We present results from analyses conducted to evaluate the performance of advanced supervised classification algorithms (decision trees and neural nets) applied to AVHRR data to map regional land cover in Central America. Our results indicate that the sampling procedure used to stratify ground data into train and test sub-populations can substantially bias accuracy assessment results. In particular, we found spatial autocorrelation in test data to inflate estimates of classification accuracy by up to 50 points. Results from evaluations performed using independent train and test data suggest that the feature space provided by AVHRR NDVI data is poorly suited for most land cover mapping problems, with the exception of those involving highly generalized classes.