An Intelligent Rice Yield Trait Evaluation System Based on Threshed Panicle Compensation

Area of Science:

• Agricultural Science
• Plant Breeding
• Computational Biology

Background:

• Traditional rice yield trait evaluation is labor-intensive, complex, and inefficient.
• Existing methods struggle with threshing difficulties and low throughput.
• Accurate phenotyping is crucial for advancing rice genetics and breeding programs.

Purpose of the Study:

• To develop a novel, high-throughput, and accurate intelligent system for evaluating rice yield-related traits.
• To address the limitations of conventional phenotyping methods in rice.
• To provide an efficient and reliable tool for rice breeding and genetic research.

Main Methods:

• Development of an integrated system including threshing, conveying, and imaging units with specialized image analysis software.
• Optimization of the Faster R-CNN architecture (TPanicle-RCNN) using Region of Interest Align, Convolution Batch Normalization-Leaky ReLU, Squeeze-and-Excitation, and optimal anchor sizes for improved threshed panicle detection.
• Implementation of AI cloud computing for cost reduction and enhanced system flexibility.

Main Results:

• The TPanicle-RCNN model achieved an F1 score of 0.929 for threshed panicle detection, demonstrating robustness across indica and japonica varieties.
• The intelligent system reduced total spikelet measurement error from 11.44% to 2.99% with threshed panicle compensation.
• Average measurement efficiency reached approximately 40 seconds per sample, representing a twenty-fold increase over manual methods.

Conclusions:

• The developed automatic and intelligent system offers a significant advancement in rice yield-related trait evaluation.
• This technology provides an efficient, accurate, and reliable tool for accelerating rice breeding and genetic studies.
• The system overcomes the challenges of conventional methods, enabling high-throughput phenotyping.