Premium
Combination of spectra and image information of hyperspectral imaging data for fast prediction of lipid oxidation attributes in pork meat
Author(s) -
Aheto Joshua H.,
Huang Xingyi,
Tian Xiaoyu,
Ren Yi,
Bonah Ernest,
Alenyorege Evans A.,
Lv Riqin,
Dai Chunxia
Publication year - 2019
Publication title -
journal of food process engineering
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.507
H-Index - 45
eISSN - 1745-4530
pISSN - 0145-8876
DOI - 10.1111/jfpe.13225
Subject(s) - hyperspectral imaging , mean squared error , tbars , pattern recognition (psychology) , biological system , residual , sensor fusion , artificial intelligence , partial least squares regression , mathematics , root mean square , coefficient of determination , lipid oxidation , texture (cosmology) , statistics , computer science , chemistry , image (mathematics) , algorithm , biology , physics , biochemistry , antioxidant , oxidative stress , quantum mechanics , lipid peroxidation
This study highlighted the potential of integrating texture and spectra information of hyperspectral imaging to predict the evolution of 2‐thiobarbituric acid reactive substances (TBARS) and peroxide value (PV) as lipid oxidation attributes in pork muscles under diverse processing conditions. Partial least square regression (PLSR) models were established on data fusion information (image and optimal spectra) and compared with other three PLSR models built on spectra (full range and optimal), and image information. Regarding TBARS measurement, model based on optimal spectra was superior with determination coefficient of prediction ( R p 2 ) of .896, root mean square error of prediction (RMSEP) of 1.034, and residual predictive deviation (RPD) of 2.311 as compared to the data fusion model ( R p 2 = .865, RMSEP = 0.994, RPD = 2.280). For PV, model on data fusion yielded the best ( R p 2 = .899, RMSEP = 0.966, RPD = 2.314) followed by that on optimal spectra ( R p 2 = .883, RMSEP = 1.073, RPD = 2.281). Then prediction maps were drawn to visualize the distribution of TBARS and PV in the samples using models built on optimal spectra. The maps revealed intrinsic differences in samples which are not readily perceptible by visual observation. Practical Applications Models based on spectra and spatial information of hyperspectral imaging have been widely used to predict and visualize food quality attributes. However, these are not enough to elucidate the impact of processing on the meat samples as it may not be able to directly measure attributes related to external properties such as texture properties. The present study investigated the feasibility of integrating texture features obtained from gray‐level co‐occurrence matrix and spectra data to predict quality changes associated with pork meat subjected to diverse processing conditions. The results obtained proved that the methods presented can be useful for designing future experiments in a bid to establishing robust models for food quality assessment.