Multivariate and machine learning models to assess the heat effects on honey physicochemical, colour and NIR data

We evaluated the effects of pre-processing thermal treatments on the physicochemical, colour and near-infrared (NIR) spectral data of 30 honey samples. The trial was settled as a bi-factorial experimental design that considered nine experimental groups according to the fixed effects of heating treatment and honey phase: none, mild (39 degrees C for 30 ') and high heating (55 degrees C for 24 h) per crystallised, bi-phase and liquid honey samples. Increasing temperatures significantly modified moisture, hydroxymethylfurfural content and lightness. The multivariate classifier models showed that NIR data of warmed crystallised and bi-phase honeys were significantly different from that of the untreated ones, while they sorted a similar assignment for all the liquid samples. The support vector machine model confirmed that the highest tested temperature represented a bias in the informative feature of NIR data, if they would be used in further analytical assessment of the intrinsic qualities of crystallised or bi-phase honey.