AD-HKFCM: A Robust Nonlinear Spectral Variability-Aware Unmixing via Intra/Inter-Class Affinity Cohesion

Spectral variability and nonlinear mixing interactions critically degrade spectral unmixing accuracy, especially in heterogeneous environments. To address these challenges, this study proposes a robust nonlinear spectral variability-aware unmixing model, AD-HKFCM, which integrates fuzzy clustering, kernel-driven nonlinear mapping, and intraclass/interclass affinity cohesion. The model introduces a hybrid kernel function combining polynomial and radial basis kernels to enhance linear separability in high-dimensional space. By replacing conventional fuzzy c-means prototypes with support vector data description-derived hypersphere centers, the model reduces dependency on pure pixels and adaptively suppresses outliers through adaptive penalty weight optimization. A physics-informed affinity distance metric is designed to explicitly quantify spectral variability by penalizing intraclass dispersion and amplifying inter-class separation, thereby enabling the precise inference of “virtual pure endmembers” from intimately mixed data. Experiments on simulated (including Orchard 2EM/3EM benchmarks and synthetic hyperspectral) and real satellite datasets demonstrate that AD-HKFCM achieves 5–26% lower abundance estimation errors compared to the best-performing comparative methods, particularly in densely mixed regions with seasonal vegetation variability. This work unifies spectral variability compensation and nonlinear unmixing into a cohesive architecture, offering a generalizable solution for robust unmixing in complex environments.