PM10 temporal behavior and predictions through a geostatistical model

Box-Jenkins methodology (1976) is commonly applied for time series analysis. Using this approach, sample autocorrelation and partial functions (ACF and PACF, respectively) are conventionally inspected in order to identify the most appropriate model which describes the temporal evolution of the process under study. The fitted model is subsequently used for prediction purposes. Opposite to the above ACF and PACF based-method, the variogram represents the basic tool in Linear Geostatistics to face a variety of inferential problems (Chilés and Delfiner, 1999; Journel and Huijbregts, 1981; Matheron, 1963). In this context, detection of a parametric model for the process under study gives way to the estimation and modeling of the variogram in order to perform predictions of the analyzed variable at unsampled points. This paper aims to illustrate the importance and convenience of variogram-based exploratory and prediction techniques to perform a complete analysis of a time series, even in presence of a periodic behaviour. In particular an extensive case study regarding the time series of PM10 daily concentrations registered at a monitoring station located in an area with high risk of particle pollution, is faced through the following steps: a) identification of trends and periodicity exhibited by data, b) estimation of missing values, c) predictions of the PM10 concentrations at time points following the last available observation, d) estimation of the distribution function. Regarding the computational aspects, a modified version of the GSLib kriging routine (Deutsch and Journel, 1998) has been used to define appropriate temporal search neighborhoods for interpolation and prediction purposes.