Interpretable multimodal fusion of near-infrared spectroscopy and color features for online monitoring of honey refining process.

Monitoring cumulative thermal reactions during honey refining remains challenging due to tightly coupled physicochemical transformations and limited interpretability of single-modality approaches. Here, we propose a physically interpretable multimodal framework integrating near-infrared spectroscopy with low-dimensional image-derived Lab color features for online quantification of glucose, fructose, moisture, pH, and 5-hydroxymethylfurfural (5-HMF). A dataset of 157 samples from seven industrial refining batches was established. Compared with single-spectral models, multimodal fusion significantly improved predictive performance, particularly for 5-HMF, achieving R2 up to 0.98. Notably, PLS outperformed nonlinear models under strong collinearity and continuous reaction progression. Variable importance and SHAP analyses revealed strong physicochemical consistency: key wavelengths corresponded to established OH and carbohydrate absorption regions, while decreasing L⁎ values consistently promoted higher predicted 5-HMF. These findings demonstrate that integrating molecular and macroscopic thermal information enhances both predictive robustness and mechanistic interpretability, providing a practical paradigm for intelligent monitoring of complex thermal processing systems.