Canopy height model derivation from spaceborne lidar
GEDI and ICESat-2 measure canopy height with centimetre-scale vertical precision but sparse footprints. Fusing those samples with Sentinel-1 SAR and Sentinel-2 optical covariates via ML regression produces wall-to-wall canopy height models, with honest uncertainty that varies by forest structure.
Sensors
- GEDI (Global Ecosystem Dynamics Investigation): Full-waveform lidar on the ISS. Three lasers split into eight ground tracks, each with a 25 m diameter footprint, spaced approximately 60 m across-track and 60 m along-track. Measures relative height metrics (RH50, RH98 etc.) to roughly ±1 m vertical accuracy in closed canopy. Restricted to latitudes 51.6°N–51.6°S by the ISS orbital inclination. No coverage above those latitudes.
- ICESat-2 ATLAS: Photon-counting lidar with six beams arranged in three pairs. Ground footprint roughly 17 m diameter, along-track spacing ~0.7 m. Reaches all latitudes including the poles. Canopy height retrieval is noisier in dense tropical canopy than GEDI because single-photon counting requires careful noise filtering, and the narrower footprint can miss sub-canopy returns in closed stands.
- Sentinel-1 SAR (C-band): C-band synthetic aperture radar at 5.405 GHz. IW mode delivers 10 m ground range resolution with a 250 km swath. Repeat cycle 6–12 days depending on latitude and acquisition plan. Backscatter and coherence in VV and VH polarisations correlate with canopy volume and structure, though C-band saturates in biomass above roughly 100–150 t/ha, limiting sensitivity in the tallest old-growth stands.
- Sentinel-2 MSI: 13-band multispectral imager. Vegetation-relevant bands at 10 m (red, NIR) and 20 m (red-edge, SWIR). Revisit 5 days at the equator. Spectral indices (NDVI, EVI, NBR, red-edge chlorophyll index) serve as proxy covariates for canopy density and leaf area, which correlate with height in structurally consistent forest types. Cloud cover is a persistent problem in humid tropical regions.
What the waveform actually measures
GEDI does not photograph the forest. It fires a laser pulse at the canopy, then records the full time-series of returning photons as a waveform. The top of the waveform marks the highest surface that reflected light; the bottom marks the ground. The vertical distance between them, corrected for terrain, is canopy height. Relative height metrics, RH98 in particular, are the standard proxy for top-of-canopy height and are published in the GEDI Level 2A product.
ICESat-2 works differently. Its ATLAS instrument counts individual photons and records their return time with extraordinary precision, achieving along-track sampling every 0.7 m. But single photons are noisy: solar background photons arrive constantly, and separating signal from noise in a dense canopy requires statistical filtering. The ATL08 land and vegetation height product applies an adaptive algorithm to classify photons as ground, low vegetation or canopy, then reports height percentiles per 100 m segment. Vertical accuracy in boreal and temperate forests is generally within 2–3 m; tropical closed canopy degrades that figure.
The sampling problem neither mission solves alone
Neither GEDI nor ICESat-2 produces a wall-to-wall map. GEDI's eight ground tracks sample perhaps 4% of a given landscape on any pass. ICESat-2's six beams are even narrower in aggregate ground coverage. Both missions accumulate samples over months to years, but the result is still a sparse point cloud, not a continuous raster. For most operational uses, a forest manager or carbon programme needs height estimates at every pixel, not just where a laser happened to point.
GEDI's latitude restriction is a hard physical constraint, not a data gap that will be filled later. The ISS orbital inclination of 51.6° means that boreal forests in Canada, Russia and Scandinavia receive no GEDI coverage at all. ICESat-2 fills that geographic gap but with lower canopy-height accuracy in structurally complex stands. Any canopy height model for high-latitude forests must rely more heavily on ICESat-2 and on the SAR and optical covariates, which introduces additional uncertainty.
Interpolating to a continuous surface
The standard approach, documented in multiple peer-reviewed studies and adopted by programmes including NASA's Global Land Analysis and Discovery (GLAD) lab, is to train a machine-learning regressor on the lidar samples and a stack of spatially continuous covariates. Sentinel-1 backscatter and coherence, Sentinel-2 spectral indices, elevation and slope from a digital terrain model, and sometimes texture metrics from SAR all enter as predictors. Random forest regression and gradient boosting are the most common model families; they handle non-linear relationships between backscatter saturation and height reasonably well.
The output is a predicted canopy height at every pixel, typically at 10–25 m resolution to match the coarser of the input layers. Crucially, a well-implemented pipeline also outputs a per-pixel uncertainty estimate, usually the standard deviation across ensemble members or a quantile regression interval. This uncertainty layer is not optional decoration. In structurally heterogeneous landscapes, such as mosaic agroforestry, secondary regrowth patches or forest edges, prediction error can exceed 5–8 m even when the model performs well on average. Reporting only the mean prediction without the uncertainty is misleading.
Where the method breaks down
C-band SAR saturates. In dense tropical forest with above-ground biomass above roughly 100–150 t/ha, the VH backscatter signal stops increasing with additional biomass or height. The model then relies more heavily on optical covariates, which are themselves obscured by cloud in the same humid tropical environments where the tallest forests grow. This is the central tension: the places where accurate canopy height matters most for carbon accounting are the places where cloud and SAR saturation conspire to degrade the estimate.
Terrain also confounds both the lidar retrieval and the regression. Steep slopes cause the GEDI footprint to intercept ground returns at angles that inflate apparent height if not corrected precisely. The GEDI L2A product applies a slope correction, but residual errors of 1–3 m persist on slopes above 30°. ICESat-2 ATL08 has similar terrain sensitivity. Any canopy height model built over mountainous forest should be treated with additional scepticism, particularly on south-facing slopes in the northern hemisphere where canopy closure varies sharply with aspect.
Calibration, validation and what a buyer should ask for
A canopy height model is only as credible as its validation dataset. The standard practice is to withhold a spatially stratified subset of the lidar samples from training, predict height at those locations, and report RMSE and bias by forest type and height class. An honest validation report will show that error is not uniform: the model is usually most accurate in closed-canopy, structurally homogeneous forest and least accurate at the extremes, both in sparse, open woodland and in the very tallest emergent trees.
Airborne lidar, where available, provides the gold standard for independent validation. Comparing spaceborne-model predictions against airborne point clouds over representative sample areas gives a far more demanding test than withheld GEDI samples alone, because airborne lidar captures the full three-dimensional structure that the regression model is trying to predict. Satellize incorporates this cross-validation step when client programmes include airborne reference data, as it did in structuring the covariate framework for the Tonga crop-estimation programme, where field-measured canopy parameters anchored the satellite-derived estimates. Buyers should ask any provider to show stratum-specific RMSE, not just a headline figure.
Practical specifications a procurement team needs
Delivered canopy height models typically come as GeoTIFF rasters at 10 or 25 m pixel spacing, with a companion uncertainty raster in the same projection. Temporal currency depends on the GEDI acquisition window used: GEDI has been operational since April 2019, so multi-year composites are available, but the ISS orbit drifts and revisit patterns are not uniform. A model built from a single season of GEDI data will reflect phenological conditions at acquisition time, which matters in deciduous forests where leaf-off heights differ from leaf-on.
Processing latency from data acquisition to delivered model is typically weeks to a few months for a national-scale product, depending on cloud cover over the optical compositing period and the size of the area of interest. Near-real-time canopy height is not a realistic expectation from this method. It is a periodic, campaign-style product, best used as a baseline layer against which change detection from other sensors is referenced.
Typical figures
| GEDI footprint diameter | 25 m |
| GEDI vertical accuracy (closed canopy) | ~±1 m (RH98 vs. airborne lidar reference) |
| GEDI latitude coverage | 51.6°S to 51.6°N only |
| ICESat-2 ATLAS along-track spacing | ~0.7 m; ATL08 product reports per 100 m segment |
| ICESat-2 latitude coverage | Global, including poles |
| Sentinel-1 IW mode resolution | 10 m ground range; 6–12 day repeat |
| Sentinel-2 MSI resolution (vegetation bands) | 10 m (red, NIR); 20 m (red-edge, SWIR) |
| Delivered CHM pixel spacing (typical) | 10–25 m |
| Interpolation RMSE (structurally homogeneous forest) | 2–5 m typical; 5–10 m in heterogeneous or edge landscapes |
| GEDI archive depth | April 2019 to present |
Analytics Satellize can run
| National or regional canopy height model | ML regression (random forest or gradient boosting) trained on GEDI L2A RH98 samples against Sentinel-1/2 and terrain covariates | GeoTIFF raster at 10–25 m resolution with companion per-pixel uncertainty layer |
| High-latitude canopy height model (boreal) | ICESat-2 ATL08 height percentiles as training labels substituted for GEDI where latitude exceeds 51.6° | GeoTIFF raster with stratum-specific RMSE report |
| Canopy height change layer (epoch comparison) | Differencing of two independently trained CHM composites anchored to different GEDI acquisition windows | Change raster with significance mask and summary statistics by land-cover class |
| Validation report against withheld lidar samples | Spatially stratified hold-out; RMSE and bias reported by forest type, height class and slope band | PDF technical report with scatter plots and error maps |
| Uncertainty-weighted height percentile map | Quantile regression forest producing P10/P50/P90 canopy height estimates per pixel | Three-band GeoTIFF (low, median, high estimate) for use in Monte Carlo carbon uncertainty propagation |
| Forest structural complexity index | Derived from GEDI waveform metrics (plant area index, foliage height diversity) combined with CHM texture statistics | GIS layer with accompanying methodology note |
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.