Convolutional Neural Networks for Evaluating Spirulina (Arthrospira spp.) Adulteration Through Digital Images

Abstract

Spirulina, a nutrient-rich product derived from cyanobacteria of the genus Arthrospira, is widely consumed as a dietary supplement. However, its high market value makes it susceptible to adulteration, particularly by adding low-cost compounds such as sodium bicarbonate. This study aimed to evaluate and classify sodium bicarbonate adulteration in Spirulina at levels of 10%, 15%, and 25% w w⁻¹ using convolutional neural networks (CNNs). A digital imaging system was developed to capture sample images, which were analyzed through their RGB channels. Traditional chemometric methods, including hierarchical cluster analysis (HCA) and principal component analysis (PCA), demonstrated limited performance as sample size increased. To overcome these limitations, deep learning techniques were implemented using ResNet-18 and ResNet-50 architectures. The CNN models achieved high classification accuracies exceeding 99%. These findings demonstrate the potential of CNNs as a robust and scalable tool for the rapid, non-destructive detection of Spirulina adulteration, representing a novel approach in food authentication.