Diffusion-Inspired Multisource Meta-Learning for Cross-Domain HSI Classification

Abstract

Cross-domain hyperspectral image (HSI) classification remains challenging in realistic deployments, where distribution shifts caused by sensor characteristics, acquisition conditions, and scene variability often coincide with scarce target-domain annotations. While domain adaptation (DA) and few-shot learning have achieved progress, point-to-point transfer tends to overfit a specific source–target pair, and multisource training may suffer from collaborative conflicts and unstable optimization. In this article, a multisource collaborative framework is presented for cross-domain few-shot HSI classification by coupling meta-learning with diffusion-inspired progressive perturbation. Specifically, episodic tasks are constructed from multiple source domains to learn transferable metric structures. A perturbation-conditioned feature extractor is then trained by modeling domain shift as progressive noise injection in the feature space via a diffusion-style forward process, where timestep conditioning guides the backbone to maintain stable class geometry across varying perturbation intensities. Here, “diffusion-inspired” refers only to progressive feature-space noise injection derived from the forward diffusion process and used as an easy-to-hard robustness curriculum; no reverse denoising process or generative diffusion model is constructed. During training, a prototype-anchored objective is adopted to jointly preserve class discriminability and encourage interdomain distribution alignment; moreover, a momentum-based prototype update is introduced to reduce high-variance fluctuations under limited samples. After pretraining, the feature extractor is frozen, and target-domain classification is performed by computing prototypes from a few labeled samples and applying nearest-neighbor inference. Experiments on multiple target datasets demonstrate consistent improvements over representative adaptation and few-shot baselines under low-label settings.