Rapid identification of the geographical origins of crops using laser-induced breakdown spectroscopy combined with transfer learning

Abstract

Identification of the geographical origins of crops using deep learning-assisted laser-induced breakdown spectroscopy (LIBS) can quickly realize food traceability and guarantee the interests of consumers. However, this technique is not suitable for practical application when the number of training samples is limited and has poor transferability. In this study, a transfer learning-assisted LIBS method was developed to identify the geographical origins of crops, which achieved a maximum accuracy of 93.81% among six different transfer combinations. To further improve the identification accuracy, Deep Adaptation Networks (DAN) was applied for the first time and demonstrated improved performance on five cases. Finally, feature visualization confirmed that the LIBS information could be transferred to other crops. Our results show that transfer learning-assisted LIBS can enhance crop traceability in cases with limited numbers of samples. The study provided new idea to identify geographical origin under sample-limited conditions.