Spectral data fusion combined with lightweight convolutional neural networks for lotus seeds geographical origin...

To address the limitations of single spectroscopic techniques and meet the demand for non-destructive geographical origin traceability of lotus seeds, this study integrated near-infrared spectroscopy (NIRS) and hyperspectral imaging (HSI) data to construct a lightweight convolutional neural network (CNNs) for accurate origin identification of lotus seeds from three core producing regions. We compared low-level and mid-level data fusion strategies, established models with multiple spectral pretreatments and feature selection algorithms, evaluated model robustness via repeated validation and independent external verification, and interpreted the model decision mechanism using gradient-weighted class activation mapping (Grad-CAM). Results showed that mid-level data fusion achieved an overall classification accuracy over 95%, with optimal combinations reaching 100% accuracy. The proposed model significantly outperformed conventional models, with the optimal overall accuracy of 98.68% for support vector machine (SVM) and 98.89% for random forest (RF). This method provides a high-precision, interpretable non-destructive solution for geographical origin identification of agricultural products.