Portable multichannel spectroscopy integrated with machine learning for the detection and classification of synthetic...

The rapid and reliable detection of synthetic dyes remains a critical challenge in food safety monitoring. Conventional analytical techniques offer high accuracy but are destructive, expensive, and limited to laboratory settings. This study investigates the performance of a portable multichannel spectroscopy system integrated with machine learning for detecting and differentiating four synthetic dyes (i.e., rhodamine b, carmoisine, ponceau 4r, and allura red) across a broad concentration range (i.e., 1 - 10 ppm, 25 ppm, 45 ppm, 70 ppm, and 95 ppm) in both water and jelly. Machine learning models were developed to quantify dye concentrations and classify dye types based on spectral data. Support vector regression (SVR) models using absorbance data demonstrated exceptional predictive performance for all dyes in distilled water, achieving R² values of 0.999 for rhodamine b and ponceau 4r, 0.996 for carmoisine, and 0.997 for allura red. In the jelly matrix, the integration of support vector regression (SVR) model with Kubelka-Munk (KM) transformation and Savitzky-Golay (SG) smoothing yielded robust quantitative reliability, with the ratio of prediction to deviation (RPD) values exceeding 32. For dye identification, the support vector classification (SVC) and k-nearest neighbors classification (KNNC) models achieved 100% accuracy by utilizing the Kubelka-Munk spectra combined with standard normal variation (SNV) preprocessing, multiplicative scatter correction (MSC) or Savitzky-Golay (SG) smoothing. These results demonstrate that the proposed portable multichannel spectroscopy platform enables fast, non-destructive, and accurate dye detection, highlighting its potential for on-site food safety monitoring and regulatory enforcement.