Deep Learning-Based Honey Adulteration Classification: A Comparative Study with Hyperparameter Tuning and Explainable AI

Honey adulterated with cheap sugar syrups continues to threaten consumer confidence in one of the most precious natural products globally. However, methods to detect such fraud can be slow, expensive, or destructive. This study developed an explainable deep-learning framework for distinguishing pure from adulterated honey using two complementary and openly available datasets: a large synthetic physicochemical dataset (D1; 247,903 observations, eight variables) and a visible-near-infrared (Vis-NIR) hyperspectral dataset of real honey samples (D2; 8,675 observations, 128 bands). A one-dimensional convolutional neural network (1D-CNN) and a bidirectional long short-term memory network (Bi-LSTM) were independently optimised for each dataset using Optuna and then evaluated on held-out test data using accuracy, precision, recall, F1-score, and ROC-AUC. The 1D-CNN outperformed the Bi-LSTM on both modalities. On D1, it reached 98.49% accuracy, an F1-score of 0.9881, and a ROC-AUC of 0.9994, compared with 92.42%, 0.9417, and 0.9832 for the Bi-LSTM; on D2, it reached 99.08% accuracy and a ROC-AUC of 0.9987, compared with 93.16% and 0.9846. SHAP analysis of the physicochemical model ranked pH and colour score as the dominant predictors, followed by density and water content. Two caveats temper these numbers. D1 is synthetically generated; therefore, the D1 results demonstrate that the framework recovers a known nonlinear decision boundary rather than detecting adulteration in real honey. Moreover, all four models were evaluated using internal validation only. Therefore, the framework is best understood as a reproducible, interpretable template for rapid authenticity screening, which requires external validation on independently collected honey spanning a range of botanical origins, adulterant types, and adulteration levels.