SpectralZero: Text-Driven Spectral–Spatial Alignment for Zero-Shot Hyperspectral Image Classification

Recently, deep learning-based hyperspectral image (HSI) classification methods have witnessed significant advancements. However, existing approaches predominantly focus on feature modeling for seen classes during the training phase, while the exploration of generalization and transfer mechanisms toward unseen land-cover categories remains limited. As a result, these methods exhibit poor adaptability and generalization performance in zero-shot scenarios. To address this challenge, we propose a novel zero-shot HSI classification framework called SpectralZero that leverages the strong generalization capability of large language models (LLMs) to compensate for the absence of semantic information in unseen classes. Moreover, our framework explicitly captures the synergistic relationship between spatial and spectral features to facilitate knowledge transfer from seen to unseen categories. Specifically, the proposed framework comprises two key modules: Semantic Prompt Enrichment (SPE) and Spectral-Spatial Union Extraction (SSUE). SPE utilizes LLMs to generate fine-grained semantic descriptions of land-cover classes, mitigating the semantic representation deficiency caused by the lack of samples in unseen categories. SSUE introduces two separate branches to independently model spatial structures and spectral reflectance characteristics, thereby enhancing the model’s capacity to extract joint spatial-spectral representations. A major advantage of our method lies in its ability to accurately classify unseen classes without requiring any samples during training or inference. Experiments on multiple public HSI datasets demonstrate that our approach consistently outperforms state-of-the-art methods in zero-shot classification tasks, exhibiting superior recognition accuracy and generalization performance. The code will be available at https://github.com/xiayang124/SpectralZero.