Enhancing Visible and Near-Infrared Hyperspectral Imaging Prediction of TVB-N Level for Fish Fillet Freshness Evaluation by Filtering Optimal Variables

Abstract

Total volatile basic nitrogen (TVB-N) is one of the most important indicators for evaluation of fish protein degradation and freshness loss. A novel algorithm of Physarum network (PN) combined with genetic algorithm (GA) was developed to select optimal wavelengths from hyperspectral images for enhancing the TVB-N level prediction in grass carp fish fillet. Partial least squares regression (PLSR) and least squares support vector machines (LS-SVM) calibration models were built using six optimal wavelengths selected by the PN-GA method and the PN-GA-PLSR model showed better performance for predicting the TVB-N value with determination coefficient in prediction (R 2 P ) of 0.956 and root mean square errors in prediction (RMSEP) of 1.737 mg N/100 g. The PN-GA-PLSR model established using the optimal wavelengths and image texture variables extracted by gray-level gradient co-occurrence matrix (GLGCM) algorithm showed higher R 2 P of 0.981 and lower RMSEP of 1.435 mg N/100 g. The results indicated that the PN-GA method was a good technique for selecting optimal wavelengths for enhancing prediction ability of hyperspectral imaging, which also demonstrated the efficiency and usefulness of this method for monitoring the freshness degree during fish cold storage.