Integrative hyperspectral imaging and artificial intelligence approaches for identifying sucrose substitutes and assessing cookie qualities

This study employed a combined approach utilizing hyperspectral imaging and artificial intelligence to identify sugar substitutes (allulose and kestose) and predict the quality characteristics of cookies formulated with these substitutes. When the hyperspectral data of sweeteners were applied to four machine learning models, the classification performance was enhanced when sweeteners were in powdered form and in cookies, compared to when they were in solution and cookie dough. Especially, the stochastic gradient descent model demonstrated noticeable performance regardless of the sample form. Furthermore, the confusion matrix revealed a tendency to classify sucrose/kestose and fructose/allulose, which have relatively similar molecular structures, suggesting the potential of hyperspectral analysis to recognize structural differences in sugars within complex food matrices. To predict the quality characteristics (spread factor and hardness) of cookies formulated with sugar substitutes, the hyperspectral data of sweeteners in solution were subjected to a multilayer perceptron model of which performance was evaluated under nine individual and combined hyperparameter conditions. The results indicated that the model performance was optimized under conditions of three hidden layers and a sigmoid activation function. Notably, the model demonstrated excellent predictive performance, achieving accuracy levels exceeding 0.93 for both spread factor and hardness.