A Hybrid Deep Learning Approach Combining CNN and Transformer for Gas Classification in Visible NIR Hyperspectral Image

Hyperspectral gas classification plays a critical role in chemical threat detection, industrial safety, and environmental monitoring. However, conventional RGB imaging is inherently limited for gas discrimination since many gases appear transparent or indistinguishable in the visible spectrum. In this study, fog, phytoncide, and H2O vapor were used to construct gas environments that cannot be reliably differentiated using RGB imagery alone. In contrast, visible near-infrared (VIS - NIR) hyperspectral imaging reveals distinctive spectral signatures that enable practical gas discrimination. To exploit these characteristics, this paper presents HyMixer, a CNN - Transformer hybrid model for VIS - NIR hyperspectral gas classification that jointly learns spatial and spectral representations while prioritizing efficient full-image inference. Experimental results show that VIS - NIR hyperspectral data provide clear separability among gas types. Compared to a strong baseline model, HyMixer achieves substantially improved inference efficiency, with a slight tradeoff in classification accuracy. This study provides a practical foundation for VIS NIR hyperspectral gas identification in time-sensitive scenarios and can be extended to the identification of hazardous gases, including real chemical, biological, and radiological (CBR) agents, in future applications.