Performance Evaluation of PLS-Based FT-NIR Calibration Models for Milk Powder Proximate Analysis

Rapid and reliable determination of the proximate composition of milk powder is essential for quality assurance, nutritional labeling, and process control in the dairy industry. Conventional analytical methods, although highly accurate, are labor-intensive, time-consuming, and require extensive sample preparation and chemical reagents. This study aimed to develop and evaluate Fourier Transform Near-Infrared (FT-NIR) calibration models based on Partial Least Squares (PLS) regression for the rapid determination of moisture, protein, fat, and ash contents in milk powder according to the recommendations of ISO 21543|IDF 201:2020. A total of 304–912 representative milk powder samples were analyzed using standard reference methods to generate calibration data, followed by FT-NIR spectral acquisition using a Bruker TANGO spectrometer. Various spectral preprocessing techniques were evaluated, and the optimal models were validated through internal cross-validation using the coefficients of determination (R²), root mean square error of estimation (RMSEE), root mean square error of cross-validation (RMSECV), bias, and residual predictive deviation (RPD). The combination of Standard Normal Variate (SNV) and first-derivative preprocessing produced the best calibration performance for all analytes. The developed models achieved calibration R² values ranging from 0.9245 to 0.9949 and cross-validation R² values between 0.8996 and 0.9933, with low prediction errors and RPD values exceeding 3. These findings demonstrated that PLS-based FT-NIR calibration models provided accurate, rapid, and non-destructive proximate analysis, supporting their application as reliable tools for routine quality control in dairy laboratories.