The rapid growth of Earth science observation and simulation data has made efficient data classification increasingly challenging, particularly under conditions of limited annotation resources and continuously evolving data semantics. Conventional classification methods rely heavily on large-scale labeled datasets, which are costly to construct and difficult to adapt to dynamic classification systems. This paper proposes a hierarchical classification framework for Earth science data that leverages large language models (LLMs) and explicitly incorporates hierarchical label relationships to constrain model inference and enhance classification consistency across complex, domain-specific semantic spaces. The framework further integrates retrieval-augmented generation (RAG) and knowledge graph (KG) techniques to introduce external domain knowledge and explicit semantic constraints, enhancing contextual understanding, interpretability, and adaptability to semantic evolution. A benchmark dataset with a two-level hierarchical label structure is constructed based on official NASA metadata. Experimental results demonstrate that by integrating few-shot learning and label space optimization strategies, the proposed framework steadily outperforms various baseline methods in hierarchical classification tasks. Compared with the Bert-BiLSTM model, it achieves an absolute improvement of 8.68% in Micro-F1 and 29.92% in Macro-F1 on the overall hierarchical paths. The framework demonstrates clear advantages in long-tailed data distributions, particularly for minority classes, highlighting its potential for scalable annotation and efficient management of large-scale Earth science datasets.
Zhao et al. (Sat,) studied this question.