Field rice panicle detection and counting based on deep learning.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Rice panicle number is a key determinant of crop yield, making accurate counting essential for agricultural research and management.
• Traditional panicle counting methods face challenges due to high plant density, occlusion, and variations in panicle appearance.
• Deep learning offers advanced capabilities for object detection but requires adaptation for large-scale agricultural imagery where small objects can be missed.

Purpose of the Study:

• To develop and validate a robust deep learning-based method for accurate rice panicle detection and counting in large-sized field images.
• To address the limitations of existing object detection methods, particularly concerning the detection of small or occluded panicles.
• To provide an accessible tool for rice researchers to facilitate efficient and precise panicle analysis.

Main Methods:

• Comparison of various deep learning object detection models, with YOLOv5 selected for its superior performance.
• Development of a novel method for eliminating redundant detections, outperforming traditional Non-Maximum Suppression (NMS) techniques.
• Implementation and evaluation of the method on diverse datasets, including varying illumination conditions, rice varieties, and image resolutions, as well as Unmanned Aerial Vehicle (UAV) imagery.

Main Results:

• The proposed method achieved a Mean Absolute Percentage Error (MAPE) of 3.44% and an accuracy of 92.77% in panicle counting.
• The novel detection removal method demonstrated superior performance compared to existing NMS algorithms.
• The system proved robust across different lighting, rice accessions, and image input sizes, and performed effectively on UAV images.

Conclusions:

• The developed deep learning approach provides a highly accurate and robust solution for rice panicle detection and counting, even in challenging field conditions.
• The method effectively overcomes the issue of missing small panicles in large images, a common limitation in existing approaches.
• An open-access web portal has been created, enhancing the accessibility and usability of this advanced panicle counting technology for the agricultural research community.