Spatio-temporal Analysis and Interpolation of PM10 Measurements in Europe

ETC/ACM Technical Paper 2011/10, Released: 2012/01/30

Benedikt Gräler, Lydia Gerharz, and Edzer Pebesma

“This study investigates the potential of spatio-temporal kriging approaches for daily mean PM10 concentrations. The methods used include separate daily variogram estimates, temporally evolving variograms, the metric model, the separable covariance model and the product-sum model, and are combined with multiple linear regression. These methods are applied to daily mean rural background PM10 concentrations across Europe for the year 2005, and incorporate daily EMEP model data and elevation data as predictors.

 Interpolated maps for daily PM10  concentration from May 1 to 9, 2005.

Interpolated maps for daily PM10 concentration from May 1 to 9, 2005.

“The air quality indicators used in this study are the daily and yearly mean PM10 concentrations and the number of days exceeding the limit value 50 µg/m³ (NOE). The quality assessment of the different techniques relies on a cross-validation. Statistical measures are used to quantify the improvement for different indicators.

“It is shown that daily interpolations can improve the statistical performance of the interpolation of annual mean PM10 concentrations. Furthermore, some advantages of daily estimates are described. Besides the improvement in annual mean PM10 concentration maps, studying the phenomenon in a wider spatio-temporal context becomes possible with daily estimates. Especially the estimation the number of days PM10 concentrations exceed certain limits can be done in a more natural way. Likewise, the detection of outliers and data inhomogeneity benefits from a daily spatio-temporal model.

“Interpolation with the simple spatio-temporal variogram models used here exploits the temporal correlations present and out performs the purely spatial interpolation methods. Based on temporal variability of the spatial short-distance variation component, a discussion is given on the suitability of this statistic to infer measurement errors, and alternative approaches are proposed.”