The estimation of area by land cover type is a key component of most large-scale forest inventories. Historically, these estimates were derived from a large sample of points, taken from aerial photos, followed by a smaller sample of ground points, which were used to correct errors in the classification of the aerial photo points. There has been interest in replacing aerial photography with satellite imagery. One problem with using satellite imagery is the registration errors between a pixel and a plot or a point on the ground. The estimators and modes of inference can differ substantially depending on whether the sample unit is a plot or point on the ground. Two terms, tessellated and point paradigm, are used to differentiate between the two approaches when the ground data consists of plots or points, respectively. The effects of registration errors are compared using the two paradigms for estimating the area of land by cover type. The effects of registration errors on the expected value and variance of the forest area estimator under both the tessellated and point paradigm were studied using a simulation study, where a percentage of the samples had a random one- or two-pixel registration error. The simulation study shows that registration errors increased the variance of the estimator of forest area from 4% to 434%. The estimator of forest area under both the tessellated and point paradigm exhibited no detectable bias. In the presence of registration errors, the estimated variance under the tessellated paradigm tended to overestimate the true variance with the achieved coverage rate for a nominal 80% confidence interval ranging from about 81% to 86%. For sample sizes of fewer than 100 ground points, the estimated variance under the point paradigm tended to underestimate both the 80% confidence interval and the true variance, regardless of whether there were registration errors. Further testing showed that sample sizes of between 100 and 250 ground points were needed before the estimator of the variance under the point paradigm converged to within 5% of the true variance of the estimator of forest area. Thus, we conclude that registration errors can drastically increase the variance of area estimators and the resulting confidence intervals will not achieve their nominal coverage rates. In this study, the forest area estimator under the tessellated paradigm was clearly more robust to registration errors than the area estimator under the point paradigm. FOR. SCI. 49(1):110–118.