CNN Based Identification of weeds in Tomato Farm

Precision farming is a technological revolution that aims to use the latest technologies to improve productivity without compromising quality. Weedicides used to control the growth of weeds need to be judiciously applied to avoid health risks that can affect ecosystems. Further early stages of weed identification can contribute to a significant reduction in herbicide usage. The present research is to investigate the machine learning algorithms that result in a faster weed detection process with better accuracy. Further, the classification of weeds by crop is also addressed. This study has considered the YOLO detection algorithm to achieve the above objective. An accessible dataset of tomato crop (Solanum lycopersicum L) and its weed species, black nightshade (Solanum nigrum L), is used to evaluate the developed approach. The images were captured using RGB cameras at several production locations in Greece under varying natural light conditions. The number of epochs is set to 200 for training 736 images. Performance comparisons between two versions of YOLO, namely YOLOv5 and YOLOv7, are made. From the results, it is concluded that YOLOv5 outperforms YOLOv7in terms of training time (17 minutes, 22 seconds). Further, the F1 score is found to increase by 2.48% when compared to YOLOv7, which accounts for the improvement in detection performance in YOLOv5. The results revealed that YoLOv5 is better for feasible real-time extension to a crop-weed detection system, although YOLOv7 has a detection time 10 times greater than that of YoLOv5.