A field rice panicle detection model based on improved YOLOv11x

Area of Science:

• Agricultural Science
• Computer Vision
• Remote Sensing

Background:

• Rice is a global staple, necessitating accurate yield prediction for food security.
• Manual rice panicle counting is inefficient and biased.
• UAV-based panicle detection faces challenges like dense distribution, scale variation, and occlusion.

Purpose of the Study:

• To develop an advanced rice panicle detection model using an improved YOLOv11x architecture.
• To enhance feature representation, spatial dependency capture, and multi-scale fusion for improved detection accuracy.
• To provide a reliable solution for intelligent in-field rice panicle detection and precise yield estimation.

Main Methods:

• An improved You Only Look Once version 11x (YOLOv11x) architecture was developed.
• Key enhancements include Bi-level Routing Attention (BRA), a Transformer-based detection head (TransHead), and Selective Kernel (SK) Attention.
• A multi-level feature fusion architecture was integrated to improve multi-scale adaptability.

Main Results:

• The improved model achieved an mAP@0.5 of 89.4%, a 3% increase over the baseline YOLOv11x.
• Achieved Precision of 87.3% and F1-score of 84.1%, outperforming YOLOv8 and Faster R-CNN.
• Panicle counting tests showed strong fitting with R² = 0.85, RMSE = 2.33, and rRMSE = 0.13.

Conclusions:

• The proposed model offers a reliable solution for intelligent in-field rice panicle detection using UAV imagery.
• This advancement is significant for precise rice yield estimation and contributes to food security.
• The model demonstrates superior performance compared to existing mainstream algorithms.