TEBS: Temperature–Emissivity–Driven band selection for thermal infrared hyperspectral image classification with structured State-Space model and gated attention

Thermal infrared hyperspectral images (TIR-HSIs) provide unique spectral insights that are often unattainable with visible imagery, making them invaluable for applications such as land cover classification and geological mapping. However, the high spectral redundancy in TIR-HSIs often leads to increased computational complexity and potential performance degradation. To address this issue, this paper proposed an unsupervised temperature–emissivity–driven band selection method (TEBS) for TIR-HSIs classification, which integrated a structured state-space model (SSM) and a gated attention mechanism (GAM). Specifically, a feature extraction (FE) module is firstly designed to separate land surface temperature (LST) and land surface emissivity (LSE) information, incorporating superpixel segmentation to extract multi-scale LST features. Subsequently, a weight computation (WC) module, leveraging SSM and GAM, is developed to generate robust band weights by sequentially leveraging multi-scale LST features. Finally, a band evaluation (BE) module is employed to assess the band selection results and optimize the model parameters. Experimental comparisons conducted on two datasets using three classic classifiers show that TEBS framework outperforms state-of-the-art (SOTA) methods in classification accuracy. These results underscore the potential of TEBS to advance land cover classification in thermal infrared hyperspectral imaging.