Online milk composition analysis with an on-farm near-infrared sensor

On-farm monitoring of milk composition can support close control of the udder and metabolic health of individual dairy cows. In previous studies, near-infrared (NIR) spectroscopy applied to milk analysis has proven useful for predicting the main components of raw milk (fat, protein, and lactose). In this contribution, we present and evaluate a precise tool for online milk composition analysis on the farm. For each milking, the online analyzer automatically collects and analyses a representative milk sample. The system acquires the NIR transmission spectra of the milk samples in the wavelength range from 960 to 1690 nm and performs a milk composition prediction afterward. Over a testing period of 8 weeks, the sensor collected 1165 NIR transmittance spectra of raw milk samples originating from 36 cows for which reference values were obtained for fat, protein, and lactose. For the same online sensor system, two calibration scenarios were evaluated: training post-hoc prediction models based on a representative set of calibration samples (n = 319) acquired over the entire testing period, with different cows in the calibration and test set, and training real-time prediction models exclusively on the samples acquired in the first week of the testing period (n = 308). The obtained prediction models were thoroughly tested on all the remaining samples not included in the calibration sets (n respectively 846 and 857). For the post-hoc prediction models, this resulted in an overall prediction error (root-mean-squared error of prediction, RMSEP) smaller than 0.080% (all % are in wt/wt) for milk fat (range 1.5–6.3%), protein (2.6–4.3%) and lactose (4–5.1%), with a coefficient of determination R2 of 0.989, 0.947 and 0.689 for fat, protein, and lactose respectively. For the real-time prediction models, the RMSEP was smaller than 0.092% for milk fat and lactose, and 0.110% for protein, with an R2 of 0.989 (fat), 0.894 (protein) and 0.644 (lactose). The milk lactose predictions could be further improved (RMSEP = 0.088%, R2 = 0.675) by taking into account a cow-specific bias. The presented online sensor system using the real-time prediction approach can thus be used for detailed and autonomous on-farm monitoring of milk composition after each individual milking, as its accuracy is well within the requirements by the International Committee for Animal Recording (ICAR) for on-farm milk analyzers and even meet the standards for laboratory analysis systems for fat and lactose. For this real-time prediction approach, a drift was observed in the predictions, especially for protein. Therefore, further research on the development of online calibration maintenance techniques is required to correct for this model drift and further improve the performance of this sensor system.