Morphological characterization and machine learning-based hyperspectral identification of naturally pigmented...

As an intangible cultural heritage, food products derived from naturally pigmented traditional starches are facing a market trust crisis due to the adulteration of dyed starch. This study aimed to develop an integrated identification system to differentiate naturally pigmented starch from commercial crop starch. The experimental design combined morphological, colorimetric, and hyperspectral analyses. Data were processed using machine learning algorithms, with model performance evaluated via five-fold cross-validation. Results showed significant differences in granule morphology, with average sizes of 19.86 μm (Eleutherine plicata), 26.41 μm (Curcuma longa), 27.29 μm (Dioscorea cirrhosa), and 18.26 μm (Castanopsis sclerophylla). Colorimetrically, naturally pigmented starch distributed in purplish-red, yellow, and brown regions, while commercial crop starch clustered in the white area. Principal component analysis indicated that the first three principal components accounted for 92.71 %, 2.32 %, and 2.10 % of the variance, cumulatively explaining 97.13 %. Using machine learning-based method, the support vector machine (SVM) model achieved perfect accuracy (100 %), outperforming the random forest (94.52 %) and artificial neural network (99.60 %) models. This multi-technology fusion system provides a non-destructive, efficient, and practical solution for authenticating and safeguarding intangible cultural heritage food products.