Predicting milk quality using portable NIR instruments

The suitability of employing two portable near-infrared (NIR) instruments operating within the ranges of 950-1550 nm (NIR#1) and 1350-2150 nm (NIR#2) was evaluated alongside chemometric techniques as a rapid method for determining milk quality. A total of 205 samples were sonicated and subsequently pre-heated to 40°C before being scanned twice in transmission mode using a cuvette with a path length of 1 mm. Spectra from 22 samples with Mahalanobis distance exceeding the third quartile were excluded. After averaging the replicate scans, spectra preprocessing was conducted to mitigate multiplicative effects caused by light scattering (Roger et al., 2022). The data sets were randomly divided into training (tr, 70%) and testing (ts, 30%) sets, and partial least squares (PLS) regression was performed on tr and validated on ts. The optimal number of principal components (PCs) for PLS was determined by minimising Root Mean Square Errors (RMSE) in cross-validation (cv) with a maximum of 5 PCs to prevent overfitting. The splitting of the data set and PLS algorithm was repeated 100 times, and average metrics, including coefficient of determination (R2), RMSE, and the ratio of performance to deviation (RPD), were reported for cv and validation (v). Milk samples were analysed for various parameters, including casein (%m/m), crude fat (%m/m), lactose (%m/m), crude protein (%m/m), and true protein (%m/m). The results indicated poor performance of both instruments, with RPDcv < 1.2 and RPDv < 1. The suboptimal performance is likely attributed to high absorbances (log 1/R), reaching values of 4 within the water bands, recorded with a noisy signal. Therefore, it can be hypothesised that the scanning method (cuvette path length, light intensity, integration time, or sample preparation) must be further optimized in order to be suitable for this specific application.