Image processing for plant determination using the Hough transform and clustering methods.

One important goal of future crop production is to reduce the usage of herbicides. In many circumstances automatic detection and determination of weeds and crops is necessary for feasible weed control. Previous approaches in the area of computer vision based plant classification in a natural environment have only limited success, especially under variable field conditions with overlapping plants. The objective of this study was to apply a technique to cluster objects classified by a computer vision system in order to recognise not only single leaves but also whole plants. Images of the crop were obtained using a device that blocked out the natural light to provide controlled artificial lighting conditions. The crop rows were calculated from the determined plant positions. Plants which were not located in the row were labelled as weeds. It has been investigated, whether information on the row position can be used to reduce the classification errors. With this approach plant classification under field conditions can be improved and a plant classification accuracy of over 90 % for cabbage and 70 % for carrots has been obtained. The results clearly depended on the growing stage of the crop, on specific parameters of the row identification process and on the question whether the type 1 error (not identified crops) or the type 2 error (weed classified as crop) should be minimised.