Hyperspectral imaging combined with convolutional neural network for accurately detecting adulteration in Atlantic salmon

Fraud frequently occurs in Atlantic salmon market and is difficult or impossible to detect through visual inspection. This study was implemented to investigate the potential of two hyperspectral imaging (HSI) systems covering the visible and near infrared range (VNIR, 397–1003 nm) and the short-wave near infrared range (SWIR, 935–1720 nm), respectively, for rapidly and accurately determining adulteration in minced Atlantic salmon. The adulterated samples containing 11 adulteration levels (0–100% (w/w) at 10% intervals) were prepared manually. Four spectral preprocessing methods and five characteristic wavelength selection algorithms were employed to combine convolutional neural network (CNN) to establish quantitative models for predicting adulteration levels. The predictive ability of the two HSI systems was compared to reveal the optimal spectral detection range. After analysis, it was found that the modeling results using VNIR data were always better than those using SWIR data. In particular, the best prediction for VNIR was from the combination model SNV-CNN with the mean RP2, RMSEP and RPD of 0.9885, 3.3526 and 9.6882, respectively. The best performance for SWIR was from the combination model SNV-VCPA-IRIV-CNN with the mean RP2, RMSEP and RPD of 0.9839, 3.9926 and 8.0251, respectively. Further, the best models were successfully used to visualize the distribution of adulterant in prepared samples. Overall, this study demonstrated that HSI combined with CNN is a promising solution for the rapid, nondestructive and accurate detection of adulteration in Atlantic salmon. In addition, VNIR-HSI was considered to be more reasonable detection range due to its low cost and better prediction compared to SWIR-HSI.