Near-Infrared Spectroscopy Combined with Explainable Machine Learning for Storage Time Prediction of Frozen Antarctic Krill

Antarctic krill (<i>Euphausia superba</i>) represents a promising sustainable protein source for human consumption. While a portion of the catch undergoes immediate onboard processing, the majority is preserved as frozen raw material, with storage duration significantly impacting product quality and safety. This study established a novel approach for rapid quality assessment through storage time prediction. Traditional chemical quality indicators of krill during a 12-month storage were first monitored and the correlation between the quality and storage time was verified. Coupled with four different regression machine learning algorithms, near-infrared spectroscopy (NIRS) was applied to develop models. Following optimal spectral preprocessing selection and hyperparameters optimization, the light gradient boosting machine (LightGBM) model yielded the best storage time prediction performance, with the R² of the test set being 0.9882 and the errors RMSE, MAE, and MAPE being 0.3724, 0.2018, and 0.0431, respectively. Subsequent model interpretation results revealed a strong correspondence between model-related NIR features and chemical indicators associated with quality changes during krill frozen storage, which further justified the model's predictive capability. The results proved that NIR spectroscopy combined with LightGBM could be used as a rapid and effective technique for the quality evaluation of frozen Antarctic krill, offering substantial potential for industrial implementation.