Multi-Path Attention Fusion Transformer for Spectral Learning in Corn Quality Assessment

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Modeling nonlinear relationships between near-infrared (NIR) spectral signatures and corn biochemical traits is challenging.
• Capturing multi-scale contextual dependencies in spectral data is crucial for predicting quality constituents like protein and oil.

Purpose of the Study:

• To develop an advanced deep learning model, SpecTran, for improved NIR spectral regression in corn quality assessment.
• To effectively capture both local absorption peaks and global spectral patterns for enhanced prediction accuracy.

Main Methods:

• Proposed SpecTran, a spectral Transformer network integrating adaptive multi-scale patch embedding.
• Implemented spectral-enhanced positional encoding to preserve wavelength order information.
• Utilized hierarchical feature fusion for robust multi-task prediction of corn traits.

Main Results:

• SpecTran achieved an average R2 of 0.483 across moisture, starch, oil, and protein traits on the Eigenvector corn dataset.
• Reduced Root Mean Square Error (RMSE) by 11.2% for protein and 10.7% for oil compared to the standard Transformer baseline.
• Demonstrated superior ability in modeling complex spectral dynamics for agricultural quality assessment.

Conclusions:

• SpecTran offers a reliable framework for NIR-based agricultural quality assessment by effectively modeling spectral data.
• The model provides interpretable insights, enhancing its utility for practical applications in corn quality analysis.