A 100 km × 200 km region of the Sahetian zone of Africa was chosen as the place to demonstrate a new method of extracting hemispherical reflectance (albedo) from directional satellite data. Five NOAA-9 AVHRR daytime scenes in band I (0·58−0·68μm) were corrected for atmospheric variations to give the directional reflectances at the ground level. Using a stratification of the Gao region of Mali (Hiernaux 1989) based on soil texture, satellite data for homogeneous areas were selected. The number of view directions and the orientation of these varied because of the variable cloud cover and the changing orientation of the scanning radiometer with each overpass. These data were used in a knowledge-based expert system called VEG. This system is designed to handle a wide variety of kinds of input data. It contains an array of different techniques for inferring hemispherical reflectance. VEG accepted the spectral directional reflectance data, determined the best techniques for inferring spectral hemispherical reflectance, applied these to the selected sites and provided a rigorous estimate of the accuracy of the inference. VEG made all decisions and inferences using the input reflectance data and internal knowledge bases. The calculated directional reflectance factors at ground level were very similar to ground measurements reported in the literature. Furthermore, it was clear that a single nadir observation technique was always less accurate for estimating hemispherical reflectance than VEG using multiple view angles. The mean errors of the hemispherical reflectance of the entire region using VEG with multiple view angles and the direct nadir technique were 3 and 10 per cent, respectively, where the errors of the individual homogeneous areas were weighted by their relative sizes.