Meta CHM v2 (ml3) — Cloud-Native Companion

A cloud-native repackaging of Meta's DINOv3 Global Canopy Height Map v2 (ml3): a global ~1.19 m canopy-height raster (213,109 Web-Mercator tiles, ~23.8 TB of COGs).

This package does not copy the 23.8 TB of pixels. Instead it adds the three things the original distribution lacks for streaming + analysis, and references Meta's COGs in place:

1s3://us-west-2.opendata.source.coop/tge-labs/meta-chm-v2/
2├── tiles.parquet            # GeoParquet 1.1 tile index (213,109 rows, bbox-indexed)
3├── stac/
4│   ├── collection.json      # STAC Collection (CC-BY-4.0)
5│   ├── items.parquet        # stac-geoparquet — one Item per tile, assets → Meta's S3
6│   └── items_sample/        # ~200 sample STAC Item JSONs (spec inspection)
7├── zarr/
8│   └── chm.zarr.icechunk    # VirtualiZarr GeoZarr (Icechunk), multiscales L1–L6,
9│                            #   zero-copy byte-range refs into Meta's COGs
10├── README.md
11└── LICENSE
1s3://us-west-2.opendata.source.coop/tge-labs/meta-chm-v2/
2├── tiles.parquet            # GeoParquet 1.1 tile index (213,109 rows, bbox-indexed)
3├── stac/
4│   ├── collection.json      # STAC Collection (CC-BY-4.0)
5│   ├── items.parquet        # stac-geoparquet — one Item per tile, assets → Meta's S3
6│   └── items_sample/        # ~200 sample STAC Item JSONs (spec inspection)
7├── zarr/
8│   └── chm.zarr.icechunk    # VirtualiZarr GeoZarr (Icechunk), multiscales L1–L6,
9│                            #   zero-copy byte-range refs into Meta's COGs
10├── README.md
11└── LICENSE

What each artifact is for

CRS & grid

• EPSG:3857 (Web Mercator) — matches the source; zero reprojection for web maps.
• 10-character Bing/Microsoft quadkey tile grid (zoom 10). Native pixel ≈ 1.19 m at the equator.
• Pixel values are canopy height in meters (uint8, 0–255). 0 is both true-zero and no-data (the source sets no explicit nodata mask).

Quickstart

Tile index (DuckDB)

1INSTALL spatial; LOAD spatial;
2SELECT quadkey, tile_size_bytes
3FROM 's3://us-west-2.opendata.source.coop/tge-labs/meta-chm-v2/tiles.parquet'
4WHERE bbox_3857.minx < 1000000 AND bbox_3857.maxx > 0;  -- AOI filter
1INSTALL spatial; LOAD spatial;
2SELECT quadkey, tile_size_bytes
3FROM 's3://us-west-2.opendata.source.coop/tge-labs/meta-chm-v2/tiles.parquet'
4WHERE bbox_3857.minx < 1000000 AND bbox_3857.maxx > 0;  -- AOI filter

STAC — serverless, cloud-native search

This is a static STAC catalog: a collection.json plus a stac-geoparquet items.parquet (213,109 Items). No STAC API server is required. The parquet is written in 54 quadkey-sorted (Z-order) row groups with per-group bbox statistics, so a spatial query reads only the matching byte ranges over HTTP — e.g. a city-scale bbox touches ~3 of 54 row groups, skipping ~94% of the file. Point any of these tools straight at the public URL:

https://data.source.coop/tge-labs/meta-chm-v2/stac/items.parquet

Collection metadata (pystac):

1import pystac
2c = pystac.Collection.from_file(
3    "https://data.source.coop/tge-labs/meta-chm-v2/stac/collection.json"
4)
1import pystac
2c = pystac.Collection.from_file(
3    "https://data.source.coop/tge-labs/meta-chm-v2/stac/collection.json"
4)

Search the Items — pick your tool, all stream + prune row groups:

1# rustac — a pystac-client-like search, in-process, no server
2from rustac import DuckdbClient
3items = DuckdbClient().search(ITEMS_URL, bbox=[13.0, 52.0, 13.4, 52.3], max_items=100)
1# rustac — a pystac-client-like search, in-process, no server
2from rustac import DuckdbClient
3items = DuckdbClient().search(ITEMS_URL, bbox=[13.0, 52.0, 13.4, 52.3], max_items=100)

1-- DuckDB (httpfs range reads + row-group pruning)
2INSTALL spatial; LOAD spatial;
3SELECT id, assets FROM read_parquet('https://data.source.coop/tge-labs/meta-chm-v2/stac/items.parquet')
4WHERE bbox.xmin <= 13.4 AND bbox.xmax >= 13.0 AND bbox.ymin <= 52.3 AND bbox.ymax >= 52.0;
1-- DuckDB (httpfs range reads + row-group pruning)
2INSTALL spatial; LOAD spatial;
3SELECT id, assets FROM read_parquet('https://data.source.coop/tge-labs/meta-chm-v2/stac/items.parquet')
4WHERE bbox.xmin <= 13.4 AND bbox.xmax >= 13.0 AND bbox.ymin <= 52.3 AND bbox.ymax >= 52.0;

1# pyarrow.dataset — predicate pushdown over the remote file
2import pyarrow.dataset as ds, pyarrow.compute as pc
3t = ds.dataset(ITEMS_URL, format="parquet").to_table(
4    filter=(pc.field("bbox", "xmin") <= 13.4) & (pc.field("bbox", "xmax") >= 13.0)
5)
1# pyarrow.dataset — predicate pushdown over the remote file
2import pyarrow.dataset as ds, pyarrow.compute as pc
3t = ds.dataset(ITEMS_URL, format="parquet").to_table(
4    filter=(pc.field("bbox", "xmin") <= 13.4) & (pc.field("bbox", "xmax") >= 13.0)
5)

Also works: geopandas read_parquet(..., bbox=...), Polars scan_parquet, and the stac-geoparquet library. For teams that specifically need the pystac-client API (Client.open(...).search(...)), serve items.parquet behind stac-fastapi-geoparquet — no database required.

Search → xarray (odc.stac): turn found Items into a lazily-loaded, dask-backed xarray cube reading the COGs directly (EPSG:3857, no reprojection):

1import os, pystac, odc.stac
2from rustac import DuckdbClient
3
4os.environ.update(AWS_NO_SIGN_REQUEST="YES", AWS_REGION="us-east-1")  # Meta's COGs are anonymous
5bbox = [13.30, 52.45, 13.45, 52.55]
6items = [pystac.Item.from_dict(d) for d in DuckdbClient().search(ITEMS_URL, bbox=bbox)]
1import os, pystac, odc.stac
2from rustac import DuckdbClient
3
4os.environ.update(AWS_NO_SIGN_REQUEST="YES", AWS_REGION="us-east-1")  # Meta's COGs are anonymous
5bbox = [13.30, 52.45, 13.45, 52.55]
6items = [pystac.Item.from_dict(d) for d in DuckdbClient().search(ITEMS_URL, bbox=bbox)]

See examples/search_and_read.py and examples/odc_stac_load.py in the chm-zarr repo.

Virtual GeoZarr (xarray + Icechunk)

1import icechunk, xarray as xr
2prefix = "s3://dataforgood-fb-data/"
3cfg = icechunk.RepositoryConfig.default()
4cfg.set_virtual_chunk_container(
5    icechunk.VirtualChunkContainer(prefix, icechunk.s3_store(region="us-east-1", anonymous=True))
6)
7repo = icechunk.Repository.open(
8    icechunk.s3_storage(
9        bucket="us-west-2.opendata.source.coop",
1import icechunk, xarray as xr
2prefix = "s3://dataforgood-fb-data/"
3cfg = icechunk.RepositoryConfig.default()
4cfg.set_virtual_chunk_container(
5    icechunk.VirtualChunkContainer(prefix, icechunk.s3_store(region="us-east-1", anonymous=True))
6)
7repo = icechunk.Repository.open(
8    icechunk.s3_storage(
9        bucket="us-west-2.opendata.source.coop",

Provenance & license

CC-BY-4.0. Underlying data © Meta / Data for Good; repackaging by Taylor Geospatial Engine Labs. See LICENSE. Cite Tolan et al. (2024), https://arxiv.org/abs/2304.07213.

The VirtualiZarr store contains only chunk references (offsets/lengths) into Meta's public COGs — if Meta's bucket moves, the Zarr reads break by design.