Improved spectrophotometric models and methods for the non-destructive and effective foodstuff parameters forecasting

In the present paper are described the results of test carried out with different calibration/prediction models in order to assess the feasibility of a visible-near-infrared sensor to monitor and forecast different parameters as firmness and soluble tannins content of persimmons fruit, lactose, dry matter and protein content of donkey milk, and Imazalil in water solution. Different types of spectra pre-processing methods (normalization, first and second derivatives) and four regression models (partial least squares, principal components regression, support vector machine, ensemble regression trees) were tested. These models were assessed by a 10-fold cross-validation with a new strategy for both outlier removal and wavelength reduction. Afterwards, their statistical significance was evaluated by 100 Monte Carlo simulation runs. The study allowed to identify the most appropriate strategy to the development of excellent regression and prediction models, defining for each parameter the most efficient model in terms of precision and predictability. The selected models could represent the core of smart, cheap, high accuracy sensors for online assessment of several food parameters, customised for the different food industry processing line requirements.