CarbonTruth

CarbonTruth turns raw satellite data into unforgeable, retire-once carbon credits — verify once, trusted everywhere.

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for CarbonTruth

Inspiration

In 2023, a study in Science examined voluntary forest-carbon credits and found that the vast majority did not represent real, additional carbon — inflated baselines, weak additionality, systematic over-crediting. That finding stuck with us: every phantom credit is a tonne someone paid to remove that is still sitting in the atmosphere.

The reflex answer in our space is "put it on a blockchain." But a ledger only proves a number never changed — not that it was ever true. Write a bad baseline on-chain and you've made the lie permanent and tradeable. That gap — the oracle problem — is the half everyone skips, and it's the half that actually matters. We built CarbonTruth to close it: the trust layer that sits in front of the chain, not just on it.

What it does

CarbonTruth turns raw satellite pixels into a Verified Climate Claim (VCC) — a receipt that is:

  • AI-measured from independent Earth-observation data,
  • consensus-verified by multiple independent oracles,
  • anchored on-chain with an evidence hash, and
  • globally unique via a geo-temporal fingerprint, so it can never be double-counted.

In the live demo you are the verifier: pick a real reforestation project, submit a claim, watch the oracle network agree (or reject an over-claim), mint the VCC, trade it, retire it once, and even try to forge the ledger — and watch it break.

Verify once, trusted everywhere.

How we built it

The pipeline is four auditable stages: observe → estimate → verify → mint.

1. Observe. We pull Sentinel-2 optical bands over the plot (before/after) and compute vegetation index from the red and near-infrared bands:

$$\mathrm{NDVI} = \frac{B_{08} - B_{04}}{B_{08} + B_{04}}$$

Scene discovery runs against Microsoft's Planetary Computer STAC through a Vercel serverless function — no API key required.

2. Estimate. The carbon number is not a hand-tuned curve — it's a trained neural network. We trained a multilayer perceptron ($6 \rightarrow 32 \rightarrow 16 \rightarrow 1$, ReLU) on GEDI-calibrated data, mapping spectral features $(\mathrm{NDVI}, \mathrm{SAVI}, \mathrm{NIR}, \mathrm{red}, \text{ratio}, \mathrm{NDVI}^2)$ to aboveground biomass. Biomass converts to CO₂e with the IPCC carbon fraction $CF = 0.47$:

$$E = \mathrm{AGB} \cdot CF \cdot \frac{44}{12}$$

The model reached $R^2 \approx 0.98$ on held-out data. Crucially, we froze the weights to JSON and wrote a pure-standard-library forward pass, so the trained net runs client-side with zero dependencies.

3. Verify. Multiple independent sources (optical, lidar, archive) each produce an estimate $m_i$ with uncertainty $\sigma_i$. We fuse them by inverse-variance weighting:

$$\mu = \frac{\sum_i w_i m_i}{\sum_i w_i}, \qquad w_i = \frac{1}{\sigma_i^2}$$

A claim clears only if confidence passes the gate $\theta = 0.90$ and a $k$-of-$n$ quorum agrees. Spoof one feed and it's rejected as an outlier.

4. Mint. Verified claims are appended to a SHA-256 hash-chained ledger, each block linked to the last:

$$h_n = \mathrm{SHA\text{-}256}\big(h_{n-1} \,|\, \text{payload}_n\big)$$

A geo-temporal fingerprint enforces retire-once uniqueness, and recomputing the chain makes any tampering evident instantly.

The prototype is deployed on Vercel (static front end + a serverless STAC function) and auto-deploys from GitHub. One primitive — the VCC — spans all four ClimateChain tracks: AI/data verification, the MRV evidence trail, on-chain integrity, and portable trust.

Challenges we faced

  • Keeping the math honest while keeping it in the browser. We wanted a real trained model, not a toy, but also zero backend for inference. Freezing the MLP to JSON and reimplementing standardization + the forward pass by hand — then calibrating the biomass backbone so a canonical plot landed at a realistic $\approx 4{,}096$ tCO₂e — took several training/validation cycles.
  • Making fraud visible, not just handled. A rejection that happens silently convinces no one. We designed the demo so an over-claim is caught live on screen and a forged block visibly breaks the chain.
  • A subtle front-end bug. Our scroll-reveal animation started the demo panel at opacity: 0 and relied on an IntersectionObserver to un-hide it — so the pipeline sometimes animated behind an invisible panel. We fixed it by force-revealing the demo the instant any run starts, so the sequence always reads continuously.
  • Framing. The hardest part wasn't code — it was articulating why blockchain alone fails. Landing the oracle-problem framing is what makes the whole project click.

What we learned

  • Trust is the product, not tonnes. The market isn't short on demand; it's short on proof. That reframing drove every design decision.
  • Verification has to be adversarial by design. Consensus, outlier rejection, and replay protection matter more than any single clever measurement.
  • Published science beats black boxes. Grounding the estimator in peer-reviewed allometry (Chave et al., 2014) with transparent uncertainty is what makes the number defensible to a buyer or a regulator.

References

  • West et al., Science (2023) — systematic over-crediting. doi:10.1126/science.ade3535
  • Chave et al., Global Change Biology (2014) — pantropical allometric models. doi:10.1111/gcb.12629

Built With

Share this project:

Updates

Submission history