Data-Driven Modeling of NIR Sensor Signals: A Comparative Study of PCR, PLSR, and Regularized Regression for Milk Composition Prediction

Abstract

The present study aims at exploring the modeling of high-dimensional near infrared (NIR) sensor signals in predicting important milk composition parameters. NIR spectroscopy is a technique that produces highly multicollinear spectral data, which is a great challenge in modeling regression problems. In this study, we have tried different regression modeling approaches such as multiple regression, linear regression, Ridge regression, Lasso regression, Elastic Net regression, Principal Component Regression (PCR), and Partial Least Squares Regression (PLSR) in predicting important milk composition parameters such as fat, protein, lactose, urea, milk yield, and somatic cell count (SCC). The NIR sensor signals with over 1000 features are analyzed using 5-fold cross-validation and root mean squared error (RMSE) metrics. From the analysis, it is observed that using dimensionality reduction approaches such as PCR is more accurate in predicting most of the important milk composition parameters. In addition, other regression approaches such as PLSR and Ridge regression are also accurate in predicting the important milk composition parameters. This study is important in understanding the selection of regression modeling approaches in NIR sensor applications.