Designing and Developing a Weed Detection Model for California Thistle

Abstract

With a great percentage of farms in New Zealand as pastures, they are mainly important in contributing to the milk and meat industries. Pasture quality is highly affected by weeds. Weeds grow fast and invade pastures by seed pollination. They consume the nutrients, water, and other minerals, and once they are bitter, cattle do not eat them. Therefore, dairy farmers have to allocate a significant portion of their budget and time to monitor and clean weeds. Unfortunately, most weed management tasks are manual with no consistent technology. Thus, the motivation behind this article was to design an object detection model for weed monitoring and control in pastures. The model was designed and tested on California thistle, a dominant and widespread weed on New Zealand pastures. Our study is one of the major model designs for identifying weeds in an in-pasture environment, one of the most complicated environments for any object detection model. A synthetic methodology was used to create three types of datasets: plant-based, leaf-based, and mixed. The trained model based on the leaf-based dataset is one of the major contributions of our work and has not been conducted by any other weed detection models. After models had been trained, tuning experimentation was undertaken to improve the model’s performance. This involved studying the model’s hyperparameters in various ranges and then recording their values at the optimum points. The improved model showed a 93% mAP accuracy in the detection of training images and over 95% accuracy for testing images. The experimentation showed that the leaf-based model was slightly better than other models. The model can automate highly any weed management system. The use of this model will save farmers time and money and help them reduce the errors of manual work.