Abstract:To accurately predict soluble solids and vitamin C content in cherry tomatoes, a backpropagation neural network model optimized using the improved cuckoo search algorithm (ICS-BPNN) is proposed. The near-infrared spectra of the samples at 1 350~1 800 nm were collected and pre-processed using different methods. The stability competitive adaptive reweighted sampling (SCARS), genetic (GA), and automatic ordinal predictor selection (Auto OPS) algorithms were then employed to extract the characteristic wavelength. BPNN and CS-BPNN models were established using machine learning methods. To further enhance accuracy and convergence of the models, an adaptive algorithm was introduced to improve the probability of cuckoo egg elimination, and the cross-border nests were newly processed via ICS-BPNN. The optimized models demonstrated ideal results. The results showed that the coefficients of determination, R2c and R2p of the soluble solid content were 0.83 and 0.85, respectively; the root mean square error of calibration (RMSEC) and prediction (RMSEP) sets were 0.85 and 0.79, respectively. The vitamin C content obtained using the optimized model had R2c and R2p of 0.91 and 0.91, respectively. The RMSEC and RMSEP values were 0.48 and 0.45, respectively. Thus, a combination of near-infrared spectroscopy and improved machine learning methods can achieve the rapid and non-destructive predictive analysis of the internal quality of cherry tomatoes.
