Weed mapping using digital image processing

In this paper an approach for automatic weed detection based on image analysis is presented. The system consists of three bispectral CCD-cameras mounted on a boom in front of a vehicle. Geocoded images were taken on the go automatically and stored on a computers hard disk. Afterwards the images were preprocessed and reduced to the outer contours of the objects within an image. Weeds and crop species were identified using shape analysis and pattern recognition. The actual weed identification was based on a comparison of multi invariant parameters, which were stored in a knowledge base. The average time for image acquisition and weed recognition for a set of three images was around 2 s. The results were documented in a file that was used for weed mapping. The average identification rate with a knowledge base based on 5 weed classes concerning to their sensitivity against herbicides was around 80%.