Determination of Total Viable Count in Rainbow-Trout Fish Fillets Based on Hyperspectral Imaging System and Different Variable Selection and Extraction of Reference Data Methods

Abstract

The aims of this study were to investigate the effect of different reference data extraction (colony-counting) and selection variable methods (Regression Coefficient: RC; Forward and Stepwise Multiple Regression: FMR and SMR) on the performance of PLSR and MLR model to predict TVC value in rainbow-trout fish fillets. TVC values were measured based on manual and digital image (OpenCFU, IMJ, and Photoshop) counting methods. The most and lowest prediction powers were obtained for Photoshop-PLSR and OpenCFU-PLSR, respectively (R2p = 0.873 and 0.815; RMSEP = 0.761 and 0.884 Log10CFU/g). In simplified-model FMR-MLR has superior performance (R2p = 0.89 and RMSEP = 0.65 Log10CFU/g). In simplified PLSR model group, RC-PLSR showed better performance (R2p = 0.866 and RSMEP = 0.782). This distribution map of TVC load was generated by transferring the FMR-Photoshop-MLR model to each pixel of the images. HSI technique revealed a great potential to determine TVC of rainbow-trout fillets and the type of colony counting method influenced on prediction power of the model.