In recent years, precision agriculture and precision weed control have been developed aiming at optimising yield and cost while minimising environmental impact. Such solutions include robots for precise hoeing or spraying. The commercial success of robots and other precision weed control techniques has, however, been limited, partly due to a combination of a high acquisition price and low capacity compared to conventional spray booms, limiting the usage of precision weeding to high-value crops. Nonetheless, conventional spray booms are rarely used optimally. A study by Jørgensen et al. (2007) has shown that selecting the right herbicides can lead to savings by more than 40 percent in cereal fields without decreasing the crop yield when using conventional sprayers. Therefore, in order to utilise conventional spray booms optimally, a preliminary analysis of the field is necessary. This analysis should determine which weeds are present in the field and the density of those weeds so that herbicides targeting those weeds may be selected.
Researchers have sought to detect and classify weeds and crops in images, but studies are limited regarding the number of plant species that can be discriminated and the flexibility of the camera setup.
In the present PhD thesis, requirements for the camera set-up are loosened, allowing the use of consumer grade cameras or even cell phones for weed species localisation and identification in images from conventionally grown fields. In total 4 537 images have been collected over three growth seasons from Danish fields. In these images 31397 plants are annotated with names, from which the 17 most frequent species are selected for automated classifiiiication. The automated classification consists of two steps: Initially, weeds are located in images after which, the weeds are classified.
Three types of weed localisation approaches are tested: Two approaches that perform a pixel-wise segmentation of plants, and one approach, that detects regions in images containing weeds. Common for all three approaches is that they aim at overcoming some of the challenges when working with images from fields: Namely changes in lighting, soil types, and plant stress due to lack of nutrition. The first of the suggested approaches segments plant material from the soil by using fuzzy C -means clustering combined with a threshold value for each pixel, which depends on the neighbourhood pixels, which helps to detect non-green stem regions. The second approach uses a fully convolutional neural network for segmenting pixels in three categories: Soil, weeds, and crops. The Neural Network is trained solely on modelled images but can segment weeds from maize with an intersection-over-union of between 0.69 and 0.93 for weeds and maize. Rather than segmenting images, the third approach produces region proposals that indicate weed locations in images. This method also uses a fully convolutional neural network, that enables it to detect weed instances in wheat fields despite occluding leaves.
The three methods for weed segmentation and localisation solve four problems in the field of camera based weed detection: handling of changing environments, handling of non-green plant stems, segmentation of weeds and crops that are overlapping, and instance detection in cereal fields with occluding leaves.
Following the detection of the weeds, the weed species are to be determined. For solving this problem, a convolutional neural network is used, which classifies the weeds with an overall accuracy of 87 percent for 17 species despite a severe degree of leaf occlusion.
Because of the ability to handle weed detection and classification in natural environments, these methods can potentially reduce the investment of farmers, and thus lead to a higher adoption rate than existing precision weed control techniques, resulting in huge potential savings regarding herbicide consumption.