Deep learning based on the Vis-NIR two-dimensional spectroscopy for adulteration identification of beef and mutton

Abstract

Rapid and accurate identification of contaminated meat products represents a global challenge, particularly with regard to minced meat products. This study reports on two methods for analyzing the adulteration of beef and mutton: the deep learning combined with two-dimensional correlation spectroscopy (2DCOS) method, and the PLS-DA method. The 2DCOS method employed in this research substantially improves the resolution of one-dimensional Vis-NIR spectra, allowing for the visualization of spectral information changes within the samples. By analyzing the impact of various proportions of chicken, duck, and pork mixed with beef or mutton, the synchronous 2DCOS images unveil distinct patterns of chemical information alteration within the spectra under different adulteration scenarios. Notably, the auto peaks and cross peaks observed within the 400–1400 nm range serve as key indicators of these changing patterns. The ResNet deep learning method possesses the advantageous capability to effectively extract 2DCOS feature information, resulting in the achievement of high accuracy models (100%). In contrast, the accuracy of the PLS-DA model test set, based on either raw or pre-processed spectral data matrices, ranged from 32.97% to 50.64%. These results substantiate the effectiveness of utilizing 2DCOS in combination with deep learning as a powerful tool for discerning beef and mutton adulteration.