Rapid quality detection in colored potatoes during alternating cold and ambient storage: Utilizing absorption and scattering properties.

The dynamic changes in sugar content during the cold-to-ambient storage stages of potato tubers pose a significant challenge to quality monitoring efforts. This study investigated the dynamic quality changes in three colored potato cultivars under a combined storage regime (4 °C for 28 days followed by 25 °C for 24 days). Utilizing a double integrating sphere (DIS) system, the absorption (μa) and reduced scattering coefficients (μ's) in the 400-1700 nm range were acquired to systematically analyze the dynamic changes in core quality attributes: moisture, starch, soluble solids content (SSC), and firmness. Various spectral pre-processing techniques, dimensionality reduction methods, and machine learning algorithms were employed to construct dynamic prediction models. The results revealed that: during low-temperature storage, starch degradation led to a decrease in μa, while μ's increased due to microstructural changes; during ambient storage, both parameters exhibited cultivar-specific fluctuations. Optimal prediction models were developed by tailoring the optical parameter, spectral range, and algorithm for each quality attribute. For moisture, the Savitzky-Golay smoothing (SGS) combined with Standard Normal Variate (SNV) preprocessing, together with Principal Component Analysis (PCA)-Partial Least Squares Regression (PLSR) model based on μa (1100-1670 nm) achieved superior performance, with determination coefficient for prediction (R2p) of 0.931, root mean square error for prediction (RMSEP) of 0.447%, and relative prediction deviation (RPD) of 3.819. For starch, the SGS combined with Multiplicative Scatter Correction (MSC) preprocessing and Competitive Adaptive Reweighted Sampling (CARS)-Back Propagation Neural Network (BPNN) model based on μa (1100-1670 nm) yielded the best results (R2p = 0.879, RMSEP = 2.207%, RPD = 2.888). For SSC, the SGS + SNV preprocessing combined with t-Distributed Stochastic Neighbor Embedding (t-SNE)-Random Forest (RF) model based on μa (1100-1670 nm) provided accurate prediction (R2p = 0.860, RMSEP = 0.124°Brix, RPD = 2.681). For firmness, the SGS + SNV preprocessing combined with CARS-BPNN model based on μ's (1100-1670 nm) was most effective (R2p = 0.841, RMSEP = 0.271 N, RPD = 2.519). In conclusion, the optical properties (μa and μ's) effectively reflect the chemical and physical changes in potatoes during storage. Coupled with chemometrics, they enable rapid, non-destructive, and dynamic quality detection for multi-variety colored potatoes throughout the complex dynamic processes of alternating cold and ambient storage.