The effect of multiplicative scatter correction (MSC) on wavelength selection for near infrared (NIR) calibration to determine fat content in raw milk was investigated. Short-wave NIR spectra (700-1100 nm) of raw milk samples were measured. The calibration equations for fat content were performed by multiple linear regression (MLR) using original, second derivative and MSC-treated spectra. It was found that first wavelength selection from the fat absorption band for a calibration equation was generally effective in all cases of original, second derivative and MSC-treated spectra. However, correlation plots did not always work well because of the multiplicative scatter effect presented in the samples. Whereas, correlation plots were still useful in the case of MSC-treated spectra and normalised second derivative spectra, even when the original spectra exhibited a multiplicative scatter effect.

A simple near infrared (NIR) transmittance measuring method for unhomogenised raw milk was developed using test tubes as sample cells and a specially designed sample holder which could control sample temperature. Shortwave NIR spectra (700-1100 nm) of 154 unhomogenised milk samples from eight cows, under different nutritional conditions, were measured at 40°C using this method. The calibration and validation for fat content in raw milk were performed by partial least squares regression in two ways, using either a specific test tube or any individual test tube as a sample cell. A functional calibration model for determining fat content in raw milk with relatively high accuracy (multiple correlation coefficient of 0.99 and SEP of 0.19%) was obtained, which is applicable even when using individual test tubes as sample cells. By examining the loading weight of each factor included in the calibration model for fat determination in milk, it was found that the model consisted of, not only the factor for fat, but also factors which worked as adjusting terms, eliminating any effects caused by the other major components of milk such as lactose, casein and water and by the sample cell.

A new measurement unit, the MilkSpec-1, has been developed to determine rapidly and nondestructively the content of fat, lactose, and protein in raw milk using near-infrared transmittance spectroscopy. The spectral range over 700 to 1100 nm was used. This unit was designed for general glass test tubes, 12 mm in diameter and 10 mL in volume. Al2O3 with a thickness of 2.5 mm was found to be optimum as a reference for acquiring the milk spectrum for this measurement. The NIR transmittance spectra of milk were acquired from raw milk samples without homogenization. The calibration model was developed and predicted by using a partial least-squares (PLS) algorithm. In order to reduce the scattering effect due to fat globules and casein micelles in NIR transmittance spectra, multiplicative scatter correction (MSC) and/or second derivative treatment were performed. MSC treatment proved to be useful for the development of calibration models for fat and protein. This study resulted in low standard errors of prediction (SEP), with 0.06, 0.10, and 0.10% for fat, lactose, and protein, respectively. It is shown that accurate, rapid, and nondestructive determination of milk composition could be successfully performed by using the MilkSpec-1, presenting the potential use of this method for real-time on-line monitoring in a milking process.

The applicability and universality of a sample combination method to develop a calibration with sample temperature compensation [J. Near Infrared Spectrosc. 3, 211 (1995)] was investigated from mathematical and theoretical points of view. From a theoretical analysis it is clear that the development of the calibration will be successful when two conditions are satisfied. One condition is that the vectors whose elements are constituent values should be orthogonal to the vector whose elements are the dot product of spectral change and calibration coefficients. The other is that the inverse vectors of the spectrum related to chemical components should be orthogonal to the vector of spectral change. The first condition is always satisfied when we make a calibration following this method, but whether the second condition is satisfied or not is completely dependent on the regression algorithm. From a mathematical analysis, we concluded that simple multiple linear regression is not completely adequate to develop a calibration with temperature compensation.