Hyperspectral Inversion Modeling of Fat and Protein Content in Milk Based on XGBoost

A hyperspectral inversion modeling method is proposed to predict the fat and protein content in milk by integrating SG, CR and XGBoost to address the drawbacks of traditional chemical methods for detecting fat and protein in milk, which require damage to the emulsion itself, complicated operation and environmental pollution. The method first combines Savitzky-Golay (SG) and Continuum-Removal (CR) to pre-process the hyperspectral data, and then uses Correlation-Analysis (CA) to screen the feature bands. Finally, the screened 10 feature bands are used as inputs to build partial least squares regression (PLSR) and XGBoost models, and the model prediction results are compared. The results show that the SG_CR algorithm is an effective hyperspectral preprocessing method to improve modeling accuracy, and the screening of feature bands can greatly reduce model complexity and improve model efficiency while ensuring model accuracy. Compared with the linear model PLSR, the machine learning integration algorithm XGBoost performed optimally. Its fat and protein models have training set coefficients of determination (R2) of 0.9449 and 0.9980, training set root mean square error (RMSE) of 0.2395 (g/100g) and 0.0409 (g/100g), training set relative analysis error (RPD) of 3.8188 and 22.2079, test set R2 of 0.8653 and 0.9273, test set RMSE were 0.4672 (g/100g), 0.2116 (g/100g), and test set RPD were 2.4625, 3.5998.