IG19145 - Artificial intelligence methods for spectral data processing to solve food fraud and authenticity issues (SensAIfood)

Non-destructive spectral sensors (NIRS, hyperspectral, Raman…) are analytical technologies of enormous potential and application for food quality control, authentication, and fraud detection along the entire food chain, due to their characteristics of speed, precision, low cost, non-contaminating nature, as well as for providing an interoperable digital signal with high connectivity. Spectral sensor applications require the use of mathematical algorithms for predictive model development, which look for the relationship between spectral and reference information in order to quantify parameters of interest or classify and discriminate products according to their quality characteristics. Traditionally, multivariate approaches based on linear, and, in some cases, non-linear regression strategies are used. However, when the aim is to develop universal models, with very large databases, the use of predictive modelling strategies based on AI (Artificial Intelligence) methods could be a breakthrough for the modelling of food processes and products. On the other hand, when working with non-destructive spectral sensor data, reference data are used in the modelling, but usually not other data that it can be denominated as “context data”, of great interest in the interpretation of the results, such as varietal, climatic, processing, producer, etc. data. The combined use of spectral data, reference data and context data for modelling by means of automatic learning, deep learning and AI with neural networks can be considered an essential advance for the development of strategies to increase the efficiency and the power of the control and authentication of products and processes in the food chain.