HDMS-YOLO: a multi-scale weed detection model for complex farmland environments.

Area of Science:

• Agricultural Technology
• Computer Vision
• Machine Learning

Background:

• Accurate weed identification is crucial for precision agriculture but challenging due to weed diversity and visual similarity to crops.
• Existing methods struggle with the complexities of weed identification in varied agricultural settings.

Purpose of the Study:

• To develop a robust weed identification model for precision agriculture.
• To enhance the accuracy and generalisation capabilities of weed detection systems.

Main Methods:

• Proposed the HDMS-YOLO model, incorporating Shallow and Deep Receptive Field Distillation (SRFD and DRFD) modules for comprehensive feature extraction.
• Replaced the C3K2 structure with a Partial Convolution-based Multi-Scale Feature Aggregation (PC-MSFA) module for improved feature representation.
• Introduced the IntegraDet dynamic task-alignment detection head to enhance localisation and classification accuracy.

Main Results:

• HDMS-YOLO achieved 74.2% accuracy, 66.3% recall, and 71.2% mAP on the CropAndWeed dataset.
• Demonstrated significant improvements (2.6%, 2.1%, 2.6%) over YOLO11 in accuracy, recall, and mAP.
• Outperformed other mainstream algorithms in overall weed detection performance.

Conclusions:

• The HDMS-YOLO model effectively extracts and represents weed features, improving identification accuracy and generalisation.
• The model shows strong potential for precision farm management and the development of intelligent weed-removal robots.
• Highlights advancements in automated weed removal for unmanned agricultural systems.