mHC-HSI: Clustering-Guided Hyper-Connection Mamba for Hyperspectral Image Classification

A research team led by Yimin Zhu has introduced mHC-HSI, a novel deep learning architecture designed specifically for hyperspectral image (HSI) classification. The model builds on DeepSeek's recently invented manifold-constrained hyper-connection (mHC) approach, which has shown improvements over traditional residual connections, but tailors it for the unique challenges of HSI data. The paper presents three key innovations: a clustering-guided Mamba module for improved spatial-spectral feature learning, a new implementation of the residual matrix that creates soft cluster membership maps for better explainability, and a physically-meaningful approach that divides spectral bands into groups used as parallel streams in the mHC framework.

The technical approach represents a significant advancement in handling HSI's complex, high-dimensional data. By decomposing heterogeneous HSI data into smaller clusters through the novel residual matrix implementation, the model creates interpretable soft cluster membership maps that enhance transparency. The physically-informed spectral band grouping leverages domain knowledge about how different wavelengths interact with materials, making the model more interpretable for applications like environmental monitoring, agriculture, and mineral exploration. Testing on benchmark datasets shows mHC-HSI not only improves classification accuracy over state-of-the-art methods but also provides enhanced model explainability—a crucial factor for scientific and industrial applications where understanding model decisions is as important as accuracy.