Detection model transfer of apple soluble solids content based on NIR spectroscopy and deep learning

Transfer and updating of near infrared (NIR) spectroscopy model of fruit internal quality has become the focus of the industrial application. Internet of Things (IoT) and deep learning (DL) were proposed to perform soluble solids content (SSC) model transfer of apple by NIR. A model transfer platform including low-power handheld internal quality terminal and interacting cloud data system had been constructed. An autoencoder (AE) neural network model was developed for the spectral correction and model transfer. The average time for transmitting detection results to the detection terminal was 1.5 to 2.0 s, with a 100% effective transmission rate. After 5000 iterations of training, the correlation coefficient of different detection terminals improved by 55%, and the root mean square error was reduced by 94%. Selected samples from the second batch of apples detected by the No. 1 detection terminal were added to the original neural network for training. After adding 30 samples, the correlation coefficient increased by 13% and the root mean square error decreased by 90%. The results demonstrated that the AE neural network for spectral correction was effective in eliminating differences between devices and significantly reducing the impact of different detection terminals on the accuracy of NIR detection of SSC in apples. Therefore, the NIR detection model transfer technique could be practically exploited for fruit quality control assessment using different detection terminals.