Machine Learning-Driven FTIR Spectroscopy for Rapid Multi-Class Detection of Adulteration in Paneer

Abstract

Paneer, an un-ripened acid- and heat-coagulated Indian cheese, is among the dairy products most frequently subjected to economically motivated adulteration, yet rapid tools that resolve several adulterants at once remain scarce. This study develops and rigorously benchmarks a non-destructive, deployment-oriented machine-learning pipeline that discriminates pure paneer from eight chemically distinct adulterant classes (corn and wheat starch, urea, caustic soda, detergent, formaldehyde, palm oil, and glyceryl monostearate) from attenuated total reflectance Fourier-transform infrared (ATR-FTIR) spectra. The spectra spanning 1702 wavenumbers (3996-501 cm-1) across nine balanced classes were processed with Savitzky-Golay smoothing, standard normal variate correction, and first- and second-derivative augmentation. Eleven state-of-the-art classifiers were compared under identical, leakage-free conditions using repeated stratified cross-validation and the Nadeau-Bengio corrected resampled t-test. Histogram-based gradient boosting was selected by a transparent composite score that balances accuracy, robustness, and inference efficiency, and it was tuned by Bayesian optimization restricted to the winning model. The selected model achieved a cross-validated macro-F1 of 0.998 and near-perfect classification on an independent hold-out set (accuracy 0.998, MCC 0.996), with a compact footprint and millisecond-scale inference. Occlusion- and SHAP-based analyses attributed decisions to chemically meaningful C-H, O-H/N-H, and carbohydrate regions. The results show that, on compact FTIR datasets, well-regularized ensemble learners deliver accurate, transparent, and readily deployable adulteration screening.