Deep convolutional neural networks for image-based Convolvulus sepium detection in sugar beet fields.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate detection of *Convolvulus sepium* (hedge bindweed) in sugar beet fields is challenging due to environmental variations and limitations of traditional machine learning methods.
• Conventional approaches often require extensive feature engineering, hindering generalization across different field conditions.

Purpose of the Study:

• To develop a high-performance deep learning model for the detection and segmentation of *C. sepium* and sugar beet.
• To improve upon existing weed detection methods by leveraging synthetic data augmentation and efficient model architectures.

Main Methods:

• A deep convolutional neural network (CNN) based on the tiny YOLOv3 architecture was developed.
• A dataset combining 2271 synthetic images with 452 field images was used for training.
• YOLO anchor box sizes were optimized using k-means clustering on the training data.

Main Results:

• The combined synthetic and field image training approach improved mean average precision (mAP) from 0.751 to 0.829 compared to using field images alone.
• The developed model demonstrated a superior balance between accuracy (APs@IoU0.5 for *C. sepium*: 0.761, sugar beet: 0.897) and speed (6.48 ms inference time).
• Performance was validated against standard YOLOv3 and Tiny YOLO models, showing enhanced capabilities.

Conclusions:

• The developed model shows promise for deployment on embedded systems (e.g., Jetson TX) for real-time weed detection and management.
• Training with a combination of synthetic and field images is recommended to enhance model performance in agricultural applications.