Estimating and Mapping Five Forest Attributes with Satellite Ancillary Data

Abstract

We investigate ordinary kriging and three cokriging procedures for making continuous maps of five forest attributes. Both ordinary kriging and cokriging use a primary variable, but cokriging, like multivariate statistics, includes secondary variables. The secondary or ancillary variables are reflectance values and calculated vegetation indices from an August 1996 Landsat Thematic Mapper satellite image. Two methods for comparing the results include examining the residuals and breaking both the estimated and sampled data into classes and then examining the resulting confusion matrix. The comparison statistics are root mean square error, overall accuracy, and kappa statistic. For the cross-validation, an additional statistic was median absolute error. A cross-validation indicated that cokriging had a higher overall accuracy and kappa statistic and a lower median absolute error, while kriging yielded a slightly lower root mean square error. Both procedures captured the same trends in Connecticut. The developed areas around New York City and the I-91 corridor running from New Haven through Hartford into Massachusetts are less forested than the less developed and higher elevation areas in the northwestern portion of the state. Kriging smoothes the maps, missing the fine-scale heterogeneity of the landscape that cokriging detects.