Identification and Classification of Lotus Root Powder Using Hyperspectral Imaging Combined with Machine Learning

Lotus root powder boasts an exceptional nutritional value but remains vulnerable to adulteration driven by economic incentives. Consequently, adulterant detection and authenticity discrimination for lotus root powder are of paramount importance. Constrained by the limitations in spatial resolution of imaging equipment as well as the fine granularity of powder particles, discerning powder mixtures containing various types of adulterants using hyperspectral imaging (HSI) technology is a challenge that needs addressing. In this study, adulterated powders were prepared using the mutual solubility mixing method, after which HSI combined with machine learning (ML) algorithms were used to classify lotus root powder and four other starch types as well as starch samples containing multiple adulterants. Anomalous data were first removed using the isolation forest (IF), and the resulting hyperspectral data were preprocessed using the Savitzky-Golay combined with multiplicative scatter correction (SG-MSC) and standard normal variate (SG-SNV). The selection of characteristic wavelengths using both the competitive adaptive reweighted sampling (CARS) and bootstrapping soft shrinkage (BOSS). Then discriminant models were built by the back propagation neural network (BPNN), partial least squares discriminant analysis (PLS-DA) and support vector machine (SVM). The results show that the SG-MSC-BOSS-SVM model demonstrated superior performance, achieving 100% accuracy for both the classification and adulteration. This study confirmed that HSI combined with ML algorithms can be effectively applied to identify the authenticity of lotus root powder.