Use of convolutional neural network (CNN) combined with FT-NIR spectroscopy to predict food adulteration: A case study on coffee

Abstract

Food systems are negatively affected by food frauds with food recalls challenging the system's sustainability and consumer confidence in food safety. Coffee, an economically important commodity is frequently adulterated for economic gains, thereby requiring fast and reliable detection techniques. Of the various tracing strategies, spectroscopic techniques have seen considerable commercial success but rely heavily on human-engineered features. Thus, this study aims to evaluate feasibility of deep chemometrics (i.e., convolutional neural network, CNN) for coffee adulterant quantification in comparison to standard chemometrics approaches (i.e., partial least squares, PLS; and interval-PLS, iPLS). Commercial ‘espresso’ coffee was admixed with chicory, barley, and maize (0–25%, w/w) and subjected to Fourier Transformed-Near Infrared (FT-NIR) spectral analysis. The results confirmed the feasibility of CNN algorithm for adulterant quantification from FT-NIR spectra with excellent performances (R 2 > 0.98). Furthermore, CNN with Data augmentation (DA) with either autoscaling (AS) and/or standard normal variate (SNV) pre-treatment showed better prediction performances with RMSEP (0.76–0.82%) and BIASP (−1.00 × 10 −2 to −1.00 × 10 −1 %) that were better to comparable to those of PLS and/or iPLS models (0.72 < % RMSEP <3.045; −7.98 × 10 −2 < % BIASP <8.63 × 10 −2 ) for the adulterants tested. The study showed that deep learning algorithms can be potential alternatives to standard methods with little to no human interference for feature extraction during real-time applications of spectroscopic tools targeted to overcome food fraud crisis. • Deep chemometric algorithm tested for coffee adulteration prediction. • Convolutional neural network led to better performance than chemometrics. • Outlier removal and data augmentation benefitted the CNN algorithm. • CNN can be a feasible alternative for FT-NIR data analysis.