Intrarow Uncut Weed Detection Using You-Only-Look-Once Instance Segmentation for Orchard Plantations

Area of Science:

• Agricultural Robotics
• Computer Vision
• Machine Learning

Background:

• Mechanical weed management in orchards is labor-intensive and poses risks.
• Intrarow weeding is challenging for autonomous systems due to GNSS signal obstruction from orchard structures.
• Existing autonomous weeders struggle with accurate weed and obstacle identification.

Purpose of the Study:

• To develop an intelligent vision module for autonomous robotic weeders.
• To enable accurate recognition of weeds and obstacles within orchard rows.
• To support robotic operations in previously inaccessible intrarow areas.

Main Methods:

• Utilized YOLO instance segmentation algorithms (YOLOv5n-seg, YOLOv5s-seg, YOLOv8n-seg, YOLOv8s-seg) trained on a custom dataset from a pear orchard.
• Collected and preprocessed 5000 images for training and testing.
• Evaluated models based on detection accuracy, complexity, and inference speed for real-time application on edge devices.

Main Results:

• Smaller YOLO models (YOLOv5-based and YOLOv8-based) demonstrated higher efficiency.
• YOLOv8n-seg was selected for its superior segmentation accuracy and acceptable performance on resource-constrained devices.
• The developed vision system achieved effective object recognition for robotic intrarow weeding.

Conclusions:

• Deep learning-based vision systems can significantly enhance autonomous robotic weeding in orchards.
• YOLOv8n-seg is a suitable choice for vision modules in robotic weeders due to its balance of accuracy and efficiency.
• The proposed system addresses key challenges in autonomous orchard management, paving the way for more efficient and precise weed control.