Chemical fingerprinting of rapeseed varieties: A BiLSTM-based fusion strategy for near-infrared and fluorescence spectra.

To achieve rapid and accurate identification of rapeseed (Brassica napus) varieties, this study proposes a classification strategy based on multi-source spectral data fusion and deep learning algorithms. Near-infrared (NIR, 4000-11,550 cm-1) and fluorescence spectra (FS, 400-850 nm) of seven rapeseed varieties were collected, and the performances of five classification models-support vector machine (SVM), random forest (RF), k-nearest neighbor (KNN), convolutional neural network (CNN), and bidirectional long short-term memory (BiLSTM)-were systematically compared. Two fusion strategies-low-level data fusion and feature-level fusion-were employed in combination with preprocessing methods such as multiplicative scatter correction (MSC) to optimize spectral quality and enhance model discriminative capability. The results demonstrated that multi-source spectral data fusion improved classification accuracy, with CNN and BiLSTM models showing the best performance, both achieving an accuracy of 98.41% after MSC preprocessing. Furthermore, the successive projections algorithm (SPA) was applied for characteristic wavelength selection, and when integrated with the BiLSTM-based feature-level fusion framework, the highest classification accuracy of 99.21% was achieved. This study supports the feasibility and potential advantages of NIR-FS fusion for rapeseed variety identification, providing an efficient and rapid spectroscopic approach for agricultural product quality assurance and cultivar authenticity verification.