It is known that primate red-green color vision is efficient at encoding the presence of red or yellow fruit or leaves against a background of green foliage. However, our observations of monkey foraging behavior in Kibale Forest, Uganda during the dry season suggested that monkeys frequently ate green leaves on trees lacking any red object. They also showed preferences for specific trees. We asked whether the neural encoding of the green leaves of such trees allows discrimination from other trees, across marked differences in illumination due to time-of-day and weather effects. We obtained 80 images of two scenes, each containing several types of tree, throughout two days at intervals of 10–20 minutes, using a calibrated digital camera system described elsewhere (Párraga, Troscianko, and Tolhurst Current Biology 12, 483–487; 2002). The camera calibration allowed the decomposition of each pixel into L,M,S cone responses, and also luminance, red-green, and yellow-blue opponent responses. We averaged the values of these responses in five separate patches for images from Day 1, and six patches for Day 2. Our first analysis replicated the approach of Nascimento, Ferreira, and Foster (2002) JOSA A 19, 1484–1490, who suggested that ratios of cone responses across different patches should be invariant against changes in illumination. This turned out not to hold when one of the patches was plunged into shadow. Importantly, if similar ratios are taken of the opponent-channel responses, particularly the red-green channel which shows invariance to shadows, these new ratios are more invariant across illumination changes by an order of magnitude. We conclude that the red-green opponent system provides information about scenes containing green leaves which is strongly invariant across changes in illumination direction, spectral composition, and intensity. In other words, for scenes containing foliage, the color constancy problem is solved at the level of the retina.