Standard-free sample model transfer of visible/near -infrared spectral model of apple ripeness under seasonal variation

The robustness of visible/near-infrared spectral models in fruit quality assessment is challenged by differences in measurement conditions from instruments, environment, and season. Model transfer is considered to be an important method to solve such problems, however, model transfer of apple ripeness classification models under seasonal variation is difficult to achieve by measuring standard samples. Two model transfer methods without standard samples were implemented for this purpose: dynamic orthogonal projection (DOP) and transfer component analysis (TCA). Model transfer was accomplished on one source domain dataset and two target domain datasets under different seasons. The results show that both DOP and TCA can effectively improve the classification performance of the model, with DOP improving the precision and recall by up to 19.7 % and 40 %, and TCA improving the precision and recall by up to 25 % and 60 %. t-distributed Stochastic Neighbor Embedding (t-SNE) results after model transfer proved effective in reducing differences between datasets by DOP and TCA, and external validation ensures the robustness and generality of both methods. In conclusion, the apple ripeness classification model under seasonal variation can be accomplished by model transfer without relying on standard samples, in which TCA is expected to be a general tool for eliminating seasonal differences in apple NIR spectra.