akar.earthAI in Carbon MRV: Where It Helps and Where It Fails
AI carbon MRV explained: where machine learning genuinely improves forest carbon measurement, where it fails, and how verifiers judge AI-based estimates.
AI helps carbon MRV most where there is a lot of image or field data to sort, such as detecting trees in drone imagery, mapping land cover, predicting canopy height and flagging bad field records. It fails when it is used as a substitute for local ground truth, when its errors are not reported, or when nobody can trace a credited tonne back to measured evidence.
Where AI genuinely helps
| Task | What AI does | Why it adds value |
|---|---|---|
| Tree and crown detection | Object detection on drone or very-high-resolution images | Counts thousands of trees fast; see drone photogrammetry |
| Land cover and change | Classification of Sentinel-2, Landsat and radar time series | Consistent maps over large areas |
| Canopy height and biomass | Regression from imagery trained on lidar and plots | Wall-to-wall predictions; see canopy height models |
| Field data QA | Outlier and duplicate detection, photo checks | Catches errors before they reach reports |
| Species identification | Image classification from leaf, bark or crown photos | Supports wood density and allometry choice, with expert review |
| Alert triage | Ranking deforestation alerts by likely cause | Focuses field verification effort |
| Document drafting and review | Language models summarising rules or checking consistency | Saves analyst time, under expert supervision |
Where it fails
- Transfer to new places. A model trained on temperate or Amazon forest can be badly biased in Kalimantan peat swamp or Javanese agroforestry.
- Saturation and extrapolation. Models predict poorly outside the range of training data, such as very tall or very high-biomass forest.
- Hidden error. Pixel predictions look precise on a map, but the uncertainty may be large and unreported.
- Label noise. Training on global maps that already contain errors reproduces those errors.
- Explainability. If a verifier cannot see why the model predicted a value, it is harder to accept.
- Generative AI hallucination. Language models can invent regulation numbers, parameters or citations. Every generated statement must be checked against source documents.
A practical framework for using AI in MRV
- Define the role. Is the model producing a credited quantity, a stratification, or a QA flag? Higher stakes need stronger validation.
- Use local training data. Field plots, RTK-located and measured to protocol. See ground truth vs satellite.
- Hold out validation data from different locations, not just random pixels from the same plots.
- Report error as bias and RMSE in the units that matter (tonnes per hectare, trees per hectare), and propagate it into the uncertainty deduction.
- Version everything: training data, code, model weights and settings.
- Keep humans in the loop for decisions with legal or crediting consequences, such as alert verification and eligibility.
- Recalibrate each monitoring period as vegetation and sensors change.
How verifiers and standards view AI
Methodologies generally specify what must be estimated and to what precision, not which algorithm to use. That leaves room for machine learning, but places the burden of proof on the project. Expect a VVB to ask for the training and validation data, accuracy statistics, and evidence the model fits local conditions. The ICVCM Core Carbon Principles emphasise robust, conservative quantification; an unvalidated AI estimate does not meet that bar.
Questions to ask an AI vendor
- What training data was used, where, and when?
- How was the model validated in conditions like ours?
- What are bias and RMSE in our units?
- Can we see predictions next to our own plots?
- How is uncertainty reported and propagated?
- Can a verifier reproduce the result with the stated version?
- Who owns the outputs and the data we provide?
A vendor who cannot answer these is selling a map, not MRV.
Common mistakes
- Marketing “AI-verified” credits. Verification is done by accredited bodies, not algorithms. See avoiding greenwashing.
- Reporting overall accuracy for a land-cover map dominated by one class, which hides poor performance on the class that matters.
- Training and validating on spatially clustered samples, inflating accuracy.
- Replacing field plots entirely to cut cost.
What good evidence looks like
A model card or technical annex: purpose, inputs, training data with dates and locations, architecture and settings, validation design and results, known limitations, and version. Plus the link from model output to methodology equations, and the uncertainty treatment.
Indonesia specifics
Indonesia’s cloud cover, peat, mangroves, steep terrain and smallholder mosaics are all conditions where off-the-shelf global models are least reliable. Local plot networks are scarce in many provinces, which is why new field data is valuable for training. AI is a multiplier on good field data, not a substitute.
Our carbon modelling service builds machine-learning models calibrated to local plots, and remote sensing biomass explains the underlying workflow. More on the dMRV hub.
Frequently asked questions
Can AI calculate carbon credits?
AI can predict biomass, map cover and detect trees, but credits are calculated with a methodology's equations and must be verified. AI outputs are inputs that need calibration, validation and uncertainty estimates.
Where does AI help most in forest MRV?
Detecting trees and crowns in drone imagery, classifying land cover, predicting canopy height and biomass from satellite data, checking field data quality, and triaging deforestation alerts.
Why do verifiers distrust some AI-based carbon estimates?
Because models trained elsewhere can be biased locally, errors are often unreported, and black-box methods make it hard to trace a credited tonne back to evidence.