Simultaneous detection of water content and ferulic acid content in Angelica sinensis using NIR-HSI and multi-task hierarchical attention CNN.

To achieve rapid, nondestructive, and intelligent quality detection of A. sinensis, this study combines NIR-HSI technology with deep learning. It proposes a multi-task convolutional neural network (MT-CNN) equipped with a multi-head hierarchical attention mechanism. First, a multi-task sample partitioning algorithm (MSPXY) was developed with two target variables to improve the representativeness and balance of sample division. Second, a multi-head hierarchical attention mechanism (4HSA + 8HSA) was added to the convolutional neural network to capture both local and global spectral features. Finally, an uncertainty-driven adaptive loss-balancing strategy (UW) was used to achieve dynamic collaborative optimization of multiple tasks. The model that combines these strategies (4HSA + 8HSA + 1DCNN(MSPXY) + UW) achieved the best performance in predicting the moisture and ferulic acid contents of Angelica sinensis. The R2 values reached 0.9925 and 0.9880, and both RPD values exceeded 9. The model performed much better than traditional PLSR and single-task CNN models. This study demonstrates that the method can effectively and accurately predict both the moisture content and ferulic acid content of Angelica sinensis. It provides a rapid, convenient, and non-destructive testing approach for quality control in industrial production. Furthermore, this method offers an efficient and accurate new approach for quality testing in other food products.