The wet deposition of nitrogen and sulfur in Europe for the
period 1990–2010 was estimated by six atmospheric chemistry transport models
(CHIMERE, CMAQ, EMEP MSC-W, LOTOS-EUROS, MATCH and MINNI) within the
framework of the EURODELTA-Trends model intercomparison. The simulated wet
deposition and its trends for two 11-year periods (1990–2000 and
2000–2010) were evaluated using data from observations from the EMEP
European monitoring network. For annual wet deposition of oxidised nitrogen
(WNOx), model bias was within 30 % of the average of the observations for
most models. There was a tendency for most models to underestimate annual wet
deposition of reduced nitrogen (WNHx), although the model bias was within 40 %
of the average of the observations. Model bias for WNHx was inversely
correlated with model bias for atmospheric concentrations of
Decreasing trends in WNOx were observed at most sites for both 11-year
periods, with larger trends, on average, for the second period. The models
also estimated predominantly decreasing trends at the monitoring sites and
all but one of the models estimated larger trends, on average, for the second
period. Decreasing trends were also observed at most sites for WNHx, although
larger trends, on average, were observed for the first period. This pattern
was not reproduced by the models, which estimated smaller decreasing trends,
on average, than those observed or even small increasing trends. The largest
observed trends were for WSOx, with decreasing trends at more than 80 %
of the sites. On average, the observed trends were larger for the first
period. All models were able to reproduce this pattern, although some models
underestimated the trends (by up to a factor of 4) and others
overestimated them (by up to 40 %), on average. These biases in modelled
trends were directly related to the tendency of the models to under- or
overestimate annual wet deposition and were smaller for the relative trends
(expressed as % yr
The fact that model biases were fairly constant throughout the time series
makes it possible to improve the predictions of wet deposition for future
scenarios by adjusting the model estimates using a bias correction calculated
from past observations. An analysis of the contributions of various factors
to the modelled trends suggests that the predominantly decreasing trends in
wet deposition are mostly due to reductions in emissions of the precursors
Atmospheric deposition of nitrogen (N) and sulfur (S) can lead to the
acidification of soils and surface waterways, resulting in damage to natural
and semi-natural vegetation, and aquatic organisms (Ulrich, 1983). Nitrogen
deposition can also lead to the eutrophication of terrestrial and aquatic
ecosystems, resulting in a reduction in biodiversity (Bobbink et al., 1998).
Most of the deposited N and S originates from the emissions of nitrogen
oxides (
Atmospheric chemistry transport models (CTMs) can be used to study the
relationships between emissions of
The variability in model performance for wet deposition is not surprising,
since wet deposition depends on many processes, such as emissions,
dispersion, atmospheric chemistry, cloud formation, cloud chemistry and
precipitation. However, despite their inherent uncertainties, CTMs are useful
tools that complement observations and study the spatial distributions of
atmospheric deposition and their evolution over time. One key question is how
well the models can simulate the trends in deposition as a result of changes
in emissions. This aspect is important, since CTMs are frequently used to
evaluate the impact of future emission control measures and so model
estimates of future deposition rates need to be reliable in order to make
well-founded policy decisions. Despite this, very few studies have evaluated
modelled wet-deposition trends in Europe with observed data. Fagerli and
Aas (2008) compared the observed trends of ammonium and nitrate in
precipitation measured by the EMEP network with those calculated from
simulations by the EMEP Unified model for the years in
which the model was run (1980, 1985, 1990 and 1995–2003). The authors
found that the modelled and observed trends in precipitation nitrate averaged
over all sites were similar (
The EURODELTA-Trends (EDT) exercise aims to assess the role of European air pollutant emission reductions in improving air quality and reducing the acidification and eutrophication of ecosystems over the period 1990–2010 (Colette et al., 2017a) using CTM simulations, as well as assess the influence of meteorological variability and long-range transport through the boundary conditions used. Wherever possible, input data (emissions, meteorology, boundary conditions, etc.) were the same for all models so that the differences in model estimates due to model formulations can be studied. Eight CTMs were used to simulate air quality over the period 1990–2010, of which six delivered estimates of wet and dry deposition of N and S, thus providing a unique data set for testing the ability of multiple CTMs to simulate deposition trends.
In this paper, we compare the EDT CTM estimates of wet deposition of S and reduced and oxidised N with observations from the EMEP network over the period 1990–2010. In order to better understand the differences between the CTM estimates of wet deposition, we also evaluate the models for atmospheric concentrations of relevant gaseous and particulate species and seasonal precipitation rates, as well as compare the model estimates for dry deposition. Due to the number of models studied and the many differences between their formulations and parameterisations, it is out of the scope of this study to provide an in depth analysis of individual model performance or inter-model differences. We also evaluate the ability of the models to estimate the absolute and relative trends in wet deposition over two 11-year periods (1990–2000 and 2000–2010) and look at the contributions that changing emissions, boundary conditions and meteorology make to the overall modelled trends. Following a discussion of the uncertainties and limitations associated with the model simulations and the observations of wet deposition, we provide suggestions on how to improve model estimates of wet deposition in the future.
Locations of the measurement sites used in the evaluation of wet
deposition and atmospheric concentrations of
Six CTMs were used to estimate wet deposition in Europe for the period
1990–2010: Chimere (Couvidat et al., 2018), CMAQ (Byun and Schere, 2006),
EMEP MSC-W (Simpson et al., 2012), LOTOS-EUROS (Manders et al., 2017), MATCH
(Robertson et al., 1999) and MINNI (Mircea et al., 2014, 2016). The shortened
model names CHIM, CMAQ, EMEP, LOTO, MATCH and MINNI are used
throughout the article. An overview of the model chemistry schemes and
parameterisations for wet and dry deposition can be found in the Supplement
(Table S1). In order to assess the differences in model estimates, due only to
model structure and parameterisations, the modelling domain and input data
used in the simulations were the same for all models, wherever possible. The
models were run on a domain that covers most of Europe (Fig. 1) with a grid
resolution of 0.25
Estimates of accumulated seasonal and annual WNOx, WNHx, WSOx (total and non-sea-salt component) and precipitation (at the same sites as the wet deposition) from the EMEP network over the period 1990–2010 were used to evaluate the model estimates. The seasonal and annual wet deposition was estimated by multiplying the volume-weighted mean concentration in precipitation by the total precipitation in the period. The concentrations for days with missing precipitation data were assumed to be equal to the volume-weighted average of the period (Hjellbrekke, 2016). For the evaluation of modelled atmospheric concentration estimates, the EMEP network data of mean annual concentrations of total nitrate, ammonium and sulfate (non-sea-salt component) were used. Although data are available for the individual gas and particulate species for many sites, the filter pack measurement methods used do not reliably estimate the partitioning of the gas and particulate N species and, therefore, the total (gas plus particulate) is used for the evaluation. Sites were selected that had data for at least 75 % of the year and had valid data for at least 75 % of the period 1990–2010, resulting in 39 sites for WNOx, 38 sites for WNHx, 36 sites for WSOx, 13 sites for TNO3, 16 sites for TNH4 and 20 sites for TSO4 (Fig. 1 and Table S2). In order to compare the trends for the two 11-year periods, a consistent set of sites was used that have valid data for both periods. However, this approach led to gaps in the spatial coverage of observations (particularly in SW Europe) and so an additional analysis was carried out using all available sites that met the selection criteria for the period 2000–2010. Note that the availability of observations for several components is strongly biased towards certain parts of Europe. For example, total TNO3 and TNH4 concentrations are mainly available for northern Europe and have very little overlap with wet-deposition sites in the centre and west of the domain. It must also be noted that the evaluation of precipitation estimates was only done at the sites with observations of wet deposition in order to assess the influence of model performance for precipitation on model performance of wet deposition. The aim was not to carry out a thorough evaluation of precipitation estimates, which would require a more detailed evaluation data set, such as E-OBS (Haylock et al., 2008).
The modelled wet deposition, precipitation and atmospheric concentration
estimates were statistically evaluated using the package “openair” (Carslaw
and Ropkins, 2012) for R (v3.3.2; R Core Team, 2016). Six metrics (proposed
by Chang and Hanna, 2004) were used to assess model performance: fraction of
model estimates within a factor of 2 of the observed values (FAC2),
fractional bias (FB), geometric mean bias (MG), normalised mean square error
(NMSE), geometric variance (VG) and the Pearson's correlation coefficient
(
The observed and modelled trends in deposition and their significance were estimated using three methodologies: Mann–Kendall (MK), seasonal Mann–Kendall (SMK) and partial seasonal Mann–Kendall (PSMK) (see Appendix A).
The six performance metrics relating model estimates (
The EURODELTA-Trends modelling experiment specifically included simulations
that can be used to determine the contribution of several factors (changes in
emissions, boundary conditions and meteorology) to the overall trends, as
described in detail by Colette et al. (2017b). The methodology assumes that
the overall trend (
Land-based
European
Land-based
The relative changes in emissions (Fig. S4) have a similar spatial
distribution to the absolute trends, although they highlight the large
relative increases in emissions in some parts of the domain (e.g.
In order to analyse the spatial distributions of modelled precipitation and wet deposition and provide a basis for the subsequent discussion of the trends for the two 11-year periods (1990–2000 and 2000–2010), this section analyses the spatial distributions of precipitation and wet deposition “snapshots” for the years 1990, 2000 and 2010 (corresponding to the years simulated by all models).
The four meteorological models (HIRLAM, RACMO2, WRF [CMAQ] and WRF [common driver]; see Table S1) estimated similar spatial distributions of precipitation for 1990, with the largest precipitation amounts on the western and north-western coasts of Norway, the western coast of Scotland, the southern coast of Iceland and the Pyrenees and Alps mountain ranges (Fig. S5). The meteorological model used by the MATCH (HIRLAM) simulations estimated the largest domain-mean precipitation while that used for the CMAQ (WRF) simulations estimated the smallest. For the year 2000 the meteorological models estimated similar distributions to those for 1990, although there was a noticeable shift southwards with less precipitation on the Norwegian coast and more in the Iberian Peninsula and central parts of the domain (the Alps, Italy, eastern Adriatic coast and the Carpathian Mountains). Domain-mean precipitation differed very little between the 2 years with the largest difference estimated by the LOTO meteorological driver (RACMO2) (7 % increase). The southwards shift in precipitation continued between 2000 and 2010. The domain-mean precipitation also differed very little between 2000 and 2010, with most meteorological drivers estimating a difference of less than 5 %. The exception was the CMAQ meteorological driver (WRF), which estimated 23 % more precipitation in 2010 than in 2000.
For 1990, MINNI estimated the smallest domain-mean WNOx and MATCH the largest.
(Fig. S6). However, in the east of the domain, EMEP estimated higher
deposition than the other models. Despite the differences between the models,
all of them estimated the highest WNOx in the centre and east of the domain,
especially on the northern and southern slopes of the Alps, the southern
coast of Norway and western Ukraine (corresponding mostly to areas with large
precipitation amounts). These deposition hotspots vary from model to
model, with LOTO and MATCH, for example, estimating higher deposition rates
on the southern slopes of the Alps compared with the northern slopes, whereas
EMEP and CMAQ estimated similar rates on both sides of the mountain range.
These differences appear to be due to the spatial distribution of
precipitation estimated by the meteorological driver. The spatial
distributions of modelled WNOx estimates for 2000 are very similar to those
for 1990 with a general decrease in deposition as a result of
Similarly to WNOx, most of the models estimated the largest values of WNHx in
1990 for the slopes of the Alps, as well as for the Netherlands and NW
Germany (Fig. S7), a well-known
The largest differences between the models, both in terms of the range of values and the spatial distributions was found for WSOx, with EMEP estimating the largest mean values and CHIM the lowest in 1990 (Fig. S8). CHIM, EMEP, LOTO and MINNI estimated the highest WSOx in NW Germany, whereas CMAQ estimated the largest values for the western coast of Norway (probably due to the inclusion of sea-salt sulfate). MATCH, on the other hand, estimated the highest deposition in Bulgaria in the south-east of the domain. In addition, both EMEP and MATCH estimated large values close to the active volcano Etna on the island of Sicily (Italy), as a result of the volcanic emissions included in these models. The spatial distributions of WSOx estimates for 2000 are similar to those of 1990, albeit with considerably lower values as a result of the large emission reductions within the domain. Domain-mean WSOx decreased between 32 % and 48 % for all models. The models estimated smaller decreases in the domain-mean WSOx between 2000 and 2010 (25 %–38 %), with decreases mostly in the north and west of the domain.
Time series of observed and modelled annual wet deposition of
Over the 1990–2010 period, all six models estimated a decrease in WNOx and WSOx when averaged (median) over all measurement sites (Figs. 2a, c and S9–S14). The model results for WNOx and WSOx follow the same pattern as the observed values, which also decreased, on average, over the same period. By contrast, the models estimated fairly constant rates of wet deposition of reduced nitrogen (WNHx) (Fig. 2b) over the same period, while the median observed deposition decreased by about 40 % between 1995 and 1996 and then remained fairly constant. This decrease occurred at several sites and corresponded with the driest year of the study period, although the largest influence came from two sites in France (FR0008R and FR0010R in Fig. S12), for which there was a decrease in both precipitation and its ammonium content. With regards to the variability between models, the estimates of WNOx are, on average, of a similar magnitude to the observed values, with the exception of MINNI, which underestimated deposition by more than a factor of 2. For WNHx, EMEP and MATCH estimated similar values to those observed, whereas CHIM, CMAQ, LOTO and MINNI tended to underestimate them throughout the time series. CMAQ and LOTO estimated similar values of WSOx to those observed, whereas EMEP and MATCH tended to overestimate deposition and CHIM and MINNI tended to underestimate it.
Performance evaluation of WNOx, WNHx and WSOx estimated by the six models that simulated the individual years 1990, 2000 and 2010. Geometric variance is VG and geometric mean bias is MG. Shaded areas and filled symbols correspond to the acceptance criteria of Chang and Hanna (2004) (blue for VG, red for MG, filled circles for FAC2). Parabolic dashed lines indicate the theoretical minimum VG for a given value of MG.
Proportion of measurement sites with increasing (pink) and
decreasing (blue) trends, and whether they are significant (dark) or
not significant (light), for the observations and model estimates for
the three wet-deposition components,
Figure S15 shows the scatter plots of modelled vs. observed WNOx, WNHx and
WSOx for the years 1990, 2000 and 2010 and Table S3 shows the performance
evaluation of the six models for each of the three deposition components
(WNOx, WNHx and WSOx). Model performance is illustrated in Fig. 3 by plotting
VG against MG for each model using a different symbol to indicate whether the
acceptability criterion for FAC2 is met. The minimum value of VG for a given
value of MG (VG
The PSMK trend calculations gave more significant trends than the other two methods (MK and SMK) for most models, periods and deposition components (Fig. S16). On average, this method gave significant trends for 57 % and 67 % of the observed and modelled time series, respectively, compared with 40 % and 52 % for MK and 45 % and 56 % for SMK. Figures S17 and S18 show that the absolute and relative trends calculated using the MK and SMK methods are similar, although there is some scatter. The only difference between the SMK and PSMK methods is the calculation of significance, and so the trends calculated by these two methods are the same. Since the PSMK method gave the most significant trends, the following analyses use the trends calculated using this method.
Maps of modelled (coloured field) and observed (circles) trends in WNOx for the periods 1990–2000 and 2000–2010.
Figure 4 shows the proportion of increasing and decreasing modelled and
observed trends for the three wet-deposited compounds over the two
11-year time periods and the full 21-year period and whether the
trends are significant (
Maps of modelled (coloured field) and observed (circles) trends in WNHx for the periods 1990–2000 and 2000–2010.
With regards to the spatial distributions of the trends, most of the statistically significant observed trends of WNOx (both increasing and decreasing) for the period 1990–2000 are located in the central and north-eastern parts of the domain (Fig. 5). The five models estimated the most significant trends (mostly decreasing) in the east of the domain, although most of this part of the domain is not covered by the observations. These trends reflect the large reported emission reductions in Ukraine, Russia and Moldova but may have been moderated by increasing trends in precipitation in this region (Fig. S19). The models, however, failed to capture the significant observed increasing and decreasing trends in the centre of the domain. Although there were also large decreases in emissions in the centre and west of the domain (e.g. Germany and the UK), the models did not estimate significant deposition trends in these regions, probably as a result of offsetting by increasing shipping emissions. CHIM estimated the largest area of significant trends (48 % of domain), whereas MINNI estimated the smallest (24 %). For the period 2000–2010, the majority of the statistically significant observed trends (mostly decreasing) are located in the central and western parts of the domain. The models also reproduce this western shift in significant trends, reflecting the spatial shift in decreasing emission trends and the lack of significant trends in shipping emissions (Fig. S3). Increasing observed and modelled trends in precipitation were also found for this region, which may have enhanced the deposition trends. Similarly to the first 11-year period, CHIM estimated the largest area of significant WNOx trends (48 % of domain), whereas MINNI estimated the smallest (30 %).
Maps of modelled (coloured field) and observed (circles) trends in WSOx for the periods 1990–2000 and 2000–2010.
For WNHx, during the period 1990–2000, there are significant observed trends
(all but one decreasing) across the domain, with the largest decrease in the
centre, whereas the models did not estimate significant decreasing trends in
this region (Fig. 6). All five models estimated the most significant
decreasing trends in the east of the domain, corresponding to the region with
the largest emission reductions but with poor coverage by observations. MATCH
estimated the largest WNHx reductions for this period. All models estimated
significant increasing trends around the English Channel despite there being
no significant increases in emissions in this area. This increase in WNHx is
probably the result of increasing trends in precipitation in the region
(Fig. S19) but could also be due to increased
Most of the observed WSOx trends for the period 1990–2000 are significant decreasing trends (Fig. 7). The models also estimate significant decreasing trends in the regions represented by the observations and estimate the largest decreasing trends in the central and eastern parts of the domain (corresponding to the regions with the largest reductions in emissions). EMEP estimated the largest trends and the largest area of significant trends (72 %) and CHIM the smallest trends and smallest area (54 %). Similarly, for the period 2000–2010, all but one of the significant observed trends are decreasing trends. Observed trends in the north-east of the domain were mostly non-significant. The models, in general, estimated significant decreasing trends in the central and western parts of the domain. All models estimated small or non-significant trends in the south and south-east of the domain, corresponding to the regions with increasing trends in modelled precipitation (Fig. S19). This suggests that the increasing precipitation partially offset the reduction in deposition in these regions during this period. LOTO estimated the largest area of significant trends (70 % of domain), whereas CHIM estimated the smallest (50 %).
Focusing on the sites with observations, the observed trends of WNOx (mostly decreasing) were larger, on average, for the 2000–2010 period than for 1990–2000 (Fig. 8a). All of the models except CHIM were able to reproduce this difference. For WNHx, there were more decreasing observed trends during the first 11-year period than during the second. By contrast, all five models estimated more decreasing trends during the second period. However, there were very few significant observed or modelled WNHx trends. This is not the case for WSOx, for which most of the observed and modelled trends were significant. Observed trends of WSOx (mostly decreasing) are largest, on average, during the first 11-year period. Although the models reproduce this difference, there is substantial variation between the models, with EMEP and MATCH estimating larger trends, on average, than those observed for the first period, in which CHIM and MINNI estimated smaller ones and LOTO estimated similar trends. This reflects the tendencies of the models to under- or overestimate annual wet deposition. The trends calculated for the period 2000–2010 using all the available sites for that period are also shown in Fig. 8. Using all sites gives slightly smaller average observed and modelled trends for WNOx, WNHx and WSOx than using the same sites as the period 1990–2000 (i.e. less sites). This is probably due to the inclusion of sites in the south-east of the domain for which the meteorological models estimated increasing precipitation trends for this period. Despite these small differences, the distribution of trends is very similar and we can conclude that the sites used in the trend analysis for both 11-year periods are fairly representative of the area covered by all sites.
Tukey-style box plots of observed and modelled absolute
Plotting the distributions of relative trends makes it possible to compare
emission trends with observed and modelled deposition trends (Fig. 8b). Total
Total
Tukey-style box plots of the contributions of the different factors
(BC: boundary conditions, Emis: emissions, Met: meteorology and Resid:
residual interactions) to the trends (Tot) of
Performance evaluation of the accumulated seasonal and annual precipitation estimates for the meteorological data used in the simulations by CMAQ, LOTO, MATCH and the common meteorological data used in the other models (OTHERS). Shaded areas and filled symbols correspond to the acceptance criteria of Chang and Hanna (2004) (blue for VG, red for MG, filled circles for FAC2). Parabolic dashed lines indicate the theoretical minimum VG for a given value of MG.
Figure 9 shows the contributions of the changes in emissions (Emis), boundary conditions (BC) and meteorology (Met) to the modelled trends (Tot) of WNOx, WNHx and WSOx at all measurement sites. For all three deposition components and both time periods, the largest contribution to the overall modelled trend is the reduction in emissions. Many of the overall trends are smaller than the trends due to emissions alone as a result of positive contributions from Met and non-linear interactions (Resid, which also could include contributions from the meteorology). However, for most of the trends, this offsetting is smaller for the second period, resulting in a stronger influence of the emission reductions for this period. The larger offsetting by meteorology and other interactions (represented by the residual component) for the first period can also be seen in the regional analysis of the land grid cells presented in Figs. S21–23, especially for England (EN), central Europe (ME) and, to a lesser extent, Scandinavia (SC), three subregions that together contain about half of the measurement sites. This difference in offsetting between the periods is not as apparent for the analysis of the land grid cells of the entire domain, since the offsetting is larger in the second period for some regions, such as the Iberian Peninsula and the Mediterranean, which are poorly represented by the observations (only one site). These effects can also be seen in the spatial distributions of the different contributing factors (Figs. S24–S29). The attribution analysis for all models shows that, for the period 1990–2000, there was a positive contribution from the Met and Resid factors in the centre and north of the domain that offset decreasing trends due to emissions alone, whereas there was a negative contribution in the Mediterranean and southern parts of the domain that reinforced them. This situation was reversed for the period 2000–2010, with negative contributions in the north and positive contributions in the south. This reflects the differences in the precipitation trends between the two periods (Fig. S19), providing further evidence that the trends in precipitation drove the contribution from the Met factor. Furthermore, the spatial distribution of the Resid factor is similar to that of Met, which suggests that Resid was also driven by precipitation trends. The offsetting and reinforcement of the trends due to emissions alone can be seen more clearly by summing the BC, Met and Resid factors (Fig. S30). The spatial distributions of the positive and negative contributions are very similar for all models despite the fact that some of them used different meteorological models, suggesting that the shift of the positive contributions from the north to the south of the domain between the two periods is a robust result.
Tukey-style box plots of observed and modelled trends in precipitation at the wet-deposition sites for the two periods 1990–2000 and 2000–2010.
Since precipitation rates have a strong influence on wet deposition, it is
useful to evaluate model performance for precipitation at the same sites with
observations of wet deposition to see if it can help to explain model
performance for WNOx, WNHx and WSOx. Model biases are very small for
accumulated annual precipitation, with three meteorological models (those
used by CHIM, CMAQ, EMEP, LOTO and MINNI) underestimating the geometric mean
precipitation (by 4 %–8 %) and one overestimating it (that used by
MATCH, by 5 %) (Fig. 10 and Table S4). Model biases are also small for
seasonal precipitation. The meteorological models used by all of the CTMs
except MATCH performed the
worst in summer with underestimations of 18 %–28 %. By contrast, the
meteorological model used by MATCH had a very small bias (2 %) for this
season. A comparison of the observed precipitation trends for the two 11-year
periods shows that the trends are small and positive, on average, and very
similar for the two periods, although the average trends for the first period
are slightly larger than those for the second (Fig. 11). CHIM, EMEP and MINNI
estimated very similar median trends to those of the observations. The HIRLAM
model used by MATCH also gave trends in the same range, although this model
estimated slightly larger median trends for the second period compared with
the first. By contrast, RACMO2 (used by LOTO) estimated larger median trends
than the other models and estimated positive median trends for the first
period and negative for the second, which could be due to the fact that the
RACMO2 simulation is not nudged towards the observed precipitation. Very few
(
Since wet-deposition estimates are also strongly dependent on atmospheric
concentrations in the air column, it is useful to evaluate model performance
for (surface) concentrations to see if it can help to explain model
performance for wet deposition. A more detailed analysis of the trends in
atmospheric concentrations estimated by the EDT simulations is provided by
Ciarelli et al. (2018). In contrast to wet deposition, for which most models
underestimated deposition rates or had a small bias (with the exception of
EMEP and MATCH for WSOx), all models overestimated mean atmospheric
concentrations of TNO3, TNH4 and TSO4 or had a small bias (Figs. 12 and S31
and Table S5). All models overestimated the geometric mean TNO3 and TSO4,
with the largest overestimation by CMAQ. Model biases were generally smaller
for TNH4, with some models overestimating concentrations and others
underestimating them. An analysis of the correlation between the performance
statistics of wet deposition and atmospheric concentrations at the same sites
shows that there is a significant (
Performance evaluation of the atmospheric concentrations of TNO3, TNH4 and TSO4 estimated by the six models that simulated the individual years 1990, 2000 and 2010. Shaded areas and filled symbols correspond to the acceptance criteria of Chang and Hanna (2004) (blue for VG, red for MG, filled circles for FAC2). Parabolic dashed lines indicate the theoretical minimum VG for a given value of MG.
Although there are no observations for evaluating dry deposition it is still useful to compare the dry deposition estimates of the models at the same sites that were used for the evaluation of wet deposition in order to determine whether the differences between the estimates of dry deposition can explain the differences between the estimates of wet deposition. Figure 13 shows that for dry deposition of oxidised N (DNOx), the median model estimates differ by a factor of about 2 for most of the time series, with LOTO estimating the lowest rates and CMAQ the highest. These high DNOx estimates by CMAQ could be due to the high TNO3 concentrations estimated by this model. There is slightly more agreement between the model estimates of dry deposition of reduced N (DNHx), with median estimates differing by about a factor of 1.5. However, MINNI estimated an increase in dry deposition between 1996 and 1999, which did not occur for the other models. Out of the other models, MATCH estimated the smallest values and CHIM the largest for most of the time series. The low DNHx estimates of MATCH could be due to the low estimates of TNH4, which in turn could be the result of the small overestimation of WNHx by this model. For dry deposition of sulfur (DSOx), the estimated median deposition values differ by about a factor of 2, mainly as a result of CHIM estimating higher values than the other models for the entire series. This could partly be due to the underestimation of WSOx by this model.
With regards to the total deposition (wet plus dry), MINNI and LOTO estimated smaller median values for oxidised nitrogen than the other models by a factor of 1.5 to 2 for the entire time series, whereas CMAQ estimated the largest values for the years 1990 and 2000 (Fig. S32). CHIM and MINNI estimated the lowest values for reduced nitrogen and EMEP and MATCH the highest for most of the time series, with a similar range of variability as for oxidised nitrogen. Summing the reduced and oxidised components to obtain total nitrogen deposition (Fig. S33) shows that MINNI and LOTO estimated the lowest median values (as a result of their low estimates of oxidised N deposition) and CMAQ and MATCH the highest (as a result of the CMAQs high estimates of oxidised N deposition and MATCH's high estimates for both the oxidised and reduced components). For sulfur, EMEP and MATCH estimated very similar rates of total deposition, as did CMAQ for 2000 and 2010. CHIM and LOTO estimated similar rates of S deposition, which were lower than those of EMEP and MATCH, whereas MINNI estimated the lowest values (Fig. S33).
Time series of modelled dry deposition of
Like any study involving observed data and/or model simulations, the results
presented here are subject to various sources of uncertainty. The national
emission data used in the simulations are based on the officially reported
values. The European Environment Agency suggests that the emission estimates
for European member states have an uncertainty of about
Another source of uncertainty is the meteorological data used in the
simulations, as well as the procedures within the models that parameterise
the atmospheric conditions from those data. Since annual wet deposition is
correlated with accumulated precipitation, it seems logical to focus the
discussion on precipitation. As shown in Fig. 10, the annual accumulated
precipitation calculated by the meteorological models used in most of the
simulations (based on ERA-Interim reanalysis data) is lower than the observed
precipitation by 4 %–8 % on average. This is consistent with the
analyses of Dee et al. (2011), who showed that the ERA-Interim reanalysis
precipitation data underestimates mean precipitation rates by
0.2–1.0 mm day
The spatial resolution used for the model simulations can also add
uncertainty, since the model estimate for a grid cell may not be
representative of the location of the measurement site. The grid cell areas
of the model domain range from approximately 425 to 1050 km
With regards to the trend analyses, the small number of significant observed trends (especially for WNHx) makes it difficult to perform a robust analysis to determine whether or not the models can reproduce the trends. Another limitation of the trend analysis is the requirement for linear trends, which is not the case for most of the trends for the period 1990–2010. A trend analysis for the 21-year period using non-linear trend estimation methods (see e.g. Venier et al., 2012) could provide a more robust evaluation. However, the linear trend analysis does allow for the assessment of trends for shorter periods, provided there are enough sites with significant observed and/or modelled trends.
Although the uncertainties highlighted above may contribute to the systematic
underestimation of wet deposition by many of the models, it is unlikely that
they account for all of the model bias. For example, in the worst case,
underestimates of 30 % in the emissions and 8 % in the precipitation
are unlikely to give an underestimate of wet deposition by 70 % (although
it is not impossible due to non-linearities) and so there are probably other
explanations for the underestimations by some models. Also, for the same
input data, some models have a very small bias, while for others it is large.
It also seems unlikely that the problem comes from underestimated emissions,
since this would be expected to also lead to an underestimate of atmospheric
concentrations (unless the errors are compensated by errors in other
variables, such as the boundary layer height), which is not the case. Another
possibility is an overestimation of dry deposition, which would leave an
insufficient amount of pollutants in the atmosphere and so wet deposition would be
underestimated. However, this would also be expected to lead to an
underestimation of atmospheric concentrations as well. The lack of
bidirectional
The most plausible explanations for large model biases are deficiencies in
the wet-deposition schemes of the models (e.g. uncertainties in the
scavenging coefficients for gases and particles) and/or errors in the
frequency and intensity of precipitation events, the vertical profiles of the
pollutants or the parameterisation of clouds and cloud chemistry. Similar
conclusions were made by Vivanco et al. (2017), who found a general
underestimation of wet deposition by several models for four campaigns over
the period 2006–2009. A comparison of model biases between their study and
ours for the models common to both studies (CHIM, CMAQ, EMEP, LOTOS and
MINNI) shows that model performances in the two studies are fairly
consistent. For example, in both studies MINNI underestimated WNOx the most,
whereas EMEP had a very low bias and the models CMAQ, LOTO and MINNI
underestimated WNHx the most and EMEP had the smallest bias. Also, in both
studies LOTO and MINNI tended to underestimate WSOx and EMEP tended to
overestimate it. CHIM had differing results depending on the study (e.g.
underestimating WNOx in this study but having a very small bias in Vivanco et
al. (2017), although that may be because of the different model version and
time periods used (annual simulations vs. month-long campaigns). Despite
these differences, the results are sufficiently consistent with the
conclusions made by Vivanco et al. (2017), namely that the tendency of models to underestimate wet
deposition and overestimate atmospheric concentrations (as is the case for
oxidised and reduced nitrogen) is likely to be due to deficiencies in
simulating wet-deposition processes, which could be related to the vertical
concentration profiles, scavenging coefficients or in-cloud processes,
including the parameterisation of clouds. The case of WSOx is slightly
different. In this study, CHIM, CMAQ, LOTO and MINNI tended to underestimate
WSOx and overestimate TSO4, which again could be due to deficiencies in the
wet-deposition processes (including vertical concentration profiles, cloud
parameterisation), while the other models (EMEP and MATCH) overestimated the
wet deposition and the concentrations, which could be due to an overestimate
of
The results presented here are also fairly consistent with studies that have evaluated individual models, despite the fact that these studies used different model versions, meteorological data and measurement sites. For example, Simpson et al. (2006) found that the EMEP model (version rv2.0) underestimated mean WNOx and WNHx by 16 %–26 % and 16 %–17 % when compared with measurement data from 160 forest sites. In the present study, EMEP (version rv4.7) underestimated WNOx and WNHx by 2 % and 14 %, respectively. However, for WSOx, Simpson et al. (2006) found that the EMEP model also underestimated mean deposition by 9 %–26 %, whereas in the current study we found an overestimate by 31 %. On a national level, Schaap et al. (2017) found that LOTOS-EUROS (version 1.10) underestimated the mean wet deposition of oxidised and reduced nitrogen by 38 % and 21 % and that of sulfur by 44 %, when compared with observations made at 150 sites in Germany. In the present study, LOTOS-EUROS (version 1.10.005) also underestimated mean deposition by 35 %, 41 % and 23 %, for WNOx, WNHx and WSOx.
A comparison of the relative trends of total domain emissions of
precursor species (
Time series of observed and bias-corrected modelled wet deposition
of WNOx, WNHx and WSOx. Points represent the median value for all measurement
sites and the shading (or error bars) represents the interquartile range. The
shaded period at the start of the time series represents the time period used
to calculate the bias correction. Note that each plot has a different
Despite the limitations and uncertainties of the analyses presented here, it
has been possible to statistically evaluate the modelled trends in
deposition. With regards to the significance and direction of the trends, the
models generally reproduce the observed larger and more significant
decreasing WNOx trends in the second 11-year period compared with the
first, despite similar relative emission reductions for both periods. The
analysis of precipitation trends, simulations with constant emissions and the
trend attribution analysis all suggest that this is due to a larger increase
in precipitation and/or other changes in the meteorology during the first
period, partially offsetting the decrease in wet deposition due to emission
reductions. This effect can be seen more clearly in Fig. 14, which shows that
the median relative trend of the observed and modelled WNOx at the
measurement sites is smaller for the first period. In fact, all models
estimate smaller average relative trends in wet deposition than those of the
emissions during the first period and larger average relative trends during
the second period due to changes in the meteorology and/or boundary
conditions. Another factor that could influence the different responses of
wet deposition during the two periods to changes in emissions is the
non-linear response of TNO3 concentrations to reductions in
The fact that the year-to-year relative changes in modelled deposition are
more reliable than the absolute changes and that model biases do not change
much over the 1990–2010 period (Fig. S34) opens up a possibility for
improving model estimates of deposition. If the model bias (MG) is calculated
for an initial period (e.g. the first 3 years of the time series), then
the bias correction necessary to remove this initial bias (multiplying the
model estimates by 1 MG
We have evaluated the wet deposition of sulfur (WSOx) and oxidised (WNOx) and reduced (WNHx) nitrogen estimated by six atmospheric chemistry transport models using observations from the EMEP monitoring network for the period 1990–2010. Most of the models met the predefined acceptability criteria for the three components, although there were some exceptions. MINNI underestimated the wet deposition of all three components by more than a factor of 2 to 3, on average. The fact that all models used the same emissions, boundary conditions and, where possible, meteorology suggests that this general underestimation is due to the parameterisation of the models, such as deficiencies in the wet-deposition scheme, the vertical concentration profiles of the pollutants or the parameterisation of clouds and cloud chemistry. The other exception is Chimere (CHIM), which underestimated WNHx and WSOx by more than a factor of 2 to 3, on average. The fact this model had a small bias for WNOx suggests that the model underestimation of WNHx and WSOx is related to the parameterisations for reduced nitrogen and sulfur compounds, such as the species-specific scavenging coefficients, the gas phase or cloud chemistry schemes or the aerosol physics. In order to understand the underestimation of wet deposition by MINNI and Chimere, a detailed study of the chemical and physical processes occurring in the model column would be required, which is out of the scope of the present study.
More than half of the observed trends of WNOx and WNHx for the two periods
1990–2000 and 2000–2010 were not significant, making it difficult to
evaluate the modelled trends statistically. For the sites with both
significant observed and modelled trends, the models tended to estimate
similar or smaller trends than those observed, with MINNI underestimating all
but two of the observed trends, reflecting the tendency for this model to
underestimate WNOx and WNHx. Despite small but significant
An analysis of the factors contributing to the modelled trends showed that
reductions in emissions contributed most to the trend estimates. However,
changes in meteorology, boundary conditions and other factors also have an
influence in the trends estimated at monitoring sites, suggesting that the
emission reduction measures had a larger effect during the second period at
these sites. Changes in atmospheric chemistry due to large reductions in
Technical details of the EURODELTA project simulations that
permit the replication of the experiment are available on the wiki of the
EMEP Task Force on Measurement and Modelling
(
The statistical significance of the trends in observed and modelled wet
deposition and precipitation, as well as in the emissions, was calculated
using the Mann–Kendall (MK) test, which assesses whether there is a
statistically significant monotonic trend in a data time series (Mann, 1945;
Kendall, 1970). This is a non-parametric test and so is suited to data sets
that are not necessarily normally distributed (unlike other methods, such as
linear regression). The method can be used to assess the significance of a
trend, even if it is non-linear, and is fairly insensitive to missing data.
The statistic
Since the temporal variability of wet deposition depends
strongly on seasonal precipitation cycles, we also applied the trend analysis
to the observed and modelled deposition for winter, spring, summer and autumn
individually and then calculated the trend significance from the sum of the
The supplement related to this article is available online at:
ACo coordinated the EURODELTA-Trends (EDT) exercise and WA was responsible for the compilation and quality control of the observations. The following modelling teams set-up, pre-processed, ran and post-processed the simulations for each model: FC, BB, MGV and ACo for CHIM; ST, MTP for CMAQ; HF and PW for EMEP; AM and MS for LOTO; CA and RB for MATCH; MM, MA, GB, ACa and MD for MINNI. Additional post-processing of model output and uploading to the AeroCom server was done by KC. All of the analyses presented in this paper were carried out by MT and MGV with assistance from GC, KM, NO, VR, YR and ACo.
The authors declare that they have no conflict of interest.
We would like to express our thanks to all those who are involved in the EMEP
monitoring efforts and have contributed through operating sites, performing
chemical analysis and by submission of data. This work was supported by the
Co-operative Programme for Monitoring and Evaluation of the Long-range
Transmission of Air pollutants in Europe (EMEP) under the UNECE. The Ineris
coordination of the EURODELTA-Trend exercise was supported by the French
Ministry in charge of Ecology in the context of the Task Force on Measurement
and Modelling of the EMEP programme of the LRTAP Convention. The GAINS
emission trends were produced as part of the FP7 European Research Project
ECLIPSE (Evaluating the Climate and Air Quality Impacts of Short-Lived
Pollutants), grant no. 282688. Meteorological forcings with the WRF model
were provided by Robert Vautard and Annemiek Stegehuis from LSCE/IPSL. We
also thank Erik van Meijgaard of the Royal Netherlands Meteorological
Institute (KNMI) for providing the RACMO2 simulations that were used by
LOTOS-EUROS. The simulations with the EMEP MSC-W model were supported by the
Research Council of Norway in the framework of the Programme for
Supercomputing: through the EMEP project (grant NN2890K) for CPU, and the
Norstore project “European Monitoring and Evaluation Programme” (grant
NS9005K) for data storage. The participation of CIEMAT was financed by the
Spanish Ministry of Agriculture and Fishing, Food and Environment. The
CHIMERE simulations were performed using the TGCC supercomputers under GENCI
computing allocation. The MATCH participation was partly funded by the
Swedish Environmental Protection Agency through the research programme
Swedish Clean Air and Climate (SCAC) and NordForsk through the research
programme Nordic WelfAir (grant no. 75007). The computing resources and the
related technical support used for MINNI simulations have been provided by
CRESCO/ENEAGRID High Performance Computing infrastructure and its staff. The
infrastructure is funded by ENEA, the Italian National Agency for New
Technologies, Energy and Sustainable Economic Development and by Italian and
European research programmes (