RSRCC: A Remote Sensing Regional Change Comprehension Benchmark Constructed via Retrieval-Augmented Best-of-N Ranking

A new benchmark called RSRCC has been constructed to push remote sensing analysis beyond simple change detection into fine-grained, localized semantic reasoning through question-answering [1][2]. The dataset, detailed in a paper posted to arXiv, contains 126,000 questions designed to require reasoning about specific semantic changes in paired satellite images [1][2]. The RSRCC dataset is split into 87,000 training instances, 17,100 validation instances, and 22,000 test instances [1][2]. Unlike prior remote sensing change captioning datasets that describe overall image-level differences, RSRCC is built around localized, change-specific questions [1][2]. The authors state this is the first remote sensing change question-answering benchmark designed explicitly for such fine-grained reasoning-based supervision [1][2]. To construct the dataset, the researchers introduced a hierarchical semi-supervised curation pipeline [1][2]. The process begins by extracting candidate change regions from semantic segmentation masks, then initially screening them using an image-text embedding model, and finally validating them through retrieval-augmented vision-language curation with Best-of-N ranking as a critical final ambiguity-resolution stage [1][2]. This approach enables scalable filtering of noisy and ambiguous candidates while preserving semantically meaningful changes [1][2]. The paper was submitted to arXiv, the open-access repository of electronic preprints that has served the physics, mathematics, and computer science communities since 1991 [6]. arXiv hosts papers after moderation but without peer review, and as of November 2024, receives approximately 24,000 submissions per month [6]. The RSRCC paper was first submitted on April 22, 2026, with a file size of 47,430 KB, and a revised version was posted on June 14, 2026, at 131,507 KB [1]. The dataset itself is available on Hugging Face at the google/RSRCC repository [2]. The paper's arXiv page also features integration with arXivLabs, a framework launched in 2020 that allows community collaborators to develop and share experimental tools directly on article record pages [4]. These tools, which include the Bibliographic Explorer for navigating citation trees and the CORE Recommender for discovering related open-access papers, operate under guidelines that require partners to share arXiv's values of openness, community, excellence, and user data privacy [4][5].