Simultaneous and non-destructive prediction of multiple internal quality characteristics in mandarin citrus with near-infrared spectroscopy and ensemble learning strategy

This study aimed to establish a simultaneous and non-destructive method for the prediction of multiple internal quality characteristics in mandarin citrus with near-infrared spectroscopy combined with ensemble learning strategy. 490 spectra were obtained over the whole picking period without destroying the citrus samples. The ensemble learning strategy was used to establish the quantitative models to simultaneously predict multiple internal quality characteristics, including soluble solids content (SSC), pH, and total acidity (TA), compared with partial least squares (PLS) method. Both validation set and independent test set obtained one month later were used to validate the models. The optimal collection points for the three characteristics were obtained. The ensemble learning strategy was better than PLS method, which can be used to improve the predictive accuracy. The best prediction models for SSC, pH, and TA were second-order derivatives (2nd)-consensus partial least squares (CPLS), 2nd-boosting-PLS (BPLS), and continuous wavelet transform-BPLS. The root mean square errors of prediction (RMSEPs) for validation set were 1.0117, 0.1924, and 0.2408, respectively, while the RMSEPs for independent test set were 1.1067, 0.2647, and 0.2563, respectively. Besides, the long-wave NIR light was more suitable for the quantitative analysis of multiple internal quality characteristics in mandarin citrus than short-wave NIR light.