Accurate Prediction of Sensory Attributes of Cheese Using Near-Infrared Spectroscopy Based on Artificial Neural Network

Abstract

The acceptance of a food product by the consumer depends, as the most important factor, on its sensory properties. Therefore, it is clear that the food industry needs to know the perceptions of sensory attributes to know the acceptability of a product. There exist procedures that systematically allows measurement of these property perceptions that are performed by professional panels. However, systematic evaluations of attributes by these tasting panels, which avoid the subjective character for an individual taster, have a high economic, temporal and organizational cost. The process is only applied in a sampled way so that its result cannot be used on a sound and complete quality system. In this paper, we present a method that allows making use of a non-destructive measurement of physical-chemical properties of the target product to obtain an estimation of the sensory description given by QDA-based procedure. More concisely, we propose that through Artificial Neural Networks (ANNs), we will obtain a reliable prediction that will relate the near-infrared (NIR) spectrum of a complete set of cheese samples with a complete image of the sensory attributes that describe taste, texture, aspect, smell and other relevant sensations.

Keywords: ANN; food quality estimation; food sensory prediction.

Figure 1

Global scheme of the prediction of 19 sensory parameters for each cheese sample. PC1, PC2 and PC3 denote the three principal components of Principal Components Analysis (PCA) on NIR spectra. Sensory characteristic i refers to each of the 19 sensory parameters of the cheese to be estimated.

Figure 2

NIR measurements (spectra) for the whole sample set in the wavelength range 1100–2000 nm. Four colors are used to distinguish the spectra according to the aging and seasons of the cheese (4 months/Winter-Green, 6 months/Winter-Magenta, 4 months/Summer-Yellow, 6 months/Summer-Cyan).

Figure 3

Loadings plot—Principal Components obtained by PCA from NIRS. The weights assigned to each one describes the relative influence for each wavelength.

Figure 4

Scores plot—cheese spectral dataset represented in the Principal Components space. It can be observed that the main variability is found along the first PC. In the third direction the variation is quite small.

Figure 5

Schematic representation of neural model for the prediction.

Figure 6

The root mean squared error of prediction for the whole set of parameters for unknown samples.

Figure 7

Squared correlation coefficient of prediction for the whole set of parameters for unknown samples.

Figure 8

Scatter plot for PCA-ANN prediction results vs tasters’ scores for all sensory attributes for the prediction dataset. The performance over the attribute span is observed for the whole set of attributes in the prediction stage. The green line is the expected score and the red lines mark the interval defined with the deviation of the tasters’ scores for the total set of samples.

Figure 9

Residual for the predictions for different atributes.