NIR and machine learning for apple freeze-drying moisture prediction

To achieve rapid, nondestructive, and precise measurement of moisture content during freeze-drying of apple slices, this study compared the performance of near-infrared spectroscopy (NIR) combined with various pretreatment-feature selection methods. Several machine learning approaches were compared as well. Savitsky-Golay (SG) and the standard normal variable transformation (SNV) were selected for the pretreatment of NIR spectra acquired during apple freeze-drying. Feature wavelengths were extracted using isometric variable selection methods such as the Competitive Adaptive Reweighted Sampling (CARS), Kmeans, Synergistic Partial Least Squares (si-PLS), and the Stochastic and Continuous Projection Algorithm (SPA). Finally, for the multivariate quantitative analysis of NIR data, 12 different machine learning algorithms were used for modeling. Results indicate that different combinations of pretreatment/feature wavelength selection and modeling methods lead to significant variations in predictive performance. Except for XGB, nonlinear regression models consistently outperformed linear regression models. Furthermore, the gain effects of pretreatment and feature wavelength selection were more pronounced in linear regression models. Among these, the SG-SNV-SPA-RFR model demonstrated optimal predictive performance (RMSEP = 0.0293, Rp2 = 0.9876). This study demonstrated that the combination of NIR technology with machine learning can be employed for intelligent nondestructive monitoring of moisture content changes during the freeze-drying process, providing a promising strategy for intelligent freeze-drying of fruits and vegetables.