A Decision Support Tool to Analyze Food Properties from Near Infrared Spectroscopy*

Near infrared spectroscopy (NIRS) is an analytical technique that is gaining popularity in the food industry due to its low operating costs, rapid analysis and non-destructive sample technique. Numerous studies have shown the relevance of NIR spectra analysis to determine certain quality attributes of food. This makes it attractive for use in quality control and continuous monitoring of food processing. However, the calibration process of NIR is difficult and time-consuming. Depending on the configuration of the NIR instrument, the sample to be analyzed and the attribute to be predicted, the analysis methods and techniques vary. This makes calibration a challenge for many manufacturers. This article aims to develop a decision support tool to assess food properties based on the analysis of selected features of NIR spectra. It intends to provide support to calibrate a predictive model based on NIR spectra. The methodology-based decision support tool was evaluated on cocoa bean samples. The tool suggested using the SG filter technique and PLSR machine learning model to predict the moisture and fat content of cocoa beans. The PLSR model with 4 components trained from 63 wavelengths obtained excellent results for the prediction of moisture content with an R2CV of 0.9 and an RMSEP of 0.28. While the PLSR model with 8 components trained from 166 wavelengths obtained satisfactory results for the prediction of fat content with an R ² CV of 0.93 and an RMSEP of 1.52.