Transfer method among water content detection models for different breeds of pork by hyperspectral imaging technique

Abstract: At present, most studies on model transfer were based on different spectrometers, and models were established using the near infrared spectroscopy. In this study, a hyperspectral detection model of water content of fresh pork was established by partial least squares regression (PLSR) method. In order to enhance model prediction applicability to different breeds of pork samples, a new model transfer method, piecewise direct standardization combined with linear interpolation (PDS-LI) was processed. In this method, the spectra of slave breed were corrected according to the spectra difference between master breed and slave breed, and then the corrected spectra of slave breed were predicted by master model. A function based on the prediction and reference values of slave breed samples was established. This function would be used to correct the prediction values of unknown test samples of slave breed. The specific steps were as followed: 1) Samples of master breed were divided into the calibration set and the test set, and the master model was built based on the calibration set by PLSR method. 2) Samples of slave breed were partitioned into standard sample selection set C2, standard sample set C2std and unknown test set T2un, and C2 was used for the selection of C2std and T2un was used to verify the transferred model. 3) Transfer matrix F was calculated by PDS algorithm according to the spectra difference between calibration set in master breed and C2std in slave breed, and then C2std and T2un were respectively corrected by transfer matrix F. 4) In order to improve the prediction accuracy of master model to the corrected spectrum of slave breed, the prediction value of T2un need to be corrected. For sample i in unknown test set T2un, symbiosis distance D(i) between sample i and every sample else in standard sample set C2std was calculated successively. D(i) was the sum of Euclidean distances between converted spectrum and absolute deviation of the prediction values. Two minimum values of D(i) were selected, so the prediction value of sample i in T2un could be corrected by the prediction and reference values of the 2 minimum samples. Three breeds, Duchangda, Maojia and Linghao pork were researched in this paper. As master breed, Duchangda samples were used to build the master model, and Maojia and Linghao were considered to be slave breeds to test the feasibility of model transfer algorithm. Equations with predictive determination coefficient (R2p) no less than 0.7 and residual prediction deviation (RPD) no less than 1.9 were considered to be applicable to predict pork quality. Model prediction results showed that for Duchangda samples, the coefficient of determination in cross-validation (R2cv) was 0.884, R2p was 0.883, root mean squared error of cross validation (RMSECV) was 0.279%, root mean squared error of prediction (RMSEP) was 0.237%, and RPD was 2.911, but for Maojia and Linghao samples, the prediction results were very poor: R2p only reached to 0.263 and 0.507, RMSEP, 1.151% and 0.857%, RPD, 1.000 and 1.214, respectively. With PDS-LI transfer method, the model prediction accuracies were substantially increased: R2p increased to 0.832 and 0.848, RMSEP decreased to 0.470% and 0.440%, RPD improved to 2.447 and 2.364, respectively, which indicated that PDS-LI transfer algorithm can achieve the model prediction transfer from Duchangda to Maojia and Linghao pork samples.