The sensitivity of the normalized difference vegetation index (NDVI) to regional variations in the bulk subpixel properties of discontinuous plant canopies is investigated within the framework of a linear stochastic-geometric landsurface reflectance model. Analytical formulas are developed for i) the rate of change of the NDVI with respect to subpixel components and ii) the variance of the NDVI for an ensemble of pixels. Atmospheric effects are not considered. The formulas are first applied to the analysis of the NDVIs of a series of five simulated scenes of Poisson distributed plants, aggregated to three different pixel sizes. Varying only a single parameter in each subsequent scene provides an understanding of the influence of that variable on the behavior of the NDVI. Next, a comparison is made between the actual NDVI of a pecan orchard calculated from aerial radiometric data obtained during the June 1988 Maricopa Agricultural Center Experiment, and the predicted NDVI using the geometric reflectance model and ground truth. Finally, the NDVI of a natural landscape that possesses randomly distributed juniper trees and random soil background reflectance is investigated. Investigations are generalized to other plant geometries and solar angles through the use of a nondimensional similarity parameter η. The functional relation among the amount of canopy cover, shadowed ground, and illuminated ground, which occurs only for spatially homogeneous plant distributions at large sampling scales, facilitates the analytical treatment of the NDVI and reflectance equations.