Abstract
This study compares the performance of several statistical methods (multiple linear regression, analysis of covariance, geographically weighted regression, regression kriging, and ordinary kriging) for deriving spatial models of soil parameters. The applications were carried out within a 186-km2 hydrographic basin situated in eastern Romania. Statistical models were computed from a sample of approximately 180 soil profiles, scattered in the eastern half of the basin. Two independent samples, each of 50 soil profiles, were used for validation inside (interpolation) and outside (extrapolation) the main sampling area. The predictors included X and Y coordinates of soil profiles, geomorphometrical parameters (altitude, slope, aspect, wetness index, terrain curvature), climate parameters (mean annual temperatures, precipitation, global radiation), the normalized difference vegetation index, the main soil types, land use, and surface lithology. For only three soil variables the geostatistical approach proved to be useful: occurrence depth of calcium carbonates, pH, and base saturation. The best spatial models were achieved using analysis of covariance, geographically weighted regression, and ordinary kriging. The most relevant continuous predictor is the mean annual precipitation, whereas the most relevant qualitative factor is the soil type.
Acknowledgments
This study was carried out with support from project POSDRU/89/1.5/S/49944, coordinated by “Alexandru Ioan Cuza” University of Iaşi and from project TE 172 / 2010, financed by the CNCSIS.