Determination of multiple bioactive components in Chrysanthemum morifolium (Hangbaiju) using hyperspectral imaging and task-incremental learning

Chrysanthemum morifolium (Hangbaiju), as a flower tea, contains various bioactive components, exhibiting numerous pharmacological effects. Establishing a rapid-adaptation model for multiple components quantification is challenging. This study utilized hyperspectral imaging (HSI) and machine learning (ML) to quantify 28 Hangbaiju components. When modeled on individual bioactive components, Partial Least Squares Regression (PLSR) showed superior overall prediction, whereas deep learning (DL) excels on specific components. For some components, Residual Predictive Deviation (RPD) values of testing set shifted by more than 6. Task-incremental learning (task-IL) was applied to achieve continual learning for various regression tasks. The model rapidly adapted to new tasks by adding and training only task-specific modules within an already trained base network, while maintaining performance on previous prediction tasks. A linear relationship exists between the performance of the task-IL model and its base network, highlighting the importance of the feature extraction and cross-task generalization capabilities of base network in task-IL. This study demonstrates the potential of DL in the field of quality detection of flower teas and provides a framework for integrating HSI and DL in handling real-world scenarios across various fields. • Chrysanthemum bioactive components were analyzed using hyperspectral imaging • Advanced machine learning predicted multiple food components from spectral data • PLSR excelled in predicting individual bioactive compounds overall • Multi-task models enabled simultaneous prediction of key food components • Task-incremental learning was promising for new food component prediction tasks