Machine Learning for Identification and Classification of Crops and Weeds

Abstract

Machine learning for computer vision tasks is an area which has grown significantly in recent years. At the same time, there has been a marked increase in demand for better, more environmentally friendly farming methods. The current study uses machine learning applied to computer vision to develop a system that is able to identify and classify plants as either crops or weeds with a view to enabling robotic appliances to perform plant cultivation and monitoring. This is performed using a state-of-the-art object detection architecture known as Faster R-CNN, trained and tested on a publicly available crop/weed dataset. Various configurations of this network were compared and tested on a Raspberry Pi 4 in order to gauge the usefulness of this detection system in field conditions, in terms of both speed and accuracy. The results show that relatively high speed predictions are possible while using low resolution images, but that significantly more accurate results can be obtained by using larger images, albeit with a significant speed penalty. Depending on the application, however, the low resolution inference may already be adequate.