A research benchmark wants satellite change detection to answer "what changed" and "why", not just "where".
JL1-CC&QA extends an existing dataset of 5,000 bi-temporal image pairs captured by China's Jilin-1 satellite at 0.5-0.75m resolution. The researchers layered two new annotation types on top of the original binary change masks: 17,021 change captions describing land-cover shifts, and 20,060 question-answer pairs spanning eight question categories. Annotations were generated by multimodal large language models, judged by a second vision-grounded LLM, then verified by human experts — a three-stage pipeline designed to keep quality high without relying entirely on manual labor.
Most satellite change detection systems still output a pixel mask: this area changed, that area did not. That answers the localization problem but leaves analysts to interpret meaning themselves. A benchmark that ties the same image set to structured language — captions, follow-up questions — creates a testbed for models that could eventually explain urban sprawl, deforestation, or infrastructure damage in plain terms, not just highlight it in red.
The dataset is public on GitHub, which is the right move; closed benchmarks in remote sensing have a history of limiting reproducibility. Whether the LLM-assisted annotation pipeline introduced systematic blind spots is a question the community will need to stress-test.