Above-ground forest biomass and carbon-density mapping
Estimating above-ground biomass at landscape scale requires combining L-band SAR, spaceborne lidar and optical indices, because no single sensor avoids saturation in high-biomass tropical forest. Uncertainty is real and must be quantified, not hidden.
Sensors
- ALOS-2 PALSAR-2 (JAXA): L-band (1.27 GHz) SAR with 25 m resolution in ScanSAR mosaic mode. L-band penetrates deeper into canopy structure than C-band, extending the biomass dynamic range to roughly 150–200 Mg/ha before backscatter saturates, compared with approximately 100 Mg/ha for C-band. Global annual mosaics are publicly released by JAXA.
- GEDI (NASA/GSFC): Spaceborne full-waveform lidar on the ISS. Delivers 25 m footprint canopy-height and vertical-structure metrics between 51.6° N and S. Not a wall-to-wall imager: footprints are sampled in parallel tracks roughly 600 m apart, so spatial coverage requires interpolation or fusion. Provides the structurally informed canopy-height input that reduces biomass prediction error in tall, dense stands.
- Sentinel-2 MSI (ESA/Copernicus): 10 m resolution visible and near-infrared bands, 20 m shortwave infrared. 5-day revisit at the equator with two satellites. Spectral indices (NDVI, EVI, NBR) correlate with biomass in low-to-medium density forest but saturate in closed-canopy tropical conditions above roughly 80–100 Mg/ha. Cloud cover in humid tropics severely limits optical acquisitions.
- Landsat 8/9 OLI (USGS/NASA): 30 m multispectral, 16-day single-satellite revisit. Decades of archive back to 1972 (MSS) enable long-term change analysis. OLI's radiometric calibration is well-characterised, making it suitable for consistent spectral index time series. Shares the same saturation ceiling as Sentinel-2 in dense tropical stands.
Why the physics of saturation is the first thing to understand
Radar backscatter increases with forest biomass up to a point, then flattens. For C-band SAR (Sentinel-1 at 5.4 GHz), that ceiling sits around 100 Mg/ha. Above it, adding more wood changes the signal very little. The sensor is effectively blind to the difference between a moderately dense forest and an old-growth stand. This is not a calibration problem or a processing artefact. It is a consequence of wavelength: shorter wavelengths interact primarily with leaves and small branches near the top of the canopy, which become opaque to the radar well before the canopy reaches its structural maximum.
L-band radar, at roughly 24 cm wavelength, penetrates further into the canopy volume and interacts with larger woody elements, trunks and major branches. ALOS-2 PALSAR-2 therefore extends the dynamic range to somewhere between 150 and 200 Mg/ha in most forest types, though published estimates vary by forest structure and moisture conditions. Above that threshold, L-band also saturates. In the world's densest tropical forests, where above-ground biomass can exceed 300–400 Mg/ha, even PALSAR-2 reaches its ceiling. No current spaceborne SAR system resolves this range unambiguously on its own.
What GEDI adds, and what it cannot do
GEDI's full-waveform lidar measures canopy height and vertical structure directly, not through a proxy relationship with backscatter. Canopy height is strongly correlated with above-ground biomass across most forest biomes, and GEDI-derived height metrics have been shown to reduce prediction error in high-biomass stands where SAR has already saturated. The combination of PALSAR-2 backscatter and GEDI height in a regression or machine-learning model typically outperforms either input alone.
The honest caveat is coverage. GEDI's footprints are not continuous. The instrument fires along discrete tracks, leaving gaps that must be filled by spatial modelling or by using GEDI-derived height products that have already been interpolated to wall-to-wall grids (such as those distributed through NASA Earthdata). The ISS orbit also excludes latitudes above 51.6°, ruling out boreal forests in northern Canada and Russia. Cloud is not a problem for lidar, but the ISS orbital precession means revisit at any given point is irregular.
Optical indices: useful context, not a substitute
Sentinel-2 and Landsat OLI contribute spectral indices, NDVI, EVI and the red-edge bands available on Sentinel-2, that correlate with leaf-area index and, indirectly, with biomass in open and semi-open canopies. They are most useful as ancillary covariates in a fusion model rather than as primary biomass estimators. In humid tropical forest, cloud cover is the more immediate problem: achieving a cloud-free composite over the Congo Basin or the Amazon can require months of acquisitions, and even then residual haze affects index values.
Optical data also carry no vertical structure information. Two stands with identical NDVI can differ by a factor of three in above-ground biomass if one is young regrowth and the other is mature closed-canopy forest. This is why structural inputs from GEDI, or eventually from future P-band SAR missions such as BIOMASS (ESA, planned launch 2025), are necessary rather than supplementary in high-biomass environments.
Calibrating against field plots: the link that holds the model together
Remote-sensing biomass estimates are only as good as the field inventory plots used to calibrate and validate them. Standard practice follows allometric equations applied to trunk diameter and height measurements in sample plots, typically 0.1 to 1 ha in size. The resulting plot-level biomass estimates carry their own uncertainty from the allometric equations themselves, which can be 10–30% even for well-studied species.
Plot placement matters enormously. Plots clustered in accessible lowland forest will produce a model that extrapolates poorly to steep terrain or waterlogged stands. Stratified sampling across forest types, elevation bands and disturbance histories is necessary for a nationally representative product. Where national forest inventory data are sparse or proprietary, publicly available datasets such as the Global Forest Biomass Database (GFBi) provide partial coverage but should not be treated as a substitute for country-specific ground truth.
Uncertainty is not a footnote, it is part of the product
A biomass map without an accompanying uncertainty layer is incomplete. In low-to-medium biomass forest (below roughly 150 Mg/ha), fusion models combining PALSAR-2 backscatter, GEDI height and optical indices can achieve relative uncertainties of 20–30% at the pixel level, improving to 10–15% when aggregated to stand or watershed scale. In dense tropical forest above 200 Mg/ha, pixel-level uncertainty can exceed 50%, and the spatial pattern of error is not random but correlated with canopy density, terrain slope and the density of calibration plots.
For carbon accounting under frameworks such as REDD+ or Article 6 of the Paris Agreement, national authorities need uncertainty estimates that are methodologically defensible, not just numerically small. Satellize's approach to biomass analytics makes uncertainty propagation explicit at each processing stage, from sensor calibration through to the delivered GIS layer. The Tonga crop-estimation programme applied a similar philosophy of transparent uncertainty quantification to a different agricultural context, and the same discipline applies here.
P-band SAR, which ESA's BIOMASS mission will introduce at roughly 70 cm wavelength, is expected to penetrate to the forest floor in most tropical stands and may eventually push the saturation ceiling high enough to resolve the dense tropical forest problem. Until that data record matures, the honest position is that current methods have a hard upper limit, and any product claiming sub-10% uncertainty in old-growth Amazonian forest should be scrutinised carefully.
Typical figures
| Primary SAR spatial resolution | 25 m (ALOS-2 PALSAR-2 global mosaic); 10 m stripmap mode available for targeted acquisition |
| SAR revisit (PALSAR-2) | 14 days single-satellite; global annual mosaics released by JAXA |
| GEDI footprint diameter | 25 m; footprint spacing ~600 m across-track, ~60 m along-track |
| GEDI latitudinal coverage | 51.6° N to 51.6° S (ISS orbit constraint) |
| Optical resolution (Sentinel-2 / Landsat OLI) | 10–20 m (Sentinel-2); 30 m (Landsat OLI) |
| SAR frequency (PALSAR-2) | L-band, 1.27 GHz (~24 cm wavelength) |
| Biomass dynamic range (L-band SAR) | Useful signal up to ~150–200 Mg/ha; saturation above this in most forest types |
| Typical pixel-level uncertainty (fusion model) | 20–30% below 150 Mg/ha; >50% in dense tropical forest above 200 Mg/ha |
| Archive depth | PALSAR-1 global mosaics from 2007; Landsat archive from 1972; Sentinel-2 from 2015; GEDI from 2019 |
| Deliverable formats | GeoTIFF biomass density and uncertainty rasters; vector summary by forest stratum; CSV plot-validation report |
Analytics Satellize can run
| Wall-to-wall above-ground biomass density map | Random forest or gradient-boosted regression fusing PALSAR-2 backscatter, GEDI canopy height metrics and Sentinel-2 spectral indices, calibrated against field plot allometrics | GeoTIFF raster (Mg/ha per pixel) with paired uncertainty layer, delivered per scene or national mosaic |
| Carbon-density layer (above-ground carbon) | Biomass-to-carbon conversion using published species-specific or biome-average carbon fractions (typically 0.47 for tropical forest, per IPCC guidance) | GeoTIFF in tC/ha; aggregated totals by administrative unit in CSV |
| Biomass change detection (annual or biennial) | Multi-temporal differencing of annual PALSAR-2 mosaics, flagged against Landsat/Sentinel-2 disturbance indicators to separate degradation from natural variation | Change raster with gain/loss classification; summary report by forest stratum |
| Saturation-zone identification and uncertainty flagging | Backscatter plateau detection in L-band time series; pixels exceeding saturation threshold flagged and assigned elevated uncertainty class | Confidence mask layer accompanying biomass raster; methodology note for REDD+ or Article 6 submissions |
| Plot-to-pixel calibration and validation report | Cross-validation of field inventory plots against model predictions; leave-one-out RMSE and bias statistics by forest type and biomass class | PDF technical report with error statistics; shapefile of validation plot locations and residuals |
| Stratified biomass summary for carbon-project boundary | Zonal statistics over project or concession polygon; stratification by canopy cover class derived from Sentinel-2 | CSV and GIS layer with mean, standard deviation and 90% confidence interval per stratum |
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.