Mapping of background air pollution at a fine spatial scale across the European Union

https://doi.org/10.1016/j.scitotenv.2008.11.048Get rights and content

Abstract

Background

There is a need to understand much more about the geographic variation of air pollutants. This requires the ability to extrapolate from monitoring stations to unsampled locations. The aim was to assess methods to develop accurate and high resolution maps of background air pollution across the EU.

Methods

We compared the validity of ordinary kriging, universal kriging and regression mapping in developing EU-wide maps of air pollution on a 1 × 1 km resolution. Predictions were made for the year 2001 for nitrogen dioxide (NO2), fine particles < 10 µm (PM10), ozone (O3), sulphur dioxide (SO2) and carbon monoxide (CO) using routine monitoring data in Airbase. Predictor variables from EU-wide databases were land use, road traffic, population density, meteorology, altitude, topography and distance to sea. Models were developed for the global, rural and urban scale separately. The best method to model concentrations was selected on the basis of predefined performance measures (R2, Root Mean Square Error (RMSE)).

Results

For NO2, PM10 and O3 universal kriging performed better than regression mapping and ordinary kriging. Validation of the final universal kriging estimates with results from all validation sites gave R2-values and RMSE-values of 0.61 and 6.73 µg/m3 for NO2; 0.45 and 5.19 µg/m3 for PM10; and 0.70 and 7.69 µg/m3 for O3. For SO2 and CO none of the three methods was able to provide a satisfactory prediction.

Conclusion

Reasonable prediction models were developed for NO2, PM10 and O3 on an EU-wide scale. Our study illustrates that it is possible to develop detailed maps of background air pollution using EU-wide databases.

Introduction

Accurate, high resolution and updateable maps of air pollution are an important information need. They are required not only to provide information on air pollution for environmental and health policy, but also to act as a basis for designing and stratifying future monitoring networks. In addition, they are needed to support the health and environmental science central to help guide and evaluate policy.

Epidemiological studies conducted for this purpose typically imply the need for large population studies, carried out at individual level, to detect the small relative risks associated with many air pollutants. In order to provide adequate exposure contrasts and to represent risks across the entire population, there is also an increasing interest in conducting epidemiological studies simultaneously in multiple centres (or across different countries). To enable this, consistent methods of exposure assessment are essential. In the absence of direct measurements of exposure, many previous studies have assigned data from the nearest air pollution monitoring site(s) to subjects (Dockery et al., 1993, Pope, et al., 2002). This may result in severe misclassification, which may dilute the strength of observed associations between exposure and health. It is now widely recognized that monitored data for point locations (i.e. monitoring sites) provide only a partial picture of the air pollution situation in any area because of the limited spatial representativeness of these sites. Monitoring is also extremely costly, so the density of these networks is limited. For many uses, the need is therefore to understand much more about the geographic variation of pollutants and to have information about pollution levels at unmonitored sites. This requires the ability to extrapolate from these monitoring stations to other, unsampled locations — and thus to model and map spatial distributions of air pollution.

In recent years, dispersion models have been used to model these spatial variations in air pollution (Nafstad et al., 2004, Penard-Morand et al., 2006). Dispersion models, however, have important limitations. One of these is the need for powerful computing facilities, especially when the models need to be applied to large areas and at high spatial (or temporal) resolution. Another is the demand for detailed input data. More sophisticated models, especially, rely on a wealth of input data, including information on source distribution, activity and emission rates, meteorological conditions and surface terrain. Suitably detailed information is often available at a local level, while coarser information can be obtained when broad, regional scale modelling is required. Problems nevertheless exist in trying to obtain these data in order to run dispersion modelling across large areas, such as the whole European Union (EU), at high spatial resolution.

As an alternative, both kriging (Cressie, 2000, Jerrett et al., 2005b) and regression methods (Briggs et al., 2000, Brauer et al., 2003, Beelen et al., 2007) have attracted considerable attention as a basis for pollution mapping in recent years.

Kriging represents a suite of techniques, all based upon the principle of regionalised variables (Journel and Huijbregts, 1978): namely, that variation in the phenomenon of interest comprises three main components — broad scale trend (or drift), local spatially-structured variation, and non-spatial random variation. Kriging thus operates by estimating these different components of variation, and using the resulting models to estimate conditions at unsampled locations. Since the methods were first developed by Krige for use in the mining industry and formalized by Matheron (1971), various forms of kriging have been devised. Following Burrough (1991), we distinguish here between two types of kriging: ordinary kriging (which takes account only of local-scale variation in the variable of interest) and universal kriging (which takes account also of long-range variation and regression relationships with external predictors). Ordinary and universal kriging have previously been used with success to model both ozone (Liu and Rossini, 1996) and particles (Cressie, 2000) at the local scale, and to model broad scale variations in background air pollution (Lefohn et al., 1988). Reasonable exposure estimates were also made in the U.S.-wide Harvard Six Cities Study and the ACS Study, which both explicitly target communities with exposure gradients (Jerrett and Finkelstein, 2005).

Stochastic modelling techniques involve developing statistical associations between potential ‘predictor variables’ (e.g. data on emission sources, topography, land cover) and measured pollutant concentrations as a basis for predicting concentrations at unsampled sites. Regression techniques are often used for this purpose and most researchers currently refer to these methods as land use regression techniques. Once developed on a training dataset (i.e. a subset of the monitored data designated for model building), the regression equations are used to predict concentrations at unsampled sites. This technique was successfully developed in a number of European cities for nitrogen dioxide (Briggs et al., 1997, Briggs et al., 2000) and has more recently been applied to other pollutants as part of EU studies (Carr et al., 2002, Brauer et al., 2003). Similar approaches have been used in the UK to develop a 1 km air quality map of the whole country (Stedman, 1998). These methods have also been used in North America. Ross et al. (2007), for example, developed a land use regression model and compared it to kriging to predict fine particulate matter concentrations in the New York City region.

Kriging and regression methods are developed and explored here by extending them to a wider range of pollutants and to the EU scale. The research was undertaken as part of the EU-funded APMoSPHERE (Air Pollution Modelling for Support to Policy on Health and Environmental Risks in Europe) project. The overall aim of this project was to assess methods to develop accurate, high resolution and updatable maps of emissions and air pollution across the EU, as a basis for environmental and health policy and environmental and epidemiological research. In this paper, we assess the feasibility of developing detailed background air pollution maps of regulated air pollutants based on consistent EU-wide databases. The main goals of the paper are to compare the validity of different techniques (regression, ordinary and universal kriging) and to illustrate the problems encountered.

Section snippets

Study design

We compared the performance and validity of three modelling methods – ordinary kriging, universal kriging and regression mapping – in developing EU-wide maps of air pollution on a 1 × 1 km scale (in total 2,854,116 grid cells). Air pollution data were obtained from European databases, described below. Rural and urban background monitoring sites were used to provide the annual average concentration (µg/m3) for the year 2001 of the major pollutants nitrogen dioxide (NO2), fine particles < 10 µm (PM10

Results

A comparison of the two kriging methods and regression method to predict concentrations at the relevant validation sites is presented in Table 3. Regression models for NO2, PM10 and O3 used to create global, urban and rural scale maps are presented in Table 4, Table 5, Table 6 respectively. The final validation results using the full 25% validation set for the composite maps are shown in Table 7. For SO2 and CO none of the three methods was able to provide a satisfactory prediction.

Discussion

Reasonable prediction models were developed for NO2, PM10 and O3 at the 1 × 1 km resolution at the EU scale. PM10 and O3 are regional pollutants and consistently the prediction of the global background contributed significantly to the satisfactory prediction of the concentrations of these pollutants in both rural and urban areas. Although NO2 is much more influenced by local sources, the predictor variables we had available gave a good representation of major sources (e.g. traffic, home heating).

Acknowledgments

APMoSPHERE was funded by the European Union Fifth Framework programme between December 1, 2002 and September 1, 2005, contract EVK2-2002-00176. The project was coordinated by Prof Briggs (Imperial College London, United Kingdom).

References (45)

  • KinghamS. et al.

    Spatial variations in the concentrations of traffic-related pollutants in indoor and outdoor air in Huddersfield, England

    Atmos Environ

    (2000)
  • KubilayN. et al.

    An illustration of the transport and deposition of mineral dust onto the eastern Mediterranean

    Atmos Environ

    (2000)
  • LefohnA. et al.

    The use of kriging to estimate monthly ozone exposure parameters for the southeastern United States

    Environ Pollut

    (1988)
  • LiuL.J.S. et al.

    Use of kriging models to predict 12-hour mean ozone concentrations in metropolitan Toronto — a pilot study

    Environ Int

    (1996)
  • PebesmaE.

    Multivariable geostatistics in S: the gstat package

    Comput Geosci

    (2004)
  • Penard-MorandC. et al.

    Assessment of schoolchildren's exposure to traffic-related air pollution in the French Six Cities Study using a dispersion model

    Atmos Environ

    (2006)
  • PutaudJ.P. et al.

    A European aerosol phenomenology—2: chemical characteristics of particulate matter at kerbside, urban, rural and background sites in Europe

    Atmos Environ

    (2004)
  • Roorda-KnapeM. et al.

    Air pollution from traffic in city districts near major motorways

    Atmos Environ

    (1998)
  • RossZ. et al.

    A land use regression for predicting fine particulate matter concentrations in the New York City region

    Atmos Environ

    (2007)
  • StedmanJ.R. et al.

    An analysis of the spatial patterns of human health related surface ozone metrics across the UK in 1995, 2003 and 2005

    Atmos Environ

    (2008)
  • BeelenR. et al.

    Long-term effects of traffic-related air pollution on mortality in a Dutch cohort

    Environ Health Perspect

    (2008)
  • BrauerM. et al.

    Estimating long-term average particulate air pollution concentrations: application of traffic indicators and geographic information systems

    Epidemiology

    (2003)
  • Cited by (0)

    View full text