A segmentation procedure using colour features applied to images of Arabidopsis thaliana

In studies of environmental effects on plant growth, the images of plants are often used for non-destructive measurements in phenotyping. In this work, a computational procedure has been developed to segment images of plants allowing an improved separation of plants and other types of objects in the frame such as moss or soil. The proposed procedure is based on colour analysis and image morphology. The red-green-blue (RGB) values are transformed into a colour space as ratios of R, G and B vs the sum of R, G, and B channels. We introduce an approach to render the training set of pixels on a Microsoft Excel two-dimensional graph and a technique to determine the discriminant regions of pixel classes. Two approaches for the classification based on colour analysis are shown: an automatic method using support vector machines and a procedure based on visual inspection. The segmentation procedure is designed to classify more than two object types utilising flexibly curved boundaries of discriminant regions that can also be non-convex. We propose a machine-vision algorithm to detect plant features - leaf anthocyanin accumulation and trichomes. The procedures of segmentation and feature detection are applied to images of Arabidopsis thaliana (L.) Heynh. that grow under either normal or drought stress conditions.