Compensating Spatial Degradation in HSI Classification: A Polarity Transformer With LBP-Guided Feature Learning

Hyperspectral images’ (HSIs) classification is inherently challenged by the degraded spatial information in complex scenarios, where existing feature extraction methods fail to realize the consistent compensation from spectral to spatial information. To address this, we propose a novel framework that integrates local binary patterns (LBPs) for texture-guided spatial enhancement and a polarity transformer to model complementary feature representations. The key innovation lies in decomposing the input features into polarity mask through the LBP texture guidance. The positive polarity emphasizes dominant spectral–spatial correlations, while the negative suppresses noise and redundant responses, jointly enabling adaptive compensation of lost spatial details. Simultaneously, multiscale LBP operators are embedded prior to the transformer to explicitly encode rotation-invariant texture features. The original HSI data and LBP-enhanced features are processed by polarity transformers and fused through a cross-polarity interaction module, ensuring the complementary advantages of global spectral context and local texture details. Experimental results on widely used HSI datasets demonstrate that our method achieves superior classification performance compared with state-of-the-art approaches, particularly in low-resolution conditions.