Calibrating the prediction model of soluble solids content and firmness in kiwifruit across years based on NIR...

Near-infrared (NIR) spectroscopy has been widely used in nondestructive detection of fruit internal quality. However, the biological variability of fruit would change their texture, which may lead to the failure of fruit quality prediction models built based on spectroscopy. Therefore, 'XuXiang' kiwifruit samples were used to investigate the calibration performance of four model calibration algorithms (model calibration (MC), slope\bias calibration (SBC), transfer component analysis (TCA) and correction-two-stage TrAdaBoost.R2 (Cor-TTB)) of biological variability brought by different years to improve the performance of kiwifruit soluble solids content (SSC) and firmness (FI) modeling based on spectral data. In addition, a wavelength-selection algorithm based on correlation (WSC) was proposed to improve the performance of kiwifruit SSC and FI prediction models in combination with model calibration. The results showed that Cor-TTB performed most optimally compared to the other algorithms. It achieved calibration when 20% of new samples were added to the SSC prediction model (RPD_p = 2.31) and 10% of new samples were added to the FI prediction model (RPD_p = 2.35). Additionally, the stability of the model calibration algorithms was confirmed with the maximum standard deviations of RPD_p for the ten repetitions calibration was 0.15. When combined with the WSC, the RPD_p increased 13.41% at most and 5.65% at least, indicating that the WSC was able to improve the accuracy of model calibration effectively. In summary, Cor-TTB performed the best in model accuracy and small-sample adaptability among the four calibration algorithms and has a greater potential to address biological variability in kiwifruit.