Two methods of developing land-cover classifications using Landsat multispectral data were compared. The first method, called P-l, uses a semi-automated approach to develop training statistics which characterize the land-cover types. The second, called multicluster blocks, depends more on analyst interaction to produce the training statistics used by the classifier. The results showed that P-l performed as well as the multicluster-blocks approach on a mountainous study area in southwestern Colorado