Goosegrass Detection in Strawberry and Tomato Using a Convolutional Neural Network.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Weed Science

Background:

• Goosegrass is a significant weed in Florida vegetable plasticulture, increasing production costs.
• Developing cost-effective weed management strategies is crucial for agricultural sustainability.

Purpose of the Study:

• To evaluate the efficacy of YOLOv3-tiny for in situ goosegrass detection.
• To compare different annotation techniques for optimizing weed detection in precision agriculture.

Main Methods:

• Utilized YOLOv3-tiny (You Only Look Once version 3 - tiny) deep learning model for goosegrass detection.
• Compared two annotation methods: entire plant (EP) and leaf blade (LB).
• Assessed detection performance using F-score in strawberry and tomato crops.

Main Results:

• The leaf blade (LB) annotation method yielded higher F-scores (0.85 in strawberry, 0.65 in tomato) compared to the entire plant (EP) method (0.75 in strawberry, 0.56 in tomato).
• LB annotation improved recall, capturing more weed targets, albeit with some over-spraying on non-target areas.
• The developed network demonstrated real-time, in situ detection capabilities.

Conclusions:

• The LB annotation technique is superior for goosegrass detection in precision spraying applications.
• The YOLOv3-tiny model shows promise for real-time weed management in field applications.
• This technology can support autonomous scouts and precision spraying for enhanced weed control.