Spatial prediction of vegetation quantities using ground and image data

Abstract

A major challenge in Earth system studies is mapping vegetation quantities over large regions. Aspatial regression is typically the empirical method applied to remotely sensed and ground data for the spatial prediction of vegetation variables. Geostatistical methods, such as cokriging and stochastic simulation, have rarely been used for this purpose. A synthetic example was constructed from imaging spectrometer data to allow an objective comparison between regression, cokriging and a new stochastic simulation method. A range of linear relations between ground sample data and image data was represented in the example. The lowest root-mean-square-error was achieved with cokriging until the correlation coefficient (r) between direct and ancillary data exceeded 0.89, at which point regression was the more accurate predictor. Probability-field simulation gave a range of possible realizations, overall less accurate than those from regression but more faithful to the histogram and spatial pattern of the variable to be predicted. The strength of the relation between ground measurements and image data was shown to be a critical factor in choosing a spatial prediction method.