An Anthocyanin Prediction Model of Blueberry Pomace Based on Stacked Supervised Autoencoders

Based on the visible and near-infrared reflectance spectroscopy technique, stacked supervised autoencoders (SSAE) in deep learning were used to model the anthocyanin content of blueberry pomace. First, preprocessing and feature screening for spectral data were performed. With the minimum value of prediction set root mean square error (RMSEP) of the preset SSAE model as the standard, 178 characteristic wavelengths were selected. The absorbance of the selected characteristic wavelength was used as the input to the SSAE model. The anthocyanin content of blueberry pomace was used as the output. By exploring the activation parameters, node number, training times and learning rate of the SSAE model, the optimal parameters of SSAE were obtained, namely, the activation function of rule, the structure of 178-60-5-1, the training times of 70, and the learning rate of 0.01. The training set root mean square error (RMSEC), prediction set root mean square error (RMSEP), and prediction set correlation coefficient (Rp) were selected as the evaluation criteria. The RMSEC, RMSEP, and Rp of the established model were 1.0500, 0.3835, and 0.9042, respectively. Compared with the classic regression prediction model extreme learning machine (ELM), least squares support vector regression (LSSVR) and partial least squares regression (PLSR) algorithm, the prediction accuracy of the SSAE model was higher. Therefore, the combination of the SSAE model with visible and near-infrared reflectance spectroscopy proved to be effective in predicting anthocyanin content of blueberry pomace.