Science & Technology
Our technological stack and deep science base, from the satellite pass to the anchored report.
Sierra de Guadarrama · Sentinel-2 cloudless by EOX
From the platform
Pinar de Peguerinos, plot 4 (Spain · 8 ha · montane conifer). Crowns detected by our pipeline on very high resolution imagery, and the metrics of each tree.
Our technological stack
The data foundation for the entire pipeline.
We capture Very High Resolution (VHR) multispectral imagery sharp enough to distinguish individual trees, down to 30 cm per pixel. This is paired with Sentinel satellites, which revisit each site frequently for rapid change detection, and LiDAR, a laser-based technology that measures tree height and forest structure.
Scalable, cost-effective monitoring.
Sentinel multispectral and radar data support recurring observations across large areas. We use these datasets to analyse historical land cover, detect forest disturbances, and monitor vegetation changes over time. Where individual-tree detail is not required, Sentinel provides a lower-cost foundation for monitoring, with more detailed observations added according to the project’s needs.
Measuring forest structure.
LiDAR provides measurements of canopy height and vertical forest structure, complementing the information captured in satellite imagery. Combined with available satellite-derived height estimates and field observations, these data help characterise forests and improve the inputs used for biomass and carbon estimation.
Revealing methane and land characteristics.
Hyperspectral observations capture information across many narrow spectral bands, allowing us to analyse features that conventional imagery cannot resolve as clearly. In our methane pipeline, these data are used to measure methane concentrations. For agricultural applications, hyperspectral observations contribute information on soil and vegetation characteristics to our soil carbon modelling.
A complementary view of land conditions.
Synthetic Aperture Radar (SAR) provides observations through cloud cover and independently of daylight. Its sensitivity to moisture and surface and vegetation structure adds information to optical imagery. We use radar observations alongside other datasets to support land monitoring and the development of models for agricultural soils.
The intelligence layer that powers every product.
Our AI identifies and delineates individual tree crowns in VHR RGB/NIR imagery, combining trained segmentation models with spectral indices, texture gradients, and watershed-based detection. Multiple calibrated runs are cross-checked through an ensemble consensus workflow, producing per-tree quality metrics (support, consensus, crown diameter, circularity, ellipse axes). Each crown is then classified by species using multispectral signatures and morphology, currently covering Pinus, Quercus, Eucalyptus, Picea, and Populus. Per-tree biomass is estimated via species-specific allometric equations.
Connecting satellite observations with the ground.
We combine complementary observations rather than relying on a single sensor. For soil organic carbon, we are developing machine learning models that fuse SAR, hyperspectral, and very high resolution imagery with ground sensors, soil samples, and environmental data. These models are calibrated to local conditions to estimate soil carbon stocks and assess changes over time, supporting the monitoring of carbon sequestration under agricultural practices.
From methane concentrations to kilograms per hour.
Our methane pipeline combines hyperspectral concentration measurements with weather data in a physics-based atmospheric plume model. By modelling how methane disperses, we infer the location of emission sources and quantify emission rates in kilograms per hour. Observations before and after mitigation help assess abatement performance and generate evidence for emissions reporting and carbon credit verification.
Assessing the reliability of every estimate.
Field observations provide the reference measurements needed to calibrate and evaluate our models. Depending on the application, these include tree diameter, height and species, laboratory soil carbon measurements, ground sensor readings, and independent methane measurements. We assess uncertainty for each application and document the conditions and limitations affecting the results.
A system designed to enhance trust.
Over 1,000 reference plots with independent ground-truth measurements (DBH, height, species) are cross-checked against NOBRAC outputs, keeping uncertainty on total biomass below 5%. Every step, from the original imagery to the model version, parameters, and final results, is anchored to a public blockchain via Merkle-tree commitments, ensuring full third-party reproducibility.
Exemplifying our approach
01
We acquire VHR multispectral imagery (30 cm GSD), supplemented by stereoscopic pairs for canopy height modelling, Sentinel-1/-2 for high-frequency change detection, and LiDAR (GEDI / ICESat-2) for height measurement and forest structure validation. Every image is recorded with date, time, sensor metadata, and acquisition geometry, and anchored on-chain from the moment it's captured.
02
We use annotated crown datasets to train, calibrate, and validate tree-detection models on very-high-resolution RGB/NIR imagery. The production pipeline combines learned crown evidence with spectral indices, texture gradients, multi-scale top-hat enhancement, local-maximum peak detection, non-maximum suppression, and marker-controlled watershed segmentation. Multiple calibrated runs are fused through a pixel-voting ensemble, generating crown outlines with support, mean consensus, diameter, equivalent diameter, circularity, and fitted ellipse metrics.
03
Neural networks classify each tree crown by species, using multispectral signatures and crown morphology. Current coverage includes Pinus, Quercus, Eucalyptus, Picea, and Populus, with tropical species coverage expanding. Each classified tree then carries the information needed to estimate its biomass.
04
Per-tree biomass is estimated using species-specific allometric equations, calibrated to local conditions (climate, soil, stand density). Above-ground biomass (AGB) is derived from crown geometry, below-ground biomass (BGB) from established root:shoot ratios, and soil organic carbon (SOC) from environmental covariates. Total carbon stock = AGB + BGB + SOC × 0.5 (carbon fraction).
05
These estimates are checked against the ground: over 1,000 reference plots provide independent measurements (DBH, height, species), cross-checked against NOBRAC outputs. Target uncertainty is below 5% on total biomass, currently being validated, with reports published openly.
06
Every step in this pipeline (imagery hash, model version, parameter set, output report) is anchored to a public blockchain via Merkle-tree commitments, ensuring full third-party reproducibility.
Beyond carbon
NOBRAC's proprietary innovation to go beyond carbon.
A composite metric combining species diversity, climate-risk exposure (drought, fire, pest), and ecosystem health into a single auditable score, designed to satisfy CRCF Article 7's sustainability criteria. Each component is independently validated and reported.
Transparent data for meaningful change
Pinar de Peguerinos · 40.6412 N 4.1992 W · PNOA © IGN