Deep learning model optimization methods and performance evaluation of YOLOv8 for enhanced weed detection in soybeans

Abstract

The neural network YOLO (You Only Look Once) is a supervised learning technique for object detection in real time. This deep learning architecture can be a tool for weed recognition and site-specific weed management. The objective of this study was to evaluate different strategies for real-time weed detection in soybean crops by means of YOLOv8 variants. A dataset for weed recognition was created which largely consisted of 10 weed species which were divided into broadleaf and narrowleaf. Three classes were designated as soybean (Glycine max), broadleaf, and narrowleaf. An experiment conducted with the original versions of YOLOv8 (nano, small, medium, large, and extra-large) showed that the transfer learning strategy was more efficient in the case of the two largest YOLOv8 architectures, where there was an increase in the mAP50 from 0.71 to 0.73. This study put forward four new variants: YOLOv8 (femto), YOLOv8 (atto), YOLOv8 (pico), and YOLOv8 (zepto), which had optimizing model parameters that led to a reduction in GFLOPs of up to 84.2 %. The YOLOv8 models (especially YOLOVv8 nano and femto) have shown a great potential for real-time weed detection in soybeans, and achieved promising results in recognizing different types of weeds in soybean crops, which are comparable to state-of-the-art methods. The use of the semi-supervised technique showed an increase in mAP50 from 0.70 to 0.73. Reducing the YOLOv8 model parameters does not affect the model accuracy and is a key factor in reducing complexity and processing time, as a means of enhancing real-time weed detection systems.