Analyst variation associated with land cover image classification of Landsat ETM plus data for the assessment of coarse spatial resolution regional/global land cover products

This study examined analyst variation associated with land cover (LC) image classification using 30x30m Landsat ETM+ data for the assessment of coarse spatial resolution regional/global LC products. The study was designed to test the effect of varying training site selections (location and number) among six analysts performing a supervised classification on a Landsat ETM + image. Design constraints maintained other aspects of the classification process constant (i.e., type of classifier, choice of band combinations, etc.). Results indicated that training site selection alone did not provide a predictive measure of classification accuracy. Only when training data selection was combined with variations in spatial resolution did significant differences occur. Differences in classification accuracies between analysts increased threefold in the aggregation process from 90x90m to 1200x1200m. Error sources (i.e. analyst differences) and the dynamics of the spatial aggregation process can potentially account for differences in environmental modeling outcomes.