A Comparison of Kriging and Cokriging for Mapping Forest Volume in Connecticut

Ordinary kriging and cokriging are two geostatistical techniques used to create continuous maps of spatially autocorrelated attributes. Both procedures use the same primary variable, but cokriging, like multivariate statistics, incorporates additional independent variables. In this study, the primary variable is cubic foot volume and the secondary variable, mean normalized difference vegetation index, is calculated from an August 1996 Landsat Thematic Mapper satellite image. Two methods of evaluating the results include comparing the residuals and breaking both the estimated and sampled data into five volume classes and then examining the resulting confusion matrix. The comparison statistics: root mean square error, overall accuracy, and kappa statistic indicate that cokriging is superior when all of the sample data points are included. The cross-validation results show that cokriging is superior for correctly classifying the observed data by volume classes as indicated by the overall accuracy and kappa statistic. However, the cross-validation root mean square error is larger for cokriging indicating the superiority of kriging. Both procedures capture the same trends in Connecticut. Low volume areas are located around New York City and along the I-91 corridor and the high volume areas are in the NW part of the state. Kriging smoothes the map missing the fine scale heterogeneity of the landscape that cokriging detects.