|
|
|
| Comparative Study of Linear and Non-linear Quantitative Analysis Models for the Moisture Content in Pork |
| WANG Dong, LI Cheng, PING Hua, LIU Jing, WANG Beihong, LI Yang |
| Risk Assessment Laboratory for Agro-products (Beijing), Ministry of Agriculture and Rural Affairs, Institute of Quality Standard and Testing Technology, Beijing Academy of Agriculture and Forestry Sciences, Beijing 100097, China |
|
|
|
|
Abstract In order to compare the difference in accuracy between quantitative predictive models for the moisture content in pork developed by linear and non-linear algorithms, the linear variable filter near-infrared (LVFNIR) spectra of whole pieces of pork and minced pork were collected by a portable LVFNIR spectrometer, and partial least square regression (PLSR) as a linear algorithm and support vector regression (SVR) as a non-linear algorithm were used to develop quantitative predictive models for the moisture contents in the two meat samples. Optimization of wavebands and data preprocessing methods were conducted for each of the four models (whole pieces of pork PLSR model, minced pork PLSR model, whole pieces of pork SVR model, minced pork SVR model). The results showed that the SVR model for minced pork based on full-band spectral data with data centralization + smoothing + standard normal variate transformation preprocessing performed best among all the models tested, with punishment coefficient, kernel parameter, determination coefficient of calibration, root mean square error of calibration, determination coefficient of cross validation, root mean square error of validation and ratio performance deviation of 14 400, 0.020, 0.778 7, 1.04, 0.871 4, 0.89 and 2.69, respectively. The predictive results for the external validation set (external blind samples) demonstrated that the SVR model for minced pork had the best performance, with root mean square error of prediction and correlation coefficient of prediction of 1.24 and 0.821 0, respectively. This study indicated that the SVR models had higher predictive accuracy than did the PLSR models for both pork samples. However, the stability of the SVR models still needs further improvement.
|
|
|
|
|
|
|
|
|
|
|
2026, Vol. 40 
