With the increasing availability of satellite imagery and the shortening revisit intervals, efficiently processing satellite hyperspectral images has become a critical task. However, in practice, a large portion of satellite hyperspectral data remains unlabeled, making it difficult to achieve satisfactory classification performance using satellite data alone. Meanwhile, UAV-based platforms offer acquisition flexibility, which facilitates the collection of rich and detailed information. To address these challenges, this paper proposes a method called Sharpness-Aware Minimization with Local-to-Global Feature Enhancement (SAMLFE), which uses UAV hyperspectral images for training to enhance the fine-grained classification performance of satellite hyperspectral images in large scenes. Specifically, a spectral dimension mapping model is first employed to unify UAV and satellite images into a common spectral dimension, thereby mitigating the impact of inconsistent feature representations. Next, a local-to-global feature extraction network is constructed to capture both local details and global semantics. Few-shot learning is applied to extract discriminative features from both the source and target domains within the shared feature space, thereby enhancing the model’s ability to utilize limited labeled data efficiently. Furthermore, a conditional adversarial domain adaptation strategy is adopted to align the feature distributions of the source and target domains, thereby alleviating spectral shift. Meanwhile, the integration of an improved Sharpness-Aware Minimization (ISAM) enhances the model’s robustness across domains. Finally, the K-Nearest Neighbor algorithm is employed to perform accurate classification. Experimental results on multiple datasets demonstrate that the proposed method achieves superior generalization and classification performance in cross-domain hyperspectral image classification. It also outperforms existing methods in terms of feature distribution alignment, robustness of feature extraction, and adaptability to small-sample scenarios.
Hyperspectral image classification typically assumes that training and test data belong to the same set of categories. However, in practical applications, unknown categories of ground objects frequently arise during the testing phase, a scenario known as open-set classification. In open-set classification, the high spectral similarity between known and unknown classes makes it challenging to effectively capture subtle differences in the frequency domain using only spatial or low-frequency features. This difficulty in delineating the decision boundary limits the effective rejection of unknown classes. We propose an open-set classification method for hyperspectral images that combines frequency domain feature enhancement with contrastive learning. First, we construct a wavelet transform enhancement model to decompose spectral features into high- and low-frequency components, capturing subtle spectral differences for refined representation. Second, we design a multi-scale feature fusion ResNet that assigns prior importance to the convolution kernel and recalibrates it using weighted convolution. This enhances the model’s ability to capture fine-grained spatial features and improves the distinction of category boundaries. Finally, we combine dynamic hard sample mining with contrastive learning. Through a progressive strategy that ensures model training stability, we achieve deeper mining and accurate identification of difficult boundary samples, thereby improving the accuracy of unknown category classification. Experimental results on the Houston 2013, GF-5 Yancheng, and Salinas datasets show that this method achieves overall classification accuracies of 90.16%, 97.24%, and 91.07%, respectively. These results outperform mainstream open-set methods, such as DTCL and CACL, fully demonstrating the method’s effectiveness and superiority in open-set classification tasks.