Abstract
Models of visual coding based on natural scene statistics depend on identifying regularities in the environment. However, the colors characteristic of natural scenes can vary widely both across different environments and within the same environment over time because of seasonal changes. We compared these seasonal variations in two very different outdoor environments: alpine forest and meadows in the northern Sierras and tropical forest and grasslands in the Western Ghats of India. Images were acquired with a digital camera calibrated by correcting the rgb values for the known values of a color palette included in a corner of each image. The palette and scene illuminant were measured with a spectroradiometer. Vegetation changes between monsoon and winter seasons in India shift the average chromaticity of scenes from green to yellow and rotate the principal axes of the color distributions from roughly an S-cone varying axis to a blue-yellow axis. Seasonal variations in the Sierra images were weaker but followed a similar pattern. In both cases the mean color shifts are largely along an LvsM axis, suggesting that this axis captures the general lushness of scenes. Rotations in the color distributions provide a further correlate of lushness, but imply that in many contexts the color distributions are not well described by independent variations along the S and LM axes. While natural color distributions and consequently models derived from them are context dependent, the fact that the variations themselves show characteristic properties suggests important regularities in the color environment that might shape color coding.