Remaining shelf-life prediction of squid using handheld Vis/NIR spectroscopy and ensemble learning

To address the limitations of traditional squid freshness assessment, which relies on destructive physicochemical indicators and neglects anatomical spatial heterogeneity—thereby compromising on-site accuracy—this study developed an integrated handheld visible and near-infrared (Vis-NIR) spectroscopy system combined with intelligent algorithms. A multi-scale decomposition of spectral signals (450-950 nm) was innovatively performed using Discrete Wavelet Transform (DWT) to extract energy spectrum features. By integrating Isolation Forest for outlier rejection and a Stacked Ensemble Learning framework (Stacking: base=Ridge/RF/HGBR, meta=GBR), a robust prediction model for storage time was constructed. The experimental results demonstrated that the proposed model achieved an excellent coefficient of determination (R²=0.9528), a Root Mean Square Error (RMSE=0.6566 d), and a Residual Prediction Deviation (RPD=4.60) on the validation set, with uniformly distributed residuals cross-verified by the Breusch–Pagan test (p=0.6944). Weibull kinetic modeling revealed distinct degradation patterns across anatomical sites; specifically, the abdominal region exhibited a maximum shape parameter ( \beta =1.787) and a massive Akaike Information Criterion advantage (| \Delta AIC|=490.6) over the first-order kinetic paradigm, indicating an accelerating, self-catalytic quality degradation profile driven by endogenous visceral enzymes. Finally, the real-time predicted storage time was successfully coupled with local Weibull parameters to output probabilistic Remaining Shelf Life (RSL) profiles via Kernel Density Estimation (KDE), which establishes a proactive digital traceability pipeline for precision grading and dynamic inventory management of high-value marine aquatic products.