FTIR spectroscopy identification of fraud in coffee powder: a study on preprocessing methods

This study investigates the use of Fourier Transform Infrared (FTIR) spectroscopy, combined with eight distinct spectral preprocessing techniques and advanced regression modeling, to detect and quantify adulteration in ground coffee. Three common adulterants-barley, chickpea, and date pit powders-were mixed into pure coffee at concentrations of 10%, 15%, 20%, and 100% (w/w). FTIR spectra were recorded in the mid-infrared range (400-4000 cm^-1), where key spectral markers-such as 1650 cm^-1 (proteins, amide I), 1740 cm^-1 (lipids, ester carbonyls), and 1000-1200 cm^-1 (carbohydrates)-enabled clear chemical differentiation between coffee and adulterants. Eight preprocessing methods were applied: Standard Scaler, Standard Normal Variate (SNV), Multiplicative Scatter Correction (MSC), Savitzky-Golay smoothing, first derivative, second derivative, wavelet transform, and median filtering. These methods aimed to minimize baseline drift, noise, and scattering effects. A suite of machine learning regression models, including Bagging, Random Forest, Gradient Boosting, and XGBoost, was trained on the preprocessed data, with hyperparameters optimized individually for each model to improve generalization. The Bagging Regressor paired with SNV or Savitzky-Golay preprocessing achieved the highest predictive accuracy across all adulterants. For date pit detection, the best model yielded R² = 0.90 and RMSE = 11.54 on the test set. For barley and chickpea, test R² values of 0.87-0.91 and RMSEs between 10.85 and 13.68 were achieved. These results demonstrate the effectiveness of the FTIR-chemometrics pipeline in accurately identifying low-level (≥10%) adulteration in coffee powder, offering a robust and non-destructive method suitable for real-time food fraud detection.