Abstract: Thirty-three samples of typical cow feed concentrate supplement without melamine-formaldehyde resin (MFR) and 99 ones adulterated with different concentrations of MFR were subjected to full wavelength near-infrared transmission and reflectance spectroscopic scanning. A discrimination model was established by BP (back propagation) neural network or PLS-LDA (partial least squares-linear discriminant analysis). The BP neural network model showed 100% accurate prediction. The minimum cross validation error rate of the PLS-LDA model was 0.0778, misclassification rate 0.0667, and prediction error rate 0.1429. Therefore, it is feasible to establish a predictive model for detecting MFR adulteration in cow feed by means of near-infrared transmission and reflectance spectroscopy.

Key words: cow feed, melamine-formaldehyde resin, near infrared spectroscopy, back propagation(BP) neural network, partial least squares-linear discriminant analysis (PLS-LDA)