MEGAN estimates the reference emission rates by emission factors of 15 PFTs, as listed in Table 1. The Global Emission Factors (version 2011) from the MEGAN website (http://lar.wsu.edu/megan/guides.html, last access: 12 March 2019) were used in this study. Emission factors of compounds are given for each of the 15 PFTs (Guenther et al., 2012). Tree-species-based emission factors and forest species composition profiles for more than 2000 ecoregions worldwide were used to generate the high-resolution (0.0083∘×0.0083∘) global emission factor dataset. On the other hand, the PSI model uses reference emission rates of typical plant species in Europe (see Table 1). The reference emission rates (µg g(dry weight)-1 h-1) of isoprene and monoterpenes from forests, pasture, and crops were calculated based on algorithms given by Lamb et al. (1993). Isoprene emissions from Norway spruce were assumed to be about 10 % of α-pinene emission rates during daytime (Steinbrecher, 1989). SQTs are the least studied among the identified BVOCs due to their high reactivity and relatively low vapour pressure (Duhl et al., 2008). Determination of their basal emission rates is therefore challenging. In the PSI model, SQT emissions were treated only as pool emissions and assumed to be 5 % (by weight) of the monoterpene emissions based on the emission rate data for 116 species compiled from various studies as given by Steinbrecher et al. (2009).

Isoprene, one of the most important BVOC species, is released after biosynthesis by volatilisation, which depends on both temperature and solar radiation. On the other hand, monoterpenes are stored in large storage pools after their production in the plant organs. Emissions of monoterpenes are mostly temperature-dependent, although there are some species that have both light- and temperature-dependent synthesis emissions of MTs (Tingey et al., 1980). In the PSI model, the isoprene emissions are corrected by light (γL) and temperature (γT) response functions based on the algorithm described by Guenther et al. (1993): γL=αCL1PAR1+α2PAR2,γT=exp[CT1(T-Ts)/RTsT]1+exp[CT2(T-TM)/RTsT], where α (=0.0027), CL1 (=1.066), CT1 (=95000 J mol-1), CT2 (=230000 J mol-1), and TM (=314 K) are all empirical coefficients determined by nonlinear fitting based on emission rate measurements, R is the gas constant (8.314 J K-1 mol-1), and Ts is the standard leaf temperature (303.16 K). The response functions of the isoprene emission in MEGAN are based on Guenther et al. (1999), an updated version of Guenther et al. (1993). The major difference of the improved algorithm is the inclusion of the influence of past temperature and light conditions. New empirical coefficients Topt and Eopt calculated by the average leaf temperature over the past 24 and 240 h are added to include the continuous influence over time, respectively (Eq. 3): γT=Eopt×CT2×exp(CT1×x)CT2-CT1×(1-exp(CT2×x)), where x=[(1/Topt)-(1/T)]/0.00831. A detailed introduction to Topt and Eopt can be found in Guenther et al. (2006).

For the light-independent response of MT pool emissions, similar exponential corrections are used by MEGAN and the PSI model, which are based on Lamb et al. (1993) and Tingey et al. (1980) as shown in Eq. (4): E=Es×exp(β×(T-Ts)), where E is the MT emission at temperature T, Es is the emission under standard conditions (Ts=30 ∘C), and β is the slope coefficient of dlnE dT-1. The slope value β has a wide range between 0.057 and 0.144 according to previous literature (Guenther et al., 1993). The value of 0.1 is used for most MT species (e.g. α-pinene, β-pinene, 3-carene, and limonene) in MEGAN2.1, while the values are between 0.065 and 0.077 for different MT species in the PSI model. The light-dependent synthesis emissions of MTs were considered in MEGAN2.1 as described in Guenther et al. (2012). Depending on different MT species, the light-dependent fraction of MT emissions ranges between 0.2 and 0.8 for MEGAN. In the PSI model, the light-dependent emissions from Norway spruce are calculated for each monoterpene species as a function of PAR based on the data of Schürmann (1993). In addition to the light and temperature response, MEGAN2.1 also covers some other factors such as leaf age and leaf area index (Guenther et al., 2012). Since the correction of soil moisture and CO2 dependence is not included in the offline version of MEGAN (Emmerson et al., 2016), we used the default parameterisation where the correction factors were set to 1.

The variation in light and temperature within the forest canopy are corrected by a canopy model in both the PSI model and MEGAN. The PSI model uses the canopy model by Baldocchi et al. (1985) combined with experiments in Hartheim forest (Germany) and central Switzerland (Joss, 1995). The detailed algorithm of the canopy correction for the PSI model was reported by Keller et al. (1995). The MEGAN canopy environmental model is based on Guenther et al. (1999), which estimates incident PAR and the temperature of sun and shade leaves at different canopy depths. Details can be found in Guenther et al. (2006, 2012). A BVOC reduction of about ∼20 % due to the canopy model was reported for the PSI model by Oderbolz et al. (2013). Although different canopy models could influence the modelled BVOC emission, such influence was within the uncertainty range of observed fluxes (Guenther et al., 2006; Lamb et al., 1996).

Three types of basic driving variables are required for both MEGAN and the PSI model, namely meteorological conditions, land use, and biomass density. The meteorological data provide hourly gridded information of temperature, solar radiation, wind speed, moisture, and surface pressure to drive the model simulation of emission response. We used the same meteorological data retrieved from the WRF-ARW model as input for both models. The main difference between the two model inputs is in the land use and leaf biomass density.

MEGAN2.1 uses the Community Land Model version 4 (CLM4) including 15 PFTs as shown in Table 1. In this study, we adopted for MEGAN the same global PFT map as in Sindelarova et al. (2014) with a resolution of 0.05∘. For the PSI model, the GlobCover 2006 inventory by the European Space Agency (http://due.esrin.esa.int/page_globcover.php, last access: 12 March 2019) was used. This inventory was developed based on MERIS (MEdium Resolution Imaging Spectrometer) FRS (Fine Resolution Full Swath product) level 1B data during December 2004 to June 2006 (Bicheron et al., 2008). The raw data have a fine resolution of 300 m and 64 categories of land use types, e.g. needle-leaf evergreen forest, broadleaf deciduous forest, and mixed broadleaf and needle-leaf forest. The grid-scale fractions of needle-leaf, broadleaf, and mixed forests were first calculated based on the GlobCover inventory data. The mixed forest was assumed to be composed of 50 % needle-leaf and 50 % broadleaf species. Different tree species in the same category may have different emission factors. For instance, although all belong to broadleaf tree species, oak (Quercus) has a high emission rate, while beech (Fagus sylvatica) and maple (Acer) are negligible BVOC emitters. Even within the same genus, there might be large differences in emissions, e.g. two oak species, where Quercus robur is a high isoprene emitter and Quercus suber a low isoprene emitter (Steinbrecher et al., 2009). Europe has a relatively low abundance of flora in both diversity and numbers, and six tree species cover two-thirds of the forest area, namely Scots pine, Norway spruce, beech, maritime pine, European oak, and evergreen oak (Simpson et al., 1999). Therefore, in the PSI model, we classified the forests into 10 typical forest species (see Table 1: vegetation classes 1–10) found in Europe based on the country-specific forest species profile from Simpson et al. (1999). The original 35 forest species in Simpson et al. (1999) were grouped into 10 classes (including five coniferous species and five broadleaf species), and the ratio of each species class to the total coniferous forest and broadleaf forest was calculated (Table S2). The ratio of “other trees” was proportionally added to the 10 tree species. As the other trees are mainly in a few Mediterranean countries, their influence on the whole domain is small. The species coverage was then generated by multiplying the forest coverage from GlobCover with the country-specific tree species profile.

The biomass density in MEGAN was calculated by the canopy environment module based on the satellite data of the leaf area index (LAI, m2 leaves m-2 projected area) with a time step of 8 days. The TERRA MODIS vegetation data products MOD15A2 were downloaded from the NASA Earth Observations website (https://neo.sci.gsfc.nasa.gov/view.php?datasetId=MOD15A2_E_LAI&year=2011, last access: 12 March 2019). The grid-scale LAI was then divided by the fraction of vegetation coverage of each grid (sum of PFT) to get the average LAI of vegetation-covered surfaces (LAIv).

As the reference emission rates of the PSI model are based on dry weight of leaf biomass, the leaf biomass density factors (grammes of dry weight m-2 projected area) of each tree species (Cannell, 1982; Satoo and Madgwick, 1982) were explicitly used in the PSI model. To simulate the vertical variation in foliar biomass in the canopy, the biomass density was scaled by the leaf area distribution in each canopy layer as described in Oderbolz et al. (2013). The temporal variation in the biomass was simulated by monthly factors for different plant types. For example, the PSI model assumes that the leaf biomass of deciduous trees, such as oak and larch, turn to zero in the winter months (November–March) and crops only have biomass in the growing season (April–August).

The modelled ozone mixing ratios from two simulations using the biogenic emissions calculated by the PSI model and MEGAN were evaluated by the measurements from the European air quality database (AirBase; Mol and Leeuw, 2005). Table 2 shows the statistical metrics of modelled average mixing ratios of afternoon (12:00–18:00 UTC) surface ozone at 537 rural background stations. The model performance in summer was generally better than in winter for all regions, but the difference between the PSI model and MEGAN was small. In winter, the two models showed similar mean bias (∼3 ppb) and RMSE (∼9.2 ppb) between modelled and measured concentrations. In summer, the PSI model showed lower (34.0 %) mean bias but slightly higher (1.3 %) RMSE than MEGAN. To investigate the difference in more detail, we compared the bias between modelled and observed O3 in different mixing ratio bins for different regions in summer (Fig. 5). In general, ozone modelled using the BVOC emission input from both models was overestimated at low mixing ratios and underestimated at high mixing ratios. A similar pattern was found in previous O3 modelling studies in Europe (Im et al., 2015; Oikonomakis et al., 2018; Solazzo et al., 2017). CAMx performed better with MEGAN emissions at most stations at the high ozone bins. Although the PSI model led to lower overall MB (Table 2), it was mostly due to compensation at the low- and high-O3 level bins.

To further explore the reasons for the different model performance in the ozone simulations, we present the spatial distributions of modelled ozone in summer calculated using BVOC emissions from the PSI model and MEGAN in Fig. 6. PSI O3 was generally lower than MEGAN O3 across all of Europe. In summer, the largest effect of using different BVOC emissions on ozone was mostly in southern Europe, especially in the Mediterranean region, with the highest relative difference between PSI O3 and MEGAN O3 reaching -14 % (7.5 ppb, in Italy). On the other hand, in the UK and Ireland, where isoprene emissions by the PSI model were higher than MEGAN (Fig. S2), a positive difference up to 3.9 ppb was found. The spatial distribution of the ozone difference, i.e. (PSI O3)–(MEGAN O3) (Fig. 6c), is very similar to that of the difference in the isoprene emissions (Fig. S2a). As an important ozone precursor, isoprene reacts with hydroxyl radicals (OH) to form peroxyl radicals (RO2, HO2), which further react with NO to generate NO2 and finally ozone (Wennberg et al., 2018). This process can be significantly affected by the availability of isoprene and NOx in the atmosphere as well as temperature (Calfapietra et al., 2013), leading to high uncertainties in the net influence of BVOC emissions. Li et al. (2007) found that increasing the isoprene emissions by 50 % resulted in an increase in the O3 mixing ratios by 5–25 ppb in urban Houston in the United States, and Zare et al. (2012) suggested that the 21 % higher annual isoprene emissions by MEGAN than GEIA (Global Emissions Inventory Activity) led to up to 10 % higher O3 concentrations in the African savannah. However, the effect of the BVOC emissions on the ozone levels in Europe was much smaller in this study. The around 3 times higher isoprene emissions in MEGAN only led to up to ∼10 % (7 ppb) higher ozone mixing ratios in summer compared to the PSI model. Similarly, an earlier study by Aksoyoglu et al. (2012) using the PSI model for BVOC emissions suggested that increasing the isoprene emissions by a factor of 4 in Europe led to an increase of less than 10 % in the afternoon ozone mixing ratios. The main reason for the weak effect of the isoprene emissions on ozone is the stronger sensitivity of ozone formation in general to NOx emissions rather than VOC emissions in Europe. An additional reason might be the rather low ozone production compared to the background ozone, where the latter is not affected by local European emissions (Oikonomakis et al., 2018; Sartelet et al., 2012). Several European studies reported that ozone formation in most regions is NOx-sensitive, except around the English Channel, Benelux, and Po Valley regions, where NOx emissions are high (due to intensive anthropogenic NOx emissions from both land and shipping or geographical characteristics leading to high accumulation of pollutants) and the response to a change in the VOC emissions is relatively strong (Aksoyoglu et al., 2012; Beekmann and Vautard, 2010; Oikonomakis et al., 2018). However, the sensitivity of ozone formation to its precursor emissions might be changing as a result of large NOx emission reductions in Europe since 1990 according to the Gothenburg Protocol. On the other hand, emissions from shipping activities are not regulated as strictly as land emissions and have been increasing continuously, especially in the Mediterranean, affecting both ozone and particulate matter concentrations (Aksoyoglu et al., 2016; Viana et al., 2014).

The effects of different BVOC emissions on organic aerosols were investigated by comparing modelled OA concentrations with measurements at nine ACSM/AMS stations. Although the OA concentrations were generally under-predicted in both cases, the model performance for OA was better with the PSI biogenic emissions (Fig. 7). About 67 % of the modelled OA concentrations were below the 1:2 line in the case of MEGAN (Fig. 7b). The mean bias between observed and modelled OA concentrations with the PSI BVOC emissions was lower than the bias obtained with MEGAN emissions (3.9 % in Paris and 83.4 % in Mace Head; see Table 3). The better model performance when using the PSI emissions was more obvious at rural or remote stations where biogenic sources play a major role in OA formation. The mean bias of OA by the PSI model was 21  % to 83 % lower than MEGAN at rural or remote stations (Finokalia, San Pietro Capofiume, Montsec, SMEAR II, and Mace Head), while the range was 4 %–12 % for Paris, Bologna, and Marseille (see Table 3). The situation of Zurich was different with an MB reduction of 67 % by the PSI model compared with MEGAN as an urban station, mostly because the station is an urban background site that is strongly affected by biogenic emissions (Daellenbach et al., 2017).

We further evaluated the model performance of the temporal variation at Zurich and Mace Head as examples of urban background and rural stations, respectively (Fig. 8, top row), because these two datasets covered almost the whole year. In spite of some underestimation, the temporal variation was well captured. At Zurich, the difference between the two cases (PSI OA and MEGAN OA) was small in February and March and they were both lower than the measurements, possibly due to underestimation of biomass burning OA (Fountoukis et al., 2014). The largest difference occurred in autumn when PSI OA reproduced the measurements quite well, while MEGAN OA showed a large underestimation. This is consistent with source apportionment studies performed for Zurich (Canonaco et al., 2013; Daellenbach et al., 2017) which reported that the contribution of biogenic sources to OA was minor in the period of January to March but significant (>50 %) in summer and autumn.

The situation was quite different for Mace Head. Located on the west coast of Ireland and 90 km away from the closest city Galway (Schmale et al., 2017), Mace Head is a remote station with low influence from anthropogenic activity (O'Dowd et al., 2014). The simulation with the PSI biogenic emission model could reproduce all the measured peaks quite well, while the simulation using the MEGAN emissions failed to capture their magnitude. To investigate the cause of the high OA concentrations during certain periods, 72 h back-trajectory analyses ending at Mace Head on 26 March (as an example for a high-OA day) and on 4 August (as an example for a low-OA day) were conducted by NOAA's HYSPLIT atmospheric transport and dispersion modelling system (Stein et al., 2015). According to the HYSPLIT results (Fig. S4), the air masses were transported from Ireland and Scotland during the high-OA period (Fig. S4a), while during the low-OA period the air masses came from the North Atlantic Ocean (Fig. S4b), suggesting that the OA peaks originated from anthropogenic or biogenic sources on land. The influence of wind direction was further studied by comparing modelled and measured OA during the two periods featured land-based wind (24–26 March) and marine-based wind (2–4 August) in Fig. S5. Measured OA in periods with dominant wind direction from land was higher than during the marine-based wind-dominant periods by a factor of ∼10. Modelled PSI OA was very close to measurements while MEGAN OA was underestimated in both periods. However, it is not possible to conclude that the good model performance for OA with PSI emissions is due to the fact that its high MT emissions are more accurate. It could also be due to the overestimated MT emissions compensating for other missing continental sources of OA, e.g. biomass burning.

The spatial distribution of the SOA difference showed a similar pattern as its main precursor, monoterpenes (Fig. 9). The PSI emissions lead to significantly higher SOA production than MEGAN (by 113 % and 109 % in winter and summer, respectively). The grid-scale difference reached up to a factor of 35 and 17 for winter and summer SOA, respectively. The largest differences occurred in central Europe, the Iberian Peninsula, and Turkey in winter, and especially in Scandinavia in summer.

The modelled primary organic aerosol (POA) was also slightly higher (6.5 % in winter and 7.8 % in summer on average) in PSI emissions compared to the case with MEGAN (Fig. S6). Unlike in the traditional CTMs, where POA is treated as inert, the VBS scheme of CAMx allows POA to evaporate and react with oxidants. According to the partitioning theory (Donahue et al., 2006; Odum et al., 1996), higher total OA concentrations led to higher partitioning to the particle phase for all compounds that are soluble in the aerosol matrix. Therefore, in our case, the high PSI OA shifted the particle–gas equilibrium of primary condensable gases towards the particle phase, resulting in higher POA.

The influence of BVOC emissions on SIA was much smaller than on SOA according to the comparison of model results with measurements (Table 3). At the nine ACSM/AMS stations, using the PSI emissions generally reduced the RMSE between modelled and measured particulate nitrate (PNO3), sulfate (PSO4), and ammonium (PNH4) by up to 15.0 %, 1.7 %, and 7.7 %, respectively, compared to CAMx simulations with the MEGAN emissions. Only Finokalia (Greece, rural) and San Pietro Capofiume (Italy, rural) had lower RMSE in the MEGAN emissions. Unlike the obvious difference in OA, the difference between the modelled temporal variations of the inorganic aerosol was negligible with the two emission estimates (Fig. 8). MEGAN PNO3 was slightly higher than PSI because lower MT emission by MEGAN led to a lower MT–NO3 reaction and therefore more NOx was available to be oxidised to PNO3 (Fig. S7).

The modelled and measured daily average concentrations match well, except for February and March at Zurich, when temperature was significantly underestimated and resulted in higher condensation (Fig. S8a). A similar effect on temperature was not observed in OA during the same period, possibly due to compensation of underestimated winter OA as a consequence of lacking sources in the model, especially biomass burning (Ciarelli et al., 2017). On the other hand, the modelled primary elemental carbon (PEC) matched the measurements at Zurich very well.

Similar to the situation of OA, the measured SIA (PSO4, PNO3, and PNH4) at Mace Head peaked during the periods with wind from land. Both biogenic models captured the peaks well but overestimated the SIA during the peak periods. The modelled elemental carbon concentrations (PEC, in Fig. 8), on the other hand, were lower than the measured equivalent black carbon (EBC) in general but followed the temporal variation very well. In a study about the aerosols at Mace Head, O'Dowd et al. (2014) reported that EBC measurements can significantly overestimate black carbon concentration by up to 50 % or more. Overestimation of SIA could result from either precursor emissions that are too high or from too much particle formation in the aqueous phase. The precursor gases SO2 and NOx from anthropogenic sources (continental, shipping) (Fig. S7) might be accumulated too highly in the surface layer since all emissions were injected into the first layer, leading to SIA formation that is too high.

The differences in the spatial distributions of PSO4, PNO3, and PNH4 between the two simulations with PSI and MEGAN emissions are shown in Fig. 10. The inorganic aerosol concentrations varied by less than 15 % on the grid scale for the different BVOC emissions. The highest PSO4 levels were predicted in central and eastern Europe in winter, where SO2 emissions are higher, while in summer the elevated sulfate concentrations were mostly along the shipping routes (Fig. 10). The PSI BVOC emissions lead to higher PSO4 than MEGAN, especially over the area from southern Poland to Turkey through the Balkan Peninsula in summer. These regions have the highest SO2 emissions in the model domain due to large combustion-based power plants and coal burning. In summer, the main pathway for sulfate formation in southern Europe is the gas-phase oxidation of SO2 with OH radical (Chrit et al., 2018; Megaritis et al., 2013). The higher sulfate concentrations predicted by CAMx with PSI BVOC emissions are consistent with the spatial pattern of the differences between PSI and MEGAN simulations for SO2 concentrations and OH radicals (Fig. S9) due to the following reason: as reaction with OH radical is the largest loss pathway for isoprene in the atmosphere (Wennberg et al., 2018), higher isoprene emissions in MEGAN consume more OH radical. As a consequence, less SO2 is oxidised to form PSO4 when MEGAN emissions are used (Fig. S9), leading to lower PSO4 formation.

Formation of PNO3 depends on the availability of NOx and NH3 emissions (Aksoyoglu et al., 2011; Wen et al., 2015). In contrast to PSO4, PNO3, and PNH4 concentrations modelled using the PSI biogenic emissions were generally lower than those using MEGAN emissions, especially in regions where the PSI model has more MT emissions (Fig. S2b). Nitrate radicals are recognised as a significant sink for BVOCs, especially monoterpenes at night (while OH oxidation is more relevant for isoprene during daytime) (Kiendler-Scharr et al., 2016; Ng et al., 2017). Higher monoterpene emissions produced by the PSI model lead to larger consumption of nitrate radicals affecting PNO3 formation from HNO3 and NH3. These results are consistent with a recent study showing the significant effect of BVOCs on ammonium nitrate (Aksoyoglu et al., 2017).