SPATIAL TEMPORAL MODELLING OF PARTICULATE MATTER FOR HEALTH EFFECTS STUDIES

Abstract

Epidemiological studies of the health effects of air pollution require estimation of individual exposure. It is not possible to obtain measurements at all relevant locations so it is necessary to predict at these space-time locations, either on the basis of dispersion from emission sources or by interpolating observations. This study used data obtained from a low-cost sensor network of 32 air quality monitoring stations in the Dutch city of Eindhoven, which make up the ILM (innovative air (quality) measurement system). These stations currently provide PM10 and PM2.5 (particulate matter less than 10 and 2.5 m in diameter), aggregated to hourly means. The data provide an unprecedented level of spatial and temporal detail for a city of this size. Despite these benefits the time series of measurements is characterized by missing values and noisy values. In this paper a space-time analysis is presented that is based on a dynamic model for the temporal component and a Gaussian process geostatistical for the spatial component. Spatial-temporal variability was dominated by the temporal component, although the spatial variability was also substantial. The model delivered accurate predictions for both isolated missing values and 24-hour periods of missing values (RMSE = 1.4 μg m−3 and 1.8 μg m−3 respectively). Outliers could be detected by comparison to the 95% prediction interval. The model shows promise for predicting missing values, outlier detection and for mapping to support health impact studies.