Hyperspectral imaging and machine learning-based classification and quantification of sodium saccharin in cat food powder

Abstract

Sodium saccharin is a synthetic sweetener occasionally added illicitly to pet food to improve palatability, yet its overuse poses health risks to companion animals. In this work, we evaluate near infrared hyperspectral imaging combined with machine learning for rapid, non-destructive quantification of sodium saccharin in ground cat food. Fifteen samples—three commercial formulations each spiked at 0, 0.05, 0.10, 0.15, and 0.20 wt% saccharin—were imaged, and mean spectra were extracted from regions of interest. Spectral preprocessing (Savitzky–Golay smoothing, standard normal variate, autoscaling) was applied to mitigate noise and scattering. We constructed qualitative and quantitative models using partial least squares regression (PLSR), random forest (RF), and a one-dimensional convolutional neural network (CNN). Among these, the SG-preprocessed PLSR model achieved the best performance on an independent test set, with calibration R² of 0.99, prediction R² of 0.98, RMSE_C = 0.08 wt%, and RMSE_P = 0.22 wt%. These results demonstrate that SG + PLSR is the optimal model across the three sample sets Therefore, SG + PLSR provides a robust approach for monitoring sodium saccharin adulteration in pet food. Our methodology offers a promising platform for inline quality control in feed manufacturing.