Seed Protein Content Estimation with Bench-Top Hyperspectral Imaging and Attentive Convolutional Neural Network...

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Wheat seed protein concentration is a critical quality trait for global food security and animal feed.
• Accurate and rapid methods for determining wheat protein content are essential for breeding programs and quality control.
• Traditional methods for protein analysis are often time-consuming and destructive.

Purpose of the Study:

• To develop and evaluate robust hyperspectral imaging methods for predicting wheat seed protein concentration.
• To compare the performance of convolutional neural networks (CNNs) with traditional machine learning models for protein estimation.
• To assess the utility of CNN classification for categorizing wheat protein levels for breeding applications.

Main Methods:

• Bench-top hyperspectral imaging in the visible, near-infrared (VNIR), and shortwave infrared (SWIR) regions was employed.
• Computer-vision-aided image co-registration was used to align VNIR and SWIR spectral data.
• CNNs with attention mechanisms, Random Forest (RF), and Support Vector Machine (SVM) regression models were utilized for analysis.

Main Results:

• The CNN with attention mechanisms achieved R² values of 0.70 (ventral) and 0.65 (dorsal) for protein content prediction.
• The RF model outperformed CNNs for direct protein content prediction, reaching an R² of 0.77.
• CNN classification effectively distinguished low, medium, and high protein concentrations with an R² of 0.82.

Conclusions:

• Hyperspectral imaging combined with machine learning offers a powerful, non-destructive approach for wheat protein analysis.
• CNNs demonstrate significant potential for automating wheat protein estimation, particularly for classification tasks in breeding.
• These techniques can advance precision breeding, optimize seed sorting, and guide targeted agricultural inputs.