A Novel Dual-Modality Deep Fusion Framework Based on VIS-NIR Hyperspectral Imaging for Soybean Classification

Multimodal spectral fusion methods have been increasingly applied in agricultural product identification, yet most remain confined to simple feature concatenation at the input or decision level and rely on basic voting schemes, which fail to fully exploit the complementary and heterogeneous characteristics of different spectral modalities. To address these limitations, this study proposes a VIS–NIR hyperspectral imaging–based framework for soybean origin classification, using more than 5,000 individual seeds collected from five countries—Argentina, Brazil, China, the United States, and Ukraine. Four innovative fusion strategies—Early-fusion, Deep-fusion, Late-fusion, and Asymmetric-fusion—were implemented within a unified convolutional neural network architecture, with visible (VIS) and near-infrared (NIR) spectral data as inputs. Among the evaluated models, the Deep-fusion strategy achieved the highest accuracy (97.64%), demonstrating superior feature representation and predictive capability; t-SNE visualization further confirmed its effectiveness by showing clear inter-class separability and intra-class compactness in the high-dimensional feature space. Notably, the Asymmetric-fusion strategy incorporates a dynamic confidence-gating mechanism to reinforce cross-modal feature integration, achieving 95.43% accuracy with a computational time of 33.09 s—substantially faster than Deep-fusion—while maintaining high classification stability. Overall, the proposed fusion strategies effectively integrate complementary VIS and NIR information, enhancing the robustness and accuracy of soybean origin classification.