Intelligent identification of processing endpoints and quality evaluation of Typhae Pollen using multimodal sensing and machine learning

Current discrimination of the processing degree of Typhae pollen (TP) relies on empirical methods and lacks objective metrics. Nondestructive, rapid, and accurate endpoint determination is essential for TP’s standardization and intelligent development. This study investigated dynamic intelligent sensory and Vis–NIR spectral changes of TP during processing. Machine learning models using single-source and fused data were developed for endpoint discrimination and validated using a zebrafish intestinal hemorrhage model. With increased processing, volatile components and color parameters decreased significantly. Single-source colorimeter and Vis–NIR models both achieved high classification accuracy (>96%), with no significant differences among different algorithms. Moreover, colorimetric and Vis–NIR data individually exhibit strong inherent class separability, enabling accurate and robust identification of processing degrees. Feature-level fusion may provide complementary information, but under the present experimental conditions it did not demonstrate a statistically significant advantage over the best single-source models. Pharmacodynamic studies showed that SP had significant hemostatic activity. Moreover, the efficacy ranking (SP ≫ raw TP/LP ≫ HP) was consistent with the processing-stage differences captured by the optimal colorimeter and Vis–NIR models without pharmacodynamic training. Colorimeter and Vis–NIR with machine learning enable rapid, non-destructive, accurate identification of TP processing degrees. The proposed workflow provides a reference framework for process optimization, quality standardization, and future online monitoring in the industrial processing of plant resources.