Data fusion of HSI and FHSI combined with improved deep learning models to identify the geographical origin of Lycium barbarum

Traditional methods for identifying the geographical origin of Lycium barbarum ( L. barbarum ) often have exhibited limited sensitivity and specificity, rendering them inadequate for meeting the increasingly stringent quality control standards required for traditional Chinese medicinal materials. Therefore, this study proposes an innovative approach that integrates Hyperspectral Imaging (HSI) and Fluorescence Hyperspectral Imaging (FHSI) to construct a deep learning-based traceability model through multi-source spectral data fusion, enabling high-precision identification of L. barbarum from different geographical origin. Specifically, Isolation Forest (IF) was employed for outlier detection alongside three different combinations of spectral preprocessing methods. Classification was then performed using Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), and Bidirectional Long Short-Term Memory (BiLSTM) models. Furthermore, an improved Sparrow Search Algorithm (SCSSA) was proposed to optimize the hyperparameters of the BiLSTM model. Experimental results showed that the SCSSA-BiLSTM model based on data-level fusion achieved the best classification performance, with training and testing accuracies of 96.5% and 93.6%. Overall, the study demonstrated the feasibility and effectiveness of combining HSI and FHSI spectral fusion with SCSSA-optimized deep learning for the geographical origin identification of L. barbarum , offering methodological support for the rapid traceability of agricultural products. ● Deep fusion of HSI and FHSI enables origin identification of L. barbarum. ● IF removes outliers and significantly improves classification performance. ● Combined preprocessing and SCSSA-optimized BiLSTM improve accuracy. ● Samples from China and Canada improve model generalization.