Filtering by upstream license¶

MillionTrees aggregates dozens of independently published datasets, and they do not share one license. Most are CC-BY 4.0, but a handful are NonCommercial, some are ShareAlike, one forbids derivative works, one is copyleft, and a few are public domain. Whether you may train on a given annotation depends on which of those it came from.

Rather than ask you to reconstruct that mapping from the source table, every loader takes a licenses= argument that keeps only the annotations whose upstream license permits your intended use:

from milliontrees import get_dataset

# Everything you can use in a commercial product.
dataset = get_dataset("TreeBoxes", root_dir="data", licenses="commercial")

# Only the most permissive terms: commercial use, derivatives, no share-alike, no copyleft.
dataset = get_dataset("TreePolygons", root_dir="data", licenses="permissive")

# Or name the licenses yourself.
dataset = get_dataset("TreePoints", root_dir="data",
                      licenses=["CC-BY-4.0", "CC0-1.0"])

What filtering does and does not change¶

It drops annotation rows. It does not re-split anything. An image whose annotations are all dropped simply leaves the dataset; every image that remains keeps the train / validation / test assignment it already had. A license-restricted run is therefore scored on a subset of the same benchmark splits, not on a new split, and per-source numbers stay comparable to the published leaderboard for the sources that survive the filter.

Because sources are dropped whole, a restricted run trains on less data and usually on fewer biomes. Report which selection you used alongside your scores.

Selections¶

licenses= accepts a preset name, a license id, an fnmatch pattern over ids ("CC-BY-*"), or a list of any of those — a list selects the union of what its entries name. licenses=None (the default) applies no filter at all.

Preset	Selects
`all`	every license, including `unknown` — the same as no filter
`known`	every license whose terms are confirmed
`commercial`	commercial use permitted (excludes the NonCommercial licenses)
`noncommercial`	only the NonCommercial licenses
`derivatives`	derivative works permitted (excludes NoDerivatives)
`no-share-alike`	no obligation to relicense derivatives under the same terms
`no-copyleft`	the license does not reach the model or code built from the data
`permissive`	commercial and derivatives and no share-alike and no copyleft
`public-domain`	CC0 only — no attribution obligation
`unknown`	only the sources whose upstream terms are unconfirmed

Licenses currently present in a release:

License	Commercial	Derivatives	Share-alike	Attribution	Copyleft
`CC0-1.0`	yes	yes	no	no	no
`CC-BY-3.0`	yes	yes	no	yes	no
`CC-BY-4.0`	yes	yes	no	yes	no
`CC-BY-SA-4.0`	yes	yes	yes	yes	no
`CC-BY-NC-4.0`	no	yes	no	yes	no
`CC-BY-NC-SA-4.0`	no	yes	yes	yes	no
`CC-BY-NC-ND-3.0`	no	no	no	yes	no
`CDLA-Permissive-1.0`	yes	yes	no	yes	no
`AGPL-3.0-or-later`	yes	yes	yes	yes	yes
`unknown`	—	—	—	—	—

These flags are a filtering aid, not legal advice. Read the license before you rely on it.

Unconfirmed sources¶

A source whose upstream terms we have not confirmed is recorded as unknown, and no preset except all and unknown selects it. Asking for commercial therefore drops it rather than assuming it is safe to use. To include those sources deliberately, ask for them:

get_dataset("TreeBoxes", root_dir="data", licenses=["commercial", "unknown"])

Inspecting what you have¶

Every loaded dataset exposes the licenses of the sources it is holding:

dataset.source_licenses   # {'Cloutier et al. 2023': ['CC-BY-4.0'], ...}
dataset.licenses          # the selection you asked for, or None

To audit a packaged release before loading it:

python -m milliontrees.common.licenses                       # the license menu
python -m milliontrees.common.licenses \
    --csv data/TreeBoxes_v0.24/within-distribution.csv       # per-license row counts
python -m milliontrees.common.licenses \
    --csv data/TreeBoxes_v0.24/within-distribution.csv --licenses commercial

Sources that mix licenses¶

Some sources are not uniform. Two version-controlled tables in src/milliontrees/common/license_data/ handle this:

sources.csv — one license per packaged source name. A mixed source records the most restrictive license it spans, so that a filter never includes a row by default that it should have excluded.
overrides.csv — per-row refinements, matched on source plus an fnmatch pattern over filename. Later rows override earlier ones, and anything no rule matches keeps the source-level license.

A release may also carry a license column, and where it is filled it wins over both tables. That is the only way to describe an image whose own annotations differ, since an fnmatch rule over filename cannot split a single image. Rows that leave it empty fall back to the tables, so a release where only one source fills it behaves exactly as before everywhere else.

Open Forest Observatory¶

OFO field 2025 is the source these tables exist for, and the only one that currently needs them. It is the one place in MillionTrees where the image and the annotation come from different upstream datasets with different owners:

the image tile is cut from an Open Forest Observatory drone mission orthomosaic — every mission in the OFO catalog is CC-BY 4.0;
the points are a contributor’s ground-reference stem map, and those plots span CC-BY 4.0, CC-BY-SA 4.0, CC-BY-NC-SA 4.0, CC0 1.0 and U.S. Forest Service public domain.

Using a packaged row means using both, so the license recorded for it is the most restrictive of the two:

Ground-reference plot	+ CC-BY 4.0 imagery	Tiles
CC-BY 4.0	`CC-BY-4.0`	2,677
U.S. Forest Service public domain	`CC-BY-4.0`	2,583
CC0 1.0	`CC-BY-4.0`	146
CC-BY-SA 4.0	`CC-BY-SA-4.0`	1,426
CC-BY-NC-SA 4.0	`CC-BY-NC-SA-4.0`	968

The CC0 and public-domain plots therefore still carry an attribution obligation — it comes from the imagery, not from the stem map. Attribute the plot’s contributor, not OFO; OFO asks to be credited as the mechanism through which the data were found.

The packaged filename is {mission_id}_ortho_{tile}_OFO_field_2025.png, so most missions need only one rule. Seven missions cover plots under two different licenses, and their tiles are split rule by rule:

source,filename_pattern,license,notes
OFO field 2025,000091_ortho_*_OFO_field_2025.png,CC-BY-SA-4.0,"mission 000091, 15 tiles; ..."
OFO field 2025,000091_ortho_340_OFO_field_2025.png,CC-BY-NC-SA-4.0,"plot 0040 (CC BY NC SA 4.0) ..."

Tiles that straddle a plot boundary¶

Four tiles — 000133_ortho_589, 000133_ortho_590, 000136_ortho_297 and 000136_ortho_298 — carry stems from two adjacent plots at once: plot 0049 (Vibrant Planet, CC-BY-NC-SA 4.0) and plot 0110 (UC Davis / TNC, CC-BY-SA 4.0). The plots are 65 m apart and share no stems, so these are genuinely two surveys of neighbouring ground, not one survey counted twice.

A filename-keyed rule cannot split one image, so in overrides.csv those four tiles take the more restrictive CC-BY-NC-SA 4.0 for all 51 of their stems. That is safe but pessimistic: it withholds 9 CC-BY-SA annotations from a commercial-use run. The license column resolves them exactly — data_prep/process_ofo_field.py stamps every annotation with its own plot’s license as it tiles, so a release built with it keeps all four images and drops only the 42 stems a commercial user may not have. Until TreePoints is next rebuilt, the conservative rules apply.

Regenerating¶

overrides.csv is generated, not hand-edited. data_prep/ofo_license_overrides.py reads the plot licenses from Zenodo record 22731179 (ofo_ground-reference_plots.gpkg, column license_short), rebuilds the tiling from each orthomosaic’s georeferencing to work out which plot each tile came from, and refuses to write the table unless the rebuilt tile set reproduces the packaged release exactly:

python data_prep/ofo_license_overrides.py \
    --validate /orange/ewhite/web/public/MillionTrees/TreePoints_v0.25

Rerun it whenever the OFO catalog is updated or the OFO tiling changes.

Young et al. 2025 unsupervised and Young et al. 2025 weak supervised are also OFO, but they need no overrides: they are drone imagery with machine-derived labels, so CC-BY 4.0 covers them end to end.

The same mechanism covers any future source that needs sub-source granularity: add rows to overrides.csv, no loader change needed.