Hyperspectral Image Classification via Cross-Domain Few-Shot Learning With Multilevel Style and Semantic Enhancement

One of the key challenges in cross-domain few-shot hyperspectral image classification lies in effectively leveraging spectral-spatial features while mitigating semantic inconsistencies between domains. To address this, we propose a Multi-Level Style and Semantic Enhancement (MSS-CFSL) framework, which enhances both low-level style features and high-level semantic representations to improve domain adaptation. Specifically, MSS-CFSL dynamically aligns spatial and spectral features using 2D spatial and 3D spectral style information from the source domain, ensuring better feature consistency in the target domain. Furthermore, a learnable class-level shared representation is introduced to bridge the semantic gap, enhancing the model’s ability to transfer discriminative knowledge under limited semantic annotations. Additionally, MSS-CFSL integrates a spectral-spatial knowledge fusion mechanism with adaptive feature enhancement, balancing the extraction of structural and spectral band information. Extensive experiments on four hyperspectral datasets show that MSS-CFSL improves the overall accuracy on the four datasets by 11.15%, 12.82%, 5.04%, and 4.42%, respectively, when compared to eight state-of-the-art cross-domain few-shot methods, underscoring its effectiveness in cross-domain few-shot scenarios.