Graph-based clustering for hyperspectral unmixing: a review of spectral graph theory, deep spectral clustering, and scalable architectures

The progression from classical graph partitioning to scalable parametric sequence models has reshaped hyperspectral image (HSI) analysis. This survey systematizes the clustering-to- unmixing pipeline, linking topological manifold learning with sub-pixel fractional-abundance estimation. We organize the screened literature into a hierarchical taxonomy that traces the field from transductive graph cuts to inductive deep frameworks. Across Vision Transformers, diffusion models, and Mamba-based architectures, a recurring design question concerns how one-dimensional sequence processing affects two-dimensional spatial coherence. Results reported on classical flightlines indicate that deep graph models can reduce sensitivity to spectral variability and noise; however, none of the surveyed clustering or unmixing methods reports a corresponding evaluation on HySpecNet-11k. We therefore treat state-space architectures as promising candidate backbones rather than an established solution, and identify protocol- matched unmixing evaluation, cross-sensor validation, and topology-aware sequence design as priorities for scalable Earth-observation workflows.