Non-destructive prediction of sucrose, proline, ash, and fructose/glucose ratio in date syrup using hyperspectral imaging and machine learning
Mohammad Hossein Nargesi, Kamran Kheiralipour
Abstract
Date syrup is one of the date fruit by-products that is nutritious rich in antioxidants and has numerous applications in the food industry. Measuring chemical compositions through laboratory methods is destructive and requires high cost and skilled operators. The aim of this research is to predict the chemical compositions of date syrup using hyperspectral imaging as a new, nondestructive, fast, and simple technique. Syrup samples were prepared and the values of sucrose, proline, ash, and fructose/glucose ratio were measured. The hyperspectral imaging system captured the emitted light from the samples within the wavelength range of 400–950 nm and stored it as hyperspectral images. To process these images, an algorithm was developed in MATLAB software. Principal component analysis was used to identify the most informative wavelengths. After extracting features from the image channels at these selected wavelengths, efficient features were selected and prediction was carried out using partial least squares regression, support vector regression, and artificial neural networks methods. The prediction accuracies of the compositions by artificial neural networks (99.99, 100, 99.99, and 100 %, respectively) were higher than partial least squares regression (98.98, 97.25, 98.98, and 96.70 %, respectively) and support vector regression (98.09, 98.92, 98.95, and 72.20, respectively) methods. The results of the present research proved the high ability of hyperspectral imaging and neural networks to estimate the chemical compositions of date syrup. • Near infrared hyperspectral imaging was used to predict compositions of date syrup. • Sucrose, proline, ash, and fructose/glucose ratio were the predicted compositions. • Prediction was done based on the features extracted from effective image channels. • The accuracy of artificial neural network model was higher than other models. • The prediction accuracies were in the range of 99–100 %.