Iformer: Irregular Spatial-Spectral Transformer for Hyperspectral Image Classification

The hyperspectral image (HSI) classification (HSIC) remains challenging due to severe class imbalance dominated by background pixels, thin and edge-adjacent structures with anisotropic context, nonstationary spectral responses across materials, and illumination variability in large HSI cubes. The existing convolutional and transformer-based methods often rely on stationary kernels and uniform spectral sampling, which limits their ability to model irregular spatial-spectral dependencies. This work introduces Iformer, a band-aware irregular spatial-spectral transformer that incorporates coordinate-constrained spectral modeling and geometry-aware spatial propagation. The proposed framework is built upon three core components. First, the irregular spectral aggregation (ISA) models the spectral axis using coordinate-aware anchors with content-gated attention, enabling adaptation to spectral nonstationarity while reducing spectral tokens (S' to R_ref ll S') to control computational complexity. Second, irregular spatial propagation (ISA_spa) introduces a coordinate-driven anisotropic kernel to capture directional spatial context, improving discrimination of fine and structurally complex regions. Third, irregular multihead self-attention (iMHSA) enables heterogeneous subspace allocation across attention heads, allowing the model to adapt its representational capacity to varying spectral-spatial complexities. A unified spatial-spectral embedding supports these components via a 3-D patchifier with channel reweighting and a calibrated CLS+GAP readout to improve prediction stability under class imbalance. Extensive experiments on four widely used benchmarks demonstrate that Iformer consistently achieves superior accuracy.