Using of spectral global H distance improves the accuracy of milk MIR based predictions


  • Zhang, L. , Li, C.F. , Dehareng, F. , Grelet, C. , Colinet, F. , Gengler, N. & Soyeurt, H. (2019). Using of spectral global H distance improves the accuracy of milk MIR based predictions. Proceedings in: ICAR Conference, Ghent, 17-21 Juin 2019,
Type Conference Proceedings
Year of conference 2019
Title Using of spectral global H distance improves the accuracy of milk MIR based predictions
Conference name ICAR Conference
Conference location Ghent
conference Date 17-21 Juin 2019
Abstract Milk MIR spectrometry predicts traits related to quality, health and environment. Standardizing MIR data enables to share prediction equations. The prediction accuracy depends partly on the calibration spectral variability. So, the calculation of spectral global H (GH) distance is of interest. The effect of GH on the accuracy of MIR predictions were studied using 198,394 milk samples collected from Chinese cows and analyzed on 3 Bentley FTS machines. The content of fat, protein, monounsaturated fatty acid (MFA), unsaturated FA (UFA), saturated FA (SFA), polyunsaturated FA (PFA) predicted by manufacturer’s models were the reference values. Fat, protein, MFA, PFA, SFA and UFA averages were 3.97, 3.43, 0.86, 0.07, 2.62 and 0.93 g/dL of milk. Bentley MIR spectra were standardized to master MIR spectra according to the method developed by the European Milk Recording network. Then the studied traits were predicted on those spectra using published MIR equations. The averaged predicted content of fat, protein, MFA, PFA, SFA, and UFA were 3.99, 3.53, 1.15, 0.15, 2.64, 1.29 g/dL of milk. GH ranged from 0 to 475. Correlation values between predicted and reference contents ranged from 0.92 to 0.98, except for PFA (0.59). Root mean square errors (RMSE) were 0.19, 0.18, 0.32, 0.09, 0.21, and 0.39 g/dL for fat, protein, MFA, PFA, SFA, and UFA. Correlation values between the squared residuals and GH were positive (0.17-0.42). This suggests a positive effect of GH on the prediction accuracy. When a threshold of GH?5 was applied, the data loss ranged from 3.83% to 9.27%. Correlation values between predicted and reference contents increased (0.94 to 0.98, 0.64 for PFA). RMSE decreased from 1.26% to 7.37%. Considering GH limits spectral extrapolation. To improve the accuracy, samples with GH > 5 must be included in the calibration set to cover the spectral variability. Moreover, the spectral standardization must be performed on a regular basis in order to check the potential spectral deviation of an instrument. In this study, the standardization was performed once.
Fichier
Authors Zhang, L., Li, C.F., Dehareng, F., Grelet, C., Colinet, F., Gengler, N., Soyeurt, H.

Team