On-line crop/weed discrimination through the Mahalanobis distance from images in maize fields

This study proposes a new automatic method for crop/weed discrimination in images captured in maize fields during the initial growth stages. The images were obtained under perspective projection with a camera installed on board at the front part of a tractor. Different approaches have addressed the problem based on crop row determination and then assuming that inter-row plants are weeds. Nevertheless, an important challenge is the identification of weeds intermixed within the crop rows. This issue is addressed on this paper by applying a minimum criterion distance based on the Mahalanobis distance derived from a Bayesian classification approach, this makes the main contribution. The identification of both intra- and inter-row weeds is useful for more accurate weed quantification for site-specific treatments. Image quality is affected by uncontrolled lighting conditions in outdoor agricultural environments. Also, different plant densities appear due to different growth stages affecting the crop/weed identification process. The proposed method was designed to deal with the above undesired situations, consisting of three phases: (i) segmentation, (ii) training and (iii) testing. The three phases are executed on-line for each image, where training is specific of each single image, requiring no prior training, as it is usual in common machine learning-based approaches, mainly supervised. This makes the second research contribution. The performance of the proposed approach was quantitatively compared against three existing strategies, achieving an accuracy of 91.8%, pixel-wise determined against ground-truth images manually built, with processing times ≤280 ms, which can be useful for real-time applications.