Semi-DST: A Relation-aware Frequency Transformer with Dual-Student-Teacher Framework for Semi-supervised Hyperspectral Image Classification.

Hyperspectral image (HSI) classification often suffers from the scarcity of labeled samples. Semi-supervised learning (SSL) alleviates this issue by jointly exploiting limited labeled samples and abundant unlabeled data. However, existing semi-supervised hyperspectral image classification methods rarely exploit unlabeled samples to effectively integrate complementary local and global information under limited supervision, resulting in less discriminative feature representations. Moreover, unreliable pseudo-labels may lead to unstable optimization and error accumulation. To address these issues, we propose a Relation-aware Frequency Transformer with Dual Student-Teacher Framework (Semi-DST) for semi-supervised HSI classification. Specifically, a Multi-level Relation-aware Frequency Transformer (MRFFormer) is developed to jointly model local spatial relationships and global frequency dependencies through multi-level relation-aware frequency modeling, thereby learning more discriminative spectral-spatial representations. Furthermore, a Dual Student-Teacher (DST) framework is introduced to enable collaborative learning between two students under the stable guidance of a teacher model, improving the reliability of knowledge propagation. To further enhance pseudo-label learning, a Stable Dual-Competition Learning (SDCL) strategy is proposed to improve pseudo-label reliability through confidenceaware collaborative learning, while a Multi-Scale Wavelet Alignment (MS-WA) module aligns teacher-student representations across multiple frequency scales to facilitate consistent feature learning. Extensive experiments on seven public HSI datasets demonstrate that the proposed Semi-DST consistently achieves superior classification performance compared with state-of-the-art methods under limited supervision. The source code will be available online https://github.com/Yali-W/Semi-DST.