Machine learning-assisted micro NIR for rapidly monitoring the variations of enzymatic activity and carbohydrates during saccharification of stored sweet potatoes

This research investigated a micro near-infrared (NIR) system (900-1700 nm) combined with machine learning algorithm (MLA) for nondestructively monitoring the dynamic variations of α-amylase activity (αAA), β-amylase activity (βAA), amylopectin content (APC), amylose content (AC), soluble sugar content (SSC), total sugar content (TSC) of during saccharification of stored sweet potatoes (SSP). Six different MLAs were executed to mine full-range NIR spectra to relate to the measured six indexes, resulting in support vector regression (SVR) models performed best with greatest values of correlation coefficients (R P ) and lowest values of root-mean-square errors (RMSE P ) in prediction (R P =0.885-0.942, RMSE P =3.998-2133.206, RPD=3.017-5.854). By applying three techniques including successive projections algorithm (SPA), recursive feature elimination (RFE), and improved whale optimization algorithm (iWOA), 25-30 optimal wavelengths were selected and enabled the optimal wavelengths-based SVR models in predicting the six indexes with similar good performance (R P =0.884-0.917, RMSE P =4.238-2261.198, RPD=2.866-3.721). The good capabilities of these optimal wavelengths-based models were further validated by a set of independent SSP with small bias. In conclusion, the micro NIR system in tandem with SVR algorithm showed a great potential in prediction of αAA, βAA, APC, AC, SSC, and TSC of SSP, providing a valuable reference to develop specialized NIR device for quality evaluation of SSP. • Micro NIR spectra were mined to relate to αAA, βAA, APC, AC, SSC and TSC • SVR-assisted NIR models was established to quantify the six indexes • SPA, RFE and iWOA were applied to select optimal wavelengths for model optimization • The αAA, βAA, APC, AC, SSC and TSC were well predicted by SVR models