Classification of unsound maize kernels using hyperspectral imaging and a dual-domain Mamba network

Accurate identification of unsound maize kernels is critical for reliable grain quality evaluation and safety management during storage. However, spectral variability among kernel components and complex morphological traits present significant challenges for conventional inspection methods in fine-grained recognition. In this study, we present a Dual-Domain Mamba Network (DDM-Net) for the identification of unsound maize kernels using hyperspectral imaging. A dedicated hyperspectral dataset, SS-M-Dataset V1, was also constructed to support model training and systematic evaluation. DDM-Net encodes hyperspectral data along both spectral and spatial dimensions to obtain structured spectral-spatial representations. The spectral branch incorporates a Hierarchical Feature Extraction Module (HFEM) to enhance inter-band correlation modeling and hierarchical spectral feature extraction. The spatial branch employs a Frequency-Aided Mamba Block (FMB) to capture long-range spatial dependencies and global contextual information. A Multi-level Feature Fusion Module (MFFM) is further introduced to integrate spectral and spatial features for cross-domain representation learning. Final classification is achieved via adaptive hybrid pooling followed by a fully connected layer. Experimental results on SS-M-Dataset V1 demonstrate that DDM-Net achieves an overall accuracy of 99.5%, precision of 99.1%, and a Kappa coefficient of 99.3%, outperforming the second-best method by 0.6%, 0.7%, and 0.5%, respectively. These results confirm the effectiveness and application potential of DDM-Net for fine-grained grain kernel identification and provide a benchmark dataset for future hyperspectral seed analysis research.