Hi-RWKV: Hierarchical RWKV Modeling for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image (HSI) classification requires models that capture complex spatial and spectral information.
• Current methods like CNNs and Transformers face limitations in scalability, receptive field, and computational complexity.
• Developing robust HSI classification models for large scenes under limited supervision remains a challenge.

Purpose of the Study:

• To propose Hi-RWKV, a novel hierarchical recurrent weighted key-value framework for hyperspectral analysis.
• To address limitations of existing models in capturing long-range spatial relations and high-dimensional spectral structures.
• To enable efficient and scalable HSI classification with improved accuracy and robustness.

Main Methods:

• Introduced a spatial structure-guided bidirectional propagation mechanism with edge-aware gating for global context integration and boundary fidelity.
• Developed a spectral identity-driven channel mixing module using learnable band embeddings and whitening transforms for enhanced cross-band discrimination.
• Implemented a multi-stage hierarchical encoder with strictly linear complexity for progressive refinement of spectral-spatial representations.

Main Results:

• Hi-RWKV achieved state-of-the-art accuracy across four benchmark datasets under various training conditions.
• Ablation studies validated the complementary contributions of each module to boundary preservation, spectral discrimination, and data efficiency.
• The model demonstrated superior performance in large-scale HSI interpretation and high-resolution remote sensing.

Conclusions:

• Hi-RWKV offers an efficient and scalable paradigm for hyperspectral image classification.
• The framework effectively unifies scalable recurrence with hyperspectral-specific structural modeling.
• The proposed approach advances the field of high-resolution remote sensing analysis.