GMFNet: A GADF–Mamba Fusion Network for Soybean Seed Hyperspectral Classification

Soybean is an important food and oil crop, and rapid nondestructive identification of seed cultivars is crucial for seed purity inspection, varietal certification, breeding management and food-quality control. However, the global spectral profiles of individual soybean seeds from different cultivars are often highly similar, making it difficult for single-representation models to simultaneously capture spectral sequential dependency and inter-band relational structure. To address this issue, this study proposes a GADF–Mamba Fusion Network (GMFNet) for soybean seed hyperspectral classification. Hyperspectral images of 24,800 seeds from eight cultivars were acquired, and reflectance spectra in the range of 900–1700 nm were collected. After preprocessing, 200 effective bands were retained. The preprocessed one-dimensional spectral sequence was fed into a Mamba-based branch to model continuous wavelength dependency and global spectral evolution, while the same sequence was transformed into a GADF image, resized to 208 × 208, and input into a ResNet18-based structural branch to extract inter-band relational features. The two heterogeneous representations were then integrated through a weighted feature fusion module for final classification. Experimental results showed that Mamba achieved the best test accuracy (0.8721) among the raw spectral models, whereas ResNet18 achieved the best test accuracy (0.8737) among the GADF-based structural models. More importantly, the proposed weighted fusion strategy achieved the best overall performance, reaching validation and test accuracies of 0.9039 and 0.9011, respectively. These results demonstrate that spectral sequential information and GADF-based structural semantics are highly complementary. Overall, the proposed framework provides an effective hyperspectral solution for single-seed soybean cultivar identification and shows potential for non-destructive automated quality control in food-industry applications.