Accounting for source location and transport direction into geostatistical prediction of contaminants.

This paper presents a variant of the well-known kriging with a trend that allows one to account for the pollutant source coordinates and information about transport process into the spatial prediction of pollutant concentration. The new technique is illustrated using lead data from a Dallas metropolitan area and cadmium data from Palmerton (PA) NPL Superfund site. Instead of modeling the local spatial trend as low-order polynomials of coordinates, it is here expressed as a function of two factors that likely control the pollution spread: the distance to the smelter and the deviation from the major wind direction. Four different combinations of these two factors are developed, and their prediction performances are evaluated for a range of wind directions using cross-validation. Comparison with two traditional algorithms (OK and KT) shows that the proposed approach leads to smaller mean square errors of prediction when the correct wind direction is determined. The best combination of trend model and wind direction is site-dependent and derived using cross-validation. Kriging of residuals from a global trend model is an alternative to the use of local trends, and both techniques are shown to outperform ordinary kriging.