Prediction of maize crude fat content based on improved conditional mutual information maximization and SHAP analysis

Area of Science:

• Agricultural Science
• Spectroscopy
• Chemometrics

Background:

• Traditional conditional mutual information maximization (CMIM) algorithms face limitations in analyzing nonlinear dependencies within continuous near-infrared (NIR) spectral data.
• Accurate and rapid prediction of crop quality parameters, such as crude fat content in maize, is crucial for agricultural management.

Purpose of the Study:

• To develop a novel framework integrating an improved CMIM with SHAP (SHapley Additive exPlanations) for efficient maize crude fat content prediction.
• To overcome the limitations of traditional CMIM in handling nonlinear relationships in NIR spectral analysis.

Main Methods:

• Implementation of a modified CMIM algorithm (CMIM_KDE) utilizing continuous variable estimates instead of discretization.
• Application of a two-stage filter-wrapper feature selection strategy.
• Integration with SHAP for interpretability of selected spectral features.
• Validation using Partial Least Squares Regression (PLSR) and Support Vector Regression (SVR) models.

Main Results:

• The CMIM_KDE method achieved high predictive accuracy, with R²p values of 0.7618 (PLSR) and 0.7531 (SVR).
• Demonstrated average performance improvements of 6.18% and 4.42% over other enhanced strategies.
• SHAP analysis identified key wavenumbers (around 5684 cm⁻¹ and 4312 cm⁻¹) linked to the CH group in maize, providing insights into the prediction mechanism.
• Validated the method's effectiveness and generalizability on external datasets.

Conclusions:

• The proposed CMIM_KDE-SHAP framework effectively addresses the limitations of traditional CMIM in NIR spectral analysis for predicting maize crude fat content.
• This approach offers a robust solution for quality detection in agricultural products, with potential applications in remote sensing and real-time monitoring.