Statistical air quality mapping

This thesis handles statistical mapping of air quality data. Policy makers require more and more detailed air quality information to take measures to improve air quality. Besides, researchers need detailed air quality information to assess health effects. Accurate and spatially highly resolved maps of air pollution levels form a basis. Since policy makers and researchers tend to focus more and more on uncertainties as well, the question is how precise these concentration maps are.To base concentration maps on measurements of air quality only, every km2 should be monitored. Measurements, however, are only taken at a limited number of locations, so between the monitoring locations relevant information will be missing or can only be predicted, i.e. interpolated, leading to uncertainty in the map. Furthermore, no information about the physical and chemical processes about the concerned component is taken into account. On the other hand, concentration maps can also be based on physical and chemical processes modeling of components only. This model output covers the full domain on a fine-mazed grid. All dispersion models are imperfect however, which may lead to biased output and uncertainties.A combination of the two approaches always results into more detailed and more accurate maps. In this thesis this is done by means of a geostatistical approach: kriging with external drift (KED). KED allows mapping of a primary variable that is accurate and precise but only available at a limited number of locations, and a secondary variable that covers the full domain on a fine-mazed grid but is less accurate.First, we focus on the use of atmospheric dispersion model output as secondary information source to compensate for the loss of spatial precision caused by a reduction in the Dutch air quality monitoring network in the mid-nineteen eighties. We compare KED with universal kriging. The impact of several parameter estimation and spatial interpolation methods, the number of observations and configuration of the network on uncertainty are quantified by cross-validation. With KED, more accurate and precise predictions are obtained where observations were sparse. However, the dispersion model output in this context was considered to be deterministic, i.e. without uncertainties, so the geostatistical model must be extended.We present a method, error-in-variable KED, which combines uncertain air quality measurements with uncertain secondary information from the atmospheric dispersion model. The new method combines KED and a measurement error model, and uses Bayesian techniques for inference. The method is flexible for assigning different error variances to both the primary information and secondary information at each location. We address actual NO2 data collected at an urban and a rural site in the Netherlands. Uncertainty assessments in terms of exceeding air quality standards are given.The error-in-variable KED procedure is further extended with a time component to assess future local NO2 concentrations near Rotterdam for the year 2010, focusing on uncertainties and exceedances of European air quality standards. The background concentration is determined by the extended error-in-variable KED. A local traffic contribution is added based on a local generic dispersion model with use of an emission scenario for 2010. This results in maps showing local NO2 concentrations, upper and lower limits, and probabilities of exceeding the air quality standard. The probabilistic measures are calculated in numbers and translated into words for easier communication to policy makers.Finally, the use of two secondary information sources is explored to map particulate matter (PM10) over Western Europe. It is almost impossible to get a consistent overview of PM10 concentrations based solely on ground based measurements because of differences between countries regarding monitoring methods used and monitoring station surroundings. We illustrate the use of statistical techniques to standardize the ground based measurements of PM10 and interpolate these standardized concentrations by combining them uncertain secondary information from a chemical transport model and from MODIS satellite observations of aerosol optical thickness. The secondary variables contain different information and a combination of both gives the most accurate and precise predictions and should therefore be preferred.