RSRCC: A Remote Sensing Change QA Benchmark via Retrieval-Augmented Ranking
Researchers have introduced RSRCC, a new benchmark for remote sensing change question-answering. It contains 126,000 questions, split into 87,000 training, 17,100 validation, and 22,000 test instances. Unlike prior datasets that describe overall image-level differences, RSRCC focuses on localized, change-specific questions requiring fine-grained semantic reasoning. This is the first remote sensing change QA benchmark designed for such reasoning-based supervision. The dataset was constructed using a hierarchical semi-supervised curation pipeline that employs Best-of-N ranking as a final ambiguity-resolution stage. The work is published on arXiv with ID 2604.20623.
Key facts
- RSRCC is a remote sensing change question-answering benchmark.
- It contains 126,000 questions total.
- Training set: 87,000 questions.
- Validation set: 17,100 questions.
- Test set: 22,000 questions.
- Questions are localized and change-specific.
- First benchmark for fine-grained reasoning in remote sensing change QA.
- Constructed via hierarchical semi-supervised curation with Best-of-N ranking.
Entities
Institutions
- arXiv