In-field Triticum aestivum ear counting using colour-texture image analysis

A colour and texture image analysis method based on the determination of a hybrid space was developed for a feasibility study for the (semi-)automatic counting of Triticum aestivum wheat ears to simplify manual counting. To detect ears, five textural and statistic features, and colour analyses were both used to give a new representation of the images within a specific space (hybrid space). This new representation was constructed with a priori knowledge about the images (especially the number of classes and training points), providing better recognition than in the standard RGB space (Red/Green/Blue). Classical methods of image segmentation and classification, combined with morphological information about wheat ears, were then applied to the new images to assist counting. Only 20 images were tested and classification accuracy ranged from 73% to 85%. The counting information, which needs to be validated on numerous images, will be in future combined with grain counting per ear and thousand-seed weight to obtain an estimation of wheat yields. The resulting information could prove to be relevant, for example, to allow French cooperatives to organise their harvest.