Selecting Milk Spectra to Develop Equations to Predict Milk Technological Traits

Area of Science:

• Chemometrics
• Spectroscopy
• Dairy Science

Background:

• Predicting milk technological traits from mid-infrared (MIR) spectra is crucial for dairy quality control.
• Global modeling approaches using all available spectral data may be suboptimal for spectra with poor representation.
• Outlier spectra can significantly impact the accuracy of predictive models.

Purpose of the Study:

• To develop and evaluate an alternative local changepoint approach for predicting six milk technological traits using MIR spectra.
• To compare the performance of the local changepoint approach against global partial least square regression (PLSR) and a fixed-neighbor LOCAL approach.
• To assess the impact of objective neighbor selection on prediction accuracy, particularly for outlier spectra.

Main Methods:

• A local changepoint approach was developed to predict milk technological traits from MIR spectra.
• Objective identification of neighboring spectra was performed using Mahalanobis distances between spectral principal components.
• Partial least square regression (PLSR) was employed, with comparisons made to global PLSR and a fixed-neighbor LOCAL approach.

Main Results:

• Global PLSR achieved the lowest root mean square error of cross-validation (RMSEV) for five out of six traits.
• The local changepoint approach outperformed the LOCAL approach for four traits and showed the lowest RMSEV for one trait.
• When analyzing outlier spectra (top 5% Mahalanobis distance), the local changepoint approach outperformed global PLSR and the LOCAL approach for two and five traits, respectively.

Conclusions:

• Objective neighbor selection in the local changepoint approach improved prediction performance over fixed neighbor selection.
• While generally not outperforming global PLSR, the local changepoint approach demonstrated advantages in handling spectral variations and outliers.
• The local changepoint approach offers a valuable alternative for milk trait prediction from MIR spectra, especially when dealing with heterogeneous spectral datasets.