A transfer learning method for spectral model of moldy apples from different origins

In this paper, two methods of compensation model and transfer learning model were employed to improve the model adaptability of moldy apple core from different origins. The spectral data of apples from two different regions (Fufeng and Lingbao) were obtained. Based on the partial least squares-discriminant analysis (PLS-DA) model and the least squares-support vector machine (LS-SVM) model, the local model of each origin, the global model and the transfer component analysis (TCA) model were established. The results showed that both the compensation model and TCA-based model could eliminate the influence of origin on model performance. Compared with the compensation model method, the specificity and accuracy of the LS-SVM model based on the TCA method using data from Lingbao origin increased by 9.09% and 4.54%, respectively. An external verification confirmed the theoretical results. This study sheds light on a practicable solution for the poor adaptability of single-origin model, and presents a reliable and general method for the spectral detection of moldy apple core. •Model performance is sensitive to sample origin.•Compensation model can eliminate the influence of origin on model performance.•Transfer component analysis can solve the model adaptability for different origins.•Transfer component analysis is superior to compensation model in model performance.