Artificial neural network analysis of VIS-NIR spectra in food and agriculture.

Optical spectroscopy in the visible and near-infrared range (VIS-NIR) has recently become increasingly important in food, agricultural and forestry sciences due to instrumental small footprint, little or no sample preparation and large penetration depth (up to few mm). Due to the difficult interpretation of VIS-NIR optical spectra, the signal is usually elaborated by means of multivariate statistics and machine learning algorithms (chemometrics). Albeit accurate, these methods are usually non-transparent and it is often difficult to trace back the discriminant frequencies characterizing two different classes of samples. In this talk, recent results obtained by the research group at CREA-IT in this field using artificial neural networks, will be presented. These algorithms appear accurate and robust. Moreover, thanks to a simple and accurate study of the features involved, it is possible to "open the box" and get a clear interpretation of the discriminating features involved.