Skin hyperspectral imaging and machine learning to accurately predict the muscular poly-unsaturated fatty acids contents in fish.

Abstract

The polyunsaturated fatty acids (PUFAs), particularly eicosapentaenoic acid (EPA) and docosahexaenoic acid (DHA), are critical determinants of the nutritional quality of fish. To rapidly and non-destructively determine the muscular PUFAs in living fish, an accuracy technique is urgently needed. In this study, we combined skin hyperspectral imaging (HSI) and machine learning (ML) methods to assess the muscular PUFAs contents of common carp. Hyperspectral images of the live fish skin were acquired in the 400-1000 nm spectral range. The spectral data were preprocessed using Savitzky-Golay (SG), multivariate scattering correction (MSC), and standard normal variable (SNV) methods, respectively. The competitive adaptive reweighted sampling (CARS) method was applied to extract the optimal wavelengths. With the skin spectra of fish, five ML methods, including the extreme learning machine (ELM), random forest (RF), radial basis function (RBF), back propagation (BP), and least squares support vector machine (LS-SVM) methods, were used to predict the PUFAs and EPA + DHA contents. With the spectral data processed with the SG, the RBF model achieved outstanding performance in predicting the EPA + DHA and PUFAs contents, yielding coefficients of determination (R² _P) of 0.9914 and 0.9914, root mean square error (RMSE) of 0.3352 and 0.3346, and mean absolute error (MAE) of 0.2659 and 0.2660, respectively. Finally, the visualization distribution maps under the optimal model would facilitate the direct determination of the fillet PUFAs and EPA + DHA contents. The combination of skin HSI and the optimal ML method would be promising to rapidly select living fish having high muscular PUFAs contents.