Spacecraft Target Visual Classification Dataset Based on LLM Prompt Annotation

The accuracy of fine-grained recognition and multi-attribute labeling for spacecraft target imagery heavily depends on dataset completeness and annotation precision. This study presents the Spacecraft Target Visual Classification Dataset, along with an optimized prompt-based annotation algorithm leveraging Large Language Model (LLM) to tackle these challenges. The dataset includes six fundamental attributes, including model, mission, components, geometry, illumination, and confidence, covering nine categories of spacecraft targets while providing detailed attribute-level annotations for complex on-orbit objects. To improve annotation accuracy and address issues in partial-attribute recognition, we propose an optimized prompt-based annotation method based on In-Context Learning (ICL), incorporating few-shot examples, Chain-of-Thought (CoT) reasoning, and Self-Consistency (SC) mechanisms. Experiments conducted on public spacecraft datasets demonstrate that our approach yields an average increase of 14.0% in Per-Image Average Accuracy (PIAA) and 21.1% in All-Attributes-Correct Accuracy (AACA), underscoring its adaptability to diverse scenarios. By significantly reducing reliance on manual labeling, this method enables precise and scalable spacecraft target classification, highlighting its broad applicability in space visual tasks.