Model updating for apple firmness prediction based on hyperspectral scattering images with semi-supervised study and auxiliary data

Abstract. Firmness is regard as one of important quality parameters used for determining apple quality. Hyperspectral scattering images combined with chemometrics is a nondestructive method for apple firmness prediction. When apple samples from the current year with the true firmness values are scarce, making the most use of the apple samples (with the true firmness values) from the same variety and previous year is important to improve the prediction accuracy of models. Regression models based on the mean spectral features extracted from hyperspectral scattering images are established using partial least square regression (PLSR). This research proposes a model updating algorithm for apple firmness prediction based on hyperspectral scattering images with semi-supervised study and auxiliary data to improve the accuracy of the model. After model updating, the RP value reached 0.76 and improved by 4.1%, compared with that of non-updated model for test set. It indicated that the proposed model updating strategy was an effective method for apple firmness prediction.