FEWheat-YOLO: A Lightweight Improved Algorithm for Wheat Spike Detection.

Area of Science:

• Agricultural Technology
• Computer Vision
• Machine Learning

Background:

• Accurate wheat spike detection and counting are vital for precision agriculture, aiding yield estimation and variety selection.
• Existing models struggle with complex field conditions, morphological variations, and small target sizes, limiting real-world applicability.
• There is a need for efficient and lightweight models deployable on agricultural edge devices.

Purpose of the Study:

• To propose FEWheat-YOLO, a novel lightweight and efficient detection framework for wheat spike detection and counting.
• To optimize the framework for deployment on resource-limited agricultural edge devices.
• To evaluate the model's performance in terms of detection accuracy, counting precision, and computational efficiency.

Main Methods:

• Developed FEWheat-YOLO, integrating FEMANet (Efficient Multi-scale Attention), BiAFA-FPN, ADown, and GSCDHead modules.
• FEMANet enhances small-target representation through mixed aggregation and Efficient Multi-scale Attention (EMA).
• BiAFA-FPN facilitates efficient multi-scale feature fusion, ADown preserves details during downsampling, and GSCDHead reduces computational cost.

Main Results:

• FEWheat-YOLO achieved a COCO-style AP of 51.11% and AP@50 of 89.8% on a hybrid dataset.
• Demonstrated high counting accuracy with R² of 0.941, MAE of 3.46, and RMSE of 6.25.
• Achieved 54 FPS inference speed with only 0.67 M parameters, 5.3 GFLOPs, and 1.6 MB storage, outperforming YOLOv11n in efficiency.

Conclusions:

• FEWheat-YOLO offers a superior balance between detection accuracy, counting performance, and model efficiency for wheat spike analysis.
• The lightweight design and high inference speed make it suitable for real-time applications on agricultural edge devices.
• The model shows significant potential for advancing precision agriculture through efficient crop monitoring and yield prediction.