Sensors
- Maxar WorldView-3: 0.31 m panchromatic, 1.24 m multispectral (8 bands including red-edge and near-infrared), on-demand tasking. Crown delineation at this resolution resolves individual shade trees down to roughly 2 m crown diameter in closed mixed canopy.
- Airbus Pléiades Neo: 0.30 m panchromatic, 1.2 m multispectral (6 bands), tri-stereo collection mode enables structure-from-motion digital surface models at roughly 0.5 m vertical precision, supporting canopy height estimation without lidar.
- Planet SkySat: 0.50 m panchromatic, 1.0 m multispectral (5 bands), collect-on-demand with same-day revisit feasible in many tropical regions. Useful for change detection between baseline and verification epochs at lower cost per hectare than Maxar or Pléiades.
- Sentinel-2 MSI: 10 m multispectral (13 bands, 5-day revisit at equator). Too coarse for individual tree segmentation but provides phenological context, NDVI time-series for canopy seasonality, and cloud-gap-fill support for the higher-resolution layers.
Why standard forest protocols miscount agroforestry trees
Most voluntary carbon methodologies, including Verra's VM0010 and the Gold Standard's agroforestry framework, require field-measured tree inventories as the primary evidence of above-ground biomass. In a cocoa or coffee system with 50 to 150 shade trees per hectare scattered across a 500-hectare project, that means tens of thousands of individual measurements. Projects routinely sample 10 to 20 per cent of plots and extrapolate, introducing uncertainty ranges that can exceed 30 per cent of claimed stock.
The spectral problem compounds the counting problem. At 10 m resolution, a Sentinel-2 pixel over a silvopastoral system contains grass, cattle-browsed shrubs and emergent timber trees simultaneously. The resulting NDVI or leaf-area-index estimate is a mixture signal that correlates poorly with any single component's biomass. Protocols that treat the pixel as forest undercount, and those that exclude it as non-forest miss real carbon entirely.
What a sub-metre image gives away about individual crowns
At 0.30 to 0.50 m ground sampling distance, the boundary between a shade tree crown and the crop canopy beneath it becomes a detectable edge. Instance segmentation algorithms, trained on manually labelled very-high-resolution imagery, routinely achieve intersection-over-union scores above 0.75 for dominant shade species such as Erythrina, Gliricidia and Cordia in published studies using WorldView and Pléiades data. Crown projected area, crown diameter and, where stereo collection is available, crown height can all be extracted per tree.
Crown area is the entry point to allometric conversion. Species-specific and pantropical allometric equations, most prominently those published by Chave et al. (2014) and collated in the GlobAllomeTree database, relate crown diameter or projected area to above-ground dry biomass for hundreds of tropical species. Uncertainty is real: a single allometric equation applied to a mixed-species system carries a coefficient of variation of roughly 20 to 35 per cent for individual trees. Aggregated across thousands of trees in a project area, that uncertainty shrinks but does not vanish. Honest MRV reporting must carry it forward into the confidence interval on total stock.
Structure-from-motion as a field-free height proxy
Stereo and tri-stereo collection modes on Pléiades Neo and WorldView-3 produce dense point clouds through structure-from-motion photogrammetry. Differencing a canopy surface model against a bare-earth digital terrain model yields canopy height at roughly 0.5 to 1.0 m vertical precision in open-canopy agroforestry, where ground returns are accessible in gaps. That is substantially coarser than airborne lidar, which achieves 0.10 to 0.20 m vertical precision, but it is sufficient to separate the 4 to 8 m cocoa shrub layer from 12 to 25 m shade trees without a single field measurement.
Height matters because the best-performing pantropical allometric models use diameter at breast height, wood density and height together. Crown area from imagery can substitute for diameter through species-specific crown-diameter-to-DBH regressions, and photogrammetric height closes the third variable. The resulting biomass estimate is noisier than a direct field measurement but covers 100 per cent of the project area rather than a 15 per cent sample, which is a different kind of accuracy.
Temporal stacking and the additionality question
A single-date tree count answers 'how much carbon is here now.' Additionality requires 'how much more carbon is here than there would have been without the project.' That demands a credible baseline epoch, typically three to five years before project start, and a change trajectory.
Planet SkySat's archive extends to 2017 for many tropical regions, and Pléiades archive imagery is available from 2012. Sentinel-2 provides continuous 10 m coverage from 2015. Combining these, a project developer can reconstruct canopy cover at baseline with sub-metre resolution for the verification epoch and coarser but consistent Sentinel-2 NDVI for the intervening years. The temporal stack also catches permanence failures: a project area where canopy cover declined between verification visits has a carbon reversal that the registry needs to see, whether the developer reports it or not.
Honest limits: cloud, species confusion and the gap between crown and root
Tropical agroforestry regions are frequently cloudy. In the humid West African cocoa belt, Sentinel-2 achieves fewer than 20 cloud-free acquisitions per year in many pixels. Very-high-resolution tasking is worse: a single overcast day wastes a collection attempt worth thousands of dollars. Practical workflows require cloud-probability screening, multi-date compositing and acceptance that some areas will need repeat tasking across two or three seasons before a clean mosaic is achievable.
Species misclassification is a persistent source of error. Erythrina and Gliricidia have similar crown shapes at 0.5 m resolution, but their wood densities differ by roughly 15 per cent, which flows directly into biomass error. Hyperspectral sensors resolve this better than multispectral ones, but no commercial constellation currently offers hyperspectral at sub-metre resolution over large areas. Multispectral species classification should be treated as a genus-level or functional-group-level estimate, not a species-level one.
Above-ground biomass is also only part of the carbon story. Roots, soil organic carbon and harvested wood products are not visible from space at all. Satellite-derived above-ground stock is a necessary input to a full MRV stack, not a complete one. Satellize builds this layer as part of a broader analytics workflow; the Tonga crop-estimation programme used a similar multi-sensor compositing approach to separate crop types in a spectrally mixed agricultural landscape, and the same logic applies here.
From segmented crowns to registry-ready numbers
The output chain runs: orthorectified very-high-resolution mosaic, instance segmentation to individual crown polygons, crown metric extraction (area, diameter, height where stereo is available), species or functional-group classification, allometric conversion to above-ground dry biomass, summation to project-area stock with propagated uncertainty, and comparison against baseline epoch to produce net carbon stock change.
Delivered as a GIS layer, each crown polygon carries its biomass estimate and confidence interval. Aggregated to the project boundary, the total becomes a single number with an honest uncertainty range that a third-party auditor can interrogate at any spatial scale. That auditability is what distinguishes satellite-derived MRV from extrapolated field sampling: the evidence is spatially explicit, reproducible and, critically, available to the registry without the project developer controlling access to the raw data.
Typical figures
| Best available spatial resolution | 0.30 m (Pléiades Neo, WorldView-3 panchromatic) |
| Multispectral resolution | 1.0 to 1.24 m (SkySat, Pléiades Neo, WorldView-3) |
| Minimum detectable crown diameter | Approximately 2 m at 0.30 to 0.50 m GSD; smaller crowns lost in adjacent canopy |
| Canopy height precision (photogrammetric DSM) | 0.5 to 1.0 m vertical RMSE in open-canopy agroforestry; degrades under closed canopy |
| Revisit for change detection | SkySat: same-day feasible; Pléiades Neo / WorldView-3: 1 to 3 days on-demand; Sentinel-2: 5 days (cloud-limited in tropics) |
| Spectral bands used | Panchromatic, blue, green, red, red-edge, near-infrared (and SWIR on WorldView-3 for canopy moisture context) |
| Allometric uncertainty (individual tree) | CV 20 to 35% per tree; reduces to roughly 10 to 15% at project-area aggregation |
| Archive depth | SkySat from 2017; Pléiades from 2012; Sentinel-2 from 2015 (all subject to tasking history) |
| Deliverable formats | GeoTIFF crown-mask raster, GeoPackage crown polygon layer with biomass attributes, PDF uncertainty report |
Analytics Satellize can run
| Individual tree crown segmentation map | Instance segmentation (e.g. Mask R-CNN class) trained on labelled VHR imagery; applied to orthorectified Pléiades Neo or WorldView-3 mosaic | GeoPackage polygon layer, one feature per detected crown with area, diameter and centroid coordinates |
| Canopy height model | Structure-from-motion photogrammetry from tri-stereo Pléiades Neo collection; DSM minus DTM differencing | GeoTIFF raster at 0.5 m resolution covering the project boundary |
| Species or functional-group classification | Random forest or support vector machine classifier on multispectral band ratios and crown shape metrics; validated against field reference points | Crown polygon layer with species or functional-group label and classification confidence score |
| Above-ground biomass stock estimate | Published pantropical allometric equations (Chave et al. 2014 framework) applied per crown using derived diameter and height; Monte Carlo uncertainty propagation | Project-area biomass summary table (Mg dry matter) with 90% confidence interval, plus per-crown GIS attribute |
| Baseline-to-verification canopy change layer | Multi-date crown segmentation comparison across archive VHR epochs; Sentinel-2 NDVI time-series for gap-fill and trend context | Change polygon layer flagging crown gains, losses and no-change, with net stock delta in Mg CO2e |
| Sentinel-2 phenological context stack | Harmonic NDVI time-series decomposition on Sentinel-2 MSI archive (2015 to present) to characterise seasonal canopy dynamics and flag anomalies | Annual NDVI amplitude and greenness-peak-date rasters; anomaly alert CSV for project monitoring periods |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.