Optimization of Hyperspectral Nondestructive Detection Model for Potato Starch Content

Abstract

To improve the detection accuracy of non-destructive hyperspectral detection of potato starch content, hyperspectral reflectance spectrum was used to scan potato, and ten kinds of spectral pretreatment methods, such as de-trend, standard normal variable transformation (SNV), Detrend + FD, were applied to the spectrum, and the partial least squares model of potato starch content was established. The results showed that the prediction model of potato starch content based on de-trend transformation combined with first derivative pretreatment was the best, and the determination coefficient of validation set was 0.7911, and the error was 0.3596. In order to further optimize the model, three methods such as CARS, RF, iterative reserved information variable (IRIV) were used to optimize the characteristic variables, and the partial least squares prediction models CARS-PLS, RF-PLS, IRIV-PLS were established. The results showed that the prediction effect of the three models for predicting potato starch content has been greatly improved. The IRIV feature wavelength extraction can simplify the model and improve the accuracy and stability. The determination coefficient of IRIV-PLS model validation set was 0.8129, and the error was 0.3516. Finally, to verify the accuracy and stability of potato starch content prediction model, 36 potato samples were selected for external validation. The determination coefficients between model predicted values and standard physicochemical values of potato starch content was 0.8125, and the error was 0.3524. The results show that model optimization is very necessary, the IRIV-PLS prediction model can improve the detection accuracy of potato starch content, and this study provides theoretical basis and technical reference for the potato starch content detection accuracy.