Rapid prediction of milk powder quality by hyperspectral imaging

Abstract

Economically motivated rapid quality testing and assurance of milk powder is a well-known issue. This is a complex problem in the dairy industry that needs to be tackled with a range of techniques including rapid quality testing using sophisticated process analysers such as near-infrared spectroscopy (NIR) and hyperspectral imaging (HSI). Instant whole milk powder (IWMP) is custom-made to exhibit good rehydration properties. Dispersibility and bulk density are among the important parameters to evaluate IWMP functional performance in dissolution, and potentially affected by variation in particle size distribution. Therefore, commercial-grade IWMP was divided into three discrete fractions of varying particle size. These three fractions were coarse (“C”, having particles of 355μm or higher); medium (“M”, having particles smaller than 355μm and larger than 180μm); and fine (“F”, having particles smaller than 180μm). These milk powder fractions were remixed with specific known ratios to prepare seven types of reconstructed milk powder samples that had an extended range of dispersibility, bulk density and particle size distribution from the commercial-grade IWMP to be used in this research. In this research, NIR spectrum (within 1000 – 2500 nm wavelength range) of reconstructed samples of IWMP was correlated to its dispersibility, bulk density and percent of fine particles with partial least squares (PLS) models. Different spectral data pre-treatment approaches such as Savitzky-Golay (SG) filtering, standard normal variate (SNV) transformation, multiplicative scatter correction (MSC), and their combinations were applied. Results showed that fine particle fraction was predicted with a coefficient iii of prediction of 0.703 for spectral smoothing by SG. For dispersibility prediction, the combination of SG & MSC gave the highest prediction correlation coefficient of 0.973 and for bulk density prediction SG treated spectra proved to be the suitable data treatment method with a prediction correlation coefficient of 0.978. Hyperspectral imaging (HSI) is an emerging technique in food quality analysis. Hyperspectral images over the 400 – 1000 nm wavelength range were recorded for the individual particle size fraction samples and reconstructed IWMP samples. Principal component analysis (PCA) discriminated HSI data of coarse, medium and fine particle fractions successfully across the first two principal components (PC1 and PC2). The partial least squares (PLS) discriminant analysis on full wavelengths successfully classified the three fractions of milk powder with a coefficient of prediction of 0.943. However, for efficient data analysis, five wavelengths were selected by the loading plot of three principal components (PC1, PC2 and PC3) from PCA and eleven wavelengths were selected from weighted regression analysis (WRC) of the HSI data of three particle size fractions. Simplified models on reduced wavelengths gave better classification results in less computation time. Average spectra of each particle size fraction were used to calculate the spectral similarity of pixels of hyperspectral images of reconstructed IWMP samples. Furthermore, based upon the similarity to coarse, medium, or fine particle fraction, their distribution in the reconstructed samples of IWMP was predicted. Partial least square (PLS) model performed well to predict the bulk density and dispersibility of reconstructed IWMP samples.