Statistical spatio-temporal modelling of PM10 in the taranto urban area from a composite monitoring network

the city of Taranto. In the present work the attention is focused on PM10 concentrations monitored by 13 stations belonging to two different networks pertaining to the regional (ARPA) and municipal (GECOM) government, with hourly, two-hourly or daily observation frequencies, for the operating period corresponding to year 2005. Preliminary analyses highlighted the necessity to tackle some problems related to the definition of the data-base: (1) a large number of missing data, due to both different working periods for groups of monitoring stations and occasional malfunction of PM10 sensors; (2) scarce comparability of the data implied by the use of two different validation techniques for the ARPA and GECOM networks. Missing data imputation and calibration were addressed by a Bayesian kriging-based technique relying on the idea of preserving and exploiting the spatial correlation of the observed PM10 concentrations, recursively estimating daily spatial interpolation models to predict missing and overestimated data (Pollice and Jona Lasinio, 2008). Meteorological data were provided by 6 monitoring stations reporting hourly measurements and they were variously combined in order to obtain a unique daily database of weather covariates at the city level. Spatio-temporal modelling of PM10 log-concentrations was performed using an approach proposed by Le and Zidek (2006): after checking for space-time separability, a preliminary analysis of the temporal structure is carried out and an AR(3) trend is removed in order to obtain time de-trended residuals. Subsequently spatial interpolation is obtained by a hierarchically specified model based on Bayesian kriging and characterised by the use of time varying covariates and a semi-parametric nonstationary covariance structure. The spatial covariance matrix among the monitoring stations is first estimated by the EM marginal likelihood maximization and then extended to new locations (interpolation grid) using a method due to Sampson-Guttorp (1992) introducing nonstationarity and anisotropy of the spatial fields. Daily estimates of the PM10 concentration surfaces, obtained after re-introducing the formerly estimated AR(3) structure, are used for the interpretation of temporal trends and for the identification of concentration peaks within the urban area.