Colour representation methods for segmentation of vegetation in photographs

Papers dealing with vegetation segmentation from RGB images confound the effect of colour representation with a clustering algorithm. In this paper, various colour representations of RGB images are compared. On a set of images, vegetation and soil pixels were manually isolated to form two known populations for each image. For each pixel in the populations, a quadratic discriminant analysis was applied. The comparison between colour representation methods was based on classification errors. Two image sets were used. One was under controlled flash illumination and the other one was under uncontrolled outdoor lighting. Images acquired under controlled illumination were transformed to simulate illuminants corresponding to other correlated colour temperatures (CCT) covering most situations that can be encountered under natural lighting in the field. Under controlled illumination, the error rates for both vegetation and soil pixels were more influenced by the site, crop or CCT than by the projection method. Based on robustness with respect to the white point definition coupled to good balance between classification errors for soil and vegetation pixels, it is concluded that the L∗a∗b∗ projection is the best projection for segmenting green vegetation against a soil background using consumer-grade charge-coupled device (CCD) cameras.