SECCT: spectral-enhanced cycle-consistent transformer for hyperspectral image classification with few labeled samples

In recent years, transformer-based methods have demonstrated strong potential in the classification of hyperspectral images (HSIs). However, most existing transformer models face challenges in HSI classification under scenarios with few labeled samples, such as insufficient focus on discriminative features, reliance on single-modal feature extraction, and lack of local spatial information. To address these limitations, we propose a spectral-enhanced cycle-consistent transformer termed “SECCT.” Unlike conventional transformers, SECCT incorporates a cycle-consistent transformer module to emphasize critical discriminative features. In addition, a spectral enhancement module is designed to extract complementary spectral features, enhancing its ability to capture fine-grained spectral patterns. A two-dimensional convolution module is also introduced for local spatial feature extraction to enhance classification performance. Comprehensive experiments on four benchmark datasets (XZ, LK, YC, and HHK) demonstrate that SECCT outperforms several state-of-the-art methods. Specifically, it achieves the highest overall accuracy (OA), average accuracy (AA), and Kappa coefficient on the XZ and YC datasets—on YC, it exceeds the second-best method by 2.05% in OA, 1.55% in AA, and 1.79% in Kappa. Even under extremely few training samples, SECCT maintains superior or competitive performance across all datasets, validating its robustness for HSI classification.