Using VIS-NIR hyperspectral imaging and deep learning for non-destructive high-throughput quantification and...

Area of Science:

• Agricultural Science
• Analytical Chemistry
• Computer Science

Background:

• Traditional methods for quantifying crop grain nutrients are slow and destructive.
• There is a need for high-throughput, low-cost, non-destructive methods for nutrient analysis.
• Accurate nutrient quantification is vital for food processing and nutritional research.

Purpose of the Study:

• To develop a high-throughput, low-cost method for quantifying wheat grain nutrients using visible-near infrared (VIS-NIR) hyperspectral imaging.
• To accurately predict nutrient content and visualize nutrient distribution in wheat grains.
• To leverage deep learning for enhanced nutrient visualization.

Main Methods:

• Utilized VIS-NIR (400-1700 nm) hyperspectral imaging for wheat grain analysis.
• Employed stepwise linear regression (SLR) for nutrient quantification, with data processed by first derivative for improved accuracy.
• Developed an improved pix2pix conditional generative network for nutrient distribution visualization.

Main Results:

• Stepwise linear regression accurately predicted hundreds of nutrients (R² > 0.6).
• First derivative processing of hyperspectral data enhanced prediction accuracy.
• Characteristic wavelengths for various nutrients were identified, primarily in the 400-500 nm and 900-1000 nm regions.
• The improved pix2pix model demonstrated superior performance in visualizing nutrient distribution compared to the original model.

Conclusions:

• VIS-NIR hyperspectral imaging offers a viable high-throughput, non-destructive method for wheat nutrient determination.
• Deep learning models, particularly the improved pix2pix network, significantly enhance the visualization of nutrient distribution.
• This approach holds substantial potential for applications in food processing, quality control, and nutritional research.