Nondestructive Detection of Eggshell Thickness Using Near-Infrared Spectroscopy Based on GBDT Feature Selection and...

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Eggshell thickness is vital for egg breakage resistance and hatchability.
• Traditional measurement methods are destructive and inefficient, hindering quality assessment.

Purpose of the Study:

• To develop a robust, non-destructive prediction approach for eggshell thickness.
• To integrate Gradient Boosting Decision Tree (GBDT) feature optimization with an improved CatBoost algorithm for enhanced accuracy.

Main Methods:

• Applied Standard Normal Variate (SNV) and Multiplicative Scatter Correction (MSC) for spectral data preprocessing.
• Utilized GBDT for nonlinear feature selection, identifying optimal wavelengths.
• Developed an improved CatBoost regression model with Ordered Boosting and anti-overfitting strategies (10-fold nested cross-validation, Bootstrap resampling).

Main Results:

• Achieved high prediction accuracy with coefficients of determination (R²) of 0.930 (calibration) and 0.918 (prediction).
• Obtained a low root mean square error of prediction (RMSEP) of 0.008 mm.
• Demonstrated superior performance compared to traditional algorithms in prediction accuracy and generalization, with errors following a zero-mean Gaussian distribution.

Conclusions:

• The proposed GBDT-CatBoost approach offers a reliable, non-destructive method for assessing eggshell thickness.
• This research provides a strong theoretical and technical basis for intelligent poultry egg quality grading.
• The method effectively mitigates challenges like the curse of dimensionality and multicollinearity in spectral data.