Land cover is a key boundary condition in weather, climate, and terrestrial biogeochemical models. Until recently, such models have used maps depicting potential vegetation, which are known to be of relatively poor quality, to parameterize land surface properties. In this paper we describe the compilation and assessment of a new map of North American land cover produced through the application of advanced pattern recognition techniques to multitemporal satellite data. This map was produced in a fully automated fashion using supervised classification methods that are robust, fully automated, and repeatable. The processing flow described in this paper is a prototype of the algorithm to be used to generate maps of global land cover using data from EOS MODIS. The superior quality and timeliness of these maps should be very useful for a wide array of sub-continental to global-scale modeling and analysis activities. ¿ 2000 American Geophysical Union