Int J Biometeorol

DOI 10.1007/s00484-016-1215-y

ORIGINAL PAPER

Modeling pollen time series using seasonal-trend

decomposition procedure based on LOESS smoothing
Jesús Rojo 1 & Rosario Rivero 1 & Jorge Romero-Morte 1 & Federico Fernández-González 1 &
Rosa Pérez-Badia 1

Received: 15 February 2016 / Revised: 12 July 2016 / Accepted: 13 July 2016

# ISB 2016

Abstract Analysis of airborne pollen concentrations provides identified more accurately than by analysis of the original
valuable information on plant phenology and is thus a useful non-decomposed data series, and for this reason, this proce-
tool in agriculture—for predicting harvests in crops such as dure has proved to be a suitable technique for analyzing the
the olive and for deciding when to apply phytosanitary treat- main environmental factors influencing airborne pollen
ments—as well as in medicine and the environmental sci- concentrations.
ences. Variations in airborne pollen concentrations, moreover,
are indicators of changing plant life cycles. By modeling pol- Keywords Seasonality . Daily variations . Grass pollen .
len time series, we can not only identify the variables influenc- Flowering phenology . Meteorological variables
ing pollen levels but also predict future pollen concentrations.
In this study, airborne pollen time series were modeled using a
seasonal-trend decomposition procedure based on LOcally Introduction
wEighted Scatterplot Smoothing (LOESS) smoothing (STL).
The data series—daily Poaceae pollen concentrations over the Airborne pollen levels are a key indicator of flowering in
period 2006–2014—was broken up into seasonal and residual wind-pollinated plants. Monitoring of pollen concentrations
(stochastic) components. The seasonal component was com- provides valuable information on plant phenology and pro-
pared with data on Poaceae flowering phenology obtained by ductivity and is thus a useful tool in botanical, environmental,
field sampling. Residuals were fitted to a model generated and agricultural research (García-Mozo 2011; Fernández-
from daily temperature and rainfall values, and daily pollen Llamazares et al. 2014). Monitoring is also of public health
concentrations, using partial least squares regression (PLSR). interest, since pollen is the leading cause of allergies
This method was then applied to predict daily pollen concen- (D’Amato et al. 2007), and the number of allergy-sufferers
trations for 2014 (independent validation data) using results has considerably increased over the last few decades
for the seasonal component of the time series and estimates of (Bousquet et al. 2011). Modeling airborne pollen concentra-
the residual component for the period 2006–2013. Correlation tions helps to identify the variables or factors influencing var-
between predicted and observed values was r = 0.79 (correla- iations over time (Aboulaich et al. 2013; Rojo et al. 2015) and
tion coefficient) for the pre-peak period (i.e., the period prior also to predict future pollen levels (Rodríguez-Rajo et al.
to the peak pollen concentration) and r = 0.63 for the post- 2006; Silva-Palacios et al. 2015).
peak period. Separate analysis of each of the components of The flowering sequence of anemophilous species deter-
the pollen data series enables the sources of variability to be mines the airborne pollen concentrations (Estrella et al.
2006), and the pollen spectrum is related to the geographical
characteristics, flora, and vegetation of a given site (Kasprzyk
* Jesús Rojo 2006; Rojo et al. 2015). Central Spain is characterized by a
[Link]@[Link] great diversity and abundance of plant communities and spe-
cies, which results in high pollen amounts and an important
1
Institute of Environmental Sciences, University of Castilla-La pollen diversity (Pérez-Badia et al. 2010). However, the daily
Mancha, 45071 Toledo, Spain pollen concentrations are also governed by the meteorological
 Int J Biometeorol

conditions influencing local pollen emission and dispersal sources of pollen emission (Frenguelli et al. 2010; León-
(Sánchez-Mesa et al. 2005). Temperature is one of the most Ruiz et al. 2011).
important variables that influence the pollen emission, and This study sought to chart the correlation between Poaceae
other meteorological variables such as rainfall or wind speed pollen data and flowering phenology data for the most abun-
and direction influence the pollen dispersal (Rodríguez-Rajo dant and characteristic species of Poaceae in a representative
et al. 2003; Aboulaich et al. 2013). area of central Spain and to model pollen time series in order
Analysis of aerobiological data series is based on the con- to identify the meteorological variables shaping each of the
tinuous measurement of airborne biological particle (e.g., pol- components in the pollen data series over the period 2006–
len and spore) concentrations, with a view to interpreting var- 2014. It also sought to generate a model to predict daily pollen
iations over time (Peel et al. 2014; Hamid et al. 2015). These concentrations, by adding the seasonal component of the se-
data can therefore be processed using time-series analysis ries to the residual component estimated on the basis of me-
techniques, and a number of methods have been developed teorological variables.
based on distinguishing between the deterministic (seasonal)
component and the stochastic (residual) component of data
series (Belmonte and Canela 2002; Aznarte et al. 2007; Material and methods
Ocaña-Peinado et al. 2008).
Research has particularly focused on Poaceae pollen (Stach Case study
et al. 2008; Kasprzyk and Walanus 2010; Fernández-
Rodríguez et al. 2015), for two major reasons: it is the most Daily airborne Poaceae pollen concentrations were recorded
common cause of pollen allergies in Europe (D’Amato et al. from 2006 to 2014 using a Hirst-type volumetric trap (Hirst
2007), and the Poaceae family comprises a large number of 1952) located in the city of Toledo (Fig. 1), following the
species with a broad range of distribution, together accounting protocol recommended by the Spanish Aerobiology
for roughly 20 % of world plant cover (Mabberley 1987). Due Network (Galán et al. 2007). Data on maximum temperature,
to the numerous grass species involved, the Poaceae pollen minimum temperature, and rainfall for the city over the study
season is both very long (Pérez-Badia et al. 2010) and highly period were provided by the Spanish Meteorological Agency
complex, with marked variations in concentrations and a con- (AEMET).
siderable randomly varying or stochastic component, which Field observations of flowering phenology were carried out
can only be estimated by reference to environmental variables in 2013. Phenological sampling was performed at ten selected
(Sabariego et al. 2012; Pérez-Badia et al. 2013). sampling sites in the vicinity of Toledo, located between
One procedure commonly used to analyze time series is 200 m (closest) and 18 km (furthest) from the pollen trap
data decomposition (Brockwell and Davis 2002), which (Table 1), at heights ranging from 450 to 686 m a.s.l. Sites
yields a number of components reflecting different sources contain various types of habitat characteristic of Poaceae spe-
of variability (Currie et al. 2011; Petropavlovskikh et al. cies, including the following: holm oak forests with clearings
2015). Seasonal-trend decomposition based on LOcally mainly covered by xeric grasslands and shrublands, anthropic
wEighted Scatterplot Smoothing (LOESS), known as STL areas of ruderal grassland in urban and periurban areas, weed
decomposition, is a filtering procedure for decomposing a communities associated to the extensive croplands surround-
seasonal times series into three components: trend, seasonal, ing the city, and finally hygrophilous grasslands with Poaceae
and remainder or residual. Advantages of this technique in- species.
clude simplicity and speed of computation, robustness of re-
sults, and flexibility in the period of the seasonal component Phenological data
(Cleveland et al. 1990). This decomposition procedure, wide-
ly used in the natural sciences (e.g., Currie et al. 2011; Vallejos Weekly phenological observations were based on estimating
et al. 2013; Bourjea et al. 2015), has recently become more flowering intensity for the Poaceae species studied, measured
popular in aerobiological research (García-Mozo et al. 2014, as the proportion of flowering individuals (%): for this pur-
2016; Aguilera et al. 2015; Rojo et al. 2015). pose, 25 individuals were selected per species, and the pro-
A number of papers report on the prediction of airborne portion of individuals in bloom each week was recorded (per-
Poaceae pollen concentrations; time-series methods have been centage of flowering individuals (PFI)). Using an adaptation
used in some cases (Moseholm et al. 1987; Voukantsis et al. of the Biologische Bundesanstalt, Bundessortenamt und
2010; Brighetti et al. 2014) but not in others (Sánchez-Mesa CHemische industrie (BBCH) scale for phenological develop-
et al. 2002; Aboulaich et al. 2013). Other studies have exam- ment (Meier 2001), flowering was deemed to begin when
ined variations in pollen concentrations as a function of the first anthers were visible (BBCH code 61) and to end when
changes in meteorological variables (Rojo et al. 2015) or in all anthers were dehydrated and all pollen had been released
the life cycle (flowering) of species regarded as the main from inflorescences (BBCH code 69). A continuous series of
Int J Biometeorol

Fig. 1 Study area. Phenological

sites in the province of Toledo
(central Spain): 1Bar Bargas, 2Oli
Olias del Rey, 3Vad Valdelobos,
4Cam Campus University, 5SAn
St. Anton park, 6Saf Safont park,
7SMa St. Martin bridge, 8Val
Tagus valley, 9Bur Burguillos,
10Nam Nambroca. The pollen
trap is located in the city of
Toledo (coordinates: 39° 51′ 55″
N, 4° 2′ 31″ W)

daily data was generated from weekly PFI values by linear (Table 1). A total of 25 species were studied: Aegilops
interpolation, as used in previous phenological research geniculata Roth, Aegilops triuncialis L., Arrhenatherum
(Oteros et al. 2013b; Rojo and Pérez-Badia 2015). elatius (L.) P. Beauv. ex J. Presl & C. Presl subsp. bulbosum
Species were selected as a function of their abundance or (Willd.) Schübl. & G. Martens, Avena barbata Link, Avena
representativeness at each sampling site, using an adaptation sterilis L., Bromus diandrus Roth, Bromus hordeaceus L.,
of the method recommended by León-Ruiz et al. (2011): 25 Bromus rubens L., Bromus squarrosus L., Bromus tectorum
individuals/m2 were selected for annuals and 25 individuals/ L., Cynosurus elegans Desf., Dactylis glomerata L. subsp.
10 m2 for perennials. As a result, the species sampled varied hispanica (Roth) Nyman, Echinaria capitata (L.) Desf.,
from site to site, depending on the type of habitat involved Elymus pungens (Pers.) Melderis subsp. campestris (Godr.

Table 1 Geographical characteristics of the phenological sites and sampled species

No. Site Coordinates Elevation Sampled speciesa

(m a.s.l.)

1 Bargas 39° 56′ 45″ N 4° 1′ 32″ W 581 AvSt, DaHi, LoRi, MiTe, TrPa
2 Olias del Rey 39° 56′ 9″ N 4° 0′ 2″ W 623 BrRu, DaHi, LoRi, RoCr, TrPa
3 Valdelobos (Toledo city) 39° 52′ 48″ N 4° 4′ 58″ W 514 ArBu, BrSq, DaHi, EcCa, ElCa, RoCr, MaTe
4 Campus (Toledo city) 39° 51′ 57″ N 4° 2′ 37″ W 450 BrDi, HoLe, PiMi, PoAn, TrPa
5 St. Anton park (Toledo city) 39° 52′ 25″ N 4° 1′ 16″ W 494 AvBa, AvSt, BrDi, BrRu, CyEl, DaHi, HoLe, LoRi, RoCr, TrPa
6 Safont park (Toledo city) 39° 51′ 28″ N 4° 1′ 3″ W 470 AvBa, BrDi, PiMi, PoAn, RoCr, TrPa
7 St. Martin bridge (Toledo city) 39° 51′ 28″ N 4° 2′ 3″ W 455 PiMi, PoAn, StCa
8 Tagus valley (Toledo city) 39° 50′ 41″ N 4° 1′ 12″ W 600 AeGe, ArBu, BrHo, BrTe, DaHi, EcCa, LoRi, StCa, VuBr, VuCi, MaTe
9 Burguillos 39° 47′ 43″ N 3° 58′ 55″ W 681 BrRu, BrTe, DaHi, RoCr, TrPa, VuCi
10 Nambroca 39° 47′ 24″ N 3° 57′ 1″ W 686 AeGe, AeTr, HoLe, LoRi
a
Abbreviations of sampled species: AeGe Aegilops geniculata, AeTr A. triuncialis, ArBu Arrhenatherum elatius subsp. bulbosum, AvBa Avena barbata,
AvSt A. sterilis, BrDi Bromus diandrus Roth, BrHo B. hordeaceus, BrRu B. rubens, BrSq B. squarrosus, BrTe B. tectorum, CyEl Cynosurus elegans,
DaHi Dactylis glomerata subsp. hispanica, EcCa Echinaria capitata, ElCa Elymus pungens subsp. campestris, HoLe Hordeum murinum subsp.
leporinum, LoRi Lolium rigidum, MaTe Macrochloa tenacissima, MiTem Micropyrum tenellum, PiMi Piptatherum miliaceum, PoAn Poa annua,
RoCr Rostraria cristata, StCap Stipa capensis, TrPa Trisetum paniceum, VuBr Vulpia bromoides, VuCi V. ciliata
 Int J Biometeorol

& Gren.) Melderis, Hordeum murinum L. subsp. leporinum Using an additive model, the daily pollen-concentration
(Link) Arcang., Lolium rigidum Gaudin, Macrochloa time series (Zt) results from the summation of three compo-
tenacissima (L.) Kunth, Micropyrum tenellum (L.) Link, nents (Harvey and Peters 1990):
Piptatherum miliaceum (L.) Coss., Poa annua L., Rostraria
Z t ¼ M t þ S t þ Rt t ¼ 1; …; T
cristata (L.) Tzvelev, Stipa capensis Thunb., Trisetum
paniceum (Lam.) Pers., Vulpia bromoides (L.) Gray, and where Mt is the trend, St is the seasonality, and Rt is the
Vulpia ciliata Dumort. residual.
Factor analysis was used to reduce dimensionality in the This structural model enables each of the components to be
flowering data for the Poaceae species (Rummel 1988; isolated and analyzed separately. Here, a seasonal-trend de-
Reyment and Jvreskog 1996), with a view to obtaining a sin- composition procedure based on LOESS (STL) was used to
gle phenological variable for each species. Given the similar- analyze daily pollen concentrations over an 8-year period
ity of flowering curves (PFI) for each species at different sites, (2006–2013).
the first principal component in the factor analysis accounted STL is a filtering procedure for decomposing time series,
in all cases for over 90 % of variance (see example in Fig. 2). based on a sequence of smoothing procedures using a locally
A single factor containing 90 % of the PFI information for a weighted regression, commonly known as LOESS (Cleveland
given species could thus be obtained by the following equa- et al. 1990). The LOESS smoother is based on fitting a
tion: weighted polynomial regression for a given time of observa-
z tion, where weights decrease with distance from the nearest
F s ¼ ∑ ai X i neighbor (Dagum and Luati 2003). LOESS produces a
i¼1
smoothed estimate (ŷj) that is defined by the following:
d
where Fs is the factor containing most PFI information for ŷ j ¼ ∑ βpj t pj j ¼ 1; …; n
each species s; a is the coefficient for each variable X, based p¼0
on the estimated PFI at each sampling site; i is the sampling
site; and z is the total number of sampling sites for the species. where βj is the (d + 1)-dimensional least squares estimate of
the weighted regression, t pj is the (d + 1)-dimensional vector of
the time of observation, j is the number of time lags up to the
Decomposition of aerobiological data maximum defined by the smoothing parameter (n), p = 0, …,
d, and d is the degree of the polynomial fitting.
The aerobiological variable (pollen concentration) was trans- The time series is fitted iteratively until trend and season-
formed by decomposition of the time series (Brockwell and ality stabilize, in a multi-step process in which moving aver-
Davis 2002). First, the structure (additive vs. multiplicative ages alternate with LOESS smoothing. At the end of the STL
models) and seasonality (stationary vs. non-stationary) of the process, the seasonal and trend components are extracted from
time series were defined using Holt-Winters filtering (Winters the data series, in accordance with the following expression:
1960) and the Ljung-Box test (Ljung and Box 1978),
ðkþ1Þ ðkþ1Þ
respectively. St ¼ St ; Mt ¼ Mt ; Rt ¼ Z t −S t −M t

Fig. 2 Factor analysis for the 100

species Rostraria cristata during
Percentage of flowering individuals (PFI)

90
the year 2013. Phenological sites:
1Bar Bargas, 2Oli Olias del Rey, 80
4Cam Campus University, 9Bur 1Bar
Burguillos, 10Nam Nambroca 70 2Oli
4Cam
60
9Bur
50
10Nam
40 Result 1st Factor

30
Results of the Factor Analysis
20 taking into account the first
principal component (96.7%
10 of the explained variance)

0
19-Mar 2-Apr 16-Apr 30-Apr 14-May 28-May 11-Jun 25-Jun 9-Jul 23-Jul
Int J Biometeorol

where k is the number of iterations in which the various phenology variable (PFI) for the grasses studied in 2013. For
steps take place. For a detailed description of the STL method, that purpose, PLSR took into account the flowering of
see Cleveland et al. (1990) and Zhou et al. (2015). Poaceae species for which a positive significant correlation
The STL decomposition method requires definition of the was observed (Spearman test).
smoothing parameter for the seasonal component (n s ) The residual or random component (Rt) was modeled using
(Cleveland et al. 1990); here, it was set at ns → ∞, thus forcing data for daily pollen concentrations and for meteorological
the seasonal component to be periodic and to be constant variables measured from 1 day prior to each time point in
across all the years in the time series, thus enabling it to be the series (x-1) up to 3 days before that time point (x-3).
used in predicting future pollen concentrations. Time-series Modeling was carried out from the PS (PS) start date, de-
analysis was performed using R software (R Core Team fined—using an adaptation of the approach recommended
2015). by Galán et al. (2001) for olive pollen—as the first day on
which at least 10 grains/m3 were recorded with ≥10 grains/m3
on the following 2 days. This definition was better suited to
Modeling pollen concentrations the aims of the present study (Jato et al. 2006), in that it
enabled the elimination of a large number of earlier days with
Decomposition of the time series into its various components small intermittent amounts of pollen, which seriously hin-
enabled each component to be independently modeled. dered modeling. The residual component was modeled sepa-
Models were developed by partial least squares regression rately for the pre-peak period (i.e., prior to the peak pollen-
(PLSR), using a non-linear kernel algorithm. This method is concentration day) and the post-peak period (i.e., after the
appropriate for models containing numerous intercorrelated peak pollen-concentration day). The model was calibrated
independent variables (Wold et al. 2001). for the period 2006–2013.
PLSR creates a block of predictors (X) from the indepen- Finally, daily pollen concentrations for 2014 were predict-
dent variables included in the model (P), using principal com- ed using data not used for model calibration. For this purpose,
ponents (T). This block is formed by orthogonal factors (i.e., the residual component was estimated using the model and the
independent of each other), commonly known as latent vari- series was reconstructed by adding the deterministic compo-
ables: nents (seasonality and trend). The model was validated by
comparing the predicted data for 2014 with real pollen con-
X ¼ TP þ Residuals
centrations (i.e., raw non-decomposed data) for that year.
Each principal component is calculated from the weight Relationships and differences between real and predicted pol-
(w) of each independent variable (x) in that factor, computed len curves were determined using the Spearman test (correla-
from the covariance between the response variable (y) and the tion between data series) and the Wilcoxon signed-rank test
independent variables: (differences between paired samples).

 Calibration and validation of the aerobiological data model
p cov x j ; y were performed using the Bpls^ package (Mevik and Wehrens
t i ¼ ∑ wia X a ; wij ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

ffi
a¼1 ∑pj¼1 cov2 x j ; y 2007) and R software (R Core Team 2015).

t 1 ¼ w11 x1 ; …; w1p xp
Results
…
t z ¼ wz1 x1 ; …; wzp xp The Poaceae pollen-concentration series for the period 2006–
2013 was an additive time series with a clear cyclic annual
where p is the maximum number of independent variables pattern (Fig. 3). The seasonality of the series did not vary over
included in the model and z is the maximum number of prin- time, and no significant trend pattern was observed; for that
cipal components selected. reason, the trend component was not subjected to separate
Next, the latent variables containing the information from analysis. The results of the Ljung-Box test confirmed that data
the predictors are used to estimate the response variable: matched a stationary seasonal series (Fig. 3).
y ¼ TC þ Error The daily pollen curve for 2013 reflected the staggered
flowering of different grass species; concentrations differed
where C is a matrix of correlation coefficients for the predictor considerably depending on the species in flower at any given
block and the response variable block. moment (Fig. 4). Low peaks were recorded in mid-April
This study sought to identify potential correlations between 2013, coinciding with the flowering of the earliest Poaceae
data for the trend and seasonality components (Mt + St) of the species such as Hordeum murinum subsp. leporinum or spe-
data series for the period 2006–2013 with the flowering cies of the genera Avena, Bromus, and Vulpia. Pollen
 Int J Biometeorol

Fig. 3 Poaceae pollen time series Characterizacion of the Poaceae pollen time series (2006-2013)
for the period 2006–2013. χ2 Structure: Aditive model
Stationarity: Stationary process ( 2 = 14104 , p < 0.001)
statistics from the Ljung-Box test
600

Original pollen data

500

400

300

200

100

0
Seasonal + Trend 250

200

150

100

50

0
400

300
Residuals

200

100

-100

-200
Time (days)

concentrations rose throughout May, due to the flowering of Modeling of the residual component, for calibration pur-
grasses such as Macrochloa tenacissima, Echinaria capitata, poses, showed that this component accounted for 58 % of
and R. cristata (Fig. 4). The highest concentrations were ob- overall variance in the pre-peak period (Fig. 6) and for 69 %
served in late May and early June (2013 peak concentration: in the post-peak period (Fig. 7). The importance of the vari-
26 May), coinciding with the flowering of late species includ- ables included in the model, as expressed by regression coef-
ing Arrhenatherum elatius subsp. bulbosum, L. rigidum, ficients, varied between pre-peak and post-peak models, al-
T. paniceum, D. glomerata subsp. hispanica, Piptatherum though, in most cases, correlations shared the same sign for
miliaceum, and various Aegilops species. both periods (Fig. 8). The most influential variable was pollen
Then, from 26 May onward, pollen concentrations began to concentration on the previous day (Poac-1). Maximum tem-
decline, as flowering gradually ended in the later grass spe- perature on the previous day (Tmax-1) also exerted an impor-
cies; by mid-July 2013, the amount of airborne grass pollen tant positive influence on pollen concentrations, whereas min-
had ceased to be significant (Fig. 4). Comparison of percent- imum temperature (Tmin) and rainfall (R) generally had a
age of flowering individuals (PFI) data over the flowering negative effect (Fig. 8).
period for each species and airborne pollen concentrations Finally, the model was validated using independent pollen
for the same period revealed that later-flowering species and data for 2014; the Spearman correlation coefficient between
Poa annua—which flowers over most of the pollen season observed and predicted data was r = 0.79 for the pre-peak
(Fig. 4)—displayed the strongest correlation with pollen con- period and r = 0.63 for the post-peak period (Fig. 9). The
centrations (Table 2). Wilcoxon signed-rank test revealed no significant difference
The PFIs for all these grass species were used to model between observed and estimated data series (Fig. 9).
pollen concentrations (Fig. 5). Models constructed using
PLSR on phenological variables accounted for 35 % of vari-
ance in the original (non-decomposed) data series for 2013 Discussion
(Fig. 5a). However, when the series was decomposed and
the residual component was removed, the model improved Aerobiological data series contain a periodic component
considerably—accounting for 79 % of variance (Fig. 5b). which enables them to be modeled as time series
Int J Biometeorol

Poaceae pollen (2013)

MARCH APRIL MAY JUNE JULY

Bromus hordeaceus
Bromus tectorum
Vulpia ciliata
Bromus rubens
Vulpia bromoides
Hordeum murinum subsp. leporinum
Poa annua
Bromus diandrus
Stipa capensis
Avena barbata
Cynosurus elegans
Micropyrum tenellum
Macrochloa tenacissima
Avena sterilis
Echinaria capitata
Rostraria cristata
Bromus squarrosus
Arrhenatherum elatius subsp. bulbosum
Aegilops geniculata
Legend (Percentage of
Lolium rigidum
flowering individuals)
Aegilops triuncialis
> 25% > 50% > 25%
Trisetum paniceum
Piptatherum miliaceum
Dactylis glomerata subsp. hispanica
Elymus pungens subsp. campestris
MARCH APRIL MAY JUNE JULY
Fig. 4 Comparison between daily pollen concentrations and PFI for the given grass species during the year 2013

(Voukantsis et al. 2010), together with a considerable random This study represents a methodological advance for the
component. In this study, airborne pollen time series were modeling of aerobiological series, with various useful appli-
modeled using a seasonal-trend decomposition procedure cations: for predicting daily pollen concentrations, for identi-
based on LOESS smoothing, which proved well suited for this fying major sources of pollen emission, and for analyzing
purpose. While the STL decomposition technique has recently variations in pollen concentrations as a function of environ-
been used by other authors to measure and analyze airborne mental factors such as meteorological variables.
pollen trends over fairly long periods (García-Mozo et al. Using STL decomposition procedures, we can isolate sea-
2014; Aguilera et al. 2015), the results obtained here confirm sonality and trend components from data series and also elim-
that decomposition of pollen data series is also of great value inate part of the random noise or residual component which
for interpreting variability in other components—seasonality cannot be accounted for by the behavior of the series itself
and residuals—over a smaller timescale (Rojo et al. 2015). (Brockwell and Davis 2002). In aerobiological data series,
 Int J Biometeorol

Table 2 Results of Spearman

correlation tests between Grass species Pollen (original data) Pollen (seasonal + trend)
phenological data (PFI) of the
species and Poaceae pollen data Aegilops geniculata 0.67* 0.72*
from original data and the pollen Aegilops triuncialis 0.61* 0.86*
time series (seasonal + trend) Arrhenatherum elatius subsp. bulbosum 0.51* 0.54*
Dactylis hispanica 0.60* 0.87*
Lolium rigidum 0.63* 0.70*
Piptatherum miliaceum 0.48* 0.80*
Poa annua 0.46* 0.45*
Trisetum paniceum 0.58* 0.88*

*p < 0.001

deterministic seasonal and trend components are associated from the plant life cycles, which vary as a function of time-
with certain environmental features and biological aspects scales and geographical factors (Petropavlovskikh et al.

Fig. 5 Results of the partial least a 0.4

square regressions developed for 600 Species
Pollen (original data)

Standardized regression coefficient

modeling the pollen data from the 0.35
phenological variables (year 0.3
2013). Pollen data: the original 500
0.25
pollen data (a) and the
0.2
summations of the seasonal and 400
Pollen grains / m3

0.15
trend components of the pollen
time series (b). MSE mean square 0.1

error, RMSE root-mean-square 300 0.05

error. Sampled species: AeGe
R2 = 0.35 0
Aegilops geniculata, AeTr MSE = 4673.0
Aegilops triuncialis, ArBu 200
RMSE = 68.4
Arrhenatherum elatius subsp.
bulbosum, DaHi Dactylis
100
glomerata subsp. hispanica, LoRi
Lolium rigidum, PiMi
Piptatherum miliaceum, PoAn 0
Poa annua, TrPa Trisetum 0 10 20 30 40 50 60 70 80
paniceum Days from begin of PS (days)

Poaceae pollen (original data) Phenology (PLS regression model)

b 250
0.3 Species
Standardized regression coefficient

Pollen (seasonal + trend components) 0.25

0.2
200
0.15
Pollen grains / m3

0.1
150
0.05

0
100 R2 = 0.79
MSE = 484.1
RMSE = 22.0

50

0
0 10 20 30 40 50 60 70 80
Days from begin of PS (days)

Poaceae pollen (seasonal + trend) Phenology (PLS regression model)

Int J Biometeorol

Fig. 6 Calibration of the model Calibration of the model for the pre-peak period (2006-2013)
for the pre-peak period, obtained
300
by adding the deterministic com-

Modeling of the residuals

ponent (seasonal + trend) to the
200
estimated residuals during the pe-
riod 2006–2013
100

-100

-200 Residuals (observed) Residuals (estimated)

200 +
Seasonal + Trend
150

100

50

600
= R2 = 0.58 ; p < 0.001
RMSE = 58.3
Modeling of the original series

500

400

300

200

100

0
Time (days)

Poaceae pollen (observed) Poaceae pollen (estimated)

2015). The present study examined the variability of each The key grass species for modeling pollen-curve seasonal-
component separately and then tested for correlations between ity proved to be later-flowering species including Aegilops
these components and environmental (meteorological and spp., D. glomerata subsp. hispanica, L. rigidum, and
phenological) parameters over the most appropriate timescale T. paniceum. Similar findings have been reported by other
in each case; if data series are analyzed without decomposi- authors in studies of the major contributors to the pollen curve
tion, certain interesting features may remain concealed, due to elsewhere in the world (Frenguelli et al. 2010; León-Ruiz et al.
the potential synergy of environmental parameters over differ- 2011). Future research, however, should focus on other as-
ent timescales. The present study focused on Poaceae pollen pects of pollen emission by different species, including pollen
series, but the same method could be used to model the be- production (Prieto-Baena et al. 2003; Aboulaich et al. 2009)
havior of the pollen spectrum for other species. and pollen grain dispersal capacity (Subiza 2003). In this
Separate processing of each component in the pollen curve sense, the major contributors to the Poaceae airborne pollen
with a view to identifying the cause of variations should start can be the species belonging to the genera Trisetum
by examining shifts in seasonality prompted by the seasonal (T. paniceum), Dactylis (D. glomerata), and Lolium
climate dynamics regulating plant life cycles (Cleland et al. (L. rigidum). These species produce a huge number of pollen
2007). Analysis of the grass flowering period in the present grains per inflorescence (5,000,000–21,000,00 pollen grains
study showed that the model generated using flowering phe- per inflorescence, Prieto-Baena et al. 2003). Otherwise, al-
nology data for a very small number of species accounted though the flowering of species of the genus Aegilops coin-
sufficiently for series seasonality. cided with the highest levels of airborne pollen
 Int J Biometeorol

Fig. 7 Calibration of the model Calibration of the model for the post-peak period (2006-2013)
for the post-peak period, obtained
300
by adding the deterministic com-

Modeling of the residuals

ponent (seasonal + trend) to the
200
estimated residuals during the pe-
riod 2006–2013
100

-100

-200 Residuals (observed) Residuals (estimated)

200 +
Seasonal + Trend
150

100

50

500
= R2 = 0.69 ; p < 0.001
RMSE = 43.8
Modeling of the original series

400

300

200

100

0
Time (days)

Poaceae pollen (observed) Poaceae pollen (estimated)

concentrations, these species can have low contribution to the Since the data series used here did not display a significant
airborne pollen concentrations because their pollen production trend pattern, i.e., it was stationary over time in probability
is below 100,000 pollen grains per inflorescence (Prieto- terms (Cryer and Chan 2008), there was no need to analyze
Baena et al. 2003). the trend component separately. In longer pollen data series,

Fig. 8 Independent variables Influence of the independent variables

taken into account when 0.4
modeling residuals. Regression Poac Tmax Tmin R
Standardized regression coefficients

coefficients of the pollen 0.3

concentrations for the previous
days (Poac) and meteorological 0.2
variables for the previous days:
Tmax maximum temperature,
0.1
Tmin minimum temperature, R
rainfall. Data for the previous day
0
(x-1), previous 2 days (x-2), and
previous 3 days (x-3) were
-0.1
considered for modeling
Tmax-1

Tmax-2

Tmax-3

P-1

P-2

P-3
Tmin-1

Tmin-2

Tmin-3
Poac-1

Poac-2

Poac-3

-0.2

-0.3

Pre-peak period Post-peak period

Int J Biometeorol

Fig. 9 Estimated Poaceae pollen Validation of the model (independent data, 2014)
concentrations obtained by 600
adding the deterministic
component (seasonal + trend) to 500
the predicted residual for 2014. r
statistics from the Spearman 400
correlation test. W statistics from Pre-peak period Post-peak period

Pollen grains / m3
the Wilcoxon signed-rank test. W = -1.95 , p > 0.05 W = -1.12 , p > 0.05
300 r = 0.79 , p < 0.001 r = 0.63 , p < 0.001
Dashed line marks the peak
pollen-concentration day
200

100

0
0 10 20 30 40 50 60 70 80

-100
Days from the start of PS (days)

Pollen (original data) Estimated pollen concentrations

however, significant trends have been observed, prompted by Poaceae pollen concentrations for 2014. The predicted pollen
variations in weather patterns (Recio et al. 2010; Makra et al. curve correlated well with the observed curve, with no signif-
2011). Reported trends are mostly attributed to changes in icant difference between distributions. The prediction model
plant phenology as a result of environmental changes for the residuals included meteorological variables and pollen
(Peñuelas et al. 2002; Gordo and Sanz 2009; García-Mozo for previous days, focusing on those accounting for the
et al. 2010). Wherever apparent trends are observed, they greatest amount of variance (Sawa 1978; Seasholtz and
should be analyzed on a long-term basis by isolating the trend Kowalski 1993). The parsimony principle should always be
component of the data series (García-Mozo et al. 2010; applied with a view to obtaining models combining simplicity
Aguilera et al. 2015). with a high degree of precision (Akaike 1974). Moreover, the
Finally, the daily pollen series contained a residual compo- calibration of pre-peak and post-peak models with determina-
nent, reflecting its stochastic nature. This residual displays tion coefficients of 0.58 and 0.69, respectively, may be con-
highly variable behavior due to the influence of short-term, sidered quite acceptable (Silva-Palacios et al. 2015) and
local-scale meteorological conditions (Petropavlovskikh et al. yielded predictions which correlated closely with observed
2015), which, in the Mediterranean region, tend to vary con- values for 2014.
siderably (García-Mozo et al. 2014). On the basis of correla- The inclusion of a large number of highly correlated pre-
tions with a range of meteorological variables (Ranzi et al. dictors may give rise to multicollinearity issues (Graham
2003), the residual component can be modeled—as here— 2003). Statistical modeling of various aspects of the pollen
with a view to identifying the meteorological variables most season often uses numerous pollen-related predictors (Stach
influencing airborne pollen concentrations (Rojo et al. 2015) et al. 2008; Veriankaitė et al. 2011). PLSR is regarded as a
and also to predicting future pollen concentrations by estimat- suitable technique for reducing the number of predictors, since
ing the residual component isolated separately from the orig- it is not sensitive to multicollinearity (Wold et al. 1984); it has
inal data series (Valderrama et al. 2010). recently been used in aerobiological modeling (Oteros et al.
The patterns of Poaceae pollen emission can vary greatly in 2013a; Brighetti et al. 2014; Rojo et al. 2015). Application of
response to meteorological changes, as it is shown in the re- this procedure for the analysis of stochastic processes yields
sults of this study. Therefore, the aerobiological processes are good results, since it reduces other statistical problems linked
greatly influenced by the meteorological variables (Sánchez- to model stability and quality of fit (Preda and Saporta 2005).
Mesa et al. 2005), and maximum temperature is the most Finally, the PS start date needs to be carefully defined,
important meteorological variable included in the predictive since it is a key aspect of modeling. Poaceae have a very long
model with respect to the standardized regression coefficients. pollen season (Pérez-Badia et al. 2010; Gucel et al. 2013),
However, other important phenomenon such as pollen dis- including numerous days at the start and end of the main
persal is conditioned by other meteorological variables such season with very low and intermittently varying pollen con-
as rainfall (Rojo et al. 2015), and thus, this variable has been centrations. These hinder modeling and, thus, the prediction of
also included in the model of our study. future airborne grass pollen concentrations. For a better fit, the
The model obtained using the STL decomposition proce- pollen season should be reduced to the period over which
dure and data for 2006–2013 was used to predict airborne pollen concentrations are reasonably high (Silva-Palacios
 Int J Biometeorol

et al. 2015). Moreover, models for predicting daily pollen J, Momas I, Nawijn MC, Neubauer A, Oddie S, Palkonen S, Pin I,
Pison C, Rancé F, Reitamo S, Rial-Sebbag E, Salapatas M, Siroux V,
concentrations—such as that constructed here—can be com-
Smagghe D, Torrent M, Toskala E, van Cauwenberge P, van
bined with models generated for predicting the PS start date Oosterhout AJ, Varraso R, von Hertzen L, Wickman M, Wijmenga
(Kasprzyk and Walanus 2010, 2014; Pauling et al. 2014). C, Worm M, Wright J, Zuberbier T (2011) MeDALL (Mechanisms
of the Development of ALLergy): an integrated approach from phe-
notypes to systems medicine. Allergy 66(5):596–604
Brighetti MA, Costa C, Menesatti P, Antonucci F, Tripodi S, Travaglini A
Conclusions (2014) Multivariate statistical forecasting modeling to predict
Poaceae pollen critical concentrations by meteoclimatic data.
Decomposition of time series using the STL procedure was Aerobiologia 30:25–33
Brockwell PJ, Davis RA (2002) Introduction to time series and forecast-
found to be a suitable technique for analyzing the phenologi-
ing, 2nd edn. Springer-Verlag, New York, USA
cal and meteorological factors influencing airborne pollen Cleland EE, Chuine I, Menzel A, Mooney HA, Schwartz MD (2007)
concentrations. Separate analysis of each of the components Shifting plant phenology in response to global change. Trends
of the pollen data series enables the sources of variability to be Ecol Evol 22(7):357–365
Cleveland RB, Cleveland WS, McRae JE, Terpenning I (1990) STL: a
identified more accurately than by analysis of the original
seasonal-trend decomposition procedure based on loess. J Off Stat
non-decomposed data series, since certain environmental fac- 6(1):3–33
tors may mask the influence of other factors operating at dif- Cryer JD, Chan KS (2008) Time series analysis with applications in R,
ferent timescales. Furthermore, the results of this work consti- 2nd edn. Springer, New York, USA
Currie KI, Brailsford G, Nichol S, Gomez A, Sparks R, Lassey KR,
tute an advance in the knowledge of the major contributors
Riedel K (2011) Tropospheric 14CO2 at Wellington, New Zealand:
(grass species) to the Poaceae airborne pollen concentrations. the world’s longest record. Biogeochemistry 104:5–22
The method can also be applied for modeling other pollen D’Amato G, Cecchi L, Bonini S, Nunes C, Annesi-Maesano I, Behrendt
types. H, Liccardi G, Popov T, van Cauwenberge P (2007) Allergenic
pollen and pollen allergy in Europe. Allergy 62:976–990
Dagum EB, Luati A (2003) Global and local statistical properties of
fixed-length nonparametric smoothers. Stat Method Appl 11:
313–333
References Estrella N, Menzel A, Krämer U, Behrendt H (2006) Integration of
flowering dates in phenology and pollen counts in aerobiology:
analysis of their spatial and temporal coherence in Germany
Aboulaich N, Achmakh L, Bouziane H, Trigo MM, Recio M, Kadiri M, (1992–1999). Int J Biometeorol 51:49–59
Cabezudo B, Riadi H, Kazzaz M (2013) Effect of meteorological Fernández-Llamazares A, Belmonte J, Boada M, Fraixedas S (2014)
parameters on Poaceae pollen in the atmosphere of Tetouan (NW Airborne pollen records and their potential applications to the con-
Morocco). Int J Biometeorol 57:197–205 servation of biodiversity. Aerobiologia 30:111–122
Aboulaich N, Bouziane H, Kadiri M, Trigo MM, Riadi H, Kazzaz M, Fernández-Rodríguez S, Adams-Groom B, Silva-Palacios I, Caeiro E,
Merzouki A (2009) Pollen production in anemophilous species Brandao R, Ferro R, Gonzalo-Garijo A, Smith M, Tormo R (2015)
of the Poaceae family in Tetouan (NW Morocco). Aerobiologia Comparison of Poaceae pollen counts recorded at sites in Portugal,
25:27–38 Spain and the UK. Aerobiologia 31:1–10
Aguilera F, Orlandi F, Ruiz-Valenzuela L, Msallem M, Fornaciari M Frenguelli G, Passalacqua G, Bonini S, Fiocchi A, Incorvaia C, Marcucci
(2015) Analysis and interpretation of long temporal trends in cumu- F, Tedeschini E, Canonica GW, Frati F (2010) Bridging allergologic
lative temperatures and olive reproductive features using a seasonal and botanical knowledge in seasonal allergy: a role for phenology.
trend decomposition procedure. Agric For Meteorol 203:208–216 Ann Allerg Asthma Im 105(3):223–227
Akaike H (1974) A new look at the statistical model identification. IEEE Galán C, Cariñanos P, Alcázar P, Domínguez-Vilches E (2007) Spanish
Trans Autom Control 19(6):716–723 aerobiology network (REA): management and quality manual.
Aznarte JL, Benítez JM, Nieto D, de Linares C, Díaz de la Guardia C, Servicio de Publicaciones, Universidad de Córdoba, Córdoba, Spain
Alba F (2007) Forecasting airborne pollen concentration time Galán C, Cariñanos P, García-Mozo H, Alcázar P, Domínguez-Vilches E
series with neural and neuro-fuzzy models. Expert Syst Appl (2001) Model for forecasting Olea europaea L. airborne pollen in
32:1218–1225 South-West Andalusia, Spain. Int J Biometeorol 45:59–63
Belmonte J, Canela M (2002) Modelling aerobiological time series. García-Mozo H (2011) The use of aerobiological data on agronomical
Application to Urticaceae. Aerobiologia 18:287–295 studies. Ann Agric Environ Med 18:159–164
Bourjea J, Dalleau M, Derville S, Beudard F, Marmoex C, M’Soili A, García-Mozo H, Mestre A, Galán C (2010) Phenological trends in south-
Roos D, Ciccione S, Frazier J (2015) Seasonality, abundance, and ern Spain: a response to climate change. Agric For Meteorol 150:
fifteen-year trend in green turtle nesting activity at Itsamia, Moheli, 575–580
Comoros. Endang Species Res 27:265–276 García-Mozo H, Oteros JA, Galán C (2016) Impact of land cover changes
Bousquet J, Anto J, Auffray C, Akdis M, Cambon-Thomsen A, Keil T, and climate on the main airborne pollen types in Southern Spain. Sci
Haahtela T, Lambrecht BN, Postma DS, Sunyer J, Valenta R, Akdis Total Environ 548-549:221–228
CA, Annesi-Maesano I, Arno A, Bachert C, Ballester F, Basagana García-Mozo H, Yaezel L, Oteros J, Galán C (2014) Statistical approach
X, Baumgartner U, Bindslev-Jensen C, Brunekreef B, Carlsen KH, to the analysis of olive long-term pollen season trends in southern
Chatzi L, Crameri R, Eveno E, Forastiere F, Garcia-Aymerich J, Spain. Sci Total Environ 473-474:103–109
Guerra S, Hammad H, Heinrich J, Hirsch D, Jacquemin B, Gordo O, Sanz JJ (2009) Long-term temporal changes of plant phenology
Kauffmann F, Kerkhof M, Kogevinas M, Koppelman GH, in the Western Mediterranean. Global Change Biol 15:1930–1948
Kowalski ML, Lau S, Lodrup-Carlsen KC, Lopez-Botet M, Graham MH (2003) Confronting multicollinearity in ecological multiple
Lotvall J, Lupinek C, Maier D, Makela MJ, Martinez FD, Mestres regression. Ecology 84:2809–2815
Int J Biometeorol

Gucel S, Guvensen A, Ozturk M, Celik A (2013) Analysis of airborne Petropavlovskikh I, Evans R, McConville G, Manney GL, Rieder HE
pollen fall in Nicosia (Cyprus). Environ Monit Assess 185:157–169 (2015) The influence of the North Atlantic Oscillation and El
Hamid N, Ali SM, Talib F, Sadiq I, Ghufran MA (2015) Spatial and Niño-Southern Oscillation on mean and extreme values of column
temporal variations of pollen concentrations in Islamabad ozone over the United States. Atmos Chem Phys 15:1585–1598
(Pakistan): effect of meteorological parameters and impact on hu- Preda C, Saporta G (2005) PLS regression on a stochastic process.
man health. Grana 54(1):53–67 Comput Stat Data An 48:149–158
Harvey AC, Peters S (1990) Estimation procedures for structural time Prieto-Baena JC, Hidalgo PJ, Domínguez-Vilches E, Galán C (2003)
series model. J Forecast 9:89–108 Pollen production in the Poaceae family. Grana 42:153–160
Hirst JM (1952) An automatic volumetric spore trap. Ann Appl Biol 36: R Core Team (2015) A language and environment for statistical comput-
257–344 ing. R Foundation for Statistical Computing, Vienna, Austria.
Jato V, Rodríguez-Rajo FJ, Alcázar P, De Nuntiis P, Galán C, Mandrioli P [Link] (accessed 11 Sept 2015)
(2006) May the definition of pollen season influence aerobiological Ranzi A, Lauriola P, Marletto V, Zinoni F (2003) Forecasting airborne
results? Aerobiologia 22:13–25 pollen concentrations: development of local models. Aerobiologia
Kasprzyk I (2006) Comparative study of seasonal and intradiurnal varia- 19:39–45
tion of airborne herbaceous pollen in urban and rural areas. Recio M, Docampo S, García-Sánchez J, Trigo MM, Melgar M,
Aerobiologia 22:185–195 Cabezudo B (2010) Influence of temperature, rainfall and wind
Kasprzyk I, Walanus A (2010) Description of the main Poaceae pollen trends on grass pollination in Malaga (western Mediterranean coast).
season using bi-Gaussian curves, and forecasting methods for the Agric For Meteorol 150(7-8):931–940
start and peak dates for this type of season in Rzeszów and Reyment RA, Jvreskog KG (1996) Applied factor analysis in the natural
Ostrowiec Św. (SE Poland). J Environ Monit 12:906–916 sciences. Cambridge University Press
Kasprzyk I, Walanus A (2014) Gamma, Gaussian and logistic distribution Rodríguez-Rajo FJ, Jato V, Aira MJ (2003) Pollen content in the atmo-
models for airborne pollen grains and fungal spore season dynamics. sphere of Lugo (NW Spain) with reference to meteorological factors
Aerobiologia 30:369–383 (1999-2001). Aerobiologia 19:213–225
León-Ruiz E, Alcázar P, Domínguez-Vilches E, Galán C (2011) Study of Rodríguez-Rajo F, Valencia-Barrera RM, Vega-Maray AM, Suárez FJ,
Poaceae phenology in a Mediterranean climate. Which species con- Fernández-González D, Jato V (2006) Prediction of airborne Alnus
tribute most to airborne pollen counts? Aerobiologia 27:37–50 pollen concentration by using ARIMA models. Ann Agric Environ
Ljung GM, Box GEP (1978) On a measure of lack of fit in time series Med 13:25–32
models. Biometrika 65:297–303 Rojo J, Pérez-Badia R (2015) Models for forecasting the flowering of
Cornicabra olive groves. Int J Biometeorol 59:1547–1556
Mabberley DJ (1987) The plant book. Cambridge University press,
Rojo J, Rapp A, Lara B, Fernández-González F, Pérez-Badia R (2015)
Cambridge
Effect of land uses and wind direction on the contribution of local
Makra L, Matyasovszky I, Deàk AJ (2011) Trends in the characteristics of
sources to airborne pollen. Sci Total Environ 538:672–682
allergenic pollen circulation in central Europe based on the example
Rummel RJ (1988) Applied factor analysis. Northwestern University
of Szeged, Hungary. Atmos Environ 45:6010–6018
Press, Evanston
Meier U (2001) Growth stages of mono- and dicotyledonous plants.
Sabariego S, Cuesta P, Fernández-González F, Pérez-Badia R (2012)
BBCH monograph. 2nd ed. Federal Biological Research Centre for
Models for forecasting airborne Cupressaceae pollen levels in cen-
Agriculture and Forestry
tral Spain. Int J Biometeorol 56(2):253–258
Mevik BH, Wehrens R (2007) The pls package: principal component and Sánchez-Mesa JA, Galán C, Martínez-Heras JA, Hervás-Martínez C
partial least squares regressions in R. J Stat Softw 18(2):1–24 (2005) The use of discriminant analysis and neural networks to
Moseholm L, Weeke ER, Petersen BN (1987) Forecast of pollen concen- forecast the severity of the Poaceae pollen season in a region with
trations of Poaceae (grasses) in the air by time series analysis. Pollen typical Mediterranean climate. Int J Biometeorol 49:355–362
Spores 2-3:305–322 Sánchez-Mesa JA, Galán C, Martínez-Heras JA, Hervás-Martínez C
Ocaña-Peinado F, Valderrama MJ, Aguilera AM (2008) A dynamic re- (2002) The use of a neural network to forecast daily grass pollen
gression model for air pollen concentration. Stoch Environ Res Risk concentration in a Mediterranean region: the southern part of the
Assess 22(Suppl 1):S59–S63 Iberian Peninsula. Clin Exp Allergy 32:1606–1612
Oteros J, García-Mozo H, Hervás-Martínez C, Galán C (2013a) Year Sawa T (1978) Criteria for discriminating among alternative regression
clustering analysis for modelling olive flowering phenology. Int J models. Econometrica 46(6):1273–1291
Biometeorol 57(4):545–555 Seasholtz MB, Kowalski B (1993) The parsimony principle applied to
Oteros J, García-Mozo H, Vázquez L, Mestre A, Domínguez-Vilches E, multivariate calibration. Anal Chim Acta 277:165–177
Galán C (2013b) Modelling olive phenological response to weather Silva-Palacios I, Fernández-Rodríguez S, Durán-Barroso P, Tormo-
and topography. Agric Ecosyst Environ 179:62–68 Molina R, Maya-Manzano JM, Gonzalo-Garijo A (2015)
Pauling A, Gehrig R, Clot B (2014) Toward optimized temperature sum Temporal modelling and forecasting of the airborne pollen of
parameterizations for forecasting the start of the pollen season. Cupressaceae on the southwestern Iberian Peninsula. Int J
Aerobiologia 30:45–57 Biometeorol. doi:10.1007/s00484-015-1026-6
Peel RG, Ørby PV, Skjøth CA, Kennedy R, Schlünssen V, Smith M, Stach A, Smith M, Prieto-Baena JC, Emberlin J (2008) Long-term and
Sommer J, Hertel O (2014) Seasonal variation in diurnal atmospher- short-term forecast models for Poaceae (grass) pollen in Poznań,
ic grass pollen concentration profiles. Biogeosciences 11:821–832 Poland, constructed using regression analysis. Environ Exp Bot
Peñuelas J, Filella I, Comas P (2002) Changed plant and animal life 62:323–332
cycles from 1952 to 2000 in the Mediterranean region. Global Subiza J (2003) Gramínes: Aerobiología y polinosis en España. Alergol
Change Biol 8(6):531–544 Inmunol Clin 18(3):7–11
Pérez-Badia R, Bouso V, Rojo J, Vaquero C, Sabariego S (2013) Valderrama MJ, Ocaña FA, Aguilera AM, Ocaña-Peinado FM (2010)
Dynamics and behaviour of airborne Quercus pollen in central Forecasting pollen concentration by a two-step functional model.
Iberian Peninsula. Aerobiologia 29:419–428 Biometrics 66:578–585
Pérez-Badia R, Rapp A, Morales C, Sardinero S, Galán C, García-Mozo Vallejos RO, Fabré NN, Batista VS, Acosta J (2013) The application of a
H (2010) Pollen spectrum and risk of pollen allergy in central Spain. general time series model to floodplain fisheries in the Amazon.
Ann Agric Environ Med 17:139–151 Environ Modell Softw 48:202–212
 Int J Biometeorol

Veriankaitė L, Šaulienė I, Bukantis A (2011) Evaluation of meteorolog- Wold S, Ruhe A, Wold H, Dunn WJ (1984) The collinearity problem in
ical parameters influence upon pollen spread in the atmosphere. J linear regression. The Partial Least Squares (PLS) approach to gen-
Environ Eng Landsc 19(1):5–11 eralized inverses. SIAM J Sci Comput 5(3):735–743
Voukantsis D, Niska H, Karatzas K, Riga M, Damialis A, Vokou D Wold S, Sjöström M, Eriksson L (2001) PLS-regression: a basic tool of
(2010) Forecasting daily pollen concentrations using data-driven chemometrics. Chemometr Itell Lab 58:109–130
modeling methods in Thessaloniki, Greece. Atmos Environ 44: Zhou J, Liang Z, Liu Y, Guo H, He D, Zhao L (2015) Six-decade tem-
5101–5111 poral change and seasonal decomposition of climate variables in
Winters PR (1960) Forecasting sales by exponentially weighted moving Lake Dianchi watershed (China): stable trend or abrupt shift?
averages. Manage Sci 6:324–342 Theor Appl Climatol 119:181–191