intelligent systems
G. Gergov, J. Cruz, E. Kirilova. A Comparative Evaluation of the Predictive Ability of PLS and RBF ANN Calibration Techniques Applied to SW-NIR Meat Data

Key Words: NIR spectroscopy; Partial Least Square method; Radial Basis Functions Neural Networks; moisture content; fat content; pork meat samples.

Abstract. In this study the performance of linear and nonlinear chemometric methods has been investigated and compared. The transmittance spectra of pork meat samples were collected by SW-NIR (short wave near infrared) analyser in the spectral range of 850 nm to 1,050 nm. Partial Least Square (PLS1) method and Radial Basis Functions Neural Networks (RBF NN) were chosen for chemometric analysis of that samples for determination of moisture and fat content. The reason for using RBF ANN is significant nonlinearity which is exhibited between the spectra and the fat and moisture content. PLS1 and RBF NN with different architecture have been combined with different pre-processing techniques such as first derivative (D1), standard normal variate (SNV), multiplicative signal correction (MSC) and the combinations of MSC and SNV with first derivative. It was found that optimal pre-processing was MSC for moisture, and the combination of D1and SNV for fat. When PLS1 was used, results showed reduction of RMSEP and REP using MSC with 15 and 13% for moisture determination. In case of PLS1 fat determination considerable reduction of RMSEP and REP was observed using a combination of D1 and SNV with 48 and 47%. Compared to PLS1 regression with suitable preprocessing, RBF ANN showed better results: reduction of RMSEP and REP using a combination of D1 and SNV with 48% for moisture, and reduction using a combination of D1 and SNV with 59% for fat determination. These improvements together with the facility of SW-NIR technology to be implemented in the process engineering made it ideal for the meat industry.