Mamba-guided lightweight capsule network for chemical hazards recognition

Deep learning has demonstrated considerable potential for terahertz (THz) spectral recognition in chemical hazardous substance (CHS) detection because of its ability to learn discriminative spectral representations directly from raw signals. However, the limited availability of large-scale, high-quality THz spectral datasets remains a major obstacle to training reliable models. Moreover, the recognition of covered CHSs in unopened inspection scenarios remains insufficiently explored, as background interference and spectral attenuation substantially increase classification difficulty. To address these challenges, this study proposes an ultra-lightweight THz capsule network guided by the Mamba mechanism (THZ-CM) for hierarchical feature extraction and robust spectral classification. In THZ-CM, hierarchical abstraction is employed to capture multilevel spectral patterns associated with characteristic absorption peaks, while Mamba-derived latent states are projected as adaptive routing priors to improve the organization of capsule features. Two THz absorption spectral datasets were constructed to evaluate the model under pure and covered CHS conditions. Experimental results show that THZ-CM achieves mean accuracies of 92.29% and 70.32% on the pure and covered CHS datasets, respectively, while requiring only 0.8-1.0 million trainable parameters. Further evaluation on an external mid-infrared spectral dataset demonstrates competitive cross-spectral generalization, while ablation experiments confirm the contributions of Mamba guidance, capsule routing, and hierarchical abstraction. Overall, THZ-CM provides a compact and effective framework for THz-based CHS recognition, particularly under covered and more challenging inspection conditions.