Effects of spectral transformations in support vector machine on predicting 'Arumanis' mango ripeness using near-infrared spectroscopy

Abstract

One of the challenges of exporting Arumanis mangoes is their accurate grading ability because the mangoes do not change color during ripening. Near-Infrared (NIR) spectroscopy is a non-destructive method for detecting the internal ripeness of fruit which is quite reliable. However, NIR absorbance bands are often nonspecific, extensive, and overlapping. Although SVM modeling is quite good in performance, it can still be improved by spectral transformation. In this study, 11 spectral transformation operations were compared with their combinations to find the best input model. Spectral transformation operations include SAVGOL, RNV, BASELINE, MSC, EMSC, NORML, CLIP, RESAMPLE, DETREND, SNV, and LSNV. In the 2 class classification model, the highest accuracy is obtained using RNV and SAVGOL. The prediction model for SSC content with the best MSE value uses 3 combinations of spectral transformation operations, namely DETREND, LSNV, and SAVGOL with parameter values: 'deriv_order': 0, 'filter_win': 31, 'poly_order': 6. As for the prediction model of mango hardness with The best MSE value uses 2 combinations of spectral transformation operations, namely LSNV and SAVGOL with parameter values: deriv_order ': 0,' filter_win ': 15,' poly_order ': 6.