Characterization and prediction of phosphate responsiveness using hyperspectral imaging in snakehead fillets during water retention.

Phosphate additives are widely used to improve water-holding capacity in fish fillets, yet their effectiveness is strongly influenced by muscle structure, particularly after freeze-thaw treatments. In this study, the potential of hyperspectral imaging (HSI) was investigated for predicting phosphate dosage levels in snakehead fillets among different freeze-thaw cycles during the water retention stage. With shortwave infrared (SWIR) spectra and physicochemical properties as predictors, ordinal logistic regression (OLR) and light gradient boosting machine (LightGBM) models were developed to predict ordered responsiveness categories. LightGBM achieved the best performance for soaking weight gain (SWG) and centrifugal loss (CL) with prediction overall accuracies (OA) of 0.77 and 0.74 (RMSE of 0.45 and 0.51) respectively, while OLR showed greater stability for bound water (BW) and immobilized water (IW), reaching accuracies up to 0.79 and 0.74 with RMSE of 0.45 and 0.51. According to confusion matrices, most samples were correctly classified, especially at lower dosage 1levels. The pixel-wise prediction maps based on LightGBM further revealed pronounced intra-fillet heterogeneity associated with freeze-thaw cycles, highlighting spatially non-uniform phosphate responsiveness. This research demonstrates that SWIR hyperspectral imaging could be used to characterize and predict phosphate responsiveness in fish fillets, providing an insight of freeze-thaw history induced heterogeneity and its impact on water retention behavior for phosphate addition.