EarthRecord holds the reference answers of a benchmark. By requesting access you agree to use the dataset for evaluation and research, not to train models on its questions or reference answers, and not to redistribute the reference answers outside this repository. The satellite, reanalysis and map data keep the licences of their providers, listed on this page.

EarthRecord

EarthRecord evaluates whether multimodal agents can turn four decades of raw, multi-sensor satellite records into reliable land-change analysis. For each question, an agent works in an offline sandbox on the record of one site as distributed by its providers: Landsat, MODIS, VIIRS and Sentinel-2 data with ancillary layers, clipped but with their values unmodified, so data selection, cloud masking, scaling, mapping and cross-sensor harmonization are left to the agent. Answers are scored against reference answers that the agent never sees.

56 sites: 48 core areas of 10 to 30 km on a side and 8 administrative regions, 1984 to 2026.
238 questions over ten analysis tasks: 153 computable from the record, 73 asking what is on the ground, and 12 with insufficient evidence.
Two controls: the closed book (the question without the record) and 22 instrument pairs, in which one satellite's data are altered while nothing changes on the ground, with 8 placebo pairs.

Paper: https://arxiv.org/abs/2611.12934

Code, the evaluation sandbox and the results of twelve models: https://github.com/ranranrannervous/EarthRecord

Path	Content
`questions.jsonl`	268 question versions: the 238 questions and the altered and placebo versions of the pairs
`references.jsonl`	Reference answers of the 238 questions
`sites/<site>.tar`	The record of each site: the files an agent sees at `/data` (56 archives, 678 GB in all)
`altered/<question>-<version>.tar`	Files replaced in an altered or placebo view
`altered/<question>-<version>.removed.txt`	Files of the site view left out of that view
`checksums/<view>.jsonl.gz`	Path, size and SHA-256 of every file of each of the 86 views

Download and unpack with the script in the code repository, which builds the altered and placebo views and checks every file against its checksum:

python scripts/download_data.py                        # everything
python scripts/download_data.py --ids a14-5,r-gao-4    # only the views these questions need

Questions

Each line of questions.jsonl is one question version:

Field	Meaning
`id`	Question id
`version`	`original`, `altered` (instrument pair) or `placebo`
`pair`	The original question whose reference scores this version
`view`	The view (directory of files) the agent works on
`question`	Question text given to the agent
`location`	The site's bounding box, given instead of the record in the closed book
`fields`	Answer fields; each one also accepts `insufficient_evidence`
`canary`	Canary string of the question
`site`, `theme`, `task`, `class`	Site, theme, analysis task, and class (computable, recognition, insufficient_evidence)
`site_role`	famous, stable, obscure or regional
`reference_source`	method, record_product, withheld_product, visual or observations
`no_change_reference`	Whether the reference conclusion is no change
`period_years`, `landsat_sensors`, `difficulty`	Period length, Landsat sensors spanned, difficulty label
`alteration`	For altered and placebo versions: the altered satellite and, for scaled data, the scale and shape

Reference answers

Each line of references.jsonl holds a question's conclusion rule and its fields with the reference value, the numeric range or accepted alternatives, and an absolute tolerance where one applies. A run is right when its conclusion field earns full credit; a question is solved when all three passes are right (pass^3). The scoring code is earthrecord/scoring.py in the code repository.

References come from global products (Hansen Global Forest Change, JRC Global Surface Water, GHSL built-up surface and MCD64A1 burned area are withheld from the record), from agreement among methods computed on the record, and from visual interpretation of the imagery.

Data sources and licences

The views hold data of the following collections. Each keeps its provider's licence; cite the products when you use them.

Collection in the views	Product	Provider	Terms and attribution
`landsat-c2-l2`	Landsat Collection 2 Level-2 surface reflectance and temperature	USGS	Public domain. "Landsat data courtesy of the U.S. Geological Survey."
`sentinel-2-l2a`	Sentinel-2 MSI Level-2A	ESA, Copernicus	Copernicus Sentinel data terms (free and open). "Contains modified Copernicus Sentinel data [2019-2026]."
`modis-13Q1-061`, `modis-13A1-061`	MOD13Q1, MYD13Q1, MOD13A1, MYD13A1 v061 vegetation indices	NASA LP DAAC	NASA Earth science data, no restrictions on use; cite the product DOIs.
`modis-11A2-061`	MOD11A2, MYD11A2 v061 land surface temperature	NASA LP DAAC	As above.
`modis-10A1-061`	MOD10A1, MYD10A1 v061 snow cover	NASA NSIDC DAAC	As above.
`static/MCD12Q1`	MCD12Q1 v061 land cover	NASA LP DAAC	As above.
`VNP13A1`, `VNP21A2`	VIIRS VNP13A1 and VNP21A2 v002	NASA LP DAAC	As above.
`era5-land-monthly`	ERA5-Land monthly averaged data	ECMWF, Copernicus Climate Change Service	Licence of the Climate Data Store dataset. "Contains modified Copernicus Climate Change Service information [2026]."
`static/esa-worldcover`	ESA WorldCover 10 m 2020 v100 and 2021 v200	ESA WorldCover consortium	CC BY 4.0. "© ESA WorldCover project / Contains modified Copernicus Sentinel data (2020, 2021) processed by ESA WorldCover consortium."
`static/cop-dem-glo-30`	Copernicus DEM GLO-30	ESA, Copernicus	Copernicus DEM licence. "Produced using Copernicus WorldDEM-30 © DLR e.V. 2010-2014 and © Airbus Defence and Space GmbH 2014-2018 provided under COPERNICUS by the European Union and ESA; all rights reserved."
`static/geoboundaries`	geoBoundaries gbOpen	geoBoundaries	Open licences of each boundary file, mostly CC BY 4.0.

The questions, reference answers, checksums and this card are released under CC BY 4.0.

Benchmark data

Do not use EarthRecord to train models. Every question carries a canary string (field canary, prefix ERC-) so that its presence in training corpora can be detected.

Citation

@misc{zhang2026earthrecord,
  title         = {{EarthRecord}: From Pixels to Land Change with Multimodal Agents over Four Decades of Satellite Imagery},
  author        = {Zhang, Haoran and Sun, Huizhen and Wen, Hanbin and Leng, Jie and Zhu, Jiaxin and Sun, Zhaohua},
  year          = {2026},
  eprint        = {2611.12934},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2611.12934}
}

Downloads last month: 30

Total file size:

689 GB

Paper for haoranranranran/EarthRecord

EarthRecord: From Pixels to Land Change with Multimodal Agents over Four Decades of Satellite Imagery

Paper • 2611.12934 • Published 6 days ago • 4

You need to agree to share your Hugging Face information to access this dataset