CoordBench

CoordBench is a benchmark suite for coordinate-only location encoders such as SatCLIP, GeoCLIP, Climplicit, and MIND. It contains 40 normalized source tables (HF configs), one per original source file or table. The accompanying submission evaluates 52 datasets and 78 prediction targets drawn from 33 of these configs; the bt_*, country, and ecoregions configs were added later and are not part of that evaluation.

Load one table

import pyarrow.parquet as pq
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "XXXX/CoordBench",
    "data/california_housing/data.parquet",
    repo_type="dataset",
    token=False,
)
table = pq.read_table(path, columns=["lon", "lat", "MedHouseVal"])
print(table.num_rows, table.column_names)

The remaining configs use the same data/<config>/data.parquet layout.

Schema

Every config's data.parquet starts with these columns:

column	type	notes
lon, lat	float64	WGS84, always non-null
timestamp	int64 (nullable)	unix ms, UTC; null when the source has no per-point time
timestamp_end	int64 (nullable)	end of the valid-time window, when the source defines one (only dm_*)
split	string (nullable)	an official train/val/test partition from the source, null when it doesn't exist
id	string (nullable)	source row identifier, when one exists

The remaining columns are the source columns kept during normalization. Each config corresponds to one source table rather than one target, so sustainbench contains all six DHS indices and cdc_places contains all twelve PLACES measures. Values are stored as released, including nodata sentinels (-999, -1, ocean index 0) and untransformed targets. The filtering and transforms applied at evaluation time are listed below. A raw/ folder alongside data/ contains unmodified copies of the original source files.

Evaluation suite

The paper's suite has 52 datasets and 78 targets in three categories. Regression targets are scored with R², and land-cover targets with accuracy.

category	datasets	targets	metric
Environmental	10	10	R²
Land cover	16	16	accuracy
Socioeconomic	26	52	R²

N is the evaluated sample count after the filtering below and before fold assignment. † marks datasets excluded from some regional splits (see Evaluation protocol).

Environmental

dataset	target	N	coverage	spatial unit	config
MOSAIKS: Elevation	Elevation	100,000	US	1 km point samples	usavars_elevation
MOSAIKS: Tree cover	Tree cover	100,000	US	1 km point samples	usavars_treecover
SatCLIP: Elevation	Elevation	100,000	Global	Point/polygon samples	satclip_elevation
ASTER GED	8.3 µm emissivity	17,636	Land, 60°S–60°N	100 m	dm_aster_ged
OpenET ensemble	Monthly evapotranspiration	35,683	Contiguous US	30 m pixels	dm_openet_ensemble
Air temperature	Air temperature	3,076	Global	Monitoring stations	air_temp
WorldClim bio1	Annual mean temperature	18,068	Global land	10 arcmin grid	worldclim_bio
WorldClim bio12	Annual precipitation	18,068	Global land	10 arcmin grid	worldclim_bio
SoilGrids: Soil organic carbon	Soil organic carbon	4,884	Global land	Raster point samples	soilgrids
SoilGrids: Soil pH	Soil pH	4,884	Global land	Raster point samples	soilgrids

Land cover

dataset	target	N	coverage	spatial unit	config
SatCLIP: Biome	Biome	32,798	Global	Point/polygon samples	satclip_ecoregion
SatCLIP: Country	Country	33,053	Global	Point/polygon samples	satclip_country
SatCLIP: Ecoregion	Ecoregion	32,798	Global	Point/polygon samples	satclip_ecoregion
Africa crop mask	Cropland	2,556	8 African countries	Reference points	dm_africa_crop_mask
Canada crops (coarse)	Crop group	16,079	Canada	Field observations	dm_canada_crops_coarse
Canada crops (fine)	Crop type	14,566	Canada	Field observations	dm_canada_crops_fine
Ethiopia crops †	Crop type	2,530	Ethiopia	10 m-radius field plots	dm_ethiopia_crops
GLanCE	Land cover	34,885	Global	30 m reference samples	dm_glance
LCMAP cover change	Land-cover change	2,320	Contiguous US	30 m reference samples	dm_lcmap_lcc
LCMAP land cover	Land cover	26,510	Contiguous US	30 m reference samples	dm_lcmap_lc
LCMAP land use	Land use	26,513	Contiguous US	30 m reference samples	dm_lcmap_lu
LCMAP use change	Land-use change	991	Contiguous US	30 m reference samples	dm_lcmap_luc
LUCAS land cover	Land cover	203,569	European Union	In-situ observations	dm_lucas_lc
LUCAS land use	Land use	226,858	European Union	In-situ observations	dm_lucas_lu
Oil palm	Plantation type	17,477	Oil-palm regions	Reference points	dm_descals
US trees	Tree genus	45,382	US	Species observations	dm_us_trees

Socioeconomic

dataset	target	N	coverage	spatial unit	config
PDFM CONUS27	27 targets (listed below)	35,157	Contiguous US	Place-level	pdfm_conus27
PLACES: Annual checkup	Crude prevalence (%)	31,435	US	ZCTA centroids	cdc_places
PLACES: COPD	Crude prevalence (%)	31,435	US	ZCTA centroids	cdc_places
PLACES: Cancer (excl. skin cancer)	Crude prevalence (%)	31,435	US	ZCTA centroids	cdc_places
PLACES: Coronary heart disease	Crude prevalence (%)	31,435	US	ZCTA centroids	cdc_places
PLACES: Current asthma	Crude prevalence (%)	31,435	US	ZCTA centroids	cdc_places
PLACES: Current smoking	Crude prevalence (%)	31,435	US	ZCTA centroids	cdc_places
PLACES: Diabetes	Crude prevalence (%)	31,435	US	ZCTA centroids	cdc_places
PLACES: High cholesterol	Crude prevalence (%)	31,435	US	ZCTA centroids	cdc_places
PLACES: Obesity	Crude prevalence (%)	31,435	US	ZCTA centroids	cdc_places
PLACES: Poor mental health	Crude prevalence (%)	31,435	US	ZCTA centroids	cdc_places
PLACES: Poor physical health	Crude prevalence (%)	31,435	US	ZCTA centroids	cdc_places
PLACES: Sleep under 7 hours	Crude prevalence (%)	32,409	US	ZCTA centroids	cdc_places
DHS: Asset index	Asset index	86,936	Survey countries	Cluster centroids	sustainbench
DHS: Sanitation index	Sanitation index	89,271	Survey countries	Cluster centroids	sustainbench
DHS: Under-five mortality	Under-five mortality	105,582	Survey countries	Cluster centroids	sustainbench
DHS: Water index	Water index	87,938	Survey countries	Cluster centroids	sustainbench
DHS: Women's BMI	Women's BMI	94,866	Survey countries	Cluster centroids	sustainbench
DHS: Women's education	Women's education	117,062	Survey countries	Cluster centroids	sustainbench
MOSAIKS: Housing	Housing	54,586	US	1 km point samples	usavars_housing
MOSAIKS: Income	Income	91,377	US	1 km point samples	usavars_income
MOSAIKS: Nightlights	Nightlights	100,000	US	1 km point samples	usavars_nightlights
MOSAIKS: Population	Population	67,968	US	1 km point samples	usavars_population
MOSAIKS: Roads	Roads	100,000	US	1 km point samples	usavars_roads
SatCLIP: Population	Population	74,512	Global	Point/polygon samples	satclip_population
California Housing †	Median house value	20,640	California	Census block groups	california_housing

PDFM CONUS27 targets: high cholesterol, poor physical health, stroke, binge drinking, physical inactivity, annual checkup, cancer excluding skin cancer, diabetes, poor mental health, coronary heart disease, high blood pressure, cholesterol screening, dental visit, asthma, chronic kidney disease, arthritis, COPD, blood-pressure medication among hypertensive adults, obesity, sleep under seven hours, smoking, median household income, median owner-occupied home value, nightlights (log10), population density (log10), tree cover, and elevation.

Evaluation-time filtering and transforms

• Rows with a missing label or coordinate are dropped.
• USAVars (MOSAIKS): rows with the -999 nodata label are dropped; income and housing records are joined to their coordinates by ID; population, nightlights, income, and housing use log(1 + y). Elevation, tree cover, and roads keep their released units.
• SatCLIP: ocean samples are dropped (country index 0; ecoregion and biome index -1). The population target is released on a log scale and used as is.
• PDFM: nightlights and population density are released as log10 values and used as is.
• All other targets keep their released units.
• Source-provided splits (the split column, including the DHS and PDFM splits) are not used; every dataset uses the common folds below.
• The six DHS datasets include deliberate displacement of survey-cluster coordinates, up to 2 km in urban and 5 km in rural areas (Burgert et al., 2013), which limits the interpretation of sub-5 km distance statistics.

Evaluation protocol

• Folds. Points are assigned to latitude–longitude blocks of size c. The unique blocks are permuted with a fixed seed and assigned cyclically to five folds. For random folds, points are assigned to folds at random; regional holdouts use 2°, 10°, 20°, and 40° blocks. Blocks have no exclusion buffer. Results average five fold assignments generated with different seeds.
• Datasets by split. A dataset is excluded from a block size when its median test-to-nearest-training distance is unchanged from the next smaller block size in the sweep (which also includes 5°), or when it has no spatial inner-validation result. Ethiopia crops is excluded at 10°, 20°, and 40°, and California Housing at 20° and 40°.
• Probes. Encoders are frozen. Features are standardized dimension-wise with training-split statistics. Ridge regression is solved in closed form in float64 over penalties from 1e-4 to 1e6 at half-decade steps. Hyperparameters (penalty, and truncation width where applicable) are tuned per target by nested cross-validation, with inner folds of the same type as the outer folds.
• Target scores. R² is computed on each held-out fold against that fold's mean, and accuracy is the fraction of correctly classified held-out points. The five fold scores are averaged per target, and each target's fold-averaged R² is floored at -1. Accuracy is not clipped.
• Macro averages. Within each fold-assignment seed, target scores are averaged within each dataset, and datasets are then averaged equally, separately for regression and classification. PDFM's 27 targets form one dataset score, and each PLACES measure is its own dataset. Reported results are the mean, and sample standard deviation, of the macro scores over the five fold-assignment seeds.
• Class-balanced metrics. For land cover, macro-F1 and balanced accuracy are also reported, computed from the same probes and aggregated the same way.

License and provenance

No single license applies to the whole repo. Most configs are CC-BY-4.0, CC-BY-SA-4.0, or public domain, but several have narrower or unconfirmed terms. Where a license is unspecified or unconfirmed, do not assume permission to redistribute, and check the original source terms before redistribution or commercial use.

config	task	license	source
air_temp	regression	not explicitly confirmed	Station air-temperature observations (Hooker et al., 2018), a SatCLIP evaluation task (source)
bt_bioclim	regression	CC-BY-SA-4.0 (data); MIT (code)	Better Together benchmark (van der Plas et al., 2026) (source)
bt_biomass	regression	CC-BY-SA-4.0 (data); MIT (code)	Better Together benchmark (van der Plas et al., 2026) (source)
bt_cropharvest	classification	CC-BY-SA-4.0 (data); MIT (code)	Better Together benchmark (van der Plas et al., 2026) (source)
bt_human_footprint	regression	CC-BY-SA-4.0 (data); MIT (code)	Better Together benchmark (van der Plas et al., 2026) (source)
bt_landcover	regression	CC-BY-SA-4.0 (data); MIT (code)	Better Together benchmark (van der Plas et al., 2026) (source)
california_housing	regression	public domain	StatLib California Housing (Pace & Barry, 1997), as distributed with scikit-learn (source)
cdc_places	regression	public US federal government data	CDC PLACES, 2023 ZCTA release (CDC, 2023) (source)
dm_africa_crop_mask	classification	CC-BY-4.0	AEF evaluation release (Brown et al., 2025); Sub-Saharan Africa cropland reference data (Kerner et al., 2024) (source)
dm_aster_ged	regression	CC-BY-4.0	AEF evaluation release; ASTER GED (Hulley et al., 2015) (source)
dm_canada_crops_coarse	classification	Open Government Licence – Canada	AEF evaluation release; Annual Crop Inventory ground truth (Agriculture and Agri-Food Canada, 2021) (source)
dm_canada_crops_fine	classification	Open Government Licence – Canada	AEF evaluation release; Annual Crop Inventory ground truth (Agriculture and Agri-Food Canada, 2021) (source)
dm_descals	classification	CC-BY-4.0	AEF evaluation release; global oil-palm reference points (Descals et al., 2024) (source)
dm_ethiopia_crops	classification	CC-BY-4.0	AEF evaluation release; EthCT2020 (Blasch et al., 2024) (source)
dm_glance	classification	CC-BY-4.0	AEF evaluation release; GLanCE training data (Stanimirova et al., 2023) (source)
dm_lcmap_lc	classification	CC-BY-4.0	AEF evaluation release; LCMAP CONUS reference data (Pengra et al., 2023) (source)
dm_lcmap_lcc	classification	CC-BY-4.0	AEF evaluation release; LCMAP CONUS reference data (Pengra et al., 2023) (source)
dm_lcmap_lu	classification	CC-BY-4.0	AEF evaluation release; LCMAP CONUS reference data (Pengra et al., 2023) (source)
dm_lcmap_luc	classification	CC-BY-4.0	AEF evaluation release; LCMAP CONUS reference data (Pengra et al., 2023) (source)
dm_lucas_lc	classification	CC-BY-4.0	AEF evaluation release; harmonized LUCAS surveys (d'Andrimont et al., 2020) (source)
dm_lucas_lu	classification	CC-BY-4.0	AEF evaluation release; harmonized LUCAS surveys (d'Andrimont et al., 2020) (source)
dm_openet_ensemble	regression	CC-BY-4.0	AEF evaluation release; OpenET ensemble (Melton et al., 2022) (source)
dm_us_trees	classification	CC-BY-NC-4.0	AEF evaluation release; iNaturalist research-grade observations (iNaturalist contributors, 2025) (source)
ecoregions	classification	CC-BY-4.0	Ecoregions2017 (Dinerstein et al., 2017) (source)
country	classification	public domain	Natural Earth 1:110m admin-0 countries (source)
pdfm_conus27	regression	Apache-2.0 (source repository); data-file terms not confirmed	PDFM CONUS27 benchmark (Agarwal et al., 2024) (source)
satclip_country	classification	unspecified	SatCLIP evaluation data (Klemmer et al., 2025) (source)
satclip_ecoregion	classification	unspecified	SatCLIP evaluation data (Klemmer et al., 2025) (source)
satclip_elevation	regression	unspecified	SatCLIP evaluation data (Klemmer et al., 2025) (source)
satclip_population	regression	unspecified	SatCLIP evaluation data (Klemmer et al., 2025) (source)
soilgrids	regression	CC-BY-4.0	SoilGrids 2.0 surface (0–5 cm) mean rasters (Poggio et al., 2021) (source)
sustainbench	regression	DHS-derived indices; redistribution terms not confirmed	SustainBench DHS indices (Yeh et al., 2021), via the TorchSpatial LocBench label pack (Wu et al., 2024) (source)
usavars_elevation	regression	CC-BY-4.0	USAVars (Rolf et al., 2021) (source)
usavars_housing	regression	CC-BY-4.0	USAVars (Rolf et al., 2021) (source)
usavars_income	regression	CC-BY-4.0	USAVars (Rolf et al., 2021) (source)
usavars_nightlights	regression	CC-BY-4.0	USAVars (Rolf et al., 2021) (source)
usavars_population	regression	CC-BY-4.0	USAVars (Rolf et al., 2021) (source)
usavars_roads	regression	CC-BY-4.0	USAVars (Rolf et al., 2021) (source)
usavars_treecover	regression	CC-BY-4.0	USAVars (Rolf et al., 2021) (source)
worldclim_bio	regression	non-commercial/academic use only; redistribution requires WorldClim's permission	WorldClim 2.1 bioclimatic variables, 10 arcmin (Fick & Hijmans, 2017) (source)

The dm_* configs come from the AlphaEarth Foundations (AEF) evaluation release. Each row also names the original data source that the release redistributes; the per-task READMEs under raw/deepmind_eval/ give the full attribution.

References

• Agarwal, M., et al. (2024). General geospatial inference with a population dynamics foundation model. arXiv:2411.07207.
• Agriculture and Agri-Food Canada (2021). Annual Crop Inventory Ground Truth Data. Open Government Portal of Canada.
• Blasch, G., et al. (2024). Ethiopian Crop Type 2020 (EthCT2020) dataset: Crop type data for environmental and agricultural remote sensing applications in complex Ethiopian smallholder wheat-based farming systems (Meher season 2020/21). Data in Brief, 54, 110427.
• Brown, C. F., et al. (2025). AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data. arXiv:2507.22291.
• Burgert, C. R., Colston, J., Roy, T., & Zachary, B. (2013). Geographic displacement procedure and georeferenced data release policy for the Demographic and Health Surveys. ICF International.
• CDC (2023). PLACES: Local Data for Better Health, ZCTA Data, 2023 release. Centers for Disease Control and Prevention.
• d'Andrimont, R., et al. (2020). Harmonised LUCAS in-situ land cover and use database for field surveys from 2006 to 2018 in the European Union. Scientific Data, 7, 352.
• Descals, A., Gaveau, D. L. A., Wich, S., Szantoi, Z., & Meijaard, E. (2024). Global mapping of oil palm planting year from 1990 to 2021. Earth System Science Data, 16, 5111–5129.
• Dinerstein, E., et al. (2017). An ecoregion-based approach to protecting half the terrestrial realm. BioScience, 67(6), 534–545.
• Fick, S. E., & Hijmans, R. J. (2017). WorldClim 2: New 1-km spatial resolution climate surfaces for global land areas. International Journal of Climatology.
• Hooker, J., Duveiller, G., & Cescatti, A. (2018). A global dataset of air temperature derived from satellite remote sensing and weather stations. Scientific Data.
• Hulley, G. C., et al. (2015). The ASTER Global Emissivity Dataset (ASTER GED): Mapping Earth's emissivity at 100 meter spatial scale. Geophysical Research Letters, 42, 7966–7976.
• iNaturalist contributors (2025). iNaturalist Research-grade Observations. Occurrence dataset via GBIF.
• Kerner, H., et al. (2024). How accurate are existing land cover maps for agriculture in Sub-Saharan Africa? Scientific Data, 11, 486.
• Klemmer, K., Rolf, E., Robinson, C., Mackey, L., & Rußwurm, M. (2025). SatCLIP: Global, general-purpose location embeddings with satellite imagery. AAAI.
• Melton, F. S., et al. (2022). OpenET: Filling a critical data gap in water management for the western United States. Journal of the American Water Resources Association, 58, 971–994.
• Pace, R. K., & Barry, R. (1997). Sparse spatial autoregressions. Statistics & Probability Letters.
• Pengra, B., et al. (2023). LCMAP CONUS Reference Data Product 1984–2021 land cover, land use and change process attributes. U.S. Geological Survey data release.
• Poggio, L., et al. (2021). SoilGrids 2.0: Producing soil information for the globe with quantified spatial uncertainty. SOIL.
• Rolf, E., et al. (2021). A generalizable and accessible approach to machine learning with global satellite imagery. Nature Communications.
• Stanimirova, R., et al. (2023). A global land cover training dataset from 1984 to 2020. Scientific Data, 10, 879.
• van der Plas, T. L., et al. (2026). Better Together: Evaluating the complementarity of Earth embedding models. arXiv:2605.18667.
• Wu, N., et al. (2024). TorchSpatial: A location encoding framework and benchmark for spatial representation learning. NeurIPS Datasets and Benchmarks Track.
• Yeh, C., et al. (2021). SustainBench: Benchmarks for monitoring the Sustainable Development Goals with machine learning. NeurIPS Datasets and Benchmarks Track.