EXP-10 · Land·Observed vs declared
Eyes in the Sky
Every hemp figure in the world today is a number someone declared. This experiment turns declared hectares into observed hectares, field by field, and puts a public counter on a pledge: one million hectares of hemp on the record by 2035.
COPERNICUS · FREE & OPENBACKGROUND IS A SYNTHETIC TILESynthetic
01The scale of the gap
A million hectares is almost six times all the hemp on Earth
The pledge sounds like a mapping problem. It is first an arithmetic problem. The world grew about 173,000 hectares of hemp in 2024; open government registries already publish about 15,000 hectares as mapped parcels, waiting to be read.
▬Drawn to scale against one million hectares
On a straight scale the open registries are 1.5% of the million and the whole world crop is 17.3%. The log view shows how many orders of magnitude separate today's zero from the pledge.
02The record ladder
Four ways a hectare can be known, and only three of them count
Companies are graded by evidence and products by provenance tier. Fields get the same ladder. A satellite can raise a question about a field; it cannot put that field on the record. An inference is never shown as a fact.
0Rung 0 · Observed
A model reading Sentinel imagery thinks this parcel behaves like hemp this season.
- Source
- Classifier output with a score
- Shown
- Analysts only. The public sees regional totals with error bars.
1Rung 1 · Declared
A government registry lists the parcel as hemp: a CAP aid claim, a licence, a THC notification.
- Source
- Open parcel data (GSAA / LPIS), cited with date and licence
- Shown
- As the registry publishes it, never more identifying
2Rung 2 · Claimed
The grower or their cooperative signs in and confirms the field belongs to a company on Hemp'in.
- Source
- Org claim plus a field boundary drawn or imported
- Shown
- On the company page, at the grower's choice of precision
3Rung 3 · Traced
A harvest from that field is linked to a lot, and the lot to a product passport.
- Source
- Origin chain stages, passport v0
- Shown
- On the passport: “grown on field EE-…, 2027 season”
A satellite can point. Only a record can count.
03Screen A · Atlas ▸ Fields
Observed versus declared, one pilot region at a time
Estonia is the calibration region: its registry is the best ground truth in Europe, every hemp parcel, grower, variety and area, open since 2008. Grand Est is the transfer test, where the hemp corridor lives. Select a parcel, move the model threshold, then switch to the public view to see what a stranger would see. Parcels on this tile are synthetic; the regional headline figures are sourced.
▦Atlas ▸ Fields ▸ Estonia ▸ 2026 season
◆Estonia · registry headline
- Declared hemp parcels
- 850
- Named growers
- 109
- Declared area
- 8,652.5 ha
- Main varieties
- Estica, CFX-2, Finola
PRIA GSAA open WFS, 2026 season, queried live. Sourced.
473 ha agree · 258 ha declared only · 148 ha observed only
- Area
- 33.6 ha
- Hemp score
- 0.83
- Declared crop
- Estica · fibre/dual
- Record tier
- Declared
Greens in late May, closes canopy fast in June, peaks in July and drops at the August cut. Matches the declared crop.
04How it would work
The labels already exist. Governments publish them as a side effect of farm aid.
The expensive part of any crop classifier is ground truth. For hemp in northern Europe it is free: six registries publish crop-coded parcels openly because CAP rules require farmers to declare every field. Each was verified live. A hemp classifier is trainable without a single field visit.
Declared parcels
EE named WFS, LV CC0, FI CC BY 4.0, DK, NL (772 fibre-hemp parcels, 2025), Wallonia (28, 2023). Every label keeps its source URL and date.
Sentinel-2 + Sentinel-1
Optical bands including red-edge and SWIR at 10–20 m, every 5 days. Radar fills the cloudy weeks of a Baltic June.
The season's shape
Per parcel: when it greens, how fast, how high it peaks, when it is cut. Hemp's timing is its fingerprint.
Parcel time-series classifier
A simple baseline first, then a temporal deep model. Trained in one country, tested blind in another.
Area with error bars
Regional hectares from good-practice sample-based area estimation, never from summing pixels.
∿The season's shape · hemp against a look-alike
◉At this date
First ten-day window where the gap stays above 0.2: 7 Jun. Whether real fields behave like this is research question Q3. Clouds in a northern June can hide exactly those weeks.
!What makes hemp hard to see
- Look-alikes. Maize, sorghum and sunflower are tall summer crops sown at a similar time. The confusion matrix against maize is the number that decides the project.
- Three kinds of hemp. Fibre is cut in August, grain and dual-purpose in September, and CBD plots are small, often under cover. The pilot targets field-scale fibre and grain only.
- Transfer. A model trained on Estonian Estica fields may fail on French Futura in a different climate. That gap is the research.
- Clouds. The weeks that separate hemp from maize are often cloudy in the north. Radar may or may not be enough.
?Research questions the partner answers
- Q1. Within one country, how well can a parcel be called hemp from one season of imagery?
- Q2. How much of that holds when the model crosses a border it was not trained in?
- Q3. How early in the season does the call become reliable: July, August, after harvest?
- Q4. Can regional hemp area be estimated with a stated error where no registry publishes parcels (France, Germany, Romania)?
- Q5. Which of these answers is good enough for a lender's siting report, and which only for research?
05Screen B · hempin.org/million
The public pledge tracker
A public page, like the Hemp Clock, that shows how much hemp land is on the record and how it got there. The counter starts at zero and says so. The calculator shows what the pledge actually requires, because it can mean two very different things.
●On the record today
of 1,000,000 · cumulative field-seasons, Declared or better
15,113 ha ready to ingest from open registries · not yet on the record
- Traced
- 0 ha
- Claimed
- 0 ha
- Declared
- 0 ha ingested
- Observed only
- not shown · does not count
ƒWhat the pledge requires
Needs 83% coverage of a 1.21M ha crop.
The million is reached in the 2034 season.
S1 and S2 are “every green light” scenarios from the impact model, not forecasts. Coverage ramps in a straight line from today's open-registry share (8.7% of 2024 area) to the share you pick. Smooth growth between scenario points is an assumption of this calculator.
≡Where the first hectares come from
| Registry | Season | Parcels | Hectares | Identity | Licence | Status |
|---|---|---|---|---|---|---|
| Estonia · PRIA GSAA (WFS) | 2026 | 850 | 8,652.5 | Company names in clear, natural persons masked | Open, no login | Ingestable, field-level |
| Latvia · LAD declared areas | 2026 | 443 | 3,176.8 | Anonymous | CC0 | Ingestable, field-level |
| Finland · Ruokavirasto GSAA | 2024 | 839 | 2,935.7 | Anonymous, organic flag | CC BY 4.0 | Ingestable, field-level |
| Denmark · FVM Marker | 2024 | 53 | 348.1 | Pseudonymous grower number | Open | Ingestable, field-level |
| Netherlands · BRP Gewaspercelen | 2025 | 772 | not summed | Anonymous | Open (PDOK) | Ingestable, area to compute |
| Wallonia · parcellaire anonyme | 2023 | 28 | not summed | Anonymous | Open | Ingestable, area to compute |
| Flanders · Agentschap L&V | 2025 | – | 540.96 | Aggregate figure only | Download service | Parcel file not yet queried |
| France · InterChanvre / RPG | 2024 | – | 24,600 (aggregate) | Aggregate; RPG parcel crop codes unverified | To check | Biggest prize · verify RPG |
| United States · USDA Hemp Public Search | 2026 | – | – | Licences, no fields | Lookup only | No field data |
▮Field-level hectares, ready to ingest
These need no research partner: a field table and a harvester. Ingesting them is how the counter moves before any satellite pilot starts.
06The pilot plan
A research partnership with dated, public gates
Hemp'in does not build a classifier. A university remote-sensing group or a Copernicus-funded team does, and publishes it. Hemp'in brings the harvested labels from six registries joined to legality and companies, a way to reach growers, and a public page where the result lives.
→Hemp'in brings
- The label set. Hemp parcels from six open registries, cleaned, with provenance, as an open dataset with a DOI.
- The legal join. Atlas legislation per country and season, so a model never treats an illegal crop as a label.
- The grower channel. Claim flows and company pages, so observed fields can become claimed fields by invitation.
- The stage. The tracker, the Atlas layer and the annual State of Hemp report.
←The partner brings
- The model and the compute. Open code, open weights where the partner allows.
- Peer review. A paper, so the accuracy figures carry someone else's name, not only Hemp'in's.
- Funding fit. The work suits EU research calls on bio-based value chains and Earth observation for agriculture.
G0 · Label inventory
Ingest the six registries with source, licence and date on every parcel. Review published cannabis and hemp classification work.
Pass: label set published with a DOI; accuracy thresholds for G1–G3 written down before any model runs.
G1 · Estonia, in-country
Train and test inside Estonia on past seasons, where every answer is known.
Pass: hemp separated from maize and other summer crops at the pre-registered threshold.
G2 · Across a border
Train on Estonia and Latvia, test blind on Finland and Denmark. Where most crop models lose accuracy.
Pass: transfer loss stays within the agreed margin.
G3 · Grand Est, blind
Estimate hemp area for Grand Est with an error bar, then compare with the declared French figure.
Pass: the declared figure falls inside the interval; the Hub Map can cite it.
07Rules that ship with the map
Mapping farms can hurt farmers. These rules come first.
Hemp fields get raided by people who think they are marijuana, and growers in many countries still carry stigma. A map of where hemp grows is useful to buyers and lenders, and also to thieves. Company data over personal data, always.
◎What the public sees
- Regional totals, not fields. Public maps show hemp hectares per 10 km hexagon with an error bar, never a parcel outline, unless the grower chose to show it.
- Registries are not made more identifying. Where a registry is anonymous, it stays anonymous. Estonia's named data is shown at company level only, never natural persons.
- Observed is labelled as a model output. Every observed figure carries the model version, season and error, in the words “estimated by satellite”.
✕What Hemp'in will never do
- Never report a farm. Observed-but-not-declared parcels are a model's guess and are not shared with police, regulators or anyone else.
- Never write observed fields into the company or place record. Promotion from research to the record is a human decision.
- Never count a satellite hectare toward the million.
- Never sell field-level location. The Hub Map and siting reports use aggregates.
08Open decisions
Three calls before this becomes a plan
?What “one million hectares” means
In one season (A), the pledge only works if the world grows hemp almost six-fold and most of it is on the record. Counted as field-seasons since 2026 (B), it depends on coverage, which can be worked on.
Leaning: B as the pledge, with the 2035 in-season figure always printed beside it.
?Which region pilots first
Estonia has perfect labels and little strategic weight. Grand Est has the corridor and the lenders but weaker public parcel data.
Leaning: calibrate on Estonia, prove on Grand Est. Verify the French RPG crop code before promising Grand Est anything.
?Ingest before the satellite
The 15,113 open hectares need no research partner, only a harvester and a field table.
Leaning: make the label inventory its own small build, so the counter moves before the pilot starts.
Sources: PRIA (EE), LAD (LV), Ruokavirasto (FI), FVM (DK), PDOK BRP (NL), Wallonia and Flanders open parcel services, each queried live; FranceAgriMer and InterChanvre via the impact model; Pôle Européen du Chanvre for world area. Map parcels, greenness curves and per-field values are synthetic and labelled so. Nothing here is a forecast or a product claim.