Color vision may be used to help identify objects and segment scenes despite variations in the spectral composition of the light. But humans, and some animals, have just three classes of cone receptors, too few to fully specify spectra. Given this limit, do normal human cones maximize the information available? This question was addressed by the present study. Images of rural and urban scenes were obtained with a high-resolution hyperspectral imaging system, which provided estimates of surface spectral reflectance over 400-720 nm at 10-nm intervals at each point in a digital representation of spatial resolution 1344 x 1024 pixels. In computational simulations, 35 scenes were illuminated by randomly selected combinations of daylights with and without filtering through a leafy canopy. Five thousand points in each scene were randomly selected and Shannon’s mutual information was calculated as a function of the spectral position of a variable number of cone pigments. The mutual information had several maxima, three of the highest being within about 10 nm of the mean normal long-, medium-, and short-wave cone positions, suggesting that trichromacy is almost optimal for extracting information about surface color in natural scenes.