Determination and Visualization of Fat Contents in Salmon Fillets Based on Visible and Near-Infrared Hyperspectral Imagery
Zhu Fengle,Peng Jiyu,Gao Junfeng,Zhao Yanru,Yu Keqiang,He Yong
DOI: https://doi.org/10.3969/j.issn.1002-6819.2014.23.040
2014-01-01
Abstract:As an important quality criterion in Atlantic salmon (Salmo salar), fat distributes heterogeneously throughout the whole salmon fillets. In this study, visible and near-infrared hyperspectral imaging was employed to determine the spatial distribution of fat in salmon fillets non-invasively and rapidly. One hundred cubed samples were cut out from different locations of different whole fillets to maximize fat content variation. After acquiring hyperspectral images with two systems operated in visible and short-wave near-infrared (Vis/SWNIR, 400-1100 nm) and near-infrared (NIR, 900-1700 nm) ranges, salmon samples were subjected to standard chemical analysis to measure their reference fat contents. Region of interests (ROI) was identified to isolate fish from the background in hyperspectral images, and the averaged reflectance spectra of the ROI image were extracted for all samples. Due to the low signal-to-noise ratio in the starting and ending spectral regions of both systems, only 740 Vis/SWNIR bands (459-1056 nm) and 151 NIR bands (947-1666 nm) were applied. The total samples were randomly divided into calibration set of 65 samples and prediction set of 35 samples. Then the extracted spectral variables with full wavelengths in two spectral sets for calibration samples were correlated with their corresponding reference fat contents using partial least squares regression (PLSR) and least-squares support vector machines (LS-SVM), respectively. And fat values for samples in the prediction set were predicted using the established models. Though good prediction results were achieved using full wavelengths, some redundant information exist among contiguous wavelengths due to the high degree of dimensionality in hyperspectral images. Competitive adaptive reweighted sampling (CARS) algorithm was employed to select effective wavelengths (EWs) for two spectral sets. Sixteen EWs of 468, 479, 728, 734, 785, 822, 863, 890, 895, 899, 920, 978, 1005, 1033, 1040, 1051 nm were selected in Vis/SWNIR spectral region, and fifteen wavelengths of 975, 995, 1023, 1047, 1095, 1124, 1167, 1210, 1273, 1316, 1354, 1368, 1575, 1632, 1661 nm were selected in NIR region. Then calibration models of EWs-PLSR and EWs-LS-SVM were established on the basis of the selected EWs of two spectral sets respectively. Improved performances for fat determination were observed for EWs-based models compared with full-spectrum models while the computational cost reduced greatly. And the linear EWs-PLSR model with NIR spectra was identified as the optimal model for fat prediction with determination coefficient of prediction (R2p) of 0.92, root mean square error of prediction (RMSEP) of 0.92%, and residual predictive deviation (RPD) for prediction of 3.50. Finally, the EWs-PLSR model was transferred to all pixels in the NIR hyperspectral images to predict their fat values for visualizing fat distribution in salmon samples. Fat distribution images for two whole salmon fillets were also generated to further explore the feasibility of hyperspectral imaging combined with the optimal model for fat visualization in whole fillets. Images like these could not only enable producers to perform proper sorting and cutting based on certain concentration thresholds but also benefit consumers to make the most appropriate decision when choosing salmon products. The overall results indicated that hyperspectral imaging coupled with chemometrics have potential for the determination and visualization of fat in salmon fillets.