Mead0w1ark/megaearth-osm-sentinel2

MegaEarth: an OpenStreetMap × Sentinel-2 dataset

MegaEarth pairs Sentinel-2 Level-2A surface-reflectance windows with area labels from OpenStreetMap (OSM). It was built to test whether an image representation carries structure that transfers to categories it was never fitted on, which needs far more classes than the usual land-cover benchmarks provide. Each window is a 2.56 km footprint stored at native 10 m and 20 m resolution with the scene classification layer. Labels are exact polygon coverages for 485 classes drawn from OSM tags, of which 188 meet support requirements. Each class has a name and, where the OSM wiki supplies one, a mapper-written description. Classes are grouped into families so that whole groups of related categories can be held out, and train, validation and test geography are separated by one-degree cells with a 5 km buffer.

Windows (accepted after quality control)	44,355
Windows retained after a cap of 150 per acquisition-target class	34,992 (train 24,574, validation 3,539, test 6,879)
Classes inventoried / admitted	485 / 188
Families among admitted classes	91
Footprint	2.56 km (256 × 256 px at 10 m); analysis footprint is the central 1.28 km
Bands	B02, B03, B04, B08 at 10 m; B05, B06, B07, B8A, B11, B12 at 20 m; SCL at 20 m
Imagery dates	1 November 2024 to 31 December 2025, one scene per window
Label source	OpenStreetMap planet snapshot of 29 December 2025
Size	188 archives and 11 tables, 22 GB

Example imagery

Nine Sentinel-2 windows from three OSM categories, with three examples per row.

Each tile covers 2.56 × 2.56 km at 10 m resolution; the study labels apply to its central 1.28 km footprint. All nine source windows were checked against their recorded pixel hashes. Select the image to view it at full size.

The gliding-aerodrome examples are outside the 188-class evaluation. Oxbow and recreation ground are included. Labels come from OpenStreetMap; the previews contain modified Copernicus Sentinel data (2024–2025).

Contents

data-00000.tar.zst … data-00187.tar.zst   windows, up to 250 per archive
tables/
  vocab/classes.parquet          485 classes: tag, refinement, name, aliases
  vocab/taxonomy.yaml            primary keys, exclusions, refinements, synonyms, name templates
  select/windows.parquet         46,833 selected windows: location, cell, split, target class, coverage
  assemble/windows_qa.parquet    44,355 accepted windows: scene, source, date, radiometry, quality
  labels_exact/labels.parquet    exact labels and coverages on four footprints
  admit/classes_final.parquet    per-class support and the admitted flag
  admit/windows_final.parquet    retained-window mask
  families/families.parquet      family and fold of every class
  families/folds.json            five family folds and five random class folds
  text/phrases.parquet           class names and OSM-wiki descriptions with source URLs
  text/attribution.json          wiki attribution and licence
manifest.json                    SHA-256 of every file

Each archive is a zstd-compressed tar with one directory per window:

File	Shape	Type	Content
b10.npy	4 × 256 × 256	uint16	B02, B03, B04, B08 at 10 m
b20.npy	6 × 128 × 128	uint16	B05, B06, B07, B8A, B11, B12 at 20 m
scl.npy	128 × 128	uint8	Sen2Cor scene classification layer
metadata.json			window_id, item_id, source, datetime, epsg, x0, y0, dn_offset, dn_scale, baseline, bad_frac, snow_frac, mgrs_tile

Arrays are stored as digital numbers (DN) on the native grid, never resampled. Reflectance is (DN + dn_offset) * dn_scale; every window in this release has offset −1000 and scale 0.0001. x0, y0 is the upper-left corner in metres in the window's UTM zone (epsg), snapped to the 20 m grid so the two resolutions nest exactly. The analysis footprint is rows and columns 64–191 of the 10 m arrays and 32–95 of the 20 m arrays.

Loading

import io, json, tarfile, zstandard, numpy as np, pandas as pd

with open("data-00010.tar.zst", "rb") as f, zstandard.ZstdDecompressor().stream_reader(f) as z, \
        tarfile.open(fileobj=z, mode="r|") as tar:
    for member in tar:
        if member.name.endswith("b10.npy"):
            b10 = np.load(io.BytesIO(tar.extractfile(member).read()), allow_pickle=False)
        elif member.name.endswith("metadata.json"):
            meta = json.load(tar.extractfile(member))

reflectance = (b10.astype("float32") + meta["dn_offset"]) * meta["dn_scale"]

qa = pd.read_parquet("tables/assemble/windows_qa.parquet")       # keyed by widx and window_id
labels = pd.read_parquet("tables/labels_exact/labels.parquet")   # labels_idx holds class_idx values
retained = pd.read_parquet("tables/admit/windows_final.parquet")
classes = pd.read_parquet("tables/admit/classes_final.parquet")

The per-window tables share the integer key widx (0 to 44,354 in window_id order). Classes are keyed by class_idx (0 to 484) with a stable string class_id such as natural=water|water=lake. To reproduce the study's inventory, keep rows with retained == True and classes with admitted == True, and take positive labels from labels_idx on the 1.28 km footprint. Apply the retained-window mask during both fitting and scoring. For a held-out-family fold, exclude from fitting any window whose labels_idx or multi_idx contains a class of a held-out family. Check these exclusions against the full 485-class taxonomy, including classes that fail admission; use the admitted classes for the evaluation gallery. The companion study scores 8,755 image-label pairs from 5,942 retained test windows. The remaining retained test windows lack an eligible scored pair under this protocol.

Classes and labels

A class is one OSM primary tag (landuse, natural, amenity, leisure, man_made, power, aeroway, historic, tourism, waterway, landcover, or place for islands), optionally refined by one secondary tag from a fixed allowlist, for example landuse=farmland|crop=rice or natural=wetland|wetland=mangrove. Administrative and legal tags are excluded. A window is labelled with a class when the union of that class's polygons covers at least 30% of the analysis footprint; classes covering at least 5% are recorded as secondary labels in multi_idx. A window can have several valid labels. Coastline, island and vegetation labels can describe the same footprint at different geographic scales. Coverage measures overlap with a mapped object; the central RGB crop can leave a regional extent, land use or subtype unresolved. Class names come from a fixed template; descriptions are the English OSM-wiki descriptions of the primary and refinement tags, fetched on 17 October 2026 with their page URL and retrieval time.

Families join classes that share a primary tag, an identical name, an alias, or a text cosine similarity above 0.90 under any of three sentence encoders. Families are defined from labels and text only, never from images. Five outer folds hold out whole families; after admission their galleries contain 17, 33, 41, 64 and 33 classes.

A class is admitted when it has at least 30 training, 5 validation and 10 test windows among retained windows, test windows from at least 5 cells and windows from at least 8 cells overall. Admitted classes range from 50 to 3,966 windows with a median of 159; the largest are island, wood and peninsula. The retention cap applies to the class used to acquire a window. Overlapping labels allow an admitted class to occur in more than 150 retained windows. Per-key counts are in classes.parquet and classes_final.parquet.

Splits and geography

A salted hash assigns each one-degree cell to train, validation or test with nominal fractions 0.7, 0.1 and 0.2, so every window in a cell shares its split. Windows of different splits are at least 5 km apart at the stored-footprint edge, and windows of the same class do not overlap. Accepted windows lie in 10,408 cells. Coverage follows OSM mapping: 35,079 windows are north of the equator and 9,276 south; Europe, North America and Asia hold 29,491 windows, Africa and South America 9,113. Each window carries a RESOLVE ecoregion and biome identifier, a Natural Earth continent name and a local mapping-density value for stratified reporting.

Imagery and quality control

Scenes were searched in one season per latitude band: all of 2025 within 23.5° of the equator, May to September 2025 north of that, and November 2024 to March 2025 south of it. Imagery comes from Microsoft Planetary Computer, with Element 84 Earth Search as a fallback for 4,702 windows whose primary catalogue search failed; both serve the same Level-2A products (processing baseline 05.11), and a 20-window pilot check confirmed pixel agreement between them. Scenes were tried in order of cloud and snow cover, at most six per window. A scene was rejected if more than 1% of the analysis footprint's SCL pixels were no-data, defective, cloud shadow, cloud or cirrus, or if any analysis pixel was zero or saturated. Of 46,833 selected windows, 44,355 passed and 2,478 were rejected on quality; there were no technical failures. Snow fraction is recorded, not filtered: 1,558 windows have more than 5% snow pixels.

A class-balanced visual review sampled 376 image-label pairs from the frozen test cohort, with two geographic cells per admitted class and 370 unique windows. One reviewer assessed the central 1.28 km RGB crop: 271 pairs were visually compatible (72.07%), 85 were unresolved (22.61%) and 20 were apparently inconsistent (5.32%). These counts describe the sampled pairs. The sampling design and two pairs per class provide limited precision and do not establish dataset-wide label accuracy.

A subsequent model-assisted source check traced the sampled labels to 1,915 objects in the frozen OSM snapshot. Recomputed polygon coverage agreed with the stored values to within 2.93 × 10⁻⁸. Of the 370 sampled windows, 178 carried multiple canonical mapped labels before admitted-class filtering. This check establishes provenance consistency; it followed the human judgements and provides no independent human annotation. All original review answers were retained. The review left the released images, labels and frozen evaluation cohort unchanged.

Licences and attribution

Component	Licence	Attribution
Labels and study tables (derived from OSM)	ODbL 1.0	© OpenStreetMap contributors
Class descriptions in text/phrases.parquet	CC BY-SA 2.0	© OpenStreetMap Wiki contributors; one source URL per phrase
Image arrays and SCL	Copernicus Sentinel data terms	Contains modified Copernicus Sentinel data (2024–2025)
Ecoregion and biome fields	CC BY 4.0	RESOLVE Ecoregions 2017, Dinerstein and colleagues
Continent field	Public domain	Natural Earth

The release is a derived database of OpenStreetMap data. Anyone who redistributes a further derived database must observe the ODbL share-alike condition.

Limitations

• OSM records what mappers have drawn. A window with no polygon of a class is not evidence that the class is absent: labels are sparse positives, not complete annotations. Polygons may be out of date relative to a 2025 scene.
• Coverage follows OSM mapping density, which is highest in Europe and North America. Per-continent reporting is advisable, and mapping_density lets users stratify by it.
• Classes are imbalanced. Class-balanced scoring is recommended.
• Some classes name a function rather than an appearance, for example college or prison, and are hard to separate at 10 m. Regional labels can extend beyond the crop, while crop species and industrial subtypes may require dated external evidence. An unresolved visual judgement alone is insufficient grounds for deleting or relabelling a pair.
• All imagery is single-date within one growing season per hemisphere. Multi-temporal structure and change are not represented.
• Cell-level splits control leakage between windows. Overlap with the pretraining data of released models remains unknown.
• A readout fitted to category text prototypes introduces category supervision even when the image encoder is frozen and was pretrained on images alone. Its performance measures transfer under that fitted readout. Causal attribution to pretraining requires a matched experimental design.
• The companion study's transfer-calibration gate remains unsatisfied. Current retrieval comparisons are descriptive for the tested cohort and fitting design; population confidence and significance claims require a separately validated uncertainty procedure that covers the intended sources of variation.

Citation

A dataset descriptor and the companion study are in preparation; this section will carry their citations once they are public. Until then, please cite the dataset by its Hugging Face identifier together with the OpenStreetMap, Copernicus and RESOLVE sources above.

@misc{ball2026megaearth,
  author = {Ball, James Edward},
  title  = {MegaEarth: an OpenStreetMap x Sentinel-2 dataset for open-vocabulary semantic evaluation of geospatial foundation models},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Mead0w1ark/megaearth-osm-sentinel2}
}