Classification of remotely sensed and ground-based data is at the heart of landscape monitoring, calculating ecological risk, and decision making. The simplest classification may be a binary decision: the land is either degraded or it is not. This research investigates several problems that arise because of the subjective nature of that final decision and suggests a process for minimizing the least-desirable errors. We use a combination of field vegetation surveys, classified remotely sensed AVHRR images, and error analysis to answer questions about the impact of classification error on management decisions. We use our data to estimate (1) the probability that the classification of an arbitrary pixel is incorrect, (2) the rates of different types of mis-classification (Type I vs. Type II statistical errors), and (3) the implication of such errors for management and risk analysis. Although our study focuses on south-central New Mexico (USA), we expect the approach will be broadly applicable and have implications for other arid regions.