecast the risk periods of Plantago pollen al-
lergy by using the ANN methodology. Aerobiologia 31, 201–211.
Ince, R., 2004. Prediction of fracture parameters of concrete by Artificial Neural Networks.
Eng. Fract. Mech. 71, 2143–2159.
Jäger, S., Litschaner, R., 1999. Aerobiological survey on Castanea pollen in Austria. Pro-
ceedings Int. Symp. Castanea sativa Allergies and Plant Diseases, Lugano, Suiza.
Jato, M.V., Rodríguez-Rajo, F.J., Aira, M.J., Tedeschini, E., Frenguelli, G., 2013. Differences in
atmospheric trees pollen seasons in winter, spring and summer in two European geo-
graphic areas, Spain and Italy. Aerobiologia 29, 263–278.
Jato, V., Aira, M.J., Dopazo, A., Iglesias, M.I., Méndez, J., Rodríguez-Rajo, F.J., 2001. Aerobiol-
ogy of Castanea pollen in Galicia. Aerobiologia 17, 233–240.
Jato, V., Rodríguez-Rajo, F.J., Aira, M.J., 2007. Use of phenological and pollen-production
data for interpreting atmospheric birch pollen curves. Annals of Agricultural and En-
vironmental Medicine 14, 271–280.
Kordon, A.K., 2010. Applying Computational Intelligence: How to Create Value. Springer.
Kos, T., Hoffmann-Sommergruber, K., Ferreira, F., Hirschwehr, R., Ahorn, H., Horak, F., et
al., 1993. Purification, characterization and N-terminal amino acid sequence of a
new major allergen from European chestnut pollen — Cas s 1. Biochem. Biophys.
Res. Commun. 196, 1086–1092.
Lee, J., Lee, K., 2000. Use of hydrologic time series data for identification of recharge mech-
anism in a fractured bedrock aquifer system. J. Hydrol. 229, 190–201.
Martínez Cortizas, A., Pérez-Alberti, A., 1999. Atlas Climático de Galicia Spain. Xunta de
Galicia.
Peeters, A.G., Zoller, H., 1988. Long range transport of Castanea sativa pollen. Grana 27,
203–207.
Puc, M., 2012. Artificial neural network model of the relationship between Betula pollen
and meteorological factors in Szczecin (Poland). Int. J. Biometeorol. 56, 395–401.
Rivas-Martínez, S., Fernández González, F., Loidi Arregui, J., Lousa, M., Penas, M.A., 2002.
Syntaxonomical checklist of vascular plant communities of Spain and Portugal to as-
sociation level. Itinera Geobotanica 14, 5–341.
Rizzi Longo, L., Pizzulin Sa uli, M., 2010. Flowering phenology and airborne pollen occur-
rence of Corylus and Castanea in Trieste (Italy), 1991–2004. Acta Bot. Croat. 69,
199–214.
Rodríguez-Rajo, F.J., Astray, G., Ferreiro-Lage, J.A., Aira, M.J., Jato-Rodriguez, M.V., Mejuto,
J.C., 2010. Evaluation of atmospheric Poaceae pollen concentration using a neural net-
work applied to a coastal Atlantic climate region. Neural Netw. 23, 419–425.
121
Rodríguez-Rajo, F.J., Jato, V., Aira, M.J., 2005. Relationship between meteorology and
Castanea airborne pollen. Belg. J. Bot. 138, 129–140.
Rudow, A., Conedera, M., 2001. Blüte und Sortenerkennung beider Edelkastanie (Castanea
sativa Mill.) auf der Alpensüdseite der Schweiz. Bot. Helv. 111, 1–23.
Sahu, P., López, D.L., Stoertz, M.W., 2009. Using time series analysis of coal mine
hydrographs to estimate mine storage, retention time, and mine-pool interconnec-
tion. Mine Water Environ. 28, 194–205.
Sánchez Mesa, J.A., Galán, C., Hervás, C., 2005. The use of discriminant analysis and neural
networks to forecast the severity of the Poaceae pollen season in a region with a typ-
ical Mediterranean climate. Int. J. Biometeorol. 49, 355–362.
Sánchez-Mesa, J.A., Galan, C., Martínez-Heras, J.A., Hervás-Martínez, C., 2002. The use of a
neural network to forecast daily grass pollen concentration in a Mediterranean re-
gion: the southern part of the Iberian Peninsula. Clin. Exp. Allergy 32, 1606–1612.
Sánchez-Monge, R., Blanco, C., López-Torrejón, G., Cumplido, J., Recas, M., Figueroa, J., et
al., 2006. Differential allergen sensitization patterns in chestnut allergy with or with-
out associated latex-fruit syndrome. J. Allergy Clin. Immunol. 118, 705–710.
Seijo, C., Jato, V., Iglesias, I., Aira, M.J., 1997. Palyological characterization of sweet chestnut
honeys from Galicia (NW Spain). Acta Botanica Malacitana 22, 73–81.
Silva, F.J., Rigueiro, A., 1992. Guía das árbores e bosques de Galicia. España, Galaxia.
Sözen, A., Arcaklioglu, E., Özalp, M., Çaglar, N., 2005. Forecasting based on neural network
approach of solar potential in Turkey. Renew. Energy 30, 1075–1090.
Tan, A.H., Pan, H., 2005. Predictive neural networks for gene expression data analysis.
Neural Netw. 18, 297–306.
Tomassetti, B., Lombardi, A., Cerasani, E., Di Sabatino, A., Pace, L., Ammazzalorso, D., et al.,
2013. Mapping of alternaria and pleospora concentrations in central Italy using me-
teorological forecast and neural network estimator. Aerobiologia 29, 55–70.
Toro, C.H.F., Gómez Meire, S., Gálvez, J.F., Fdez-Riverola, F., 2013. A hybrid artificial intel-
ligence model for river flow forecasting. Applied Soft Computing Journal 13,
3449–3458.
Venkatasubramanian, V., Rengaswamy, R., Kavuri, S.N., Yin, K., 2003. A review of process
fault detection and diagnosis part III: process history based methods. Comput. Chem.
Eng. 27, 327–346.
Yu, B., Lam, W.H.K., Tam, M.L., 2011. Bus arrival time prediction at bus stop with multiple
routes. Transportation Research Part C: Emerging Technologies 19, 1157–1170.