Application of instance-based transfer learning for quantitative detection of honey adulteration

Honey adulteration poses a significant challenge to market quality regulation, as spectral variability among honey varieties limits the stability of single-domain models. Raman spectral data of syrup-adulterated lychee honey were used as the source domain to investigate cross-honey-type adulteration quantification. Three categories of transfer learning approaches were constructed and compared, including instance-based transfer learning models (TrAdaBoost + SVR, TrAdaBoost + RF, KMM + SVR, and KMM + RF), a parameter-based transfer learning model (TL-CNN), and a feature-based transfer learning model (DANN). The results show that KMM + SVR exhibits the best predictive performance across all target domains, achieving an average test-set R² of 0.9778, RMSE of 0.0202, and MAE of 0.0149, outperforming TL-CNN, DANN, and non-transfer models. Additional validation using acacia honey and Manuka honey adulterated with multifloral honey demonstrated the potential applicability of the proposed approach. Overall, the results demonstrated the potential of transfer learning for cross-domain honey adulteration quantification, although further validation under more diverse conditions is required.