NIR-Based Detection of Starch Adulteration in Soft Cheese Using Machine Learning Models

Abstract

The control of adulteration in dairy products requires rapid, accurate, and non-destructive methods to ensure authenticity and quality. This study aimed to detect starch adulteration levels in soft cheese using NIR spectra (700–1000 nm) acquired through hyperspectral imaging (HSI) and machine learning models. A total of 20 averaged spectra were collected across five adulteration levels, and eight representative models were trained, ranging from penalized regression approaches to non-linear methods. Results show that Elastic Net achieved the highest performance (R2_test = 0.986, RMSE = 0.628), outperforming Lasso and Ridge, while more complex models such as Random Forest, Gradient Boosting, and MLP exhibited overfitting. After hyperparameter optimization via Grid Search, Ridge reached R2_test = 0.981 and RMSE = 0.739, confirming the robustness of penalized linear methods when working with reduced datasets. These findings indicate that the combination of NIR spectroscopy and regularized regression provides an efficient and feasible tool for the rapid detection of starch adulteration in soft cheese, enabling future applications in in-plant quality control and routine monitoring.