Comparison of ordinary and lognormal kriging on skewed data of total cadmium in forest soils of Sweden

Spatial statistical analysis of georeferenced data of total cadmium (TCd) in forest soils of Sweden was assumed to provide more advantageous maps than traditional interpolated maps. However, 264 measurements of TCd in O-horizon of forest soils displayed skewed frequency distribution. Since atypical observations affect badly the variogram, outliers were identified, different data transformations were tested and ordinary (OK) and lognormal kriging (LK) scenarios were compared based on cross-validation. Results were compared using overall measures of predictors, e.g. traditional mean squared prediction error (MSPE), mean of kriging variances, variance ratio, median of internally standardised residuals, and assessments of classification accuracy, such as percentage of correctly predicted samples and within-class MSPE. One outlier was identified based on the absolute value of skewness of value differences less or equal to one in data pairs separated at certain lag classes. Mapping categories characterised by percentage of correct classification and within-class MSPE were found to be essential in comparison of kriging results additionally to the overall measures. In comparison of kriging methods, OK predicted high values more accurately and LK was more effective to predict low and medium values. Thus, OK was suggested for mapping high concentration of TCd and other pollutants. Percentage of correctly predicted samples and within-class MSPE were found to be dependent on kriging method, as well as on the number and limits of categories.