# Carbon-Eye > Carbon-Eye is a geospatial decision-support system that estimates above-ground biomass and carbon stock across Kenyan counties - combining satellite and environmental data with machine-learning ensembles in Google Earth Engine, as interactive maps and county statistics. Carbon-Eye is free to use in the browser at https://carbon-eye.org. Public pages (landing, methodology, explorer, guide) need no account; running a live analysis needs a free sign-in. ## Pages - [Home](https://carbon-eye.org/): What Carbon-Eye is, how the pipeline works, data sources, models, use cases and FAQ. - [State of Kenya's carbon 2026](https://carbon-eye.org/state-of-kenyas-carbon): Annual open data release: above-ground carbon in all 47 counties, densest and sparsest counties, tree-cover loss, mangroves, downloadable CSV (CC BY 4.0). - [Methodology](https://carbon-eye.org/methodology): Data, predictors, machine-learning ensemble, carbon conversion, validation and limitations. - [Kenya county carbon atlas](https://carbon-eye.org/atlas): Above-ground carbon for all 47 Kenyan counties from one national analysis: mean density, total carbon, CO₂e and tree-cover loss, with a page per county. - [Pricing](https://carbon-eye.org/pricing): Free county screening; project-developer and API tiers with methodology eligibility, dynamic baselines, field calibration and the API. - [Developers: REST API and MCP server](https://carbon-eye.org/developers): Personal API keys, the REST API (analyses, zone statistics, GeoJSON, GeoTIFF, county reports) and the Carbon-Eye MCP server for AI assistants. - [Scientific validation report](https://carbon-eye.org/validation-report): Measured accuracy at seven Kenyan test areas: spatial-block validation, median baseline, errors by carbon density, 90% range coverage, GEDI lidar check and a model benchmark. - [Explorer](https://carbon-eye.org/explore): Rank and filter Kenya's 47 counties by measured carbon density, total carbon and tree-cover loss (national analysis), plus measured per-country biomass and tree-cover loss worldwide (ESA CCI, Hansen). - [Guide](https://carbon-eye.org/guide): Step-by-step walkthrough: create an account, run a live analysis, read every results tab. - [About](https://carbon-eye.org/about): Mission and vision, and who the platform is built for. - [Changelog](https://carbon-eye.org/changelog): Dated record of what has shipped and what is coming next. ## How it works 01. **Select 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). 02. **Build predictor stack** - Earth Engine assembles a cloud-filtered, multi-source stack of optical, radar, terrain, climate and soil layers. 03. **Sample & split** - Reference carbon points are sampled from the stack and split into training and testing sets. 04. **Train models** - Random Forest, Gradient Tree Boosting and SVM each learn the link between predictors and carbon. 05. **Map & 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. ## Models - **Random Forest** - Bagged decision-tree ensemble - **Gradient Tree Boosting** - Sequentially boosted trees - **Support Vector Machine** - Kernel regression The output is the mean of the three models' carbon maps. Carbon stock = above-ground biomass × 0.47 (IPCC default carbon fraction); CO₂e = carbon × 3.67. ## Data sources (all via Google Earth Engine) - [ESA CCI Above-Ground Biomass v7.0](https://climate.esa.int/en/projects/biomass/) - Reference target (2007, 2010, 2015–2024 (v7.0, May 2026)) - [Sentinel-2 Surface Reflectance (Copernicus)](https://sentiwiki.copernicus.eu/web/sentinel-2) - Optical imagery (2015–present, 10–20 m) - [Sentinel-1 GRD & JAXA ALOS PALSAR](https://sentiwiki.copernicus.eu/web/sentinel-1) - Radar (2014–present / 2007–present) - [Google Dynamic World](https://dynamicworld.app/) - Land cover (2015–present, 10 m) - [SRTM DEM](https://www.earthdata.nasa.gov/sensors/srtm) - Terrain (2000, 30 m) - [WorldClim BIO](https://www.worldclim.org/data/bioclim.html) - Climate (1970–2000 normals, ~1 km) - [Meta Canopy Height & MODIS LST](https://sustainability.atmeta.com/blog/2024/04/22/using-artificial-intelligence-to-map-the-earths-forests/) - Structure & temperature (2019–2022, 1 m / 2000–present, 1 km) - [Hansen Global Forest Change (UMD)](https://glad.earthengine.app/view/global-forest-change) - Forest change (2001–latest release, 30 m) - [geoBoundaries ADM1](https://www.geoboundaries.org/) - Boundaries ## Features - **Interactive map** - Switch model, toggle Carbon / AGB view, and inspect any location. - **Model comparison** - RF-vs-GTB difference map and the three-model spread as an uncertainty proxy. - **Validation** - Held-out test data with RMSE, MAE, R² and actual-vs-predicted charts. - **County insights** - Zonal mean, min, max and total carbon per county - downloadable as CSV. - **Feature importance** - The most influential predictors for the Random Forest and boosting models. - **Learning guide & export** - A beginner-friendly guide plus reporting and export tools. ## Who it is for - County environment & forestry teams - Conservation & restoration organisations - Researchers & students - Carbon-project teams (early screening) - Land-use planners & NGOs Supports: Climate-action planning, Restoration screening, Landscape monitoring, County prioritisation, Evidence-based conservation dialogue. ## Validation Models are evaluated on spatial blocks of land the models never trained on against ESA CCI AGB v6 (2020): RMSE, MAE, R² and actual-vs-predicted plots, with the three-model spread as a spatial uncertainty proxy. Measured on 2026-10-15 with the default settings: - Error removed: 30%-74% (vs. a median guess, at 6 Kenyan test areas) - Ensemble RMSE: 4.7-14.5 t C/ha (on held-out spatial blocks) - 90% ranges held: 90% (of held-out points, against a 90% target) ## 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. ## Roadmap - **Transformer & deep-learning models** - Attention-based models over multi-temporal Sentinel stacks, added alongside RF / GTB / SVM in the ensemble. - **LLM-assisted interpretation** - A language-model guide that explains maps, metrics and county results in plain language and drafts reports. - **Beyond Kenya** - Other countries, counties and states, using the same Earth Engine pipeline. - **Deeper validation** - Local field-plot data for calibration and independent accuracy assessment. ## FAQ ### What does Carbon-Eye produce? County-level maps and statistics of above-ground biomass (t/ha) and carbon stock (t C/ha), from an ensemble of machine-learning models trained on the ESA CCI AGB reference layer. Carbon stock is AGB × 0.47 (the IPCC default carbon fraction). ### How do I learn to use the platform? The Guide page (/guide) walks through the full flow - creating an account, configuring and running an analysis, and reading each of the twelve results tabs. The dashboard itself also has a short interactive tour the first time you sign in, covering the New Analysis form; replay it anytime via the "Take the tour" link on that card. ### Which data does it use? Everything runs in Google Earth Engine: Sentinel-2 optical, Sentinel-1 and ALOS PALSAR radar, SRTM terrain, WorldClim climate, high-resolution soil carbon and clay, Meta canopy height, MODIS land-surface temperature, Google Dynamic World land cover and geoBoundaries. The reference target is ESA CCI AGB v7.0 (May 2026, maps to 2024). Each run also reports land context for its area: the share of rainfed and irrigated cropland from Kenya-specific cropland data, the Dynamic World land-cover mix and topsoil carbon. ### Which models are used? Random Forest, Gradient Tree Boosting and a Support Vector Machine. The output blends their three carbon maps, weighted by how accurately each predicted held-out test points (a model that does no better than the average gets almost no weight), which reduces reliance on any single method. Transformer and deep-learning models are on the roadmap. ### How is accuracy reported? On a held-out test split: RMSE, MAE, R² and actual-vs-predicted plots. The three-model spread is shown as a spatial uncertainty proxy, and Random Forest / boosting feature importance is available. ### Is this a verified carbon inventory? No. Carbon-Eye is a screening and prioritisation tool. Results are model estimates - not field inventories, verified carbon stocks or certification decisions. Validate with field data before any crediting or investment decision. ### Does it work outside Kenya? The pipeline is Kenya-first (47 counties) but generalises to other countries, counties and states using the same Earth Engine workflow. ### Who is it for? County environment and forestry teams, conservation and restoration organisations, researchers and students, early-stage carbon-project teams, and land-use planners and NGOs - including the Kenya Forest Service and KEFRI. It is designed for non-specialists as well as technical users. ## Citation Carbon-Eye (2026). Above-ground biomass & carbon stock across Kenyan counties - a geospatial decision-support system.