A transfer learning method for near infrared models of potato starch content and traceability from different origins

The robustness of near-infrared (NIR) models in detecting agricultural product quality is challenged by the differences in statistical conditions such as territory, variety, and collection time, and geographical differences in sample sources. This study aimed to use global and local migration models to improve the generalization ability of prediction models of potato starch content from different sources. The results showed that both models could eliminate the influence of samples from different sources on the model performance; the transfer component analysis (TCA)-based model was superior to the global model. The correlation coefficient (R_P), root-mean-square error of prediction (RMSEP), and relative percent deviation (RPD) of the prediction model of starch content in the target domain of Whale Optimization Algorithm (WOA)–Radial Basis Function (RBF) based on the TCA method reached 0.931, 0.763 %, and 2.740, respectively. After the second model transfer, the model still had an extremely reliable performance (RPD = 2.050 > 2). The precision and accuracy of the WOA–RBF traceability model reached 91.25 % and 95 %, respectively. This study provided a feasible solution to the problem of poor generalization ability of a single-source model and proposed an effective, stable, and universal method for nondestructive testing of potato traceability.