Advanced interpretable data-driven Bi-Mamba model for accurate classification of soybean seeds via hyperspectral...

The rapid and accurate identification of soybean seed varieties is of critical importance for agricultural production and seed quality assurance. This paper proposes a hyperspectral soybean seed classification method based on the Bi-Mamba bidirectional state-space model to further improve identification efficiency and accuracy. Hyperspectral imaging technology is employed to acquire spectral data from 12 soybean varieties, and classification is performed using the Bi-Mamba model, which leverages a bidirectional state-space architecture to effectively enhance the modeling of long-range dependencies in spectral sequences. In addition, a gradient-aware weighting mechanism combined with a sliding window strategy is introduced to adaptively reweight spectral responses, thereby enhancing the representation of locally discriminative spectral features. Furthermore, SHAP-based interpretability analysis is incorporated to quantitatively evaluate the contribution of different spectral bands to the model output, thereby enhancing model transparency and interpretability. Experimental results demonstrate that the Bi-Mamba model achieves excellent performance in the soybean seed cultivar classification task. Under the 2× data augmentation condition, the model achieves a classification accuracy of 99.87% on the calibration set and 98.61% on the prediction set, indicating strong generalization capability. Moreover, SHAP-based interpretability analysis reveals that the model primarily focuses on characteristic spectral bands concentrated in the 1300-1600 nm region, which are closely associated with OH, NH, and C-H-related absorption features in soybean seeds, ensuring consistency between classification decisions and key spectral characteristics. Overall, this method offers a fast, reliable, and non-destructive solution for soybean variety identification.