SDTN and TRN: Adaptive Spectral-Spatial Feature Extraction for Hyperspectral Image Classification

Hyperspectral images (HSIs) classification has become a foundational tool in the field of precision agriculture, enabling the extraction of detailed information critical for applications such as monitoring crop conditions, identifying plant diseases, and assessing soil characteristics. Nevertheless, conventional approaches are often constrained by the inherently high dimensionality of hyperspectral data, redundancy across spectral-spatial domains, and the limited availability of annotated training samples, resulting in less-than-optimal performance. In response to these limitations, we propose a novel Self-Adaptive Tensor-Regularized Network (SDTN), which integrates tensor decomposition with regularization strategies, allowing for dynamic adjustment of tensor rank in accordance with data complexity. This adaptive mechanism facilitates more expressive and compact feature representations. Expanding upon SDTN, we further introduce the Tensor-Regularized Network (TRN), a lightweight structure that incorporates the spectral-spatial features derived from SDTN and enhances them through multi-scale processing. TRN not only achieves high classification accuracy but also reduces computational complexity, making it highly suitable for real-time deployment in resource-constrained conditions. Experiments on the PaviaU benchmark datasets demonstrate improvements in accuracy and reduced model parameters compared to state-of-the-art methods.