Performance of genetic algorithm in optimization of NIRS PLS models to predict apple fruit quality

Near infrared spectroscopy (NIRS) has become a versatile tool for probing the properties of biomaterials such as fruit. Employing chemometric methods has allowed the effective use of NIRS to study complex materials including foodstuff and horticultural products. However, the complexity of the NIR spectral information may require that selective methods are used to improve the accuracy of quality attribute prediction models. A genetic algorithm (GA) was used as a selective method to determine the most relevant wavelengths (800-1150 nm) in PLS models for predicting apple fruit quality. A comparison of GA with alternative variable selection methods was used to measure the effectiveness of GAs in producing optimal NIRS prediction models. Prediction model performances were measured using the coefficient of determination in PLS (R²) and the root mean square error in cross validation (RMSECV), and relative model performances, were used to compare wavelength selection methods. When applied to predict both the external (bruise damage) and internal (total soluble solids and titratable acids) quality of apple fruit, GA performed relatively better than orthogonal PLS loadings and variable importance in projection (VIP). Based on these findings, GA is recommended as a proven tool for producing optimal models for NIRS prediction of quality in fruit and vegetables.