Quantitative analysis and visualization of live spotted seabass (Lateolabrax maculatus) flesh texture using hyperspectral imaging and machine learning

Flesh texture is one of the most important quality traits determining consumer satisfaction and perception of fish. Traditional methods for measuring flesh texture characteristics present several limitations. This study developed a rapid, efficient and non-invasive approach based on hyperspectral imaging (HSI) technology for evaluating texture profile analysis (TPA) parameters in live spotted seabass. Three types of surface hyperspectral data (skin with scales, SWS; skin without scales, SOS; and reversed skin without scales, RSOS) were collected from the dorsal region of 300 live fish across 400–1000 nm wavelength range. After applying five spectral pre-processing methods, five machine learning algorithms, including partial least squares regression (PLSR), least square support vector machine regression (LS-SVR), random forest (RF), convolutional neural network (CNN) and back propagation artificial neural network (BP-ANN), were performed to construct optimal prediction models for seven TPA parameters. Gumminess was predicted most effectively, with the RF model achieving a prediction set coefficient of determination (R 2 P ) of 0.923, a ratio of performance deviation (RPD) of a mean absolute percentage error (MAPE) of 6.378%. The optimal models enabled quantitative prediction of gumminess, chewiness, and hardness (RPD > 3.0 and MAPE