A field assessment of a potential method for weed and crop mapping on the basis of crop planting geometry

This paper describes a computer vision based technique for automatically estimating the number of crop and weed plants and the area that they cover. Images were acquired under natural lighting using a CCD video camera mounted on a small toolframe tractor equipped with a PC. This allowed mapping to be conducted in real time in the field. The mapping techniques operated on monochrome images that exhibited a good contrast between living plant material and the background. In this work, images were obtained in colour and transformed into monochrome prior to analysis. Crop plants were segmented from weeds using a grid representing prior knowledge of their planting geometry. An extended Kalman filter (EKF) based tracking algorithm was used to track the grid from image to image, thus improving precision and robustness to weeds and missing crop plants. The EKF also yields a covariance matrix describing the confidence in grid position, allowing plant features to be classified on a probabilistic basis. A preliminary field assessment covering 30 m of 1.8 m wide transplanted brassica beds was conducted at 0.3 m s(-1). Automatically generated images viewed a three row bed that was sampled every 2 m, such that 64% of total area was examined. Subsequently high quality plan view photographs of the assessment areas were taken, and analysed through a combination of automatic and manual techniques to obtain comparative ground truth measurements. Automatically generated weed counts provided a consistent relative measure of weed populations, but increasingly underestimated weed numbers as populations increased. The results for weed area showed an underestimation by 0.48% of total image area with a standard deviation (S.D.) of 1.6%, though this result neglects two frames that exhibited gross errors due to misclassification of crop plants as weed. The automatic system correctly identified an increase in average crop plant size from run one to run two, though this difference was put at 29% compared with 37% derived from ground truth. Average absolute crop plant areas were underestimated by 1.12% of total image area. (C) 2001 Published by Elsevier Science B.V.