Using a competition model to quantify the optimal trade-off between machine vision capability and weed removal effectiveness

The algorithm of an optical detection system was first investigated for its ability to correctly classify transplanted crops and weeds during the critical early stages of crop establishment and its robustness over a range of different crop species. The trade-off was then examined between increasing the sensitivity of the detection system vs. the possibility of, in doing so, misclassifying some crop plants as weeds and inadvertently removing them. This was achieved by running a competition model using parameters derived from the image analysis and assessing the outcome of scenarios in terms of yield. The optimum parameter values to maximize the detection of the crop and the optimum parameter values to maximize the detection of the weed appeared relatively insensitive to time of image capture or weed density. They also appeared insensitive for different crop species where the crop had similar growth habit. However, competition scenarios indicated that the detection system parameter settings to achieve optimum yields were sensitive to the competitive ability of the weed species. For Veronica persica, crop yield was more sensitive to accidental crop removal than from competition. In contrast, in the presence of Tripleurospermum inodorum, yield loss was more attributable to weed competition. Importantly, linking the detection system with the competition model illustrated the principle that optimum yield may not necessarily be obtained by maximizing weed removal or minimizing crop removal. This first example of combining a detection system with a competition model presents a new opportunity to quantify the sensitivity of image classification in terms of yield.