An uncertainty sampling strategy based model updating method for soluble solid content and firmness prediction of apples from different years

Visible/near-infrared spectroscopy combined with chemometric methods has been widely used in fruit quality detection. In order to ensure the performance of the prediction model, the calibration samples used to train model should cover the range of variability anticipated in prediction samples. However, this requirement is difficult to meet in practice, especially, for quality detection of fruit from cross-years with variational cultivation conditions, climate conditions, as well as growing management. In this study, a model updating method based on uncertainty sampling strategy was proposed to accommodate the sample from different periods. The proposed method firstly selected feature wavebands to form new feature space by using PLS projection analysis (PLS-P) algorithm. The representative samples which have high feature similarity to updating samples were then selected from original calibration set and used to train the initial partial least squares regression (PLSR) model. The uncertainty sampling strategy combined spectral similarity and predicted value from initial PLSR model was applied to obtain the samples with most informative in updating set, these samples were added to the training set. Finally, the PLSR model was updated iteratively until the demand is met. Three cultivars of apples, namely, 'Jonagold', 'Golden Delicious' and 'Red Delicious', harvested in 2009 and 2010 were used for evaluating the performance of the proposed method. Compared with random sampling (RS), traditional Kennard-Stone (KS) and joint x-y distances (SPXY), the proposed method achieved the best performance. It demonstrated that the proposed updating method is an effective way to improve the prediction accuracy of samples from cross-years.

Automated summary

Visible/near-infrared spectroscopy combined with chemometric methods is widely used in fruit quality detection. A model updating method based on uncertainty sampling strategy was proposed to accommodate samples from different periods, improving prediction accuracy of samples from cross-years. The method achieved better performance compared to random sampling, traditional Kennard-Stone, and joint x-y distances methods.