Study on Detection of Potato Starch Content by Optimum Hyperspectral Characteristic Wavelength Method

In this paper, using hyperspectral imaging technology combined with feature variable optimization algorithm, the method of rapid detection of potato starch content was explored. The spectral pretreatment uses 10 methods such as de-trend and first derivative, second derivative and so on. Through comparative analysis, it is determined that the combination of de-trend and first derivative is the best. The spectral data preprocessed by the de-trend method combined with first derivative method were used to optimize the characteristic wavelength variables by four algorithms: Competitive adaptive reweighted sampling (CARS), Random-frog (RF), Iterative reserved information variable (IRIV), and Interval combin optimization (ICO). The partial least squares regression models CARS-PLS, RF-PLS, IRIV-PLS and ICO-PLS were established one by one for the selected characteristic variables. The results show that the number of characteristic wavelength variables optimized by ICO algorithm is at least 16, the prediction accuracy of ICO-PLS model is the highest, the determinant coefficients of prediction set is 0.8255, and the error is 0.1697. ICO algorithm is a new algorithm to improve the accuracy of spectral detection prediction model. It provides theoretical methods and technical support for the study of the detection accuracy of potato starch content.