Geometric algebra transformer with spectral-spatial-volumetric feature extraction for hyperspectral image classification

• Weight-free multivector embedding of spatial, spectral, and volumetric data. • Geometric Algebra Transformer enabling relational reasoning in hyperspectral data. • Compact and interpretable geometric representation with reduced model complexity. • State-of-the-art accuracy with fewer parameters than baseline models. • Opens Geometric Algebra Transformer to multimodal fusion and broader sensing tasks. Current approaches to hyperspectral imaging classification have achieved important results, mostly depending on convolution-based preprocessing that adds complexity to the approach. At the same time, these methods often do not directly integrate all the data’s spatial, spectral, and volumetric dependencies at once. This article introduces Hyperspectral Geometric Algebra Transformer (H-GATr), an approach that addresses these limitations by using a Geometric Algebra (GA) architecture to integrate hyperspectral information into multivector representations, without additional trainable parameters. Thus, exploiting the intrinsic geometric structure of the data is possible to obtain an efficient integration of all hyperspectral sample dependencies. Experiments carried out on various benchmark datasets demonstrate that H-GATr is capable of achieving performance comparable to and even superior to current models, offering a compact solution for remote sensing applications.