Search

Knowledge of the energy and macronutrient content of complex foods is essential for the food industry and to implement population-based dietary guidelines. However, conventional methodologies are time-consuming, require the use of chemical products and the sample cannot be recovered. We hypothesize that the nutritional value of heterogeneous food products can be readily measured instead by using hyperspectral imaging systems (NIR and VIS-NIR) combined with mathematical models previously fitted with spectral profiles.118 samples from different food products were collected for building the predictive models using their hyperspectral imaging data as predictors and their nutritional values as dependent variables. Ten different models were screened (Multivariate Linear regression, Lasso regression, Rigde regression, Elastic Net regression, K-Neighbors regression, Decision trees regression, Partial Least Square, Support Vector Machines, Gradient Boosting regression and Random Forest regression). The best results were obtained with Ridge regression for all parameters. The best performance was for estimating the protein content with a RMSE of 1.02 and a R2 equal to 0.88 in a test set, following by moisture (RMSE of 2.21 and R2 equal to 0.85), energy value (RMSE of 21.84 and R2 equal to 0.76) and total fat (RMSE of 2.17 and R2 equal to 0.72). The performance with carbohydrates (RMSE of 2.12 and R2 equal to 0.61) and ashes (RMSE of 0.25 and R2 equal to 0.38) was worse. This study shows that it is possible to predict the energy and nutrient values of processed complex foods, using hyperspectral imaging systems combined with supervised machine learning methods.

Moisture content and its distribution is a critical parameter in the production of cereal bars. Inappropriate control of this quality parameter can lead to non-conforming products and excess waste on production lines. In the field of hyperspectral imaging, the search for alternative light sources to stabilised-halogen (cheaper and emitting less heat) is a growing need for the application of this technology in industry. This study compares three different illumination systems for moisture prediction in the visible-near infrared (vis-NIR) range (from 400 nm to 1000 nm). The hyperspectral images were acquired using three illumination systems including two halogen-based systems (stabilised-halogen and conventional-halogen) and an LED-based illumination system. The results showed that halogen-based illumination systems combined with a partial least squares model better predicted moisture in bars. Lower accuracies were obtained when the experiment was performed with an LED-based illumination system, which showed double the error of the halogen-based systems. It was concluded that this is a consequence of the information lost in bands appearing above 850 nm that may be revealing information about the moisture in bars since the second overtone of the water O–H is found at 970 nm. The results demonstrate that conventional halogen-based light systems in the vis-NIR range are a promising method for moisture prediction in cereal bars.