Common ragweed (
The model is then applied and evaluated on a European domain for the period 2000–2010. To reduce the large uncertainties notably due to the lack of information on ragweed density distribution, a calibration based on airborne pollen observations is used. Accordingly a cross validation is conducted and shows reasonable error and sensitivity of the calibration. Resulting simulations show that the model captures the gross features of the pollen concentrations found in Europe, and reproduce reasonably both the spatial and temporal patterns of flowering season and associated pollen concentrations measured over Europe. The model can explain 68.6, 39.2, and 34.3 % of the observed variance in starting, central, and ending dates of the pollen season with associated root mean square error (RMSE) equal to 4.7, 3.9, and 7.0 days, respectively. The correlation between simulated and observed daily concentrations time series reaches 0.69. Statistical scores show that the model performs better over the central Europe source region where pollen loads are larger and the model is better constrained.
From these simulations health risks associated to common ragweed pollen
spread are evaluated through calculation of exposure time above
health-relevant threshold levels. The total risk area with concentration
above 5 grains m
One of the goals of the project “Atopic diseases in changing climate, land
use and air quality” (ATOPICA) (
Presently a number of regional models, mostly designed for air quality prevision, incorporate release and dispersion dynamics of pollen (Helbig et al., 2004; Sofiev et al., 2006, 2013; Skjøth, 2009; Efstathiou et al., 2011; Zink et al., 2012; Prank et al., 2013; Zhang et al., 2014). Methods for producing ragweed pollen emission suitable for input to regional scale models have been developed in recent studies (Skjøth et al., 2010; Šikoparija et al., 2012; Chapman et al., 2014). Due to a lack of statistical information related to plant location and amount within a given geographical area, the bottom up approach to produce plant presence inventories is unpractical for most herbaceous allergenic species like ragweed. Quantitative habitat maps for such species are often derived from spatial variations in annual pollen sum, knowledge on plant ecology and detailed land cover information by top-down approach (such as Skjøth et at., 2010, 2013; Thibaudon et al., 2014; Karrer et al., 2015). Lately, an observation-based habitat map of ragweed has been published in the context of the ENV.B2/ETU/2010/0037 project “Assessing and controlling the spread and the effects of common ragweed in Europe” (Bullock et al., 2012). This inventory is further calibrated against airborne pollen observations to reproduce the ragweed distribution with high accuracy, according to Prank et al. (2013). Recently Hamaoui-Laguel et al. (2015) used the observations collected in Bullock et al. (2012), combined with simplified assumptions on plant density and a calibration using observations to obtain a ragweed density inventory map. This approach made use of the Organising Carbon and Hydrology in Dynamic Ecosystems (ORCHIDEE) and the Phenological Modeling Platform (PMP) for obtaining daily available pollens (potential emissions) in Europe.
On average, one ragweed plant can produce 1.19
The timing of the emission can be estimated from a combination of phenological models and the species specific pollen release pattern driven by short-term meteorological conditions (Martin et al., 2010; Smith et al., 2013; Zink et al., 2013). Ragweed is a summer annual, short-day plant. Before seeds are able to germinate, it requires a period of chilling to break the dormant state (Willemsen, 1975). The following growth and phenological development depends on both temperature and photoperiod (Allard, 1945; Deen et al., 1998a). Flowering is initiated by a shortening length of day but could be terminated by frost (Dahl et al., 1999; Smith et al., 2013) or drought (Storkey et al., 2014). A number of phenological models have been developed for ragweed, either based on correlation fitting between climate and phenological stages (García-Mozo et al., 2009) or explicitly represented by biological mechanisms (Deen et al., 1998a; Shrestha et al., 1999; Storkey et al., 2014; Chapman et al., 2014). The mechanistic models take into account the responses of development rates to temperature, photoperiod, soil moisture, or stress condition (frost, drought, etc.). Mostly they are based on growth experiments but have to enforce a standard calendar date or a fixed day length for the onset of flowering when they are used in real conditions. While the airborne pollen observations from European pollen monitoring sites have a high year to year, site to site variability. Therefore it might be practical to combine the mechanistic model with correlation fitting when the knowledge of plant physiology and local adaptation of phenology are not sufficiently known at the moment.
In this paper, we present a pollen emission scheme that incorporate plant distribution, pollen production, species-specific phenology, flowering probability distribution, and pollen release based on recent studies. By combining the emission scheme with a transport mechanism a pollen simulation framework within the Regional Climate Model (RegCM) version 4 is then developed to study ragweed pollen dispersion behaviours on a regional scale. In Sect. 2 we provide a description of the RegCM-pollen simulation configuration, emission parameterization details, the processing of plant spatial density and observations data used for calibration in the study. In Sect. 3 we define the model experiment, explain the method used to calibrate ragweed density, present the simulation results of pollen season, evaluate the performances of the coupled model system over a recent period covered with observations, and finally present the climatological information about the ragweed pollen risk over European domain on a decadal timescale. Summary and conclusions appear in Sect. 4.
The development of RegCM-pollen model is based on the Abdus Salam International Centre for Theoretical Physics (ICTP) regional climate model, i.e. RegCM4, which has been used for a number of years in a wide variety of applications (Giorgi et al., 2006, 2012; Meleux et al., 2007; Pal et al., 2007). In this framework, we develop a pollen model for ragweed which calculates (i) the seasonal production of pollen grains and (ii) their emission and atmospheric processes (transport and deposition) determining regional pollen concentrations. As detailed hereafter pollen emission and transport are developed in the preexisting framework of the RegCM atmospheric chemistry module (Solmon et al., 2006, 2012; Zakey et al., 2006; Tummon et al., 2010; Shalaby et al., 2012). Pollen production is developed in the framework of the Community Land Model (CLM) version 4.5 (Oleson et al., 2013), which is the land surface scheme coupled to RegCM. Figure 1 gives an overview of such development framework. In the following subsections, we give details about the important data and steps of the development.
Ragweed pollen modelling within online RegCM-pollen simulation framework.
Pollen observations are central for calibration and validation of the pollen
module as discussed further. The pollen data are provided by the European
Aeroallergen Network (
Model domain and the observation sites with topography.
General information (2000–2010) for pollen observation sites. The annual pollen sum is calculated from 15 July to 31 October. Only years with data available exceeding 67 % between 20 July and 2 September are used to determine the observed start date and years with data available exceeding 56 % between 3 September and 18 October are used to determine the end date.
Ragweed pollen simulations are carried out for a European domain ranging from
approximately 35 to 70
Ragweed spatial distribution is obtained through a procedure discussed in
Hamaoui-Laguel et al. (2015). For country where
observations are available and of sufficient quality, ragweed distribution is
assumed to result from habitat suitability combined with infestation (not all
suitable habitats are populated). The habitat suitability is assumed to scale
as the product of the fraction of suitable land use surface
Pollen emission patterns on regional scale depend on plant density,
production, and meteorological conditions. The parameterization of pollen
emission flux is a modified version of Helbig et al. (2004). The vertical
flux of pollen particles
The characteristic concentration
Annual pollen production
In Eq. (2),
For simulating the timing of the flowering season, we adapt the mechanistic
phenology model of Chapman et al. (2014), which is based on growth
experiments (Deen et al., 1998a, b, 2001; Shrestha et al., 1999). Phenology
is simulated using BD accumulated for the current year of simulation and from
the first day (
where
The response of development rates to photoperiod is simulated using a
modified version of function presented by Chapman et al. (2014)
The response of development rates to soil moisture is assumed to occur from
the germination to seedling emergence stage. We use a linear function similar
to the one used to account for soil moisture impact on biogenic emission
activity factor in MEGAN (Guenther et al., 2012)
According to this phenology model, a total of about 25 BD are theoretically
needed to reach the beginning of pollen season BD
First guess
Average (2000–2010) annual pollen sum for first guess
Experimentally, pollen season can be defined in a number of ways from
observed pollen concentrations and listed for example in Jato et al. (2006).
A widely used definition is the period during which a given percentage of the
yearly pollen sum is reached. Another definition refers to the period between
the first and last day with pollen concentrations exceeding a specific level.
Looking at the temporal distribution of observations, particularly long
distribution tails can be found in some cases at the beginning and the end of
the pollen season, especially in stations where pollen levels are moderate.
This makes the definition of pollen season rather imprecise, while it is in
general more constrained in areas with high yearly pollen sum. In our
approach, we define the start of the pollen season from 44 observation
stations (described in Sect. 2.1) as the following: the first day of a series of 3 days
in a weekly window for which the pollen concentrations exceed 5
grains m
In Eq. (2), the
A first pollen run is performed using the first guess ragweed density
described in Sect. 2 and displayed in Fig. 3a. First guess density map shows
maxima of ragweed in the south-east of France, Benelux countries, and central
Europe regions. When comparing the resulting field to observation, simulated
concentrations obtained with the first guess distribution are generally
overestimated over France, Switzerland and Germany, underestimated in parts
of central Europe, and have comparable order of magnitude over some Italian
and Croatian stations (Fig. 4a). These important biases are in large part due
to assumptions made in the construction of the first guess plant density
distribution. In order to reduce these biases we perform a model calibration
by introducing a correction to the first guess ragweed distribution. For each
station, calibration coefficients are obtained by minimizing the yearly root
mean square error (RMSE) after constraining the decadal (2000–2010) mean
simulated pollen concentration to match the decadal mean observed
concentrations (2000–2010) within an admissible value. Calibration
coefficients obtained over each station are then interpolated spatially on
the domain using ordinary Kriging technique. Then a calibrated simulation
using the calibrated density distribution is carried out and repeated several
times. After three iterations, the correlation of yearly totals across
observation stations increase from 0.23 to 0.98 and the patterns are
clustering around the
The final calibrated ragweed distribution (Fig. 3b) shows high density in central Europe including Hungary, Serbia, Bosnia and Herzegovina, Croatia, and western Romania, northern Italy, western France, and also in the southern Netherlands and northern Belgium. The calibration adjusts the density over all the grid cells with ragweed presence by a factor ranging between 0.1 and 4.4 with an average of 0.98.
To estimate the error and sensitivity of this calibration method to the individual stations we implement a five-fold cross validation. The 44 sites are randomly divided into 5 groups. Five calibration experiments are conducted each time with one group left and used for validation respectively. The results of five validation groups are then combined to assess the final performance. With this approach a model measurements Pearson correlation of 0.54 is obtained together with a normalized root mean squared error of 21 % (Fig. 4c). Without surprise, this is less than when using the full data sets for calibration. In particular a few stations with particularly high concentrations protruding from surrounding sites (for example, ITMAGE and ROUSSILLON) have a large impact on the results of validation. We compared our cross validation (eight or nine sites left out each time) with three papers about ragweed pollen source estimation over the Pannonian Plain, France and Austria (Skjøth et al., 2010; Thibaudon et al., 2014; Karrer et al., 2015). Their cross validations (one site left out each time) show corresponding correlations of 0.37, 0.25, 0.63 and root mean squared error of 25, 16 and 3 %, respectively. Our results are within this range. We agree that caution should be taken in areas without a decent number of station coverage where the calibration cannot be done.
Note that through correction, other systematic sources of errors possibly affecting the modelling chain might also be implicitly corrected, leading to undesirable error compensations. However, after running additional tests (not shown here), for example varying model dynamical boundary conditions, a relatively small impact on pollen model performance is found when compared to the ragweed density distribution impact.
The simulated starting dates, central dates, and ending dates of pollen season are averaged from 2000 to 2010 and presented in Fig. 5. The pollen season generally shows a positive gradient from the south to the north and from low altitude to high altitude, resulting from the combined effects of temperature, day length, and soil moisture. The starting date varies between 21 July and 8 September. Flowering starts in the central European source regions earlier than in west and north of source regions. The central dates are reached between 1 August and 27 September, without noticeable difference between central and west source regions. Flowering ends in the central later than in the west of source regions. The pollen season is longest in the central main source regions.
Average pollen season (day of the year) from 2000–2010: start dates
Table 2 lists the statistical correlation between simulated and observed pollen starting, central, and ending dates. The model can reproduce starting and central dates better than ending dates. Goodness-of-fit tests show that the models account for 68.6, 39.2, and 34.3 % of the observed variance in starting, central, and ending dates. The RMSE is 4.7, 3.9, and 7.0 days for the pollen starting, central, and ending dates, respectively. The model reproduces the pollen season in the main source regions fairly well (Table 1), where the averaged differences between the simulated and observed pollen season progression are less or equal to 3 days and RMSE is lower than 6 days. For the areas with lower ragweed infestation the results vary widely. The starting dates and central dates are still reproduced well for a majority of the stations while the ending dates are more problematic with averaged differences above 6–10 days and RMSE over 8–12 days at some stations. This might result from patchy local ragweed distribution and the effect of long-range transport of pollen, which contributes to the determination of pollen season dates and are assumed to be representative of local flowering in our approach. Some stations also stop pollen measurement before the actual end of pollen season which leads to a lower accuracy of season ending date.
Statistical correlation between simulated and observed ragweed pollen season for fitting 2000–2010 and prediction (2011, 2012).
This phenology model is further tested for years of 2011–2012 and compared to observations (Table 2). Despite lower correlations, starting dates in both years and ending dates in 2012 are predicted reasonably well with 38.5, 28.7, and 26.1 % of the explained variance. The model however fails in predicting central dates in 2012 with low correlations to experimentally determined dates. Even so the prediction errors of RMSE for all dates in both years are well controlled and the differences between fitting and prediction RMSE are kept within 1.6 days, which means degradation of model performance has limited effects on the prediction of pollen season. Extending the fitting to several years of observation may contribute to improve the stability and robustness of the fitted threshold and further improve the phenology modellings of ragweed.
The evaluation of the model performance is made by comparing the modelled to
observed airborne pollen concentrations over the 2000–2010 period. In the
Taylor diagram on Fig. 6, we present an overview on how the models perform in
terms of spatio-temporal correlations, standard deviations, and RMSEs
compared to observations. The statistics are given for different timescales
of variability: daily, annual, or for the full 11-year period (in this case,
it is equivalent to spatial statistics only). Different variables are
analysed: the daily concentrations, the annual concentration sums, means, and
maxima, and the 11-year concentration sum, mean, and maxima. To plot all the
statistics on a single diagram, standard deviation and RMSE are normalized by
the standard deviation of observations at the relevant spatiotemporal
frequency: observations are thus represented by point OBS on the diagram
(perfect correlation coefficient, RMSE
Normalized Taylor diagram showing spatial and temporal correlations coefficients, standard deviations and RMSEs between simulations and observations for the period 2000–2010. Standard deviation and RMSE are normalized by the standard deviation of observations at the relevant spatiotemporal frequency.
Statistical measures between simulated and observed daily pollen
time series for each site: correlation coefficients
From the diagram, we can see that the model tends to perform very well when the variability is purely spatial and concentrations averages over the 11-year period (dots 5, 6 are very close to OBS). Not surprisingly it means the uncertainties are reduced to a large extent by the calibration procedure. However, the calibrated simulations do not capture the concentration maximum as well and tend to underestimate the measured spatial standard deviation (decade maximum dot 7 and also for the annual maximum dot 4). The model does not perform that well, but still shows some realism when the variability is involved in both spatial and temporal correlations. The yearly statistics, which reflect the interannual variation of pollen concentrations over the stations, are captured well with correlation coefficients all above 0.80 and normalized standard deviations of 0.89, 0.88, and 0.61 for concentration sum, mean, and maximum respectively. When scores are calculated for daily concentrations over all the stations, the overall spatial-temporal correlation coefficient reaches 0.69 for a relative standard deviation of 0.80.
Categorical statistics at thresholds of 5 (left
column), 20 (middle column), and 50 grains m
Daily variability is obviously the most difficult to simulate but is at the
same time the most relevant in terms of pollen health impact. To investigate this point
further, the model performance is regionally evaluated with both
discrete and categorical statistical indicators as listed in Zhang et
al. (2012). The discrete indicators considered in this study include
correlation coefficient, normalized mean bias factors (NMBF), normalized mean
error factors (NMEF), mean fractional bias (MFB), and mean fractional error
(MFE). NMBF
Model performance on simulation of daily average concentrations for 2000–2010.
The spatial distributions of correlation coefficient, NMBF, NMEF are shown in
Fig. 7. The correlations between simulated and observed daily time series are
above 0.6–0.7 in the central Europe source region and are mostly above
0.5–0.6 in the source regions of northern Italy and eastern France, while
the correlations are low in areas without strong local emission where the
majority of observed pollen may originate from long-range transport or
sporadic ragweed sources. Overall 56.8 % of the stations show an NMBF
within
A categorical evaluation is done by classifying the values of pollen
concentration with regard to the thresholds of 5, 20, and 50
grains m
With a reasonable confidence in model results, risks region can be identified
over the domain. Risk is defined from certain health-relevant concentration
thresholds. First we can consider minimum ragweed concentrations triggering
an allergic reaction. These thresholds are based on experiments involving
short exposure time to pollen and then extrapolated in order to define health
thresholds in terms of daily average concentrations. It is not known, whether
a short-time exposure to a large pollen concentration is equivalent to the
same dose when less pollen is inhaled over a longer period. Furthermore,
these thresholds vary largely between different regions and ethnic groups. The
likely range of such daily thresholds is 5–20 grains m
Annual footprint of ragweed pollen at the surface, obtained by selecting the maximum from daily averaged concentrations during the whole pollen season.
Footprints of ragweed pollen at the surface in each month during pollen season, average from 2000–2010, obtained by selecting the maximum from daily averaged concentrations in each month.
Number of days when the daily average concentration exceeding certain risk levels. Ground-based measurement locations are indicated with circles coloured by the measured number of days (left half) and corresponding simulated number of days (right half).
On this basis, simulated surface concentrations are post-processed to produce
24 h average concentrations. The footprints of ragweed pollen risk are then
obtained by selecting the yearly and monthly maximum from daily averaged
concentrations. The yearly and monthly maximums are averaged over the decade
(2000–2010) to produce footprints depicted in Figs. 9, 10). The risk is
divided into 16 levels to reflect the range of health relevant threshold used
in different countries and regions as listed in Table 4.3 of Bullock et
al. (2012). The numbers of grid cells at different threshold risk levels are
given in Table 4. Hereafter we select some of the representative risk levels
to be discussed in more detail. From annual footprint of ragweed pollen
spread risk, the area with concentration
Percent area with the surface concentration of ragweed pollen at different risk levels, average for 2000–2010.
Temporally, the pollen risk is determined by seasonal evolution (Fig. 10).
August is in general the month contributing the most to the annual risk
footprint, with an average concentration of 25.6 grains m
Besides the triggering of allergic reactions at a certain threshold, the time
of exposure above a certain threshold might also be important, e.g. in terms of
sensitization to ragweed pollen. To assess a risk based on this criterion,
exposure time, expressed as the decadal average of the number of days per
season above a certain threshold, is calculated and reported in Fig. 11.
Relevant thresholds are 5, 10, 20, 50 grain m
The longest exposure times occurs in Pannonian Plain at all thresholds,
reaching for example about 30 days above 20 grains m
This study presents a regional-climatic simulation framework based on RegCM4
for investigating the dynamics of emissions and transport of ragweed pollen.
The RegCM-pollen modelling system incorporates a pollen emission module
coupled to CLM4.5 and a transport module as part of the chemistry transport
component of RegCM. Because climate, CLM4.5 and chemistry components are
synchronously coupled to the RegCM model, this approach allows dynamical
response of pollen ripening, release, and dispersion to key environmental
drivers like temperature, photoperiod, soil moisture, precipitation, relative
humidity, turbulence, and wind. Through the pollen production link to NPP,
other environmental and climate relevant factors as atmospheric CO
The RegCM-pollen framework is applied to the European domain for the period 2000–2010. Comparing with the observed flowering season, the model can reproduce starting dates and central dates well, with 68.6 %, 39.2 % of the explained variance and 4.7, 3.9 days of RMSE in starting date and central date, respectively. The pollen season in the main source regions are reproduced fairly well while in the areas with lower ragweed infestation the deviations are evident. The model in general captures the gross features of the pollen concentrations found in Europe. Statistical measures of NMBF, MFB, and MFE over the domain fall in the range of recommendation for a good performance while NMEF is a bit large with a value of 0.83. The model performs better over the central European source region, where the daily correlations at most stations are above 0.6–0.7 and NMEF lie within 1.0. Performance tends to degrade in France and northern Italy. Still, the values of NMEF for pollen simulation are generally consistent with what is expected from operational air quality models for aerosols for example. Categorical evaluation reveals the model tends to give better predictions for high threshold while giving more false alarms for low threshold. A better performance is also shown over the central European source region at all levels, with correct predictions above 80 % and false alarms within 20 %.
The multi-annual average footprints of ragweed pollen spread risk are
produced from calibration simulations. The pollen plume with concentration
The modelling framework presented here allows simultaneous estimation of ragweed pollen risk both for hindcast simulations (including sensitivity studies to different parameters) and for study of potential risk evolution changes under future climate scenarios as illustrated in Hamaoui-Laguel et al. (2015). Still a long list of uncertainties hinders an accurate estimate of the airborne pollen patterns and risk within the presented framework. Caution should also be taken while interpreting the results in areas without a dense observational network and where calibration is weaker. In this regard, challenging research efforts should focus on a better characterization of ragweed spatial distributions and biomass, in addition, a better understanding of phenological process and the dynamic response of release rate to meteorological conditions will help to reduce these uncertainties and improve model performance. An accurate and diverse observation of ragweed phenology is therefore essential to better represent local flowering and there is also a need for experimental observations to better constrain the release model. In parallel, systematic ragweed pollen concentrations should be further developed as part of air quality networks and public access to data should be promoted.
CRUNCEP atmospheric forcing data are available at
This research is funded by the European Union's Seventh Framework Programme
(FP7/2007–2013) under grant agreements no. 282687 Atopica (