Identification of Maize Kernel Varieties Using LF-NMR Combined with Image Data: An Explainable Approach Based on...

Area of Science:

• Agricultural Science
• Biotechnology
• Data Science

Background:

• Accurate maize variety identification is crucial for germplasm management, genetic diversity, and agricultural optimization.
• Current methods for maize variety identification may lack speed, non-destructiveness, or precision.

Purpose of the Study:

• To develop a rapid, non-destructive, and interpretable machine learning model for precise maize kernel variety identification.
• To integrate low-field nuclear magnetic resonance (LF-NMR) data with morphological image features for enhanced classification.

Main Methods:

• Collected LF-NMR signals and morphological image features from eleven maize kernel varieties.
• Employed recursive feature elimination (RFE) and principal component analysis (PCA) for feature selection and data distribution analysis.
• Developed an optimized support vector machine (SVM) model using an improved differential evolution algorithm for hyperparameter tuning.

Main Results:

• Achieved a final classification accuracy of 96.36% for maize variety identification.
• Demonstrated strong robustness and precision in the developed classification model.
• Utilized Shapley values to reveal the interpretability and key feature contributions (e.g., Max Signal) to classification.

Conclusions:

• The integrated LF-NMR and image-based machine learning approach offers an innovative solution for efficient maize variety identification.
• This method supports refined germplasm resource management and provides a foundation for genetic improvement.
• The interpretable nature of the model enhances trust and applicability in agricultural settings.