Vigour testing for the rice seed with computer vision-based techniques.

Area of Science:

• Agricultural Science
• Biotechnology
• Computer Vision

Background:

• Rice is a global staple, with seed vigour critical for yield.
• Traditional seed vigour testing is destructive, slow, and labor-intensive.
• Developing non-destructive, efficient testing methods is essential.

Purpose of the Study:

• To propose and evaluate a non-destructive, multispectral image-based approach for rice seed vigour testing.
• To develop an automated pipeline (MsiFormer) for enhanced efficiency and accuracy.
• To compare MsiFormer with existing deep learning methods and explore alternative image analysis techniques.

Main Methods:

• Collected multispectral data (19 wavebands) for six rice varieties.
• Developed MsiFormer, integrating a YOLO object detector and a vision transformer.
• Compared MsiFormer's performance against three other deep learning methods.
• Analyzed CIELab color space for tetrazolium chloride staining and normalized canonical discriminant analysis (nCDA).

Main Results:

• MsiFormer achieved 94.17% accuracy, outperforming other methods by 2.5%-18.34%.
• CIELab L* and A* values from TTC staining showed strong correlations with vigour indices.
• Significant vigour variations were observed among rice varieties, highlighting sampling risks.
• Wavelengths between 780 nm-970 nm showed a strong relationship with vigour variations.

Conclusions:

• MsiFormer offers a highly accurate and efficient non-destructive method for rice seed vigour assessment.
• Multispectral imaging and deep learning integration provide a robust solution for seed quality evaluation.
• Further research into specific wavelengths and image analysis techniques can refine seed vigour prediction.