Related Experiment Video
Updated: Jul 4, 2025
06:41
Imaging and Analysis for Quantifying Maize (Zea mays) Abiotic Stress Phenotypes
Published on: March 28, 2025
815
Evaluation of two deep learning-based approaches for detecting weeds growing in cabbage
Hu Sun1, Teng Liu2, Jinxu Wang2
1Shandong Provincial University Laboratory for Protected Horticulture, Weifang University of Science and Technology, Weifang, China.
Pest Management Science
|February 7, 2024
Summary
Indirect weed detection using machine vision is more effective than direct detection for precision agriculture. This approach reduces herbicide use and costs by identifying weeds outside crop bounding boxes, simplifying dataset creation.
Area of Science:
- Agricultural Engineering
- Computer Vision
- Machine Learning
Background:
- Precision weed management using machine vision can significantly lower herbicide use and control expenses.
- Two deep learning methods for weed detection in cabbage were evaluated: direct weed detection and indirect detection via crop bounding boxes.
Purpose of the Study:
- To compare the performance of two deep learning approaches for detecting weeds in cabbage fields.
- To assess the efficacy of direct versus indirect weed detection strategies.
Main Methods:
- Utilized You Only Look Once (YOLO) v5 and YOLOv8 for object detection tasks.
- Implemented an indirect detection method by identifying green pixels outside predicted crop bounding boxes as weeds.
- Employed a segmentation procedure for improved weed extraction.
Main Results:
- Direct weed detection models (YOLOv5, YOLOv8) showed limited performance (metrics < 0.891) due to weed diversity.
- Indirect detection via crop bounding boxes and segmentation achieved higher effectiveness.
- Crop detection metrics for YOLOv5 and YOLOv8 were high (precision, recall, F1-score, mAP).
Conclusions:
- Indirect weed detection requires less manual effort and smaller training datasets compared to direct detection.
- Challenges remain in detecting weeds closely associated with crops within bounding boxes.
- The developed models are suitable for integration into smart sprayers and mechanical weeders.
