Detection of commercial crop weeds using machine learning algorithms.

This work investigates the YOLOv5 object detection algorithms for classifying commercial crops such as tomatoes, chili, and cotton. The data sets comprise 707 images of green chillies, 200 images of tomato crops and 130 images of weeds from Ponnandagoundanoor farms in western agro climatic Zones (WAZ) of Tamil Nadu. The objective of this research is to explore the determination of weed present in the crops and further the machine learning (ML) algorithms that have deployed for computing the F1 score, detection time, and mAP of each machine learning algorithms. As a result, a tomato dataset contains an F1 score of 98%, a mAP of 0.995, and a detection time of 190 ms; a cotton dataset with an F1 score of 91% and a mAP of 0.947; and a chilly dataset with an F1 score of 78% and a mAP of 0.811. A Further investigation has been carried out for the same crops; improving YOLOv5 accuracy includes adaptively spatial feature fusion (ASSF) blocks to its architecture head. An enhanced YOLOv5 algorithm using ASFF modules on the same datasets achieved an F1 score of 99.7% in the tomato dataset and 79.4% in the chilly dataset, resulting in a 1.14% improvement in the F1 score. With a 93.53% F1 score, were able to obtain a 2% enhancement over the cotton dataset. The extended YOLOv5 increased the mAP by about 0.5% and resulted in an insignificant drop in the number of computations carried out, rendering the model more compact.