Deforestation-free supply-chain verification (EUDR) & forest carbon MRV

Importers of cattle, cocoa, coffee, oil palm, rubber, soy and wood products must show that goods come from plots not deforested after 31 Dec 2020. Satellite-based AI checks every supplier polygon against forest-cover and change maps, flags risk and produces the evidence behind the due-diligence statement. The same remote-sensing stack, extended with LiDAR-based canopy-height and biomass models, supports measurement, reporting and verification (MRV) of forest carbon projects.

Problem & value

• Today: operators collect plot coordinates from thousands of smallholders in spreadsheets and PDFs, eyeball maps, or rely on certificates. Geodata is often missing, invalid (swapped lat/lon, wrong precision, self-overlapping polygons) or overlaps other plots. Carbon projects rely on sparse field plots and periodic third-party audits.
• AI value: automated geodata validation, repeatable per-plot deforestation screening against multiple independent maps, continuous monitoring instead of a one-off check, and audit trails. For carbon: wall-to-wall biomass and change estimates that make baselines and leakage checks cheaper and more transparent.
• Regulation timing: Regulation (EU) 2023/1115 (EUDR) was postponed again by Regulation (EU) 2025/2650: application from 30 Dec 2026 for large and medium operators and 30 Jun 2027 for micro and small ones, with simplified duties for downstream operators and traders. Further review or scope changes (e.g. by delegated act) should be verified on EUR-Lex before relying on these dates [1][2].
• No vendor-independent figures on cost savings are cited here; measure your own baseline.

Variants

• Importer / trader (EUDR operator): supplier onboarding, plot validation, DDS preparation for commodities such as cocoa, coffee, palm oil, soy, rubber, cattle/leather, timber.
• Downstream operator / SME: consumes upstream DDS references; verifies and retains them under the simplified regime.
• Commodity producer / cooperative: collects polygons by mobile app, pre-checks before shipping.
• Financial institutions and investors: portfolio-level deforestation exposure (links to DATA-012).
• Competent authorities: risk-based checks on operators, building on VIS-038.
• Forest carbon MRV: project developers, ratings agencies and verifiers estimating above-ground biomass, degradation and baseline leakage (VCS/Verra, Gold Standard, EU CRCF).
• Private/landowner variant: smallholder forest owners documenting stock for carbon schemes.

How it works

1. supplier data intake (GeoJSON / KML / shapefile / app capture, plus trade documents)
2. geodata validation (CRS, topology, area, overlaps, plots <4 ha as point vs polygon, country match)
3. overlay with forest-cover baseline (31 Dec 2020) and land-cover maps (consensus of several products)
4. change detection on Sentinel-2 / Landsat / SAR time series (cloud-robust, optical + radar)
5. per-plot risk score + evidence (imagery dates, map versions, confidence)
6. human analyst review of flagged plots (very-high-resolution imagery, field verification)
7. risk assessment & mitigation record
8. DDS data
9. EU information system (TRACES-based)
10. continuous monitoring and alerts (e.g. new clearing after onboarding) carbon branch: canopy height / biomass regression (LiDAR + optical + SAR)
11. stock & change + uncertainty
12. MRV report

• AI components: semantic segmentation of forest/non-forest; change-detection models on time series; geospatial foundation models for embeddings; regression of canopy height and biomass from GEDI LiDAR calibrated optical/SAR features; LLM/document parsing for supplier paperwork (DOC-006).
• Human-in-the-loop: the operator remains legally responsible. Analysts review each flagged or low-confidence plot, resolve definition gaps (forest vs. tree plantations, agroforestry, "degradation"), and decide on supplier actions. Map output is evidence and screening, never a legal conclusion.

Data preparation

The central data is supplier plot geolocations (GeoJSON, KML, shapefiles, app captures) plus supplier declarations, certificates and trade documents (HS codes, volumes). These live in the ERP or procurement system, supplier portals, e-mail attachments, Excel lists and paper certificates. Satellite and reference layers are external and clean.

Typical gaps: coordinates in wrong formats or reference systems, points where polygons are required, overlapping or implausible plots, missing links between plot, supplier and delivery batch, scanned certificates, and supplier data that is simply not delivered.

Preparation steps: 1. Build a supplier and plot register with a stable ID and an owner per supplier (PREP-004, PREP-011). 2. Extract existing coordinates from ERP and spreadsheets (PREP-005, PREP-003). 3. Digitise paper certificates and parse documents (PREP-006, DOC-006). 4. Automatic geodata validation: CRS, topology, area, overlaps, country match. 5. Link plots to batches for traceability. 6. Build a test set of known cases with analysts (PREP-018) and record map versions and imagery dates as evidence.

Effort share: about 40-50 percent; supplier onboarding and data quality dominate, not the AI.

Implementation paths

Which path for whom

• smb (A): EU-based EUDR platforms (Satelligence, Meridia and others) cover validation, satellite checks and DDS submission at manageable prices; the process is standard and deadline-driven.
• enterprise (B): large traders with thousands of plots need their own scoring logic, ERP and trade-document integration and full control of evidence, using Copernicus, Sentinel Hub and open reference layers.
• developer (B): build on openEO, Copernicus Data Space and open forest layers; self-hosting the heavy compute (C) only pays at large scale.
• public (B): authorities and agencies use sovereign EU cloud with Copernicus data to keep evidence under their own control and meet procurement rules.

Path A — Buy (recommended for most operators)

• EUDR compliance platforms with satellite monitoring: Satelligence (NL, EU), Meridia (NL, EU), Preferred by Nature / NEPCon tools (DK, EU), Starling (Airbus, FR/EU), Land Sense, Orbify (EU), plus global players such as Planet, Descartes Labs/EarthDaily and supply-chain suites that integrate geo-risk (e.g. Everstream, Sourcemap, TraceX). Check: EU hosting, DDS/API submission support, map provenance, evidence export.
• Carbon MRV and ratings: Pachama, Sylvera, Renoster, Chloris Geospatial, and open projects; verify methodology acceptance by the relevant standard.
• Choose A when: < several thousand plots, no in-house GIS team, deadline pressure.

Path B — Build on cloud APIs

• Data: Copernicus Data Space Ecosystem (EU), Sentinel Hub / openEO, Google Earth Engine, Microsoft Planetary Computer, Planet and Airbus APIs for VHR imagery.
• Reference layers: the JRC global forest cover map for 2020 (EUDR baseline reference), Global Forest Watch / Hansen tree-cover loss, ESA WorldCover, JRC TMF (tropical moist forest), Forest Data Partnership layers, RADD alerts, Meta/WRI high-resolution canopy height.
• Orchestration: Python (geopandas, rasterio), workflow engine, PostGIS; EU information system API for DDS submission.
• Choose B when: you need your own scoring logic, ERP/trade-document integration and full control of evidence.

Path C — Self-host (open source)

• Stack: STAC catalogues (Copernicus), openEO, Open Data Cube, xarray/stackstac, PostGIS, QGIS; segmentation with TorchGeo/segment-geospatial; geospatial foundation models such as Prithvi-EO (IBM/NASA), Clay, TerraMind (ESA/IBM) and AlphaEarth Foundations embeddings.
• Hardware: HW1 for polygon validation; HW3 for national-scale inference; storage is the larger cost driver than compute.
• Ops effort: needs GIS and remote-sensing engineers; ongoing map-version management.

Path D — Custom training

Worth it for local forest definitions (e.g. cocoa agroforestry vs. forest, shade coffee, oil-palm vs. natural forest), regional biomass models, or low-cloud-cover radar fusion. Needs labelled reference polygons (visual interpretation of VHR imagery, field plots, GEDI/airborne LiDAR for biomass) in the hundreds to thousands per region; validate with independent samples.

State of the art (2026)

• Global maps agree poorly at the margin. EUDR-relevant forest definitions (FAO: >0.5 ha, >10% canopy, >5 m) are hard to reproduce; best practice is consensus of several maps plus VHR review of disagreements instead of trusting one product.
• Radar and optical fusion (Sentinel-1/2) gives near-weekly clearing alerts under cloud, e.g. RADD and GLAD systems.
• Geospatial foundation models reduce label needs for segmentation and change detection; benchmark gains are real but regional transfer should be tested locally.
• Canopy height and biomass: GEDI-based and deep-learning canopy-height maps (Meta/WRI), ESA CCI Biomass and NASA/ESA NISAR and BIOMASS missions (launched 2025) improve stock mapping; uncertainty at plot level remains large, particularly in dense tropical forest and above ~300 Mg/ha (saturation).
• Limits and failure modes: degradation and selective logging are hard to detect; small plantations and agroforestry are confused with forest; post-2020 clearing on neighbouring land is not attributable to the supplier plot without correct polygons; polygon fraud (submitting a clean decoy plot) is not detected by imagery alone, so combine with trade-flow and volume plausibility checks (ANOM-011). Carbon credits: over-crediting from poor baselines is the main documented integrity risk, so treat satellite output as input to a verifier, not as the credit itself.

Shared resources & combinations

• Same remote-sensing stack as VIS-021 (change detection), VIS-038 (CAP area monitoring), VIS-042 (methane plumes), VIS-058 and VIS-061: one STAC/openEO data lake plus GPU segmentation service (HW3) serves all.
• Upstream/downstream: DOC-006 (trade documents, bills of lading), AGENT-005 (procurement and supplier outreach), ANOM-011 (supplier risk), DATA-012 (CSRD/ESRS reporting), DATA-021 (product carbon footprint and Digital Product Passport), SCI-010 (earth-system models).

Effort, cost & KPIs

• Path A: weeks to onboard; pricing typically per plot, per supplier or per hectare monitored (vendor-specific, request quotes).
• Path B: 3-6 months with a small geo-data team; imagery from Copernicus is free, VHR imagery is the main cost.
• Path C/D: 6-12 months incl. validation; ongoing map maintenance.
• KPIs: share of volume with valid geolocation, share of plots screened and cleared, false-positive rate on analyst review, time from flag to resolution, share of suppliers with high-risk plots remediated, DDS rejection rate; carbon: bias and RMSE of biomass vs. independent plots, uncertainty reported.

People & change

Management and legal/compliance carry the legal responsibility as the operator, so they sponsor the project. The procurement or sustainability lead owns supplier onboarding; IT integrates ERP and trade-document systems; data owners of supplier and ERP data must release exports. Suppliers are key stakeholders: they have to deliver plot data, so explain requirements, formats and deadlines early and offer simple capture tools. A small analyst team reviews flagged plots in a pilot with one commodity and a few suppliers.

Works council: depends. Mostly no employee monitoring, but if the platform logs analyst decisions per person, §87(1) no. 6 BetrVG may apply; inform under §90(1) no. 3 if processes change.

Data protection: plot coordinates of smallholders can identify individuals, so personal data may be involved; involve the DPO, sign an AVV with the platform vendor and check legal basis and purpose (PREP-025).

AI literacy: support analysts with training on map limits, definitions (forest, plantation, degradation) and why map output is screening, not a legal conclusion (GUIDE-006).

Concerns: "burden on suppliers" - phased onboarding and support; "false flags" - human review of every flagged plot before supplier action (GUIDE-008, GUIDE-009, GUIDE-010).

Risks, regulation & compliance

• EU AI Act: minimal risk (no Annex III category); AI literacy duty (Art. 4). Not a safety component.
• EUDR: legal duty stays with the operator (information collection, risk assessment, mitigation, DDS and 5-year record retention). Satellite screening does not replace risk assessment of legality (land tenure, labour, indigenous rights).
• GDPR: plot polygons of smallholders and farmers are personal data when linked to an identifiable person; use data processing agreements, minimise attributes, and secure cross-border transfers to producer countries. Publishing polygons risks reprisal and competitive exposure.
• Fairness: smallholder exclusion is a documented side effect; provide support for data collection and avoid blanket supplier delisting based on map noise.
• CSRD/ESRS and carbon claims: claims need verifiable method documentation; the EU Green Claims and CRCF rules affect carbon marketing.
• Mitigations: multi-map consensus, human review, versioned evidence, accuracy reports, grievance channel for suppliers.

Getting started

1. Inventory commodities, HS codes, origins and volumes in scope; confirm your role (operator, trader, downstream) and deadlines (30 Dec 2026 / 30 Jun 2027).
2. Collect and validate polygons/points for a pilot origin; fix topology and coordinate errors first.
3. Run a 2-3 vendor bake-off (Path A) on 200 known plots, including disputed ones, and compare with analyst labels.
4. Define flag triage, escalation and supplier remediation process; link to your ERP/trade documents.
5. Set up monitoring and DDS data export; for carbon, start by comparing public biomass products against your own plot data.

Sources

1. Regulation (EU) 2023/1115 on deforestation-free products — https://eur-lex.europa.eu/eli/reg/2023/1115/oj
2. European Commission Access2Markets — Delay until December 2026 and other EUDR developments — https://trade.ec.europa.eu/access-to-markets/en/news/delay-until-december-2026-and-other-developments-implementation-eudr-regulation
3. Council of the EU — targeted revision to simplify and postpone the EUDR (18 Dec 2025) — https://www.consilium.europa.eu/en/press/press-releases/2025/12/18/deforestation-council-signs-off-targeted-revision-to-simplify-and-postpone-the-regulation/
4. European Commission — Deforestation Regulation (policy page and EU Forest Observatory) — https://environment.ec.europa.eu/topics/forests/deforestation/regulation-deforestation-free-products_en
5. Global Forest Watch — data and tools — https://www.globalforestwatch.org/
6. Copernicus Data Space Ecosystem — https://dataspace.copernicus.eu/
7. Satelligence — EUDR compliance — https://satelligence.com/
8. Meridia — https://www.meridia.land/
9. Tolan et al. (2024), Very high resolution canopy height maps from RGB imagery (Meta & WRI), Remote Sensing of Environment 300:113888 — https://doi.org/10.1016/j.rse.2023.113888
10. NASA GEDI mission — https://gedi.umd.edu/
11. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj