D-BLGNet: Dynamic Biaxial Local-Global Alignment Network for Hyperspectral Image Classification

Hyperspectral image (HSI) classification demands joint modeling of fine-grained local spectral-spatial structures and long-range contextual dependencies. While convolution and self-attention provide complementary inductive biases, their fundamentally different weighting mechanisms often lead to poorly coordinated representations, especially when integrated in parallel architectures. To address this issue, this letter proposes a Dynamic Biaxial Local-Global Alignment Network (D-BLGNet), which explicitly aligns convolutional and attention-based feature extraction through dynamic consistency. At its core, a Dynamic Biaxial Feature Extraction Module (DBFEM) jointly employs spectral-mapping depthwise dynamic convolution and centeraware self-attention, enabling adaptive, spectral-dependent local responses that are better aligned with attention-based global aggregation. Furthermore, a Dual-Direction Multi-Scale Feature Fusion mechanism reuses the learned dynamic kernels across multiple dilation rates to capture multi-scale contextual information without introducing additional parameters. Extensive experiments demonstrate that D-BLGNet achieves stable and competitive classification performance compared with representative CNN-based, Transformer-based, and hybrid methods.