Improving the prediction performance of soluble solids content (SSC) in kiwifruit by means of near-infrared spectroscopy using slope/bias correction and calibration updating.

Soluble solids content (SSC) is particularly important for kiwifruit, as it not only determines its flavor, but also helps assess its maturity. Visible/near-infrared (Vis/NIR) spectroscopy has been widely used to evaluate the SSC of kiwifruit. Still, the local calibration models may be ineffective for new batches of samples with biological variability, which limits the commercial application of this technology. Thus, a calibration model was developed using one batch of fruit and the prediction performance was tested with a different batch, which differs in origin and harvest time. Four calibration models were established with Batch 1 kiwifruit to predict SSC, which were based on full spectra (i.e., partial least squares regression (PLSR) model based on full spectra), continuous effective wavelengths (i.e., changeable size moving window-PLSR (CSMW-PLSR) model), and discrete effective wavelengths (i.e., competitive adaptive reweighted sampling-PLSR (CARS-PLSR) model and PLSR-variable importance in projection (PLSR-VIP) model) respectively. The R values of these four models in the internal validation set were 0.83, 0.92, 0.96, and 0.89, with corresponding RMSEV values of 1.08 %, 0.75 %, 0.56 %, and 0.89 %, and RPD values of 2.49, 3.61, 4.80, and 3.02, respectively. Clearly, all four PLSR models performed acceptably in the validation set. However, these models performed very poorly in predicting the Batch 2 samples, with their RMSEP values all exceeding 1.5 %. Although the models could not be used to predict exact SSC, they could still interpret the SSC values of Batch 2 kiwifruit to some extent because the predicted SSC values could fit a specific line. To enable the CSMW-PLSR calibration model to predict the SSC of Batch 2 kiwifruit, the robustness of this model was improved by calibration updating and slope/bias correction (SBC). Different numbers of new samples were randomly selected for updating and SBC, and the minimum number of samples for updating and SBC was finally determined to be 30 and 20, respectively. After calibration updating and SBC, the new models had average R, average RMSEP, and average RPD values of 0.83 and 0.89, 0.69 % and 0.57 %, and 2.45 and 2.97, respectively, in the prediction set. Overall, the methods proposed in this study can effectively address the issue of poor performance of calibration models in predicting new samples with biological variability and make the models more robust, thus providing important guidance for the maintenance of SSC online detection models in practical applications.