The NIR side of lentil

The interest towards bioeconomy concepts has been considerably growing during the last years and, in particular, the development of sustainable and renewable bio-based technologies for food production is becoming increasingly important and studied. One of the most interesting applications of bioeconomy in the “food” area is the use of enzymes for the modification of food materials [1], to improve safety and to optimize the overall treatment processes. In this perspective, the present study was focused on two processes for treating lentil flour: extraction and hydrolyzation, aimed at making protein available in solution. For both processes, an initial extraction phase with Ca(OH)2 at controlled pH = 8 and fixed temperature of 60 °C was performed. Regarding only the hydrolyzation process, 0.2 % of protease enzyme was also added. The two processes were then carried out with two experimental runs each, differing by stirring rate (60 rpm and 120 rpm). A total of 32 samples of the processed solutions were collected at fixed time points in the range 0–300 minutes and stored frozen upon spectroscopic analysis. All samples were then analysed with a FT-NIR spectrometer (MPA by Bruker) and a Visible spectrometer (Carey by Agilent). The acquired data were imported under MATLAB environment to undergo data quality assessment aimed at removing clear outliers and at choosing the proper preprocessing, specific for each dataset. The NIR data were preprocessed using standard normal variate (SNV) and mean centering, while the Visible data were only mean centered. The noisy regions with also low variance were removed from both datasets. The datasets were explored by means of PCA, with the aim of obtaining information related to the type of process (extraction only vs extraction + hydrolyzation) and its evolution in time. Curiously, the NIR data provided less clear information compared to the Visible data, with which interesting trends could be more easily identified. For this reason, a low-level data fusion approach was also put in place, by directly joining the two spectral datasets, after proper preprocessing and after the application of block scaling to give both blocks the same variance. The detected trends were confirmed, suggesting that both datasets provided useful information that could be efficiently combined and extracted.