Weeds pose a significant challenge in Australia, costing the livestock industry an estimated AUD 2,409 million annually due to control measures and lost production. The application of computer vision for weed detection in agriculture is an expanding area of research, offering the potential to develop autonomous, site specific weed management (SSWM) solutions that minimize herbicide use and reduce the labour needed for weed control. Using images collected by autonomous ground vehicles in Southeast Australian pastures infested with a range of weeds, several computer vision techniques were considered.
In one study, a pipeline was developed using local binary patterns, histograms of oriented gradients, and colour features, and an extreme learning machine, to produce a fine, patch-based segmentation of images containing serrated tussock, thistle species, white horehound, and Bathurst burr, yielding a mean Intersection over Union (mIOU) of 87.6, 81.6, 79.5 and 87.1 respectively. The fine-grained segmentation produced by this model enabled more accurate localization of individual weed plants, which should enable a greater precision in herbicide application.
In another study, a dual-purpose network for plant segmentation and spray point detection was developed using a novel convolutional neural network (CNN) architecture incorporating ConvNeXt modules. This network was tested on a dataset of serrated tussock and achieved a mean Intersection over Union (mIOU) of 80.7 for segmentation and an F1-score of 79.6 for spray point detection, even under field conditions not encountered during training. These results outperformed state-of-theart dual-purpose networks. Furthermore, when a novel normalization technique, termed HistMatch, was applied prior to input, the network’s performance improved to an mIOU of 85.4 and an F1-score of 80.6, respectively.
The dual-purpose network was then tested on a multispecies dataset of serrated tussock, white horehound, variegated thistle and stinging nettle, yielding mIOU values of 88.8 and 82.6 for the plant and spraypoint tasks respectively. This demonstrated that the network was able to generalize well to the situation in which multiple species of weed are present in a field.
The studies presented in this thesis have advanced the understanding of computer vision techniques for weed localization in pastures in several ways. Firstly, the ELM-based segmentation approach demonstrated greater precision in weed localization compared to previous methods that relied on classification or bounding box techniques. Additionally, the development of a dual-purpose network for plant segmentation and spray point localization in pastures represents the first of its kind, enabling more flexible treatment options depending on whether herbicide is applied to the entire plant or just its center. HistMatch normalization proved to be a superior technique compared to other normalization methods, particularly in scenarios where illumination and growing conditions in the test set differ from those encountered during training. Further testing of this dual-purpose network showed that it remains robust even in complex scenarios where multiple weed species are present in the field.