Methodology
Carbon-Eye turns satellite and environmental data into county-level above-ground biomass and carbon-stock estimates, entirely inside Google Earth Engine. This page documents each step.

The pipeline
- 01Select area & year. Choose one or more Kenyan counties, or draw a custom boundary, and an ESA CCI AGB reference year (or a range, averaged across every snapshot in it).
- 02Build predictor stack. Earth Engine assembles a cloud-filtered, multi-source stack of optical, radar, terrain, climate and soil layers.
- 03Sample & split. Reference carbon points are sampled from the stack and split into training and testing sets.
- 04Train models. Random Forest, Gradient Tree Boosting and SVM each learn the link between predictors and carbon.
- 05Map & validate. Estimates are mapped in carbon or biomass units, checked against held-out data (RMSE, MAE, R²), and summarised by county alongside tree-cover loss.
Data sources
All accessed and processed through Google Earth Engine.
| Role | Dataset | Coverage |
|---|---|---|
| Reference target | ESA CCI Above-Ground Biomass v7.0 | 2007, 2010, 2015–2024 (v7.0, May 2026) |
| Optical imagery | Sentinel-2 Surface Reflectance (Copernicus) | 2015–present, 10–20 m |
| Radar | Sentinel-1 GRD & JAXA ALOS PALSAR | 2014–present / 2007–present |
| Land cover | Google Dynamic World | 2015–present, 10 m |
| Terrain | SRTM DEM | 2000, 30 m |
| Climate | WorldClim BIO | 1970–2000 normals, ~1 km |
| Structure & temperature | Meta Canopy Height & MODIS LST | 2019–2022, 1 m / 2000–present, 1 km |
| Forest change | Hansen Global Forest Change (UMD) | 2001–latest release, 30 m |
| Boundaries | geoBoundaries ADM1 |
Predictors
Vegetation greenness & moisture
Spectral indices describing how green, dense and hydrated the canopy is.
Sentinel-2 bands · NDVI · EVI · SAVI · NDMI · NDRE
Vegetation structure
Radar and canopy-height signals describing the physical structure and volume of vegetation.
Sentinel-1 VH · SAR texture (contrast) · PALSAR HH/HV · Canopy height
Site conditions
Terrain and climate context that shapes where and how much biomass can accumulate.
Elevation · Slope · Aspect · Temperature · Rainfall · Soil organic carbon · Land-surface temp
Models & carbon conversion
Three models are trained on reference points sampled at 100 m (the reference map's own resolution) from the predictor stack and scored on held-out points; the output blends their predictions, weighted by each model's held-out error, with a model that does no better than the average given almost no weight. Support Vector Machine inputs are standardised first.
- Random Forest Bagged decision-tree ensemble
- Gradient Tree Boosting Sequentially boosted trees
- Support Vector Machine Kernel regression

Carbon pools
The IPCC counts carbon in five pools. Carbon-Eye maps the two that carbon standards make mandatory, shows soil carbon beside them, and leaves out the two that satellites cannot see.
| Pool | In Carbon-Eye | How |
|---|---|---|
| Above-ground biomass | Counted | Mapped pixel by pixel by the model; carbon = biomass × 0.47 (IPCC 2006 default, also used by the UN CDM afforestation manual). |
| Below-ground biomass (roots) | Estimated | Above-ground carbon × a root-to-shoot ratio from the NASA/ORNL harmonised biomass maps (Spawn et al. 2020); 0.24 where they have none, close to the 0.235 of Mokany et al. (2006) used in the Winrock guidance. |
| Soil organic carbon | Shown separately | Top 30 cm from a high-resolution African soil map, cross-checked with SoilGrids. Never added to the vegetation totals: soil changes over decades after land-use change and is accounted for separately (Winrock). |
| Deadwood | Not counted | Not visible to satellites. The CDM afforestation rules allow leaving it out where a project will not reduce it; where it matters, measure it in the field (Winrock's line-intersect method). |
| Litter | Not counted | Not visible to satellites, and optional under the same CDM rule; measured with small clip plots in the field when needed. |
Field plots are planned with the UN CDM afforestation sample-size formula (±10% at 90% confidence by default), and tree measurements are turned into plot carbon with Chave et al. (2014): biomass (kg) = 0.0673 × (wood density × diameter² × height)0.976, with Winrock's default wood density of 0.6 g/cm³ where a tree's is not known.
Validation
Each run holds out spatial blocks of land the models never trained on and reports RMSE, MAE and R² against ESA CCI AGB v6 (2020), plus a predicted-vs-observed plot. Measured on 2026-10-15 with the previous model settings. The tuned settings in use since 2026-10-16 had lower error at the same test areas (Random Forest 8.4%, boosting 4.4% on average), so these figures are, if anything, conservative. Across the test areas:
- Error removed 30%-74%
- Ensemble RMSE 4.7-14.5 t C/ha
- 90% ranges held 90%
The spread between the three models is mapped as a spatial uncertainty proxy, and Random-Forest / boosting feature importance is reported per run.
Transparency: what we measured
Held-out R² with the default settings, measured on 2026-10-15. R² here means agreement with the ESA reference map, not with field plots - so every run also reports an independent lidar check (below). The full scientific validation report has every site, a baseline comparison, errors by carbon density and a model benchmark.
| Area | R² (RF / GTB / SVM) | Typical error |
|---|---|---|
| Kitui smallholder farm (277 ha) | 0.53 / 0.54 / 0.50 | ±8 t C/ha |
| 0.25 ha plot, Kitui | 0.57 / 0.56 / 0.55 | ±8 t C/ha |
| Mau farmland (~970 ha) | 0.50 / 0.47 / 0.40 | ±7 t C/ha |
| Kakamega rainforest (~2,500 ha) | 0.89 / 0.89 / 0.91 | ±14 t C/ha |
| Kirinyaga county | 0.90 / 0.90 / 0.89 | ±8 t C/ha |
| Makueni county | 0.86 / 0.88 / 0.78 | ±5 t C/ha |
Small, uniform areas score lower R² because there is little variation to explain, even though their error in t C/ha is small.
- Independent lidar check. Every run is compared with NASA GEDI spaceborne lidar biomass footprints the model never trained on (correlation, bias, typical difference; the most extreme 2.5% of lidar readings are dropped as outliers). The ESA reference - and so any model trained only on it - can sit well above or below the lidar: about 44% above in Kirinyaga, about 43% below on a Kitui farm. An opt-in “Calibrate with GEDI lidar” setting also trains on lidar from the training blocks and weights the models on lidar from separate calibration blocks; checked against lidar in the untouched test blocks, that cut the gap to about +15% and -18%. A straight-line correction to lidar was tested and rejected: it flattened the map towards the average, hiding real differences between places.
- Below-ground carbon. Root carbon is the run's above-ground map multiplied by root-to-shoot ratios from NASA/ORNL's harmonised biomass carbon maps (Spawn et al. 2020); runs report above, below and total carbon.
- Risk screen. Recent tree-cover loss in and within 5 km of the area (Hansen), fire history over 5 years (MODIS burned area) and protected-area overlap (WDPA) - the permanence, leakage and additionality questions a carbon buyer asks. A screening, not due diligence.
- Provenance. Each run lists the exact data versions, imagery window, sample settings and service version behind it, so results can be checked and reproduced.
Why a county's figures can differ
A run you start trains its three models on your area alone, at the run's own resolution. The public county atlas comes from one national run that trains on all of Kenya at a coarser grid, so the same county can read a little differently in each - typically within 10-20%. Inside one run, the Map tab, Zonal Stats table, PDF report and briefing all read the same stored figures.
Which to quote: for a decision about one area, quote your run's ensemble figures. To compare counties with each other, quote the atlas, where every county comes from the same analysis.

Limitations
- Signal saturation - optical and C-band radar lose sensitivity in dense, tall closed-canopy forest, so the highest biomass can be under-estimated.
- Sparse vegetation - very low-cover drylands sit near the noise floor; small absolute errors are large in relative terms. Land ESA leaves blank (below its detection limit: bare ground, grass, low shrub) is counted as 0 t/ha, so the models learn bare land too - before that (to 15 Nov 2026) drylands were over-estimated.
- Cloud & data gaps - persistent cloud or missing scenes force wider compositing windows and raise uncertainty.
- Reference layer - trained against a modelled global product (ESA CCI AGB), not local field plots, so it inherits that product's biases until co-calibrated.
- Carbon fraction - a single IPCC default (0.47) is applied everywhere; species- and tissue-specific fractions vary by roughly ±0.03.
- Screening only - outputs support comparison and targeting, not verified carbon accounting or certification.
Before acting on the numbers
Use Carbon-Eye to screen, compare and target areas. Before any investment, crediting or certification decision:
- Validate with local field data
- Assess uncertainty
- Check land tenure and baselines
- Check leakage and permanence
- Follow the relevant methodology
Glossary
- AGB
- Above-ground biomass - the dry mass of living vegetation above the soil, in tonnes per hectare (t/ha).
- Carbon stock
- Carbon held in that biomass; AGB × 0.47 (IPCC default carbon fraction). Reported in t C/ha.
- CO₂e
- Carbon-dioxide equivalent; carbon stock × 3.67 (44/12), the mass of CO₂ that carbon represents.
- MRV
- Measurement, Reporting and Verification - the process of quantifying and checking carbon changes for climate reporting.
- FREL / FRL
- Forest Reference (Emission) Level - a national baseline against which REDD+ results are measured.
- REDD+
- Reducing Emissions from Deforestation and forest Degradation, plus conservation, sustainable management and enhancement of carbon stocks.
- NDVI / EVI / SAVI
- Vegetation indices from red and near-infrared reflectance that track greenness and canopy density (EVI and SAVI reduce soil and atmosphere effects).
- NDMI / NDRE
- Moisture and red-edge indices - sensitive to canopy water content and chlorophyll.
- SAR backscatter
- The fraction of a radar pulse returned to the sensor; VH (cross-polarised) responds to vegetation volume and structure.
- Ensemble
- The mean of the Random Forest, Gradient Tree Boosting and SVM predictions, used to reduce reliance on any single model.
References & further reading
- ESA Climate Change Initiative - Biomass
- IPCC 2006 Guidelines for National GHG Inventories, Vol. 4 (AFOLU)
- UNFCCC - CDM A/R methodological tools: field measurement manual
- Winrock International - Guidance on calculating carbon stocks
- Chave et al. 2014 - improved allometric models for tropical tree biomass
- Spawn et al. 2020 - harmonised global above- and below-ground biomass carbon
- Mokany et al. 2006 - root:shoot ratios in terrestrial biomes
- GFOI - Integration of remote sensing and ground data
- Copernicus Sentinel missions
- Google Earth Engine
- Global Forest Watch