Comparing Machine Learning Models for Adulterant Detection in Sago Via Visible Near-Infrared Hyperspectral Imaging

Abstract

Sarawak is the largest producer of sago flour in Malaysia. The assessment of flour quality is generally made by its noticeable color. The excessive use of whitening chemicals, such as calcium carbonate, to alter the color and nutritional value of the product has increased the potential for food fraud. Traditional laboratory methods used to detect adulteration are expensive and time-consuming; hence, this study employed a visible near-infrared (Vis-NIR) hyperspectral camera combined with machine learning models to quantify calcium carbonate levels in sago flour rapidly. The imaging was carried out in the 400-1,000 nm regions, and the calcium carbonate concentrations used ranged from 2 w/w% to 5 w/w%. Machine learning models considered in this study were Principal Component Regression (PCR), Partial Least Square Regression (PLSR), and Multiple Linear Regression (MLR). The mean reflectance from the spectral data was used to train and test these machine learning models. Upon optimizing the hyperparameters, the PLSR model outperforms both MLR and PCR models, where its training had R², RMSE, and MAE values of 0.99981, 0.00008, and 0.00006, respectively. These indicate that a visible near-infrared (Vis-NIR) hyperspectral camera coupled with PLSR has the potential to be deployed in detecting adulterants in sago flour.