Pollen samples are obtained using volumetric suction pollen trap based on the impact principle (Hirst, 1952), the standardized method in European aerobiological networks (Galán et al., 2014). The Hirst sampler (Hirst, 1952) is calibrated to handle a flow of 10 L of air per minute, thus matching the human breathing rate. Pollen grains are impacted on a cylindrical drum covered by a Melinex film coated with silicon fluid (LANZONI srl^®) as trapping surface. The drum rotates at 2 mm per hour; so, each 48 mm represents 24 h of continuous sampling. The drum is changed weekly, and the exposed tape is cut into pieces, with each one corresponding to 1 d. Pollen grains are counted under a light microscope at 600× magnification. Daily average pollen counts are obtained following the standardized Spanish method (Galán et al., 2007), consisting of running four longitudinal sweeps along the 24 h slide for daily data and identifying and counting each pollen type found. To obtain the hourly concentrations, 24 continuous transversal sweeps separated every 2 mm along the daily sample slide are analyzed, since the drum rotates at a speed of 2 mm per hour. Daily and intra-diurnal (hourly) pollen concentrations are obtained by converting the pollen counts into particles per cubic meter of air, taking into account the proportion of the sample surface analyzed and the air intake of the Hirst pollen trap (10 L min-1).

The profiles of the particle backscatter coefficient and the volume and particle depolarization ratios were measured with the Barcelona micro-pulse lidar (MPL) system (Sigma Space Corporation; model MPL-4B). The system is part of the MPLNET (Micro-Pulse Lidar Network; http://mplnet.gsfc.nasa.gov/, last access: 1 December 2021; Welton et al., 2001) network. The MPL system is a compact, eye-safe lidar designed for full-time unattended operation (Campbell et al., 2002; Flynn et al., 2007; Welton et al., 2018). It uses a pulsed solid-state laser emitting low-energy pulses (∼ 6 µJ) at a high pulse rate (2500 Hz) and a co-axial transceiver design with a telescope shared by both transmit and receive optics. The Barcelona MPL optical layout uses an actively controlled liquid crystal retarder which makes the system capable of conducting polarization-sensitive measurements by alternating between two retardation states (Flynn et al., 2007). The signals acquired in each of these states are recorded separately and called “co-polar” and “cross-polar”. In a nominal operation, the raw temporal and vertical resolutions are 30 s and 15 m, respectively.

The total lidar signal, PT, as a function of the altitude, z, is reconstructed from the co-polar, Pco, and cross-polar, Pcr, signals as follows (Flynn et al., 2007): 1PTz=Pcoz+2Pcrz. The particle backscatter coefficient, βp, was retrieved with the two-component elastic algorithm (also known as the Klett–Fernald–Sasano method; Fernald, 1984; Sasano and Nakane, 1984; Klett, 1985), with a constant lidar ratio of 50 sr, and applied to the total lidar signal, PT.

By adapting the notations of Flynn et al. (2007) to ours, one can formulate the linear volume depolarization ratio, δV, for the MPL system as follows: 2δVz=PcrzPcoz+Pcrz. The linear particle depolarization ratio, δp, can then be determined by Eq. (4) of Sicard et al. (2016a). Finally, the extraction of the total pollen backscatter coefficient, βpol, out of the particle backscatter coefficient is made thanks to the pollen depolarization capabilities (see Sicard et al., 2016a, and references therein).

We also calculated the vertical height, hpol, up to which the pollen plume extends. As shown in Sicard et al. (2016a), the pollen plume is characterized during the entire pollination event by a near-constant or slightly decreasing profile of βpol. From this aspect, the structure of the pollen plume is much simpler than the atmospheric boundary layer structure usually found in Barcelona (Sicard et al., 2006) and allows us to use a simple threshold method (Sicard et al., 2016a).

All the MPL retrievals presented in this work were performed with in-house algorithms and not with the MPLNET processing, since the system only entered the network in 2016.

The MONARCH model implements a mass-based aerosol scheme that has been extended to pollen bioaerosols. Table 2 summarizes the main characteristics of the implementation. Pinus pollen is represented as a spherical particle with a geometric diameter, DPinus, of 59 µm (Jackson and Lyford, 1999) and a dry mass density, ρPinus-dry, of 560 kg m-3 (Jackson and Lyford, 1999). The shape of Pinus grains is not completely spherical, but for the sake of simplicity, we make this assumption in the model. The aerosol life cycle is strongly affected by the water uptake. In our implementation, we consider Pinus pollen as being a hydrophilic particle. Griffiths et al. (2012) describe the impact of relative humidity on the pollen density and radius. Following their results, we have implemented an increase in the Pinus pollen density with the relative humidity (RH), while the diameter of the particle is considered constant and independent of the water uptake. This effect is introduced in the model by considering RH-prescribed mass-based growth factors (ϕ=massPinus/massPinus-dry), derived from Fig. 1 of Griffiths et al. (2012) and shown in Table 3. Thus, the density of the particle is computed as follows: 3ρPinus=ρPinus-dry⋅ϕ, where ρPinus is the mass density of the pollen with water uptake. ρPinus is used in the calculation of sedimentation, dry deposition, and wet deposition. The numerical schemes for all these processes are described in Pérez et al. (2011).

The emission scheme implemented in MONARCH is based on the concepts of the parameterization of Helbig et al. (2004), with some modifications. It computes the vertical emission flux of Pinus pollen grains per grid cell x and unit time t, as follows: 4EPinusx,t=Px⋅Ft⋅Rx,t⋅C, where EPinus is the vertical emission flux of Pinus pollen (kilograms per square meter per second; hereafter kg m-2 s-1), P is a characteristic mass concentration of the pollen grains available in the canopy of the pine trees (kilograms per cubic meter; hereafter kg m-3), F is the unitless phenology function scaling P through the pollination event, with values ranging from 0 to 1, R is the weather-dependent release scaling function (meters per second; hereafter m s-1) which depends on instantaneous meteorological conditions, and C is a global calibration factor accounting for the uncertainty of the model schemes. Note that P is not the actual pollen released in the atmosphere by the tree; instead, this will depend on the environmental conditions that may favor the release of the pollen grains from the tree described by R.

The available number of pollen grains per tree during a season is a highly uncertain parameter. Several works provide estimates of this parameter for pines. Tormo Molina et al. (1996) calculated the annual production of pollen grains for three individual pine trees (P. pinaster) in good health and shape located in southwestern Spain, obtaining values of 20.9 billion, 32.3 billion, and 22.2 billion pollen grains. Parker and Blush (1996) reported, for an open-grown P. taeda tree, a production of 8 billion pollen grains during a season. And Williams (2008) emphasized the fact that pollen production depends on the tree age and the size of the crown and increases with both. In this work, we use the average of Tormo Molina et al. (1996) for P. pinaster as a representative estimate of the emission of type of pine trees present in the area of interest, though this is a highly uncertain parameter. Thus, we formulate P as follows: 5Px=Grain⋅Mγ⋅Hs, where Grain is the production of pollen grains in mass per tree during the season (kilograms per tree), M is the number of trees in a grid cell, Hs is the average canopy height for Pinus trees (meters), and γ is the area of a grid cell (square meters; hereafter m2). Grain is the average of the results presented by Tormo Molina et al. (1996), with 25.1 billion pollen grains, transformed into mass using the Pinus grain size and density assumed in Sect. 3.1. Different values of Hs are found in the literature for pine trees. The Forest Inventory of Catalonia (Gracia et al., 2000–2004) provides an average height of 11.3 m for the area of study, which is the value used in our work. No detailed information is available to describe Mγ for the domain of study; thus, we approximate the term with available information of the Pinus tree density as follows. The spatial distribution of the pine species of interest in our study, which pollinates from February to April in Catalonia (P. halepensis; P. pinea), is obtained from the cartography of habitats of Catalonia (Carreras et al., 2015). The cartography has been remapped to 1 km resolution data set and then combined with the pine tree density reported by the Forest Inventory of Catalonia (Gracia et al., 2000–2004). The pine density in Catalonia ranges between 484 and 957 trees per hectare. With this information, a database with the number of pine trees per grid cell of the model is derived to estimate M (see Fig. 1a). This data set is complemented with the inventory of pine trees of the Barcelona city council that provides specific georeferenced information of the trees of the city. Note that the available pollen grains for the season are reduced by the already emitted amount of pollen grains at each model time step.

Once P is estimated, the emission scheme needs a function that describes how the availability of pollen grains in the tree evolves during the pollination season. This function is known as the phenology function, F. Not all the pollen that the tree will release during a season is available on the first day of the pollination; this will evolve during the season, depending again on the meteorological and climatological conditions that have affected the plant during the last year. Several efforts have been made to develop phenology functions for different types of plants based on relationships between pollen counts and key meteorological factors (Jones and Harrison, 2004; Schueler and Schlünzen, 2006; Laursen et al., 2007; Linkosalo et al., 2010; Marceau et al., 2011) and have been used in regional models to predict airborne concentrations of different types of pollen (Helbig et al., 2004; Sofiev et al., 2006, 2013; Zink et al., 2013; Zhang et al., 2014; Sofiev et al., 2017). All current schemes show limitations in the prediction of the onset and duration of the pollen season and may have significant biases for very localized and specific pollen events, similar to the one we are studying in this work. In this sense, we decided to implement a much simpler function that is mainly constrained by the observed pollen counts in the aerobiological station of Barcelona. The evolution of the pollen concentrations during the period 1 March–30 April 2015 has been fitted with a Gaussian function that describes the evolution of the event (Fig. 2). A similar approach has been used in other works (e.g., Wozniak and Steiner, 2017). The resulting normalized function has been selected as the phenology function F of the emission scheme. With this simple approach, we are assured that the emission scheme will release an amount of pollen in the atmosphere that is reasonable to reproduce the observed pollen concentrations in the atmosphere retrieved by the lidar and the Hirst collectors in the region of Barcelona.

The weather-dependent function R implemented in the model accounts for the mobilization effect of the wind and turbulence of the near-surface layer of the atmosphere as follows (Helbig et al., 2004): 6Rx,t=Ke⋅u∗, where Ke is a wind effect scale factor between 0 and 1 (unitless), and u∗ is the friction velocity (m s-1). Ke is parameterized using a threshold friction velocity, u∗te, as follows: 7Ke=1-u∗teu∗,u∗>u∗te0,u∗≤u∗te8u∗te=α⋅u∗t. The threshold friction velocity is the product of a resistance term, α, and a standard threshold friction velocity, u∗t (m s-1). The expression of u∗t is the regression formula of Greeley and Iversen (1985), based on wind tunnel data for sand erosion. The resistance term α is used to distinguish the different natures of sand erosion on the ground and the pollen release above the canopy height, and it is defined as follows: 9α=u10eu10, where u10e is an empirical threshold wind speed for the type of tree (m s-1), and u10 is the 10 m wind speed of the model (m s-1). No specific values of u10e are found in the literature for pine trees; here we set u10e to 2.9 m s-1, using the value of Helbig et al. (2004) for alder. Zhang et al. (2014) did not find significant sensitivity to different values of u10e.

To study the dispersion of Pinus pollen in the region of study, the MONARCH model has been configured with four domains, using the telescoping nest capability of the system (see Fig. 1b). A parent domain covering central Europe, and centered over the region of interest, is set with 27 km horizontal resolution and 48 vertical layers, with the top of the atmosphere at 50 hPa. A total of three nested domains are defined to increase the model resolution up to 1 km in the innermost domain, with a ratio of 1 : 3 from one nest to the other.

The simulations cover the period 20 March to 2 April 2015, which comprises the pollination event under study in this work. The meteorological initial and boundary conditions are obtained from the ERA-5 reanalysis at 30 km horizontal resolution and 6 h frequency. No boundary conditions for the Pinus pollen are prescribed in the parent domain, as the contribution from long-range transport is considered minor compared with the local emissions. The model outputs the Pinus pollen mass concentration over the domain of study with an hourly resolution that is converted to number concentration as a diagnostic tool.

Note that the Pinus pollen mask used in all domains is the one described in Sect. 3.2 and only considers the trees of the Catalonia region pollinating during March and April. The main objective of the work is to assess the vertical distribution of pollen nearby the Barcelona site where the lidar is available. We consider that the Pinus pollen detected by the lidar and collected by the Hirst instruments is mainly of local origin, as the size of the pollen is significantly coarse and limits the long-range transport of this type of bioaerosol. To explore this initial hypothesis, the following two different runs were conducted: (1) a base case simulation, using the model parameters described in Sect. 3.1, and (2) an enhanced simulation, where the lifetime of the Pinus pollen is increased by assuming that the sedimentation and dry deposition of the pollen is half of what is estimated in the base case. A global calibration factor C of 2.9 was used to minimize the error of the base case simulation with respect to the Hirst observations. Other works use C as a 2D spatial factor to calibrate the emission flux better due to the large uncertainties of most pollen emissions schemes and input data sets (e.g., Kurganskiy et al., 2020). In Sect. 5, the model results for both scenarios are discussed with the Hirst and lidar measurements.

From the analysis of the samples of the Hirst collectors, the concentration of the pollen type i, C#,iH (pollen per cubic meter; hereafter Pollen m-3), is obtained. The manual procedure to do so includes neither annotating information on the size of the pollen grain, nor any kind of information on its optical properties. This observation is taken as the reference value against which the lidar retrieval is constrained and to which the model output is compared.

The lidar instrument measures the particle backscatter coefficient from which the total pollen contribution, i.e., the total pollen backscatter coefficient, βpol (meters per steradian; hereafter m-1 sr-1), can be extracted as in Sicard et al. (2016a). The total pollen mass concentration, Cmlid (kg m-3), is obtained using the following relationship: 10Cmlid=βpolLRσ∗, where LR (steradian) is the lidar ratio, and σ∗ (square meters per gram; hereafter m2 g-1) is the specific extinction cross section. The latter is actually not a fixed constant, as it will depend on the taxa present, their concentration, and weight, as well as their optical properties. It is thus virtually unknown.

The mass concentration of Eq. (10) can be converted into number concentration of the pollen grain i, C#,ilid (Pollen m-3), as follows: 11C#,ilid=Cmlidwfi4π3Di23ρi=βpolLRwfiσ∗4π3Di23ρi, where for each pollen grain iwfi is the weight factor, Di (meters) is the diameter, and ρi (kg m-3) is the mass density. It is important to note that, in this last conversion, the volume of a pollen grain is calculated assuming that the grain has a spherical shape, in agreement with the assumption made in the model. wfi represents the fraction of the mass of the pollen grain i to the total pollen mass. It is calculated on a daily basis, using the daily concentration measurements of the Hirst collectors as follows: 12wfi=C#,iH4π3Di23ρi∑uC#,uH4π3Du23ρu=C#,iHD3iρi∑uC#,uHD3uρu. In Eq. (11), the product Cmlidwfi represents the mass concentration estimated from the lidar-measured total backscatter coefficient for the pollen grain i and is directly comparable to the model output. This implies that the choice of Di and ρi has no impact on the comparisons of the lidar and the model.

Finally, the model provides mass concentration for the pollen grain i, Cm,imod (kg m-3). The number concentration of the pollen grain i is then calculated as follows: 13C#,imod=Cm,imod4π3Di23ρi. The next step is to optimize all parameters to minimize the difference between the two observations, C#,iH and C#,ilid, and then apply these optimum parameters to compute C#,imod and compare it to both C#,iH and C#,ilid.

The lidar number concentration, C#,ilid, for the pollen grain i depends on the following four parameters, namely LR, σ∗, Di, and ρi, while the model number concentration for the pollen grain i depends only on two parameters, namely Di and ρi. The conversion of the lidar retrieval into number concentration has 3 degrees of freedom, while the conversion of the model output has only 1.

Each day, for the 24 h (t=1, 2, 3, …, 24) between 00:00 and 23:00 UTC, we use the least squares method to obtain the optimal specific extinction cross section by minimizing the sum of squared residuals expressed by the following: 14∑tC#,iHt-C#,ilidt2=∑tC#,iHt-βpoltLRwfiσ∗4π3Di23ρi2. This minimization can be done in terms of LR, σ∗, Di, or ρi. As the greatest unknown is without any doubt σ∗, all other parameters (LR, Di, and ρi) are fixed. However, we also perform a sensitivity analysis on these parameters to quantify the range of uncertainty associated to σ∗ (see Sect. 4.3). The minimization of Eq. (14) yields the best estimate solution, in terms of σ∗, as follows: 15σ∗=LRwfi4π3Di23ρi∑tβpolt2∑tC#,iHtβpolt. The weight factor wfi for Pinus pollen is calculated as the daily mean of the hourly fractions of mass concentration presented in Sicard et al. (2019). Daily values of wf(i) are reported in Table 4. The mean value over the whole event is 0.901, which is equal to saying that Pinus pollen represents 90.1 % of the mass of total pollen. The main assumption made to calculate the Pinus mass fraction is that the pollen mixture is assumed to be made only from Pinus and Platanus taxa, whereas Cladosporium and Cupressaceae were also present (Sicard et al., 2016a). However, in terms of number concentration, one sees from Sicard et al. (2019) that Pinus and Platanus represent more than 90 % of the amount of total pollen. Our assumption is equivalent to neglecting the remaining ∼ 10 %. As Cladosporium and Cupressaceae are much smaller in size than Pinus, our assumption is equivalent to neglecting, in terms of mass, probably less than 1 % of the total mass. LR is set to 50 sr, which is a reasonable value for pollen, as justified in Sicard et al. (2016a). The Pinus pollen diameter DPinus and mass density ρPinus are taken from Jackson and Lyford (1999) and are set to 59 µm and 560 kg m-3, respectively, in agreement with the model (see Sect. 3.1). With these values set, we calculate σ∗ with Eq. (15) and find the daily values reported in Table 4. Values of σ∗ oscillate between 0.78 and 1.67 m2 g-1. A factor of almost 2 between the minimum and maximum values is not that surprising, taking into account the lengthy methodology followed and, in particular, the extraction of the contribution of pollen out of the total backscatter coefficient and the hypothesis made that the first lidar measurement (at 225 m) is a good proxy of what it would be at ground level. As an example, if we compare our pollen values of σ∗ with the ones of other large particles, we find that our values are larger than mineral dust observed in southern Europe, which has values of the order of 0.5–0.6 m2 g-1 (Pérez et al., 2006), and they are smaller than values for sea salt surrogate estimated to be of the order of 6 m2 g-1 (Radney et al., 2013). These results are consistent, since pollen grains are much less dense in mass (and, thus, have a larger σ∗) than mineral dust and because of the strong scattering properties (and, thus, large σ∗) of sea salt.

As an illustration of how the lidar and Hirst number concentration agree with the values of σ∗ that are found, a scatterplot of the lidar vs. Hirst concentration on 27 March (σ∗=1.21 m2 g-1) and the daily cycle of both concentrations are shown in Fig. 3. The dates of 27 March is chosen because, according to Sicard et al. (2016a), it is the day when all column properties correlate the best with the Pinus pollen surface concentration. The scatterplot is rather disperse around the 1 : 1 line (Fig. 3a); hence the correlation coefficient, equal to 0.55, is not very high. One sees that both lidar and Hirst concentration daily cycles agree relatively well (Fig. 3b), except in the very first hours of the day when no pollen is detected in the column, whereas it is present at the surface.

In order to quantify for the pollen type of interest in this study, Pinus, the sensitivity of σ∗ to the rest of parameters, LR, DPinus, and ρPinus, we vary these latter values and resolve Eq. (15). LR, DPinus, and ρPinus are varied around their nominal value (see Sect. 4.2) as follows: 50±10 sr, 59±10 µm, and 560±200 kg m-3, respectively. Although the estimation of σ∗ is made on a daily basis, it is not necessary to perform the sensitivity study for the 5 d of the event, since the relative variations in σ∗ as a function of the same ΔLR, ΔDPinus, and ΔρPinus from 1 d to another will not be different. The sensitivity analysis is, thus, performed for 27 March, which is the day when all column properties correlate the best with the Pinus surface concentration (Sicard et al., 2016a). The dependency of σ∗ on the rest of parameters is shown in Fig. 4. Obvious results are that σ∗ decreases with decreasing LR and increasing DPinus and ρPinus. The strongest sensitivity of σ∗ is observed for DPinus (for LR and ρPinus are equal to their nominal values; σ∗ varies between 0.76 and 2.12 m2 g-1 when DPinus=59±10 µm), while the lowest one is observed for LR (for DPinus and ρPinus are equal to their nominal values; σ∗ varies between 0.97 and 1.45 m2 g-1 when LR=50±10 sr). In these conditions, when all parameters are fixed to their nominal values, σ∗ will vary by -0.45/+0.91 m2 g-1 (-37/+75 %) around its nominal value of 1.21 m2 g-1 if DPinus differs from its nominal value by ±10 µm, it will vary by -0.32/+0.68 m2 g-1 (-26/+56 %) around its nominal value if ρPinus differs from its nominal value by ±200 kg m-3, and it will vary ±0.24 m2 g-1 (±20 %) around its nominal value if LR differs from its nominal value by ±10 sr. In the worst-case scenario, when all parameters are relaxed, σ∗ could vary down to 0.45 (-63 %, with regard to its nominal value) and up to 3.95 (+226 %, with regard to its nominal value). This worst-case scenario is, however, very unlikely to happen for various reasons. First, all parameters (LR, DPinus, and ρPinus) should take their extreme values simultaneously, and this is rather improbable. Second, the probability of LR taking values of 40 (50 - 10) or 60 (50 + 10) sr is very low, as suggested by Noh et al. (2013), who found a mean columnar lidar ratio of 50 sr and a standard deviation not greater than 6 sr during a 6 d pollination event (mostly dominated by Pinus and Quercus pollen) in South Korea. Third, in terms of grain diameter, although there is a large range of Pinus grain diameter measured in different regions of the planet, i.e., 46 µm for Japanese black pine in Japan (Hirose and Osada, 2016), 51±7 µm for P. sylvestris in China (Song et al., 2012), 52 µm for P. sylvestris in the USA (Durham, 1946), 55±4 µm for P. sylvestris in Finland (Varis et al., 2011), and 51–100 µm for P. sylvestris in Austria (Halbritter, 2016), the species considered in this study (P. halepensis and P. pinea) have a narrower range of mean diameters. Roure (1985) measured the diameters of these species in the Iberian Peninsula as being 41–55, 43–65, and 47–65, respectively. The range explored in our sensitivity study, 49–69 µm, corresponds to ±10 µm around the value of 59 µm measured for P. sylvestris by Dyakowska and Zurzycki (1959) and referenced in Jackson and Lyford (1999). Fourth and last, Pinus grain density measurements are given pretty seldom in the literature. Our nominal value of 560 kg m-3 was taken from Jackson and Lyford (1999). Other values of 450 kg m-3 have been estimated for P. sylvestris in the USA (Durham, 1946) and of 550 kg m-3 for Japanese black pine in Japan (Hirose and Osada, 2016). The range explored in our sensitivity analysis, 360–760 kg m-3, is thus conservative, and extreme values of 360 and 760 kg m-3 are rather unlikely to be reached.

The event of interest took place between 27 and 31 March 2015, the most intense period of the Pinus pollination season in the region (Sicard et al., 2016a). The 5 d of the event are denoted as the following: 27M, 28M, 29M, 30M, and 31M. A detailed analysis of the event is presented in Sicard et al. (2016a). In the second half of March 2015, a strong anticyclone positioned in the Atlantic Ocean west of the Portuguese coast generated northwesterly winds in the northeastern part of the Iberian Peninsula. The event was characterized by medium–strong winds in the medium troposphere (500 hPa) veering from north to northwest during the 5 d of the period and resulting in northwesterly winds near the surface and clear skies in Barcelona most of the time. Figure 5a shows the hourly number concentration of the total pollen and Pinus pollen measured in the Barcelona downtown with the Hirst instrument during the event. Total pollen hourly concentrations reached values higher than 6000 Pollen m-3 on 31M. On every day, the Pinus pollen shows a diurnal cycle with several concentration peaks of major intensity. Figure 5b represents the fraction of the mass concentration of Pinus pollen, calculated according to Sicard et al. (2019). Pinus pollen represents 90 % of the mass of total pollen. As Pinus pollen grains are much larger than the rest of the pollen types present during this event, their contribution in mass dominates. In this work, we focus the analysis on the Pinus pollen because it is the taxon with the major contribution to the total mass of pollen measured in Barcelona, and the emission source is widely distributed across the domain of interest (see Fig. 1a). The latter allows a better analysis of the role of short-to-medium range transport and vertical mixing processes in the dispersion of this type of pollen.

Figure 5c also shows the quicklook plot, also called time–height plot, of the volume depolarization ratio of the lidar. It is not a quantitative plot, since both the molecules and the particles contribute to the volume depolarization, but it is an indicator of the presence (or not) of depolarizing particles (here, pollen grains). In Fig. 5c, the dark green areas represent the molecular level (only detectable at the high temporal resolution of the quicklook during nighttime because of the reduced background signal – as opposed to daytime), light green areas represent low-depolarizing particles (urban background), and yellow/orange areas represent high-depolarizing particles, i.e., pollen. The vertical distribution of the airborne pollen also shows a clear diurnal cycle, with usually no or weak nighttime activity in the upper layers, where the pollen grains are probably staying very close to the ground. The diurnal cycle is marked by an increase in the amplitude and height of the volume depolarization ratio, starting around 10:00 UTC, and a decrease, starting before 16:00 UTC. This diurnal pattern is observed on each single day of the pollination event. On the first 4 d, the volume depolarization ratio has come back to its background value (<0.08) before 20:00 UTC. Most of the aerosol load is usually found below 1.5 km. While the peak of Pinus pollen hourly concentration at the surface (1266 Pollen m-3) is reached on 30M at 08:00–09:00 UTC, the peak in the column (quantified by the pollen aerosol optical depth (AOD) in Sicard et al., 2016a) is reached 4 h later at 12:00–13:00 UTC. Note that, even though weak or no nighttime activity is detected by the lidar, the ground measurements indicate that periods with significant nighttime pollen concentrations (e.g., the first hours of 28M and 31M) are present. This is a good indication of the stratification of the pollen in the lowermost layers during the nighttime.

In order to understand the dynamics of Pinus pollen transport during the event, we present here the model results over the innermost domain (D04) at 1 km horizontal resolution of the base case configuration described in Sect. 3. The Pinus number concentration at the surface (first model layer, around 24 m a.g.l. – above ground level), from 27M to 31M, is shown in Fig. 6. The entire event is characterized by a dispersion of pollen grains towards the Mediterranean Sea, mainly driven by the northwesterly winds. The highest concentrations are in the southwestern region of the domain. This area is dominated by the natural channelization of the Ebro valley, and a characteristic northwesterly wind, called cierzo, develops under west–northwest advection situations. Regular outbreaks of plumes with high concentrations of pollen are transported hundreds of kilometers to the Mediterranean Sea. This effect is less pronounced towards the north of the domain. Note that the regions with pines are found within the first 150 km inland from the coast (see Fig. 1a); beyond that distance, Pinus trees are scarce. The following three main areas have been identified from the pine tree density map used in the model: (1) in the south, a wide area with medium–high pine density, (2) in the central coast, the area with pines is even more extensive, with high densities of trees, and (3) in the northern coast, the presence of pines is more reduced, but there are still some localized regions with high densities of trees present. This tree distribution and the strong winds in the southwest of the domain explain the spatial gradient in concentrations simulated by the model.

To assess the representativeness of the model results, we have compared them with the hourly meteorological and daily aerobiological observations described in Sect. 2. Here, we focus on surface observations. The quality of the results is quantified by means of classical statistics (Pearson correlation coefficient, r, root mean square error, RMSE, and bias), using seven aerobiological stations of the XAC network and seven meteorological stations of the XEMA network (details on the localization are in Fig. 1a and Table 1). The variables evaluated are hourly 2 m air temperature, 10 m wind speed and direction, irradiance, 2 m relative humidity, and daily Pinus number concentration at the first model layer.

Regarding the meteorological results, Table 5 presents the statistics for the 9 km base case run and all seven stations (results for 3 and 1 km in the Tables S1 and S2 in the Supplement), and Fig. 7 shows the temporal evolution of the same variables (except concentration) at the three model resolutions for Barcelona and Bellaterra/Sant Cugat (results at the other five stations in Fig. S1 in the Supplement). The statistics indicate a good agreement for the temperature, irradiance, and relative humidity in most of the stations (correlations above 0.9 and low bias), while higher errors are observed in the wind. Overall, the results are within the typical performance range of mesoscale models in the area of study (i.e., Jiménez-Guerrero et al., 2008), with surface winds being one of the most difficult variables to reproduce in coastal regions with a complex terrain like the one under study. On the one hand, the model results in coastal sites may show higher errors in temperature, winds, and relative humidity. These sites are close to the sea–land interface, and small inaccuracies in the representation of the coastline may have a strong impact in the results. The Tarragona site is a clear example where the temperature results are degraded with the 1 km domain compared with the upper nests (see the Supplement) as a result of the land and sea grid cells represented in each model resolution surrounding the site. An excessive influence of marine air masses will result in a lower thermal amplitude and overestimated relative humidity. On the other hand, better statistics are obtained at inland sites where the topography is properly captured by the model resolution. Most sites show RMSE and bias below 2.6 and ±1.2 ∘C, 2 and ±0.5 m s-1, 90 and ±11∘, and 15 % and ±6 % for temperature, wind speed, wind direction, and relative humidity, respectively.

The Barcelona site is representative of a coastal station, located few kilometers off the shore, while Bellaterra/Sant Cugat is an example of an inland site not affected by major mountain ranges (i.e., the Pyrenees). At both stations an improvement is detected in most meteorological variables with the increase in resolution (see Fig. 7). The model reproduces the daily cycle of the temperature and relative humidity and captures the cloudiness observed during 29M and 30M. As noted in the statistics, more disagreement is seen in the wind speed, and the different resolutions show larger variability compared with the other variables. The wind results in the Barcelona site are consistent among the model resolutions. Some systematic bias is observed in the wind speed at the end of the day and during the first hours of the following day. The model overestimates the calm winds observed during nighttime, and it tends to underestimate the morning peak (i.e., transitions from 27M to 28M, 29M to 30M, and 30M to 31M). This results in lower relative humidity and higher temperatures during the calm periods compared with observations, although this does not impact the wind direction that is well captured most of the time. The results at the Bellaterra/Sant Cugat site show a clear improvement with the resolution. The site is located in a long valley surrounded by two mountain ranges that are better represented in the model by increasing the resolution. The overestimated wind during nighttime in Bellaterra/Sant Cugat results in the model wind veering from southwest to northwest (i.e., 27M and 29M), while observations show a calm, stagnated situation. Such an error leads to a model underestimation of the relative humidity and an overestimation of the temperature there. On the night of 31M, the observations show weak easterly winds, while the model develops northwesterly winds. Albeit that the limitations identified, the model captures the meteorology of the event under study reasonably well.

The main driver of the pollen release in the atmosphere is the wind. The period of study is not characterized by intense surface winds. This means that wind speed peaks are below 10 m s-1 (see Figs. 7 and S1). During periods of moderate wind (around 5 m s-1), the release of pollen is significantly stronger, concentrations of 300–1500 Pollen m-3 are simulated 200 km downwind the emission sources (see Fig. 6), and concentrations over the Mediterranean Sea above 5000 Pollen m-3 are not unusual (e.g., 29M at 12:00 UTC). Under weak or calm conditions (wind speed below 1 m s-1), very small concentrations are simulated downwind of the sources (<100 Pollen m-3), while localized high concentrations are still present close to the regions with higher pine density. The strength of the wind decreases during nighttime, which reduces the dispersion of pollen and could stop the emission process. This might be explained by the presence of weak but non-negligible winds that promote some emission, an important reduction in the dispersion, and an accumulation of grains in the lower boundary layer due to a stable atmosphere.

The daily average of the Pinus surface concentration is evaluated in the XAC sites. Table 5 presents the statistics for the period 25 March to 2 April 2015 (all stations; 9 km domain), and Fig. 8 shows the temporal evolution for the same period for the Barcelona, Tarragona, and Bellaterra sites (all three domains). The model performance is significantly different at those three sites compared with the rest of the locations. The three sites are characterized by an important upwind tree density (see Figs. 1a and 6). On the one hand, the correlation is 0.5 in Barcelona, 0.6 in Tarragona, and 0.9 in Bellaterra, with a negative bias that is more pronounced in Bellaterra (-366 Pollen m-3). At this site, measurements show very high concentrations, particularly on 30M, that the model does not capture in intensity, although it reproduces the evolution of the event. In fact the evolution of the event is captured in the three sites; the concentration increases from 25M to 28M and reaches a maximum on 30M in Barcelona and Bellaterra and 31M in Tarragona, which is followed by a sudden decrease on 1 and 2 April when the pollination event vanishes. On the other hand, the model does not capture the measurements in the other XAC sites (Girona, Lleida, Manresa, and Roquetes; not shown). In some cases, the pine trees did not pollinate during the event but 2 weeks later (e.g., Manresa) or the week before (e.g., Roquetes; see https://aerobiologia.cat/pia/en/); in other places, the distribution of trees imposed in the model (Fig. 1a) might not be accurate enough (e.g., Girona). Even in a relatively small geographical region, the behavior of the pine trees is significantly heterogeneous due to the types of trees present, different micro-climates, and meteorological conditions. It is clear that more efforts are needed to develop accurate phenology functions and detailed maps to improve our current results, but the performances of the model over the Barcelona site are good enough to study the vertical structure of pollen there.

In this section, we discuss the hourly resolution results of the surface concentration at the Barcelona XAC site where specific high-temporal resolution measurements were available for the period 27M to 31M. Figure 9 shows the time series of the hourly Pinus surface concentration of the model at 9, 3, and 1 km resolution (colored lines) and the observations (black line) for the Barcelona site. Complementarily, we plot the daily averages that extend beyond the period where high-temporal resolution measurements are available. As discussed in the previous section, the model has a delay in the increase in the pollen concentration observed from 25M to 27M. During that period, although the model emits pollen grains (see Fig. 6), the concentrations in the air are strongly underestimated. The model starts to reproduce the observations on the morning of 28M, and the agreement with the observations is maintained for the rest of the event. The three model resolutions follow the trend of the measurements reasonably well. The 1 km run tends to simulate sudden rises and falls in the concentrations compared with the 9 and 3 km, which are able to maintain some background concentrations in the air. According to Fig. 9, the model is capable to reproduce the temporal variability in the observations with sudden peaks on 30M and 31M or sustained high concentrations at midday on 28M or the afternoon of 31M. Some model underestimations can be explained by the biases in the winds. For example, the strong underestimation during the first hours of 28M is attributed to the underestimated wind peak observed in Fig. 7a, or the underestimation during 29M in the morning is associated with a wrong wind direction (northerly wind in the model versus easterly wind in the observations). In this sense, the 1 km simulation is more sensitive to the errors in the wind than the upper domains. Note the need for some wind intensity in the model to capture the concentrations; under weak conditions, the model simulates extremely low concentrations pointing to the need for a minimum mobilization term in the emission scheme as suggested by other studies (e.g., Sofiev et al., 2013).

To compare the performances of the model at different horizontal resolutions, Table 6 presents the day-by-day statistics for the hourly Pinus number concentration at the Barcelona site. None of the domain resolutions is doing significantly better than the others, although, overall, the domain with the best statistical indicators (lowest RMSE and bias and highest r) is the 9 km domain. The start of the event is better captured by the 9 km domain, but the 1 km results improve significantly afterwards, resulting in lower biases. The correlations are low in most cases (around 0.2–0.3), highlighting the complexity in reproducing the hourly variability in pollen models.

The study of the mechanisms responsible for the pollen transport and the analysis of the model performances for predicting the vertical distribution of Pinus grains are performed in terms of two vertically integrated statistical indicators, namely the fractional bias (FB), and the Pearson correlation coefficient (r). FB and r are both calculated for the number concentration. For each vertical 1 h profile, the vertical extension considered starts at the lowest pair of simultaneously available model and observed values (fixed at 225 m, which is the height of the first valid lidar measurement) and ends at the pollen top height, hpol, taken from Sicard et al. (2016a). Both the base case (100 % deposition; 100 % sedimentation) and the enhanced (50 % deposition; 50 % sedimentation) simulations are considered, both at the three domain resolutions of 9, 3, and 1 km. We recall that the model has been calibrated with the surface measurements (see Sect. 3.3), and thus, that the model vertical resolution of Pinus is independent from the column (lidar) measurements.

Before analyzing the results, we present the latitudinal and longitudinal cross sections of the model Pinus number concentration at the coordinates of the Barcelona lidar site (2.112∘ E, 41.389∘ N; see the red dashed lines in Fig. 6) for the 1 km resolution on 30M at 12:00 UTC (Fig. 10). A more vertical structure is observed in the longitudinal cross section compared to the latitudinal one where the orography is more pronounced (Pyrenees). In agreement with the emission scheme, most of the pollen dispersion occurs downwind of the second mountain range from the shore (latitudinal cross section). Above the lidar site, the model predicts a thick Pinus pollen layer up to ∼ 0.8 km and then a decrease up to ∼ 1.3 km. As it will be shown later, the model vertical structure is not always in agreement with the measurement and presents more variability with increased horizontal resolution, particularly from noon onwards.

In order to avoid misleading results caused by the averaging of low (nighttime) and high (daytime) concentration profiles, a first analysis is made with the diurnal (defined as the average of the hourly values between 09:00 and 17:00 UTC) statistical indicators. In Table 7, we report the diurnal average of FB and r for all 5 d of the event for both simulations at all three domain resolutions (9, 3, and 1 km). In order to investigate the variability in the model at each resolution, we also include the standard deviation ratio (SDR) defined as the ratio of the model standard deviation to the lidar standard deviation measured at the profile and averaged over the diurnal hours (09:00–17:00 UTC). Figure 11 shows that, for each of the 5 d of the event, the results of the base case simulation at 9 km resolution (the resolution resulting in the lowest statistical error at the surface; see Sect. 5.2) have daily mean (defined as the average of the hourly profiles between 00:00 and 23:00 UTC) vertical profiles of the Pinus number concentration, simulated by the model and measured by the lidar (top panels), a temporal evolution of the hourly FB and r (middle panels), and the plot of hourly FB vs. r (bottom panels). According to Table 7, the model significantly underestimates the Pinus number concentration in the column during the diurnal hours on the first (27M; FB<-60 %) and last (31M; FB<-27 %) day of the event, which is a result similar to what is found at the surface (Sect. 5.2). This is partly due to the definition of the phenology function which is fitted to the surface concentration regardless of the vertical distribution. While the Pinus daily mean number concentration increases by 50 % between 27M and 28M, the pollen AOD remains constant between both days (Sicard et al., 2016a). In these conditions, our phenology function will yield an underestimation of the vertical distribution of Pinus pollen in the column on 27M, as it is observed. The same occurs on 31M. While the Pinus daily mean number concentration decreases between 30M and 31M, the pollen AOD remains constant between both days, hence, causing the underestimation of the vertical distribution of Pinus pollen in the column observed on 31M. This result suggests that the phenology function, fitted to the surface concentration, works relatively well to also reproduce the quantity of pollen transported in the column, at least during the 3 most intense days of the event. The diurnal FB is positive (+11.4<FB<+102.3) on 28M, 29M, and 30M.

The diurnal correlation coefficient is higher than 0.08 in all cases (Table 7) and varies between 0.08 and 0.92. The lowest values of r are usually reached for the 1 km resolution. At 9 km resolution (the resolution chosen for Fig. 11), r varies between 0.69 and 0.91. The middle plots of Fig. 11 show that, on each day, many values of r approach 1.00 very closely, especially on 28M, 29M, and 30M. On these 3 d, the diurnal correlation coefficient is greater than 0.83 and 0.52 for the 9 and 3 km resolutions, respectively. This result suggests that the model reproduces the shape of the Pinus number concentration vertical distribution during the diurnal hours quite well, at least during the 3 most intense days of the event (r>0.52). The bottom plots of Fig. 11 summarize the score of the model well, as far as vertically integrated FB and r are concerned. There is a large variability in the hourly values of FB varying from negative (underestimation; nighttime) to positive (overestimation; daytime), and r values are close to the ideal value of 1 on 28M, 29M, and 30M.

A lot of variability in FB is also observed as a function of the domain resolution considered for the simulation (Table 7). In all cases, the simulations at 1 km resolution give lower diurnal FB than at 9 and 3 km resolution. Except on 29M, the ay for which the 9 and 3 km resolution yields similar FB, the 3 km resolution gives a slightly lower FB than the 9 km one. The correlation coefficient can also vary significantly from one resolution to another; a difference of up to 0.80 (enhanced simulation on 29M) is observed between the 9 (r=0.88) and 1 km (r=0.08) resolutions (Table 7). Differences between 9 and 3 km resolutions are not higher than 0.36 (enhanced simulation on 29M). On 30M, and independently of the simulation version, the correlation coefficient is remarkably constant and equal to 0.91 for the two resolutions of 9 and 3 km. In agreement with the surface results (Sect. 5.1 and 5.2), errors in the meteorology have a major impact on the dispersion of pollen at a higher model resolution. While the structure of the plumes downwind emission sources are more well defined at 1 km resolution, small errors in the wind speed or direction result in larger errors at such a high mesoscale resolution that are smoothed at coarser resolutions due to the numerical diffusion. This is a classical problem when mesoscale models approach the 1 km horizontal resolution.

The diurnal standard deviation ratio (Table 7) also illustrates the underestimation of the model during the diurnal hours on 27M and 31M (SDR < 1 in general) and the overestimation on 28M, 29M, and 30M (SDR > 1 in general). On 28M, 29M, and 30M, the variability in the model is larger for the 3 km resolution and generally smaller for the 9 km resolution. At this resolution, the diurnal SDR varies between 1.06 on 28M (variability of the model is similar to the atmospheric variability) and 2.84 on 29M (variability of the model is ∼ 3 times larger than the atmospheric variability). From the above discussion, we conclude that the 9 and 3 km resolutions are probably the most suitable to minimize bias and maximize correlation between Pinus pollen forecast and observations in the column. The 9 km resolution also presents the advantage of reducing the model variability. The 9 km resolution is chosen for plotting the hourly and daily mean variations shown in Fig. 11.

The effect of the deposition and sedimentation on the quantity of pollen grains transported vertically is studied by means of the two simulations defined in Sect. 3.3 as the base case (100 % deposition; 100 % sedimentation) and enhanced (50 % deposition; 50 % sedimentation) simulations. The results are reported diurnally in Table 7 and daily in the top plots of Fig. 11. The study on the sedimentation is motivated by the high sedimentation velocity of large Pinus grains (3–4 cm s-1; Jackson and Lyford, 1999). As an example, particles of mineral dust with a diameter smaller than 10 µm have a sedimentation velocity lower than 1 cm s-1 (Li and Osada, 2007). Pollen sedimentation velocity is partly controlled by the grain size, density, and shape, all of which are highly variable. Consequently, the sedimentation velocity may be one of the main factors regulating the pollen vertical dispersion. In the case of Pinus pollen, the sacci, or inflated air bladders on each side of the grain, play an additional role in such a mechanism in that they act as aids for aerial dispersal (Wodehouse, 1935; Proctor et al., 1996; Schwendemann et al., 2007). As seen in Table 7, the combined effect of deposition/sedimentation on FB is quite notable. In the cases for which the base case simulations result in an underestimation (negative diurnal FB; see Table 7), the enhanced simulations, as expected, always reduce this underestimation and, thus, improve the model score. On the contrary, when the base case gives a positive diurnal FB, the enhanced simulation increases this value and, thus, worsens the overestimation. In terms of the fractional bias, the increase in diurnal FB, when the deposition/sedimentation is set to half of its nominal value, varies between 4.7 % (31M; 1 km) and 13.3 % (28M; 3 km). On average, over the 5 d of the event, the mean increase in diurnal FB is almost the same for all three domain resolutions, where ΔFB is 9.1 %, 10.0 % and 8.4 % for 9, 3, and 1 km resolutions, respectively. The effect of the reduction in the deposition/sedimentation is also visible on the top plots of Fig. 11 in which the results of enhanced simulations are also reported. In all cases, the concentration increase due to the reduction in deposition/sedimentation is significant. As expected, larger differences are observed in the lowermost layers. The daily relative increase in concentration (vertically integrated up to 2 km height; see the top plots of Fig. 11), due to a 50 % decrease in deposition/sedimentation, is 16.1 %, 14.1 %, 16.5 %, 13.0 %, and 9.7 % on 27M, 28M, 29M, 30M, and 31M, respectively. The deposition/sedimentation has a small effect on the capability of the model to reproduce the shape of the Pinus pollen vertical distribution (Table 7) because, from the base case to enhanced simulations, the least significant figure of the correlation coefficient does not change by more than ±0.04. To conclude this analysis, reducing deposition/sedimentation has virtually no impact on the way the model reproduces the shape of the vertical distribution of the pollen number concentration, and this suggests that other processes are more relevant to explain the vertical structures observed.

To have a closer look to the performance of the model in the column with respect to time, we plot, in Figs. 12, 13, and 14, the hourly evolution of the profiles of Pinus number concentration for the base case simulation at all three domain resolutions and for the observations on 28M, 29M, and 30M, respectively. On 28M, 29M, and 30M, the 3 most intense days of the event, a bimodal diurnal cycle of the fractional bias is visible in Fig. 11 (middle plot), showing an underestimation during the nighttime and two peaks of overestimation at 06:00–10:00 and at 15:00–16:00 UTC. The morning peak is probably due to an excessive mobilization of pollen during nighttime associated to a possible lack of ventilation during the first hours of the day.

As seen in Sect. 5.1 and in Fig. 7 in particular, the model has some systematic error that induces an overestimation of nocturnal wind and underestimation of the increase in wind during the morning. The afternoon peak is likely due to the fact that pollen stays aloft longer than it does in reality, or that there is an excessive emission flux upwind during the previous hours. This behavior of the model is presented in Figs. 12, 13, and 14. According to the lidar, the pollen layer grows between 09:00–14:00 UTC, following a typical development of the convective boundary layer. During that period, the top of the layer is well reproduced by the model with excessive mobilization of pollen grains, depending on the horizontal resolution and hour. Note the good agreement in some specific profiles, see, e.g., on 30M at 12:00–13:00 UTC for 3 km (Fig. 14). The sink of the layer is observed between 14:00 and 15:00 UTC on 28M and 29M (the model predicts it 1 h later) and between 13:00 and 14:00 UTC on 30M (the model predicts it 2 h later). The delay of the model in predicting the pollen layer drop in the afternoon might be linked to the fact that the model delays the decay of the convective boundary layer. In the afternoon, and especially for the 1 km resolution, the model predicts a strong and steep decrease in concentration towards the surface. The same result is observed with the surface concentration in Sect. 5.2, where the simulation at 1 km resolution shows a decrease in concentration (which leads to an underestimation of the model; see Fig. 9), followed by sudden peaks that correlate well with the wind speed. Again, it shows the higher sensitivity to wind speed and direction of the 1 km simulation compared to the 3 and 9 km simulations. Also, more structures are visible at 1 km resolution than at the other resolutions. The finer the resolution, the more vertical structures can be seen. However, the structures visible at the 1 km resolution are not always reproducing reality (see, e.g., on 29M at 14:00 UTC; Fig. 13).

The Hirst observations are much more variable than the model concentration and the meteorology. Although following a general trend, the Hirst concentrations oscillate up and down most of the time and throughout the whole day (see the black circles in the hourly plots of Figs. 12, 13, and 14, and the black line in Fig. 9), while the model concentration usually steadily increases in the morning and decreases in the afternoon after 15:00–17:00 UTC. Such a difference between model and observations is not visible in the meteorological variables (Fig. 7). In fact, for some variables, the opposite happens, e.g., the model wind speed is more variable than the observation. The high variability in the observed Pinus concentration, independent of the meteorology, emphasizes again the difficulty in predicting airborne pollen grain concentration and the need to develop emission schemes especially designed for bioaerosols, the emission of which is much more complex than that of the rest of natural particles.