Simultaneous detection of citrus internal quality attributes using near-infrared spectroscopy and hyperspectral imaging with multi-task deep learning and instrumental transfer learning

Simultaneous determination of multiple quality attributes of citrus fruits using hyperspectral imaging (HSI) and near-infrared (NIR) spectroscopy and successfully transferring the models among different instruments are two main challenges. In this study, a HSI system and a portable NIR spectrometer were employed to determine the soluble solid content (SSC) and pH value of two varieties of citrus (Quzhou Ponkan and Xiangxi Changye). The single-task and multi-task convolutional neural network (CNN) models for citrus quality inspection were developed. The feasibility of transferring the single-task and multi-task models from HSI to NIR was explored. For the two citrus varieties, the correlation coefficients of optimal models for SSC and pH were over 0.8 and 0.9, respectively. This study demonstrated the potential application of multi-task learning and instrumental transfer learning in citrus quality inspection, which could facilitate the real-world applications of HSI and NIR for accessing the quality citrus and other fruits.