Sensors
- GEDI (NASA/ISS): Full-waveform lidar at 25 m footprints spaced 60 m apart along eight ground tracks. Provides canopy cover fraction (rh metrics), plant area index and foliage height diversity directly, without spectral ambiguity. ISS orbital inclination limits coverage to roughly 51.6° N/S; tropical and subtropical agroforestry zones are well served. No night-time or cloud penetration limitation for lidar returns, though dense cloud prevents acquisition.
- Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands, 20 m in red-edge and shortwave infrared. Five-day revisit at the equator with two satellites. Red-edge bands (B5, B6, B7 at 705, 740, 783 nm) are sensitive to canopy chlorophyll density, useful for separating sparse agroforestry tree cover from bare soil and herbaceous understorey via spectral unmixing. Free and open archive from 2015.
- Planet SuperDove: 3 m resolution, eight spectral bands including two red-edge channels, near-daily revisit globally. Resolves individual tree crowns above roughly 5 m diameter, enabling crown delineation and density counting in parkland and silvopastoral systems where GEDI footprint spacing is too coarse for stand-level estimates. Commercial licence required.
- Maxar WorldView-3: 31 cm panchromatic, 1.24 m multispectral, 3.7 m shortwave infrared. Eight VNIR bands plus eight SWIR bands allow fine-grained spectral discrimination of canopy types. At this resolution, individual crowns are countable in open parkland systems. Tasking is on-demand; archive depth varies by location. High cost per km² limits use to validation strips rather than national-scale wall-to-wall mapping.
The gap that neither statistician sees
National forest inventories typically apply a canopy cover threshold, often 10 or 30 per cent, combined with a minimum area and tree height criterion. Agroforestry systems sit awkwardly around those thresholds. A shade-grown cocoa plot at 40 per cent canopy cover with 8 m trees may qualify as forest under one national definition and as cropland under another. Agricultural surveys, meanwhile, record the cash crop and ignore the trees entirely. The result is a systematic blank in both datasets.
The carbon accounting consequences are material. The Intergovernmental Panel on Climate Change Tier 1 default biomass factors for agroforestry are acknowledged to carry large uncertainty, and many countries simply omit these systems from their national greenhouse gas inventories or lump them into a residual land-use category. Satellite methods offer a way to produce spatially explicit, annually updatable estimates that are defensible to third-party auditors.
What a floating roof gives away: canopy cover from GEDI waveforms
GEDI fires laser pulses at 242 Hz and records the full vertical distribution of returned energy. From that waveform, the algorithm extracts canopy cover fraction as the ratio of canopy returns to total returns, corrected for ground reflectance. In agroforestry contexts the metric of most immediate value is the cover fraction at each 25 m footprint, which can be spatially interpolated or fused with wall-to-wall optical data to produce continuous cover maps.
Published validation work using airborne lidar as reference shows GEDI canopy cover estimates within roughly ±5 percentage points in closed canopy conditions, with larger errors in sparse and heterogeneous canopies precisely the conditions common in parkland agroforestry. That honest caveat matters: in Sahelian parkland systems where Faidherbia albida trees are scattered at 10 to 30 per cent cover, GEDI footprint-level estimates carry meaningful uncertainty, and crown-level optical counting from Planet or WorldView provides a useful cross-check.
GEDI's ISS orbit means footprints are not on a regular grid. Fusion with Sentinel-2 spectral layers is the standard approach for producing continuous national-scale maps: GEDI provides the calibration signal for canopy cover, Sentinel-2 provides the spatial continuity.
Spectral unmixing: reading the mixture, not the dominant class
A Sentinel-2 pixel at 10 m in a parkland system contains a mixture of tree crown, herbaceous ground cover, bare soil and shadow. A standard supervised classifier assigns the pixel to whichever class wins the majority vote, which tends to produce a map that calls the whole area grassland. Spectral unmixing instead estimates the fractional contribution of each endmember to the pixel's reflectance. The tree fraction image that results is a continuous variable, not a binary class, and it correlates well with field-measured canopy cover.
Red-edge bands improve separation between sparse woody cover and dense herbaceous vegetation, which have similar broadband NDVI values but different chlorophyll absorption depth signatures. The shortwave infrared bands in Sentinel-2 and WorldView-3 add sensitivity to canopy water content and woody structure. For shade-grown coffee and cocoa, the multi-date approach matters: acquiring imagery during the dry season, when understorey grasses senesce, sharpens the spectral contrast between persistent tree canopy and seasonal ground cover.
Crown counting and density: where very high resolution earns its cost
Canopy cover fraction tells you how much sky is blocked. Crown density, meaning the number of individual trees per hectare, requires a different approach. In open parkland systems with crown diameters of 5 to 20 m, Planet SuperDove at 3 m can resolve individual crowns well enough for automated detection using local maxima filtering or deep learning object detection trained on manually labelled samples.
WorldView-3 at 31 cm panchromatic takes this further, resolving crown shape, gap structure and even species-level texture in some cases, though the cost means it is used for validation transects rather than national coverage. A practical national workflow uses WorldView-3 or Planet to train and validate a crown detector, then applies that detector to Planet's near-daily archive for annual density estimates across the full mapping area.
Minimum detectable crown diameter is roughly 3 to 5 m at Planet resolution and under 1 m at WorldView-3 resolution. Below those sizes, individual tree detection gives way to sub-pixel fraction estimation.
Distinguishing agroforestry from closed forest and open cropland
The classification problem has two boundaries. At the upper end, a shade-grown cocoa system at 70 per cent cover begins to resemble degraded forest spectrally and in GEDI waveform shape. At the lower end, a sun-grown coffee plot with scattered shade trees at 5 per cent cover is nearly indistinguishable from open cropland in 10 m imagery. Both boundaries require explicit treatment.
The most defensible approach combines three signals: GEDI canopy cover fraction to establish the vertical structure, Sentinel-2 spectral unmixing to estimate tree fraction continuously, and phenological trajectory from the Sentinel-2 time series to detect the seasonal behaviour of the understorey. Closed forest has a stable canopy signal year-round. Agroforestry systems show a stable upper canopy layer but a seasonally varying understorey. Open cropland shows seasonal variation throughout the pixel. That three-way phenological separation is not perfect, but it is physically grounded and auditable.
Satellize applies this fusion workflow operationally, having developed comparable spectral unmixing pipelines for the Tonga crop-estimation programme. The agroforestry mapping problem adds the GEDI lidar dimension and the need to handle multi-storey canopy structure, but the open-data infrastructure is the same.
Honest limits and what to do about them
Cloud cover is the most persistent operational constraint. Sentinel-2 optical data over humid tropical agroforestry zones can have fewer than 20 usable clear observations per year in some locations. Time-series compositing across the full annual archive reduces but does not eliminate cloud contamination. Sentinel-1 SAR backscatter, which penetrates cloud, adds a structural signal that can partially substitute in heavily cloud-affected areas, though SAR-optical fusion for canopy cover retrieval carries its own calibration complexity.
GEDI's sampling density means that in a 1 km² area, you might have 20 to 40 footprints, depending on latitude and acquisition history. That is sufficient for statistical characterisation of a landscape unit but not for mapping individual field parcels. For parcel-level work, Planet or WorldView is the necessary instrument.
Finally, the distinction between agroforestry and forest degradation is genuinely ambiguous in some systems. A heavily shaded cocoa farm established inside former forest may have canopy cover and structure metrics nearly identical to a lightly degraded forest. Ground-truth data, ideally from national agricultural surveys or community mapping, remains necessary to resolve that ambiguity at the boundary. Remote sensing narrows the problem; it does not eliminate the need for local knowledge.
Typical figures
| GEDI footprint diameter | 25 m, footprints spaced 60 m along-track |
| GEDI latitudinal coverage | Approximately 51.6° N to 51.6° S (ISS orbital constraint) |
| Sentinel-2 spatial resolution | 10 m (VNIR), 20 m (red-edge, SWIR) |
| Sentinel-2 revisit | 5 days at equator (2-satellite constellation) |
| Planet SuperDove resolution | 3 m, 8 spectral bands, near-daily revisit |
| WorldView-3 resolution | 31 cm pan, 1.24 m multispectral, 3.7 m SWIR |
| Minimum detectable crown (Planet) | Approximately 3–5 m diameter |
| Canopy cover fraction accuracy (GEDI) | Typically ±5 percentage points in closed canopy; larger uncertainty in sparse systems |
| Sentinel-2 archive depth | From 2015 (Sentinel-2A launch) |
| GEDI archive depth | From April 2019 (ISS deployment) |
Analytics Satellize can run
| National canopy cover fraction map | GEDI-calibrated Sentinel-2 spectral unmixing; GEDI rh metrics as training labels for a regression model predicting per-pixel tree fraction | GeoTIFF raster at 10 m, continuous 0–100% cover fraction, annual update |
| Agroforestry system extent and class map | Three-class separation (closed forest / agroforestry / open cropland) using GEDI cover fraction, Sentinel-2 spectral unmixing and phenological trajectory analysis of annual time series | Vector polygon layer with class label, cover fraction and confidence score; compatible with QGIS, ArcGIS and PostGIS |
| Tree density estimate (crowns per hectare) | Automated crown detection on Planet SuperDove or WorldView-3 imagery using local maxima filtering or object-detection model trained on manually labelled crowns | Point layer of detected crowns plus summary statistics per administrative unit or land parcel; CSV and GeoJSON |
| Phenological separation index | Harmonic analysis of Sentinel-2 NDVI and red-edge time series to decompose stable canopy signal from seasonal understorey variation | Raster stack of harmonic coefficients plus derived agroforestry probability layer; annual delivery |
| Carbon accounting input layer | Canopy cover fraction and crown density combined with IPCC Tier 2 or country-specific biomass factors to produce spatially explicit carbon stock proxy; methodology documented for national inventory submission | Tabular summary by administrative unit with uncertainty ranges; supporting methodology note for UNFCCC reporting |
| Change detection: agroforestry expansion or contraction | Bi-temporal or annual time-series comparison of canopy cover fraction maps; change pixels flagged by threshold and reviewed against Sentinel-1 SAR to reduce cloud-driven false positives | Change polygon layer with direction, magnitude and date of detected change; annual GIS layer |
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.