An Efficient Deep Learning Technique for Detecting and Classifying the Growth of Weeds on Fields

The agriculture cycle needs to be expanded in the next decades to meet the demand of the world population. Weeds are one of the main challenges that severely affect the agricultural production and its quality. An accurate, automatic, low cost, little environmental impacts and real-time weeds detection technique is required to control weeds effectively on fields. In addition, automating the classification process of weeds based on their growth stages is crucial for using appropriate weeds-controlling techniques. In this paper, we fly a drone to collect a dataset of four different weed (Consolida Regalis) growth stages. As well, we developed and trained deep learning object detector (YOLOv5) to detect weed (Consolida Regalis) and to classify its four growth stages in real-time with a sufficient accuracy. The results show that the generated YOLOv5 small model succeeds to detect and classify the weed’s growth stages in real-time with highest recall 0.794 at 156 FPS. However, YOLOv5 large model depicts efficient detection and classification precision of 0.827 at 70 FPS.