Climate Startup Idea: Satellite Deforestation Attribution

A U.S. importer of flooring buys oak from 100 small suppliers in three countries. A Seattle coffee trader sources his beans from 4,000 smallholder plots. Under EU rules, each has to prove that no plot was deforested since December 31, 2020, with coordinates, documentation and a risk assessment that passes an audit. Today, they either buy a dashboard that turns red when any tree cover disappears, or they hire a consultancy to hand-check plots. Both fail in the same way. An alert is not proof. In 2022 a cyclone flattened 30 hectares of a supplier’s forest; a legal plantation harvest looks exactly like illegal clearing from 10 metres up; a fire that spreads from a neighbour’s land is not the supplier’s doing. Every false positive means a supplier suspended, a shipment held up, and a procurement team angry. A false negative is a compliance violation.

The unfair insight: the hard problem was never about detecting change — public alert systems do that for free already. The hard problem is attribution, and the signals for that are also public and largely unused together: fire hotspot timing tells you whether flames preceded the loss, radar backscatter and patch geometry separate a clean mechanical clear-cut from a ragged storm blowdown, road and settlement proximity indicates access, seasonality and regrowth over the following 18 months distinguish a harvest rotation from a permanent conversion to pasture, and elevation and river data flag landslides and flooding. Combine those into a cause classification with a confidence score and the imagery strip that justifies it and you have taken a red dot and turned it into a document an auditor will accept. That’s the document people pay for.

Who you’re building for: Dana, sustainability lead at a mid-size US hardwood and furniture exporter with 120 supplier forests. Her question: “I have eleven alerts this quarter. Which of them are actually my problem, and can I prove the rest are not?”

Why this stack: the product is a geospatial pipeline with a thin web surface, so Next.js handles the plot map, dossier viewer and supplier risk views while the heavy work sits in Python. PostgreSQL with PostGIS on Neon stores plots, loss patches, attribution outputs and model versions, and supports the spatial joins against roads, slope and hotspot layers. A FastAPI service with rasterio and xarray runs the time-series extraction and classifier, scheduled as batch jobs on a GPU-or-CPU-on-demand provider such as Modal or Railway; pull imagery through Google Earth Engine or Sentinel Hub rather than downloading scenes yourself, and cache per-plot time series so a dossier regenerates in seconds. Mapbox draws plots and loss patches. Stripe bills per plot and per seat. Keep every model and ruleset versioned with the dossier that used it — reproducibility is the product for regulated buyers — and store dossiers immutably with retention long enough to satisfy record-keeping rules.

Budget breakdown: geospatial pipeline, imagery access and per-plot time series $5k–$11k; attribution engine — feature extraction, labelling of 1,500–3,000 loss patches with a remote-sensing contractor, classifier and calibration $5k–$12k; plot onboarding, validation and supplier risk views $2.5k–$6k; dossier generator with imagery strips and audit trail $2k–$5k; design $1.5k–$3k; analyst review queue and billing $1k–$2.5k; legal and expert review — remote-sensing scientist sign-off, disclaimers, data-licence review, terms $2k–$5k. Monthly run rate at launch: $400–$1,200 (imagery API tiers, compute for batch jobs, Neon, Mapbox, Vercel).

Sprint plan: Week 1 — PostGIS schema, plot upload and validation, imagery access, baseline forest cover, alert ingestion. Week 2 — per-plot Sentinel-1 and Sentinel-2 time series, patch delineation, feature extraction (hotspots, geometry, roads, slope, regrowth), start labelling. Week 3 — classifier training and calibration, confidence bands, analyst review queue, supplier risk rollup. Week 4 — dossier PDF with imagery strips and audit trail, monitoring alerts, billing, validation against 200 known events (documented fires, storms, licensed harvests, confirmed illegal clearing), then a pilot with two importers and one carbon-project developer.

Select 30 actual forest-loss occurrences (independently verifiable, e.g., documented storms and wildfires, authorised harvests of public timber land, clearing cases already documented by NGOs) and manually identify them from free photography. If your manual approach can’t do cause separation on 24 of 30, the product does not exist yet and no amount of engineering fixes that. In parallel, call 15 US operators subject to the EU deadline (hardwood and furniture exporters, pulp and paper, coffee and cocoa dealers) and ask three questions: how many source plots do you have, who is mapping them, and what do you do today when an alert triggers. Ask two carbon project developers and one rating agency how much they would pay for independent baseline and reversal verification. Pick a remote sensing expert as the advisor – it is their name on the approach that gives the dossier credibility.

Kill criteria: manual attribution accuracy <80% on the 30-event set, or no operator prepared to conduct a paid pilot on 100 plots — stop. And narrow to 1 commodity and 1 biome, or to fire-vs-clearing only.

Sources of revenue: recurring base from per-plot monitoring at $1-5/plot/year with volume tiers (a coffee trader with 4,000 plots lands at $8-20k a year) platform subscription for operators at $15-75k a year, including seats, supplier risk views, record retention and unlimited dossiers paid verification dossiers for disputed or high-value cases at $500-2,500 each carbon-market products - baseline, leakage and reversal verification for project developers, registries and ratings agencies at $10-100k per engagement API licencing to auditors, certifiers and trading platforms later.

Targets: 85%+ attribution accuracy published per cause in launch biomes clearing-versus-natural, false-positive rate low enough that operators act on alerts (track 'alerts dismissed as noise' under 15%), pilot-to-contract conversion 40%, net revenue retention 120%+ as operators add plots each season, gross margin 75%+ after imagery and compute, CAC recovered inside one annual contract. So $2.5M–$4M ARR at 500k plots monitored + 30 subscriptions for operators.

The reality of contracts is that compliance budgets are annual and approved before the quarter of the deadline. Sell now for the December cycle and expect procurement and security review to add 60-90 days at larger operators.

0-10 customers: Two paid pilots with mid-size US exporters in one commodity and one carbon-project developer. Publish a methodology paper and an open validation set. In this market, credibility is distribution. Content aimed right at the deadline (“What evidence an EU deforestation audit actually accepts”) and presence at trade bodies: hardwood and furniture associations, coffee and cocoa trade groups, and forestry certification communities.

10-50 customers: partner with the EUDR consultancies, certification bodies and traceability platforms already inside these supply chains – they need an evidence layer and don’t want to build a satellite pipeline. Quarterly attribution reports by region, driving press and inbound. Expand to 2nd commodity and 2nd biome.

50+: registries and ratings agencies as channel partners, insurers and lenders for forest-loss triggers, api licencing to auditors and trading platforms, and public sector work (agencies and NGOs that want independent attribution rather than a black box)

• You are the evidence provider, not the certifier. Filing of due diligence statement is done by the operator and legal liability rests with them. Please write this in terms of , and never market the product as EUDR certification or a compliance guarantee.

• Imagery and data licences: Copernicus Sentinel and USGS Landsat data are open and commercially usable; a number of high resolution basemaps and programme datasets have non-commercial or limited use conditions. Audit every layer in the pipeline before charging for output. Record the licence per dossier.

• Methodology transparency: publish the method, model versions and accuracy by cause. Keep an immutable audit trail linking each dossier to the exact model and data vintage. Allow re-running an old dossier unchanged.

• Accuracy claims: FTC rules on environmental marketing apply to your own claims and those made by your customers using your output; avoid absolute language such as 'deforestation-free verified'.

• Liability: professional indemnity and errors-and-omissions insurance; contractual limits; documented dispute process for a supplier to dispute an attribution, with analyst review.

• Sensitive geographies: plot coordinates can disclose smallholders and indigenous lands. Keep coordinates confidential . Respect free , prior and informed consent principles . Never publish plot - level locations without consent from operator and community .

• Data residency and record keeping: retain records for the periods specified by the regulation and offer EU hosting for data with EU operators. GDPR applies when plot data is linked to identifiable farmers.

• If you are in carbon markets: registry methodology rules, separation from conflict-of-interest of project developers you verify, no revenue tied to project’s approval.

• Delaware C-corp with EU entity or hosting partner once EU operators are customers. Trademark the name of the methodology.

Minimum team: founder-CTO (pipeline, product, web), remote sensing or geospatial ML engineer (critical hire or contractor for the build), remote sensing scientist advisor with peer-reviewed forest-change work (1-2% equity) who owns and defends the methodology, labelling contractor for the training set. First hires: A compliance-side salesperson who has sold into supply-chain sustainability teams, then an FT analyst for the review queue as plot volume grows.

Metrics that matter Attribution accuracy per cause and per biome on a held-out validation set Share of outputs requiring analyst review (target under 20% and falling) Time from alert to dossier (target under 24 hours) Operator alert-dismissal rate Supplier disputes upheld (should be rare and always published internally) Plots under monitoring Net revenue retention Pilot-to-contract conversion

Biggest risks: attribution is really hard — cloud cover, smallholder mosaics and degradation all degrade accuracy. So launch narrow (one biome, clearing vs natural) and expand only with validation data. Regulatory timing risk is real: the EU deadline has already moved twice, so build products that stand on their own for carbon and lending buyers rather than betting the company on one date. Incumbents and adjacent players — Global Forest Watch and academic alert systems (free, detection only), traceability platforms (no imagery science), and satellite analytics firms selling to governments — mean your position must stay squarely on attribution plus audit-grade evidence. A wrong, confident label damaging a supplier relationship is the reputational risk; calibrated confidence, analyst review and a dispute process are the mitigations. Finally, imagery and compute costs scale with plot count — cache aggressively and price per plot accordingly.

Funding: bootstrap the MVP from $18k–$40k, then chase non-dilutive money that fits precisely: NASA and NOAA SBIR topics on Earth observation applications, NSF SBIR, USDA Forest Service and state forestry innovation programmes, and ESA or Copernicus accelerator support if you add an EU entity. Pre-seed of $500k–$1.5M from climate and geospatial investors (Congruent, At One Ventures, Overture, Seraphim for space data, plus Techstars Sustainability) once two paid pilots show accuracy and one operator has signed an annual contract. Seed after $500k ARR and a registry or ratings-agency partnership.

90 days: Days 1–14 hand-validate the 30-event set, interview 15 operators, secure the science advisor and two pilot commitments. Days 15–42 build the MVP and label the training set. Days 43–60 run pilots on 100–500 plots each, measure accuracy against known events, tune confidence thresholds and the review queue. Days 61–90 publish the methodology and validation results, convert pilots to annual contracts ahead of the December cycle, submit one SBIR application, and open conversations with a certification body and a carbon registry.