Segmentation of row crop plants from weeds using colour and morphology

Differentiating between crop plants, weeds and soil in digital images is a complex task especially when the crop plants and the weeds have reached an advanced growth stage. Algorithms based on a single feature such as hue or intensity are rarely successful because there is considerable overlap between the three classes in any one dimension of the commonly used colour spaces. In this work, a segmentation algorithm was developed that combined colour with knowledge about the planting grid to increase the classification potential of colour alone. Morphological filtering and line fitting were used to find the planting grid centre points. The likelihood distribution of crop plant pixels about the grid points was modelled as a bivariate Gaussian distribution. Morphological postprocessing was applied to the image to take advantage of the textural differences between weed and crop to reduce the misclassification of weed in areas close to plants. Algorithm performance was assessed on 12 images of cauliflower plants, which represented a range of plant and weed growth stages. The assessment involved measuring changes in the performance of the segmentation algorithm as the parameter values changed. On average, over 12 images, the highest crop plant classification rate was 96% and the lowest was 82%. The highest weed classification rate was 92% and the lowest weed classification rate was 68%.