Scholarly article on topic 'Assessing ozone exposure for epidemiological studies in Malmö and Umeå, Sweden'

Assessing ozone exposure for epidemiological studies in Malmö and Umeå, Sweden Academic research paper on "Earth and related environmental sciences"

CC BY-NC-SA
0
0
Share paper
Academic journal
Atmospheric Environment
Keywords
{"Land use regression" / Ozone / "Air pollution modelling" / Epidemiology / "Risk assessment"}

Abstract of research paper on Earth and related environmental sciences, author of scientific article — E. Malmqvist, D. Olsson, A. Hagenbjörk-Gustafsson, B. Forsberg, K. Mattisson, et al.

Abstract Ground level ozone [ozone] is considered a harmful air pollutant but there is a knowledge gap regarding its long term health effects. The main aim of this study is to develop local Land Use Regression [LUR] models that can be used to study long term health effects of ozone. The specific aim is to develop spatial LUR models for two Swedish cities, Umeå and Malmö, as well as a temporal model for Malmö in order to assess ozone exposure for long term epidemiological studies. For the spatial model we measured ozone, using Ogawa passive samplers, as weekly averages at 40 sites in each study area, during three seasons. This data was then inserted in the LUR-model with data on traffic, land use, population density and altitude to develop explanatory models of ozone variation. To develop the temporal model for Malmö, hourly ozone data was aggregated into daily means for two measurement stations in Malmö and one in a rural area outside Malmö. Using regression analyses we inserted meteorological variables into different temporal models and the one that performed best for all three stations was chosen. For Malmö the LUR-model had an adjusted model R 2 of 0.40 and cross validation R 2 of 0.17. For Umeå the model had an adjusted model R 2 of 0.67 and cross validation adjusted R 2 of 0.48. When restricting the model to only including measuring sites from urban areas, the Malmö model had adjusted model R 2 of 0.51 (cross validation adjusted R 2 0.33) and the Umeå model had adjusted model R 2 of 0.81 (validation adjusted R 2 of 0.73). The temporal model had adjusted model R 2 0.54 and 0.61 for the two Malmö sites, the cross validation adjusted R 2 was 0.42. In conclusion, we can with moderate accuracy, at least for Umeå, predict the spatial variability, and in Malmö the temporal variability in ozone variation.

Academic research paper on topic "Assessing ozone exposure for epidemiological studies in Malmö and Umeå, Sweden"

Accepted Manuscript

Assessing ozone exposure for epidemiological studies in Malmo and Umea, Sweden M

E. Malmqvist , D. Olsson , A. Hagenbjörk-Gustafsson , B. Forsberg , K. Mattisson , E. Stroh , M. Strömgren , E. Swietlicki , L. Rylander , G. Hoek , H. Tinnerberg , L. Modig

PII: S1352-2310(14)00385-9

DOI: 10.1016/j.atmosenv.2014.05.038

Reference: AEA 12982

To appear in: Atmospheric Environment

Received Date: 2 October 2013

Revised Date: 9 May 2014

Accepted Date: 13 May 2014

Please cite this article as: Malmqvist, E, Olsson, D, Hagenbjörk-Gustafsson, A, Forsberg, B, Mattisson, K, Stroh, E, Strömgren, M, Swietlicki, E, Rylander, L, Hoek, G, Tinnerberg, H, Modig, L, Assessing ozone exposure for epidemiological studies in Malmö and Umeâ, Sweden, Atmospheric Environment (2014), doi: 10.1016/j.atmosenv.2014.05.038.

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Assessing ozone exposure for epidemiological studies in Malmo and Umea, Sweden

Malmqvist Ea, Olsson Db, Hagenbjork-Gustafsson Ab, Forsberg Bb, Mattisson Ka, Stroh Ea, Stromgren Mc, Swietlicki Ed, Rylander La, Hoek Ge, Tinnerberg Ha, Modig Lb

aDivision of Occupational and Environmental Medicine, Department of Laboratory Medicine, Lund University, Sweden

bDivision of Occupational and Environmental Medicine, Department of Public Health and Clinical Medicine, Umea University, Sweden

cDepartment of Geography and Economic History, Umea University, Sweden

dDivision of Nuclear Physics, Department of Physics, Lund University, Sweden

eInstitute for Risk Assessment Sciences (IRAS), Utrecht University, Utrecht, The Netherlands

Abstract

Ground level ozone [ozone] is considered a harmful air pollutant but there is a knowledge gap regarding its long term health effects. The main aim of this study is to develop local Land Use Regression [LUR] models that can be used to study long term health effects of ozone. The specific aim is to develop spatial LUR models for two Swedish cities, Umea and Malmo, as well as a temporal model for Malmo in order to assess ozone exposure for long term epidemiological studies. For the spatial model we measured ozone, using Ogawa passive samplers, as weekly averages at 40 sites in each study area, during three seasons. This data was then inserted in the LUR-model with data on traffic, land use, population density and altitude to develop explanatory models of ozone variation. To develop the temporal model for Malmo, hourly ozone data was aggregated into daily means for two measurement stations in Malmo and one in a rural area outside Malmo. Using regression analyses we inserted meteorological variables into different temporal models and the one that performed best for all three stations was chosen. For Malmo the LUR-model had an adjusted model R of 0.40 and cross validation R2 of 0.17. For Umea the model had an adjusted model R2 of 0.67 and cross validation adjusted R of 0.48. When restricting the model to only including measuring sites from urban areas, the Malmo model had adjusted model R of 0.51 (cross validation adjusted R2 0.33) and the Umea model had adjusted model R2 of 0.81 (validation adjusted R2 of 0.73). The temporal model had adjusted model R 0.54 and 0.61 for the two Malmo sites, the cross validation adjusted R was 0.42. In conclusion, we can with moderate accuracy, at

least for Umeâ, predict the spatial variability, and in Malmo the temporal variability in ozone variation.

1. Introduction

1.1 Ozone and health

Ozone is present in the stratosphere but also in the troposphere (ground level up to 10km) where it can be detrimental to health. Ground level ozone [hereafter referred to as ozone] is due to its reactive nature a potent oxidant and thus harmful to human health, vegetation and materials. It is one of the most harmful air pollutants in Europe today (EEA, 2011). Ozone can enter the lung and due to its reactive nature cause oxidative stress (Mudway and Kelly, 2004; Koike and Kobayashi, 2004, Sousa et al., 2013). There is evidence that ozone can cause both acute and chronic effects to human health (OECD, 2008; WHO 2013). There are several studies on the short term effect of ozone on respiratory diseases and mortality (Sousa et al., 2013; WHO, 2008). There is less evidence on long term effects due to lack of accurate exposure data. However some data suggest that there is a long term effect of ozone on respiratory mortality as well and that the effects of ozone have a very low threshold (Brunekreef et al., 2012). It has been suggested that for conducting long term health effect studies of ozone it is crucial to know the local spatial variation to reduce exposure misclassification (WHO, 2008). Although large parts of the ozone measured in Sweden are generated elsewhere, there is a regional and local variation in concentrations. The regional variation of ozone is well known, while the local variation is much less studied (WHO, 2008; SMHI, 2010). WHO have set a health based limit for ozone with a maximum 8-h mean concentration of 100 |g/m3 (WHO, 2006).

Recent studies have focused on the development of more precise exposure assessment models for primary pollutants such as nitrogen oxides [NOx] and particulate matter [PM] (Beelen et al., 2013; Noth et al., 2011; de Hoogh et al., 2013). Less focus has been put on developing models for ozone exposure assessment in epidemiological studies, even if the health risks of ozone exposure is considered at least as serious as primary pollutants (Brunekreef et al., 2012). It has been found that there is a need of ozone modelling since ozone exposure can vary substantially within the area covered by the measurement station (WHO, 2008).

To summarize, there is an obvious and important need to investigate the spatial variation of ozone on the local scale to make it possible to study the long term health effects of ozone.

1.2 Ozone variation in space

Ozone is a secondary pollutant and is formed through chemical reactions including NOx, carbon monoxide [CO] and volatile organic compounds [VOC], collectively called ozone precursors, in the presence of sunlight. The formation of ozone acts over extensive areas, and levels of ozone might be caused by combustion related emissions and photochemical reactions thousands of kilometres away (WHO, 2008). In urban areas on the other hand, emissions of pollutants can reduce ozone levels close to the source, such as road traffic areas (WHO, 2008; Intergovernmental Panel on Climate Change, 2007). This can be explained by freshly emitted nitrogen monoxide [NO] which reacts with ozone and converts it into nitrogen dioxide [NO2] and oxygen [O2]. The implication is that ozone near roadside are scavenged often referred to as the urban decrement (Stedman and Kent, 2008). Ozone precursors can also be carried away from the source and also cause ozone formation in the downwind plume from the city (WHO, 2008; Intergovernmental Panel on Climate Change, 2007). In Europe emissions of ozone precursors have decreased with about 30% from 1990-2004, however levels of ozone have not decreased as much (Garcia et al., 2010). An international policy review has concluded that there is strong evidence that background ozone concentrations in the northern hemisphere have instead increased by up to 10 pg/m per decade over the last 2030 years (Raes and Hjorth, 2006). Also in many Swedish cities, the trend over the last years is that background ozone is slightly increasing which can partly be explained by a change in the composition of vehicle exhaust (Johansson and Forsberg, 2005). However, episodes of high levels of ozone seems to be fewer and less severe (Jenkin, 2008).

The ozone processes are not only affected by traffic and other sources of combustion gases. During the growing season plants emit highly reactive hydrocarbons, such as isoprene or terpenes, especially when temperatures are higher (Sanderson, Jones, Collins, Johnson and Derwent, 2003). These hydrocarbons are so called biogenic VOCs [hereafter referred to as BVOCs] that increase ozone formation. However, in the Scandinavian region it is common with certain tree types that emit BVOCs that could have an opposite effect and decreasing levels of ozone through ozonolysis, i.e. ozone causing the breakage of molecules thru oxidation (Curci et al., 2009).

Spatially, altitude could also be a predictor for ozone concentrations, at lower altitudes ozone levels tend to be lower due to local destruction of ozone with NO (titration) and dry deposition. Whereas in higher altitudes and in coastal regions winds tend to be stronger and there is no formation of shallow boundary layers which affects the atmospheric mixing rates of pollutants (Stedman and Kent, 2008).

To summarize, ozone formation are depending on ozone precursors, mainly emitted by traffic. Near roadside however, traffic exhaust can lower levels of ozone. BVOCs can contribute to both formation and destruction of ozone.

1.3 Ozone variation in time

Ozone has an annual variation with a typical summer maximum in major cities due to local photochemical productions in stagnant meteorological conditions and a spring maximum in other areas due to increased solar radiation (Simpson, 1995; Monks, 2000). Ozone has a daily pattern driven by levels of ozone precursors and solar radiation and tends to build up and peak in the afternoons (Garcia et al., 2010).

2. Aim and outline for this study

The main aim of the study is to develop a Land Use Regression [LUR] model that can be used in epidemiological studies on health effects of ozone exposure. The specific aim is to describe the local spatial variation of ozone in two Swedish cities (Malmo and Umea) with a LUR model. To be able to use this model in other time windows than the studied period adjustments can be made with stationary measurements. But this data have continuously short periods of missing data. In addition we will therefore also develop a model accounting for temporal variations based on meteorological data (only Malmo) in order to assess ozone levels when there are missing data from the stationary measurements of ozone.

3. Method

3.1 Study areas

We selected two study areas in different parts of Sweden for the measuring campaigns; the Malmo region in the south and the Umea region in the north of Sweden with a north-south distance of 1250 km between the two areas (Figure 1). In the Malmo area we included the city of Malmo with surroundings (Figure 2), with approximately 300 000 inhabitants in the area. The Umea region includes the city of Umea and nearby villages with about 150 000 inhabitants. The meteorological definition of spring is when there is a daily temperature above 0°C during 7 executive days. This period is expected to arrive in the middle of February in the Malmo area and in the middle of April in the Umea area. However, during the studied year 2012 spring arrived in late February in Malmo and in late March in Umea. Summer (above 10°C as a daily temperature for 5 executive days) arrived in late April in Malmo (expected middle of May) and in late May in Umea (expected beginning of June) (SMHI, 2013). Although there is no strong north-south gradient in the average ozone levels,

episodes with very high levels of ozone are more frequent in the southern parts of Sweden (SMHI, 2010). The Malmo area has a more traffic and thus higher levels of ozone precursors in comparison with the Umeá area.

Figure 1, Study areas

3.2.1 Measurements for the LUR-model

The week of the measurements was coordinated between the two areas to get as comparable

data as possible. The ozone measurements were performed as weekly averages at 40 sites in

each study area, during three seasonal measuring periods in 2012;

Spring: 16th - 24th of April

Summer: 28th of May - 4th June and

Autumn: 20th - 27th August.

The measurement stations were categorized, based on their location, as: regional background, urban background and traffic sites according to the criteria by EUROAIRNET (1999). The criteria say that a regional background site is located in an area which is not continuously built up and without influence of nearby industry or traffic. An urban background site is located in a continuously built up area but not heavily influenced by nearby traffic or industry. A traffic site is located so their pollution level is mainly influenced by emission from a nearby road. The distribution of sites between these three classes should also represent the distribution of the population, with most of the sites located in urban background areas and a smaller part in rural areas and in areas heavily influenced by traffic (Table 1). We later on had to exclude one site in Malmo and one in Umeá due to measurement errors.

The Ogawa diffusive sampler (Ogawa & Company, Pompano Beach, FL, USA) was used for all measurements. The sampler consists of a Duralon cylinder containing two nitrite-coated filters. The coated filters were supplied by the manufacturer (Ogawa). Ambient ozone oxidizes nitrite [NO2-] on filters to nitrate [NO3-]. The filters are extracted and the accumulated NO3- is analyzed by ion chromatography. The time-weighted average ozone concentrations for one-week measurement are estimated by the amount of NO3- on the filter, the uptake rate of the sampler, and the exposure time. A comparison between parallel measurements with the reference ozone method (UV-monitors) and the Ogawa samplers during eight sets of replicate measurements in Umeá 2011 showed good precision of six

samplers (RSD=1-4%) and the uptake rate was determined to 27.3 ml/min. (Hagenbjork-Gustafsson, 2012)

The samplers were placed inside an Ogawa rain and wind shelter at approximately 2-3 meters from the ground. All samples were analyzed at the Division of Occupational and Environmental medicine, Umea University, Umea. Two field blanks were deployed for each study site and at each sampling period. These blanks were run together with the samples, and the blank signal in each run was subtracted from the samples. The ozone concentrations obtained with the Ogawa samplers were compared to continuous measurements by active UV-absorption monitors at two sites included in the Regional Air Quality Monitoring in Malmo during two of the measurement periods (Table 2).

3.2.2 Land Use Regression model

Briefly, the basis of developing a LUR-model is to collect a large number of digitalized geographical variables, such as road networks and land use, surrounding each of the measuring sites and to include these into a Geographical Information Systems [GIS]. The collected variables should previously be known to have an impact on the pollutant of interest. LUR is a statistical approach to model air pollution, by trying to explain the variation in measured levels between different sites based on differences in geographical variables, hereafter referred to as predictor variables. Levels of pollution may then be predicted for any location, such as individual homes, using the parameter estimates derived from LUR-model. GIS analyses were conducted to derive the values for the predictor variables at each monitoring site.

3.2.3 Predictor variables

We used ArcGIS version 10 (ESRI) to extract the predictor variables. The geographical coordinates of each measurement site was obtained using GPS and checked using digital maps. Around each measurement station we created buffers. The buffer zones used to study influence from traffic were: 25, 50, 100, 300, 500 and 1000 m to correspond for both near source emissions and larger areas of traffic density. We intersected the buffers with a digital road network holding information on traffic intensity attached. The digital road network was developed by the Swedish Transport Agency from 2010 for both areas, complemented with road data from the local municipalities. With this data it was possible to collect total traffic

load within each buffer zone (calculated as: ([length of the road segment] x [traffic intensity])/road segment for all roads within a buffer zone). In Malmo this was also done separately including only heavy (such as trucks and buses) traffic. The same calculations were also done only including major roads (>5000 vehicles/24 hours). Traffic intensity and distance to the nearest road and the nearest major road and other traffic variables was also calculated.

The buffer zones used to study the influence of land use were: 100, 300, 500, 1000 and 5000 meters. Within each buffer zone, areas of different land cover were calculated in m . Land use data was collected from the EU initiated Coordination and Information on the environmental programme database [CORINE] from 2006 for Malmo area and 2000 for Umea area. It compromises 44 land cover types and was regrouped into 6 classes (High density residential areas, Low density residential areas, Industry, Port, Urban Green and Semi natural andforested areas [the latter two hereafter referred to as Natural]. In addition we used locally derived population density data for 2009 in Malmo. In Umea the population data was derived from Statistics Sweden for the year 2008. We also calculated distance to the sea and altitude.

3.2.4 Model development

The predictor variables are combined in a supervised step wise linear regression model to try to find the combination of variables explaining as much of the spatial variation in pollution levels as possible. We developed a mean of ozone for the three measurement periods (spring, summer and autumn), which were used in the analyses. Regression models that maximises the R2 values, adjusted percentage explained variance were developed to explain the variance in ozone concentration at different sites. First, univariate regression analyses for all predictor variables were conducted. The variable with the highest R was chosen as the start model and

then the remaining variables were added consecutively. The predictor variables with the

22 highest additional increase in R were maintained if they also increased in adjusted R more

than 1%. If a variable had buffer sizes that were overlapping in the model, the variable was

rewritten (and analyses redone) using concentric adjacent rings (so called doughnut rings).

Diagnostic tests were applied to the models such as; multi co-linearity between included

variables by Variance Inflation Factors and influential observations by Cooks D and QQ-

plots were investigated to verify normality assumptions of the residuals. The statistical

procedure is further described in the ESCAPE exposure manual

(http://www.escapeproject.eu/manuals/ESCAPE Exposure-manualv9.pdf) and in (Beelen et al., 2013). Depending on the performance, the final model can then be used to predict the levels of ozone at any given location within the area for which the model was developed, by collecting and combining the geographical variables according to the model estimates. There was one major difference in our model building compared to the ESCAPE model; we chose not to have a predefined direction of effect in the model building as ozone is a secondary pollutant involved in many chemical reactions and for all variables we had no clear a priori defined direction of effect. For each area, Malmo and Umea analyses were conducted separately resulting in local specific predictor variables. The statistical analyses were performed using R (version 2.13.0).

As additional analyses we developed a separate LUR-model, excluding the predefined regional background sites, to see if the fit improved when only including measuring sites in areas where NOx emissions are higher and more local ozone reactions takes place. In the analyses we excluded 12 stations from Umea and 7 for Malmo and these models are referred to as the urban models.

3.2.5 Model validation

We validated with leave-one-out cross validation, in which a model is developed for all sites and the model structure (with the predictor variables) are applied to n-1 (where n is the number of sites).

3.3.1 Measurements for the temporal model (only Malmo)

Hourly data from 3 measurement stations that continuously measure levels of ozone were collected and aggregated into daily means. The data was collected from one urban background site and one traffic site in Malmo and one regional background site (situated 50 km north of Malmo). The measurement method was active UV-absorption. We used data for the same three weeks as in the LUR-analyses and had thus 21 daily measurements from each of the three stations. Meteorological data are collected from the Swedish Meteorological and Hydrological Institute [SMHI] site Heleneholm in Malmo.

3.3.2 Regression analyses for the temporal model (only Malmo)

To investigate the important variables for predicting ozone variations in time, meteorological variables and ozone levels were inserted into a multiple linear regression model. The model that performed best for all three stations was chosen. The meteorological variables were: wind

speed, wind direction, wind vector, global radiation, net radiation, temperature, vertical temperature difference (between 2 and 8 meters) and vertical temperature difference (between 24 and 8 meters). The latter two variables are measures of temperature inversions which can affect atmospheric convection rates.

3.3.3 Validation for the temporal model (only Malmo)

The model was validated by selecting 4 random weeks of daily measurements from the urban background station between the years of 2000 and 2012. The same model (with the same predictor variables) was applied to those measurements. The adjusted R2 for these analyses were used as a validation R .

4 Results

4.1 Measurements LUR-

Higher levels of ozone were reported from the Malmo area than the Umea area. Levels were also highest during spring than during summer and autumn, which is in accordance to cities of this magnitude. The mean levels of ozone for the entire study period in Malmo were 68 |ig/m (range 60-83 |ig/m3) and in Umea 52 |ig/m3 (range 36-63 |ig/m3).

4.2 LUR-model

For Malmo the best LUR-model included the predictor variables; low density residential areas within a 300 meter buffer, urban green areas within a 100 meter buffer, traffic load within a 1000 meter buffer and distance to sea.

All variables had a negative direction of effect which means that the variables explained the ozone decay. This model gave an adjusted R of 0.40. The correlations between variables were moderate (r<0.4). The model was validated with "leave one out validation" and had a cross-validation adjusted R of 0.17. Levels of ozone in Malmo had a smaller variation than Umea (Table 1) and this can be a reason why the LUR-model explained less of the variation.

The urban model for Malmo included the following predictor variables: low density residential areas within 100m buffers, urban green areas within 100m buffers and

heavy traffic load within 1000m buffers.

All variables had a negative direction of effect and the model adjusted R was 0.51 with cross validation adjusted R of 0.33.

For Umea, the model that performed best was a model including the following predictor variables;

traffic load within a 300 meter buffer,

high density residential areas within a 5000 meter buffer,

urban green areas within 500m buffers and

natural areas in a 5000 meter buffer.

All directions of effects were negative except urban green areas which explained some of the increase in ozone and the model performed with an adjusted R of 0.67. The correlations between variables were high, (r<0.6), and the model had a cross-validation adjusted R of 0.48.

In the urban model for Umea the following predictor variables were included; traffic load within 300m buffers, natural areas within 500m buffers, population density within 5000 and

population density within1000m buffers.

22 This model gave an adjusted R of 0.81 and cross validation R of 0.73. Population density

and natural areas had positive and traffic load negative direction of effect.

Each variables contribution to the model building is summarized in table 3 with more in-depth data and residuals in Supplemental materials.

4.3 Temporal model for Malmo

The temporal model was the model that best fitted all the sites in the Malmo region. The

predictor variables included in the temporal model for best fit were:

vertical difference in temperature between 24 meters and 8 meters,

net radiation and

global radiation.

For the urban background station the adjusted R was 0.54, for the traffic site the adjusted R2 was 0.61 and for the regional background site (situated 50 km from Malmo) adjusted R was 0.22.

The factors included in the temporal model seem to be factors that are important for local

production of ozone such as solar radiation variables (net radiation and global radiation), rather than for in transport such as wind variables. The variables also seem to predict better at the traffic site, where local ozone reaction is higher, than in the urban background site where local ozone production is non-existent due to the lack of ozone precursors, the model performs poorly. When validating the temporal model we used four randomly selected weeks between the years 2000 and 2012 and the model seems to works even in this extended time span with an adjusted validation R of 0.42. Thus the model showed that levels of ozone could be predicted by knowing meteorological factors. However the temporal model will not be applicable for the winter season since this period was not studied, yet ozone levels is not of a health concern in this region during winter due to low solar radiation.

5. Discussion

5.1 Previous studies

The LUR-model approach is a way of assessing exposure intended for use in epidemiological studies with increased spatial resolution. The intention is to have a better exposure assessment than merely attributing exposure levels from nearest measurement stations which can be several kilometres away. LUR have mostly been used for primary pollutants to explain the relationship between measured levels and different land uses (Beelen et al. 2013; Eeften et al. 2012; Noth et al. 2011; de Hoogh et al. 2013) and was originally initiated by Briggs et al. (1997). To put our results into perspective Gulliver et al. (2011) considered model R of 0.47 to be an acceptable level when they modelled PM10 using LUR. Ryan and LeMasters (2007) found R in the range of 0.54-0.81 in a review of 12 LUR models for primary pollutants. Hoek et al. (2008) found most R2 of NO2 to be around 0.6-0.7. Beelen et al. (2008) developed a background ozone model for most of Europe (not Sweden) using annual concentration and found R for the LUR models between 0.38-0.54. Also, applying meteorological variables to explain ozone variation by multiple linear regression analyses have been done previously with R2 of 0.83 (Gvozdic et al., 2011).

A previous study in Sweden found that the urban emissions of NOx and VOCs do not affect the levels of ozone outside the cities (SMHI, 2010). This study (SMHI, 2010) also showed that high levels of ozone in Sweden tend to come from sources outside of Sweden. Another study conducted at the Swedish west coast showed regional background levels of ozone to be relatively similar to levels measured in urban background, while the levels at heavy traffic sites were considerably lower (Piikki et al., 2008). Decreasing levels of ozone with increasing levels of vehicle exhausts were also seen in the same study when comparing two housing

areas situated at different distances from heavy traffic (Piikki et al., 2008). In accordance to this, our results showed that traffic lowered the levels of ozone, and that the model fit improved dramatically in the urban models. The extent of residential areas can also be a measure for urbanity and higher levels of NOx-emissions and thus went into the direction of lowering levels of ozone. All measures of residential areas had in accordance negative effects. In contrary all measures of population density went in the opposite directions with positive direction of effects. In the area population density is not closely linked to traffic density and people tend to live outside the commercial city centre where traffic is dense.

5.2 Modeling experiences

The full Malmo LUR-model performed poorer than the urban LUR-model and corresponding Umea models. The contribution from the long distance transported ozone is the most likely explanation for the relatively poor LUR-model fit in Malmo. Another possible explanation could be that it was hard to find a good LUR-model since the variation of ozone was smaller in Malmo than in Umea. The Malmo urban area model performed better than the full model.

One finding that could be investigated further was the variety of ozone levels in Natural areas (Figure 2). We found low levels of ozone in a regional background station located in a near urban forested area, which can be due to dry deposition to the trees. Another regional background site in the Malmo area was located under a wind power plant situated in an agricultural field which, especially during spring, had high levels of ozone which could not be attributed to higher general wind speed (and thus causing measurement errors by an excess of air generated by movement from the wind power plant entering the sampler) during that period.

We also found that the predictor variable Natural areas both increased and decreased the levels of ozone in Umea, depending on full or urban model. In the urban model it increased the ozone levels. In the full model it decreased while urban green areas increased levels of ozone. However, in Malmo "urban green areas" decreased the levels of ozone in the urban LUR-model. The role of BVOCs in the local formation of ozone formation in urban areas has been studied recently and it has been shown that contribution can be of great impact but that it is largely dependent on: meteorological factors, if there are high levels of ozone precursors such as NOx present, as well as type of tree species and thus which kind of BVOCs are emitted (Curci et al. 2009; Oderbolz et al., 2013). An example of the latter is that it has previously been found that green areas in Scandinavia lowered levels of ozone due to the

ozonolysis of the BVOC terpenes commonly found in Scandinavian forests (Curci et al., 2009). This is especially true in areas with a lot of Betula which have highly reactive terpenes (Oderboltz et al., 2013). In the Umea area Birch trees from the family Betulaceae is very common and the city is even known as the city of birches. A suggestion for future studies would be to carefully consider regional background sites and surrounding vegetation and to have more detailed information on the BVOC related variables. Altitude did not have any predicting value for our LUR-models even if it can be a predictor in other areas where variation in altitude is larger.

Limitation 5.3 Limitations

We validated our spatial models using leave one out cross validation. This is the most common validation for LUR-models and standardised in the ESCAPE project (Beelen et al. 2013). It should however be mentioned that this validation method has been criticised for overestimating the predictive ability of LUR-models (Wang et al., 2013). Another possible caveat is the timing of measurements. We decided to measure simultaneously in spring, summer and autumn at both study sites. It can be argued that instead of measuring simultaneously we should have measured according to when the areas are equally green due to spring coming later in the North than in the South. The two studied areas are situated geographically apart (1250 km) and seasons can vary substantially in time. However, in the studied year the spring and summer season in Umea came earlier than usual which minimized this error. As we developed separate models it should not influence results either. The benefits from using simultaneous measurements should however exceed the limitations as we are studying time periods with similar large scale weather patterns. It should also be noted that we used a standardized model (with the exception that we had no predefined direction of effect), ESCAPE, which was originally developed for primary pollutants. One caveat using this, is that we had very coarse predictor variables for natural areas (all fields and forests lumped into one variable) and thus do not capture fully the role of different BVOCs. There were also pros and cons of using a predefined direction of effect in the model building. We choose not to have a predefined direction of effect as we wanted to explore in what direction each variable went. We did however find that variables went into the direction of effect that could be expected based on previous knowledge.

Another possible limitation is that we use land use data from CORINE for 2000 in Umea and 2006 in Malmo and it can be argued that the land use might not always be accurate for 2012. By visiting the sites to place and collect the samplers we could at least to some extent make sure that no major change in land use had been done in the vicinity of the measuring sites. The models were built separately for the two areas; using different years of CORINE-data should therefore not have influenced the results.

6. Conclusion

We studied the spatial variation of ozone at two heterogeneous cities (Malmo and Umea) in Sweden using LUR-models. The LUR-model in Umea had the best fit with adjusted model R2 values of 0.67 whereas the LUR-model in Malmo had adjusted model R of 0.39. When restricting the models to only urban settings we had a better fit, here the Malmo model had adjusted R2 of 0.51 and Umea a R2 of 0.81.A temporal model was also developed for Malmo which performed best in a traffic site with adjusted R of 0.61. To summarize, this study shows that it is possible to predict the spatial variability of levels of ozone, at least for Umea. This means that we can assess exposure in Umea. In order to have exposure data for other time periods than the modelled period we can build our LUR-model using input data from stationary measurements but careful consideration has to be done in order for input data for GIS-analyses still being accurate. We could also predict the temporal ozone variation, adjusted R of 0.54-0.61 in the Malmo area. Thus, in shorter time periods of missing stationary data on ozone levels we can use meteorological data to estimate these levels. Our results should be put in the context of our cities being considerable small and levels of ozone precursors moderate.

References:

Beelen R., Hoek G., Pebesma E., Vienneau D., de Hoogh K., and D. J. Briggs D.J,. 2009. Mapping of background air pollution at a fine spatial scale across the European Union. Science of the Total Environment, 407, 1852-1867.

Beelen R., Hoek G., Vienneau D., Eeftens M., Dimakopoulou K., Pedeli X., 2013. Development of NO2 and NOx land use regression models for estimating air pollution exposure in 36 study areas in Europe the ESCAPE project. Atmospheric environment. 72, 10-23.

Briggs D.J., Collins S., Elliott P., Fischer P., Kingham S., Lebret E., Pryl K., Van Reeuwijk H., Smallbone K., Van der Veen A., 1997. Mapping urban air pollution using GIS: a regression-based approach. International Journal of Geographic Information Science. 11, 699-718.

Brunekreef B., Annesi-Maesano I., Ayres J.G., Forastiere F., Forsberg B., Kunsli N. et al. Brunekreef 2012. Ten principles for clean air. European Respiratory Journal 39, 525-528.

de Hoogh K, Wang M, Adam M, Badaloni C, Beelen R, Birk M, et al. 2013. Development of land use regression models for particle composition in twenty study areas in europe. Environ Sci Technol 47:5778-5786.

Curci, G., Beekmann, M., Vautard, R., Smiatek, G., Steinbrecher, R., Theloke, J., Friedrich, R., 2009. Modelling study of the impact of isoprene and terpene biogenic emissions on European ozone levels. Atmospheric Environment. 43, 1444-1455.

Eeftens M, Beelen R, de Hoogh K, Bellander T, Cesaroni G, Cirach M, et al. 2012. Development of land use regression models for pm(2.5), pm(2.5) absorbance, pm(10) and pm(coarse) in 20 european study areas; results of the escape project. Environ Sci Technol 46:11195-11205.

EUROAIRNET, 1999. Criteria for EUROAIRNET, the EEA Air Quality Monitoring and Information Network. Technical Report 12, European Environment Agency.

European Environment Agency [EEA], 2011. Air Quality in Europe 2011. Technical report nr 12/2011. Copenhagen Denmark.

García F., Valiente González V., and López Rodríguez F., 2010. Geostatistical Analysis and Mapping of Ground level Ozone in a Medium Sized Urban Area. International Journal of Civil and Environmental Engineering 2, 7182.

Gulliver J., de Hoogh K., Fecht D., Vienneau D., Briggs D.J., 2011. Comparative assessment of GIS-based methods and metrics for estimating long-term exposures to air pollution Atmospheric Environment 45 70727080.

Gvozdic V., Kovac-Andric E., Brana J., 2011. Influence of meteorological factors NO2, SO2, CO and PM10 on the concentration of O3 in the urban atmosphere of Eastern Croatia. Environmental Modeling & Assessment 16, 491-501.

Hagenbjork-Gustafsson A. 2012. Jämförande mätning av ozon utomhus med Ogawa diffusionsprovtagare och referensmetoden UV-fotometri. Report 2012:2 from Occupational and Environmental Medicine. Umeá University. Yrkes- och miljömedicin i Umeá, Umeá.

Hoek G, Beelen R, de Hoogh K et al., 2008. A review of land use regression models to assess spatial variation of outdoor air pollution, Atmospheric Environment, 42, 7561-7578.

IPCC [International Panel of Climate Change], Pachauri RK, Reisinger A. (Eds.) 2007. Contribution of Working Groups I, II and III to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change Climate Change. Synthesis Report, IPCC, Geneva, Switzerland.

Jenkin M. E., 2008. Trends in ozone concentration distribution in the UK since 1990: Local regional and global influences. Atmospheric Environment 72, 5435-5445.

Johansson C., Forsberg B., 2005. Kvävedioxid och ozon i tätortsluften. Halternas samspel samt konsekvenser för hälsan [Nitrogen dioxide and ozone in urban air. Correlation between levels and consequences for health]. In: Naturvárdsverket [Swedish EPA] Report 5519.

Koike E., Kobayashi T., 2004. Ozone exposure enhances antigen-presenting activity of interstitial lung cells in rats. Toxicology 196, 217-227.

Mudway I.S., and Kelly F.J., 2004. An Investigation of Inhaled Ozone Dose and the Magnitude of Airway Inflammation in Healthy Adults. American Journal of Respiratory and Critical Care Medicine 169, 1089-1095.

Noth EM, Hammond SK, Biging GS, Tager IB. 2011. A spatial-temporal regression model to predict daily outdoor residential pah concentrations in an epidemiologic study in fresno, ca. Atmospheric Environment 45:2394-2403.

Oderbolz D.C., Aksoyoglu S., Keller J., Barmpadimos I., Steinbrecher R., Skj0th C.A, et al. 2013. A comprehensive emission inventory of biogenic volatile organic compounds in Europe: improved seasonality and land-cover Atmospheric Chemistry and Physics 13, 1689-1712.

1 OECD, 2008. OECD Environmental outlook to 2030. Organization for Economic Co-Operation and

2 Development www.oecd.org/dataoecd/29/33/40200582.pdf (accessed October 2011).

4 Paul S., Monks P.S., 2000. A review of the observations and origins of the spring ozone maximum. Atmospheric

5 Environment 34, 3545-3561.

6 Piikki K., Karlsson P., Klingberg J., et al. 2008. Mätningar av marknära ozon och meteorologi vid kustnära och

7 urbana miljöer i Halland, Skâne och Västra Götalands län. Report to Länsstyrelsen Västra Götalands län [County

8 Administrative Board of Västra Götaland].

9 Raes F., Hjorth J., 2006. Answers to the Urbino questions. ACCENT Secretariat, Universita di Urbino, Italy,

10 ISBN 92-79-02413-2.

11 Ryan P.H., LeMasters G.K., 2007. A review of land-use regression models for characterizing intraurban air

12 pollution exposure. Inhalation Toxicology, 12, 127-133.

13 Sanderson, M.G., Jones, C.D., Collins, W.J., Johnson, C.E., Derwent, R.G., 2003. Effect of climate change on

14 isoprene emissions and surface ozone levels. Geophysical Research Letters 30, 1936.

15 Sjöberg, K., Karlsson, P.E., Haeger-Eugensson, M., 2010. Förslag till nationellt övervakningsprogram för

16 marknära ozon. [Proposal for a national surveillance programme for ground level ozone] SMHI Report U2325

17 (in Swedish).

18 SMHI [Swedish Meteorological and Hydrological Institute] Engardt M., Andersson C., Bergström R., 2010.

19 Modellering av Marknära Ozon [Modeling of ground level ozone], METEOROLOGI Nr 145.

20 Sousa S.I., Alvim-Ferraz M.C., Martins F.G., 2013. Health effects of ozone focusing on childhood asthma: what

21 is now known--a review from an epidemiological point of view. Chemosphere. 90, 2051-2058.

22 Stedman J.R., Kent A.J., 2008. An analysis of the spatial patterns of health related surface ozone metrics across

23 the UK in 1995, 2003 and 2005. Atmospheric Environment 42, 102-1716.

24 Wang M., Beelen R., Basagana X., Becker T., Cesaroni G., De Hoogh K., et al. 2013. Evaluation of land use

25 regression models for NO2 and particulate matter in 20 European study areas: The ESCAPE project.

26 Environmental Science and Technology. 47, 4357-4364.

28 WHO, 2006. Air Quality Guidelines - Global Update 2005. World Health Organization, Regional Office for

29 Europe, Copenhagen, Denmark (2006).

31 WHO [World Health Organisation]. 2006. Air Quality Guidelines, Global Update 2005. World Health

32 Organisation Regional Office for Europe, Copenhagen, Denmark.

33 WHO [World Health Organisation]. Derwent D., Forsberg B., Amann M., Hänninen O., Hurley F.,

34 Krzyzanowski M. (Eds), 2008: Health Risks of Ozone from Long-range Transboundary Air Pollution. WHO

35 Europe, Copenhagen Denmark Print ISBN: 9289042893.

36 WHO, 2008. Health Risks of Ozone from Long-Range Transboundary Air Pollution World Health Organization,

37 Regional Office for Europe, Copenhagen, Denmark (2008).

39 WHO [World Health Organisation] Europe, 2013. Review of evidence on health aspects of air pollution -

40 REVIHAAP Project: Final technical report, WHO Europe, Copenhagen, Denmark.

Figure 2, Study areas and mean ozone concentration at the measuring sites.

Mean ozone (|jg/m3)

O <65 O 65-<70 O 70-<75 O >75

Land use

Agriculture Artificial surface Forest Water Road

Vannas

Mean ozone (|jg/m3)

O <50 O 50-<55 , O 55-<60

- Municipality border

Land use

Artificial Agriculture Forest Wetland Water Road

Kristoffer Mattisson, OEI

1 Table 1, Ozone concentrations (^g/m3) and number of stations by station type.

City Site type N Mean SD Min Max

Malmö Regional background 7 69.1 5.9 60.4 79.0

Urban background 15 66.7 3.2 59.9 73.0

Traffic 18 68.4 8.0 60.0 83.1

Umeâ Regional background 5 54.9 18.3 27.5 93.3

Urban background 23 52.5 14.2 29.4 71.8

Traffic 12 49.3 13.5 26.5 69.7

3 Table 2, Measurements of ozone (pg/m3) by active versus passive sampling in Malmo.

Measurement site by type and period Active sampling (pg/m3) 28th of May-4th of June Passive sampling (pg/m3) 28th of May-4th of June Active sampling (pg/m3) 20th -27th of Aug Passive sampling (pg/m3) 20th -27th of Aug

Urban Background 73.3 71.5 64.6 68.6

Traffic 65.6 67.6 57.2 59.5

1 Table 3, Predictor variables contribution to model.

Umea Umea urban Malmo Malmo urban

Predictor variables (buffer) Adj R2 Predictor variables (buffer) Adj R2 Predictor variables (buffer) Adj R2 Predictor variables (buffer) Adj R2

Traffic load (300) 0.44 Traffic load (300) 0.67 Low density residential areas (300) 0.19 Low density residential areas (100) 0.13

High density Residential areas (5000) 0.52 Natural areas (500) 0.75 Distance to sea 0.20 Urban green areas (100) 0.24

Natural areas (5000) 0.61 Population density (5000-1000) 0.79 Urban green areas (100) 0.26 Heavy traffic load (1000) 0.51

Urban green areas (500) 0.63 Population density (1000) 0.81 Traffic load (1000) 0.40

Highlights

• We measured levels of ozone at 80 sites during three weeks.

• We modeled spatial variation in ozone in two areas of Sweden.

• We modeled the temporal variation in ozone levels in South of Sweden.

• We can, for Umea and urban areas, predict the spatial variability of ozone.

• We can predict the temporal variability of ozone in Malmo.

Table with more in depth values of the models (should this table be included as a supplemental material)

Model Predictor variables (buffers) beta se standardized coefficients p-value vif Adjusted R2

Malmö

(Intercept) 77.4 2.2 77.4 <0.001 1

Low density residential areas (300) 0.00000 2.75e-5 7.91e-6 -3.2 0.001 1.4 0.19

Distance to sea -5.43e-4 2.56e-4 -1.9 0.04 1.3 0.20

Urban Green (100) -2.34e-4 1.16e-4 -1.7 0.05 1.1 0,26

Traffic load (1000) -1.36e-06 4.57e-7 -2.7 0.005 1.3 0.40

Malmö Urban

(Intercept) 77.4 1.9 77.4 <0.001 1

Low density residential areas (100) -2.62e-4 6.11e-5 -3.28 0.0002 1.4 0.13

Urban Green (100) -3.0e-4 1.01e-4 -2.26 0.006 1.4 0.24

Heavy traffic load (1000) -2.30e-7 5.63e-8 -2.67 0.0003 1.0 0.51

(Intercept) 61.3 2.0 61.3 <0.001 1

Traffic load (300) -5.47e-9 8.79e-10 -2.8 <0.001 1.1 0.44

High density residential areas (5000) -8.86e-6 1.95e-06 -2.8 <0.001 2.1 0.52

Natural areas (5000) -1.28e-7 3.60e-08 -2.1 0.001 1.8 0.61

Urban Green (500) 6.53e-6 3.73e-06 0.9 0.09 1.3 0.63

Umea Urban

(Intercept) 48.1 1.2 48.1 <0.001 1

Traffic load (300) -5.76e-9 6.26e-10 -3.3 <0.001 1.4 0.67

Natural areas (500) 9.89e-6 2.20e-6 1.6 0.0002 1.4 0.75

Population density (5000 - 1000) 4.08e-5 2.03e-5 0.7 0.06 1.4 0.79

Population density (1000) 3.1e-4 1.4e-4 0.8 0.04 1.6 0.81

\\ '/ V/

Vannas

Residuals Umea (|jg/m3)

O <-4 to -5,4 O -4-<0 O 0-<4 O >4 to 8,9

- Municipality border

- Road

Kristoffer Mattisson, OEI^f Lund University

Vannas

Residuals Umea Urban ([jg/m3)

O <-4 to -5,4

O -4-<0 N

O 0-<4 w

O >4 to -8,9

-Municipality border

- Road

Kristoffer Mattisson, OEI^f Lund University

Figure Maps with residuals on measured and modelled levels of ozone in the different areas and models.