From Generation to Adaptation: Gradual Diffusion Transitions for Cross-Domain Hyperspectral Image Classification
Abstract
Cross-domain hyperspectral image (HSI) classification via unsupervised domain adaptation (UDA) is practical for Earth observation, yet its performance is often limited by severe distribution shifts. Existing methods typically seek a one-step nonlinear mapping in a highly nonconvex optimization space, which risks representation collapse and semantic distortion. To address this issue, we propose a diffusion-guided manifold evolution adaptation (Diff-MEA) framework. Instead of using generative models merely for data augmentation, the Diff-MEA reformulates the UDA as a continuous manifold evolution process that decouples macroscopic domain gaps into tractable microshifts via gradual diffusion transitions. To facilitate this evolution, we introduce a synergistic semantic-boundary coevolution (SSBCE) mechanism, which alternates between a semantic anchoring (SA) subphase to penalize generative drift and a boundary adaptation (BA) subphase to accommodate target-shifted manifolds. To realize these phases, a memory-transformer domain adapter (MTDA) integrates salient peak response selection (SPRS) and local feature mean aggregation (LFMA) to construct persistent memory tokens, which serve as robust semantic anchors and support dynamic decision boundary expansion. Extensive experiments on three benchmark datasets demonstrate that the Diff-MEA consistently outperforms the state-of-the-art methods in handling severe domain gaps. The code is available at https://github.com/B-Xi