Weed target detection at seedling stage in paddy fields based on YOLOX

Area of Science:

• Agricultural Engineering
• Computer Vision
• Deep Learning

Background:

• Weeds pose a significant threat to rice crops, especially during early growth stages, leading to substantial yield losses.
• Traditional large-area spraying methods for weed control are inefficient, cause herbicide waste, and contribute to environmental pollution.
• Precision spraying requires rapid and accurate detection of weed distribution in rice fields.

Purpose of the Study:

• To develop and evaluate a deep-learning-based computer vision method for detecting weed targets in rice fields.
• To address the challenge of identifying small, dense weed targets during the rice seedling stage.
• To enable the transition from broadcast spraying to precision weed management in agriculture.

Main Methods:

• A weed target detection method based on the YOLOX model, incorporating a CSPDarknet backbone, feature pyramid network (FPN), and YOLO Head detector.
• Feature extraction using CSPDarknet at multiple scales (80x80, 40x40, 20x20 pixels).
• Performance comparison of YOLOX-tiny against other models like YOLOv3, YOLOv4-tiny, YOLOv5-s, and SSD.

Main Results:

• The YOLOX-tiny model demonstrated superior performance, achieving a mean Average Precision (mAP) of 0.980, F1 score of 0.95, and recall of 0.983.
• YOLOX-tiny requires only 259.62 MB of intermediate variable memory, making it suitable for deployment on intelligent agricultural devices.
• The proposed method effectively improves detection of small, sheltered, and dense weeds in early-stage rice fields.

Conclusions:

• The YOLOX-tiny model is highly effective for detecting weeds in rice fields, particularly small and dense targets at the seedling stage.
• The developed weed detection model is suitable for embedded systems, paving the way for autonomous targeted herbicide spraying by agricultural robots.
• This research contributes to sustainable agriculture by enabling precision weed management and reducing environmental impact.