Forest soil organic carbon proxy mapping from surface indicators
Direct soil organic carbon measurement from space remains beyond current sensors. Proxy-based mapping, using SAR backscatter, L-band moisture and optical indices in forest gaps, constrains REDD+ baseline uncertainty and tells field teams where to dig.
Sensors
- ALOS-2 PALSAR-2: L-band SAR (1.27 GHz) penetrates forest canopy to interact with the litter-soil interface. Single-look complex data at 3–10 m resolution; 14-day repeat. HV cross-polarisation backscatter correlates with litter moisture and surface roughness, both proxies for organic horizon depth.
- Sentinel-1 SAR: C-band SAR (5.405 GHz) at 10 m resolution, 6-day revisit over most land surfaces. C-band penetration is shallower than L-band, so Sentinel-1 is more sensitive to canopy structure and litter surface conditions than to deeper organic horizons. Useful for change detection and gap identification.
- Sentinel-2 MSI: 10–20 m multispectral imagery, 5-day revisit. In forest gaps and clearings where mineral or organic soil is exposed, bands in the red-edge (705 nm, 740 nm) and SWIR (1610 nm, 2190 nm) support organic matter indices such as the Soil Organic Matter Index and bare-soil composites. Cloud cover is the primary constraint in humid tropical forests.
- SMAP (NASA): L-band passive radiometer at 36 km effective soil moisture resolution, 2–3 day revisit. Too coarse for stand-level mapping but provides the regional moisture state needed to normalise SAR backscatter: a dry litter layer and a wet one produce very different signals for the same carbon content.
Why the soil carbon problem is harder than the biomass problem
Above-ground biomass has a well-established SAR and lidar signal. Soil organic carbon does not. The carbon is buried. Radar at C-band rarely penetrates more than a few centimetres of moist forest litter; even L-band, which can reach the litter-soil interface under dry conditions, is reading a proxy of a proxy. Published studies using ALOS PALSAR data over boreal and temperate forests find statistically significant but noisy correlations between HV backscatter and forest floor organic horizon depth, with explained variance typically in the 30–55% range depending on stand type and moisture conditions. That is useful for stratification; it is not a direct measurement.
The honest framing for any buyer is this: space-based proxies reduce the uncertainty envelope around soil carbon estimates and improve the efficiency of ground-truth campaigns. They do not replace field cores. A REDD+ project that claims soil carbon accounting without field validation is on shaky methodological ground regardless of how many satellites were involved.
What a floating roof gives away: reading litter through L-band
The forest floor organic horizon, sometimes called the O-horizon, accumulates decomposing litter over decades. Its depth and moisture content affect L-band backscatter in ways that are separable, at least in principle, from the overlying canopy signal. ALOS-2 PALSAR-2 HV backscatter is sensitive to volumetric scattering within the litter layer. Studies in boreal Scandinavia and temperate European forests have shown that thicker, wetter organic horizons produce measurably higher HV returns than thin, dry mineral-soil surfaces. The effect is modest, typically 1–3 dB across realistic organic horizon depth ranges, which means it is detectable only after careful moisture normalisation.
SMAP provides the regional soil moisture state needed for that normalisation. The workflow is: acquire SMAP surface moisture for the acquisition date, use it to flag or weight PALSAR-2 scenes, then model the residual backscatter variation as a function of litter and organic horizon properties. This is an active area of published research rather than an operational standard, so uncertainty bounds must be reported explicitly in any deliverable.
Gaps are the windows: optical indices where the canopy opens
Sentinel-2 cannot see through a closed forest canopy to the soil below. But forests are not closed everywhere. Natural treefalls, harvesting, road edges, and regenerating patches all expose soil or low-litter surfaces. In those gaps, Sentinel-2 SWIR bands are genuinely informative. The ratio of Band 11 (1610 nm) to Band 4 (665 nm) is one published indicator of organic matter content in exposed soil; the Bare Soil Index and Modified Soil Adjusted Vegetation Index are others used in land degradation monitoring.
The practical constraint is that gap pixels are sparse in dense tropical forest, perhaps 2–8% of total area in undisturbed stands, and cloud cover in humid tropics can make even those pixels unavailable for months at a time. Building a gap-soil composite from two or three years of Sentinel-2 imagery, selecting only scenes below 10% cloud cover and using the darkest available SWIR values as an organic matter indicator, is a workable approach. It produces a patchy map that samples the forest, not a continuous surface. Spatial interpolation between gap observations is necessary and introduces its own uncertainty.
From proxy layers to a sampling design: where the map earns its money
The most defensible use of a proxy-based soil carbon map is not to report a carbon stock number. It is to stratify the landscape so that field sampling is allocated efficiently. A forest with high spatial variability in predicted organic horizon depth needs more field plots than one that is homogeneous. A proxy map that captures even 40% of the spatial variance in soil carbon allows a stratified random sampling design to achieve the same statistical precision with roughly half the field effort of a purely random design. That translates directly into cost savings on REDD+ baseline establishment.
Satellize structures its soil carbon proxy outputs as stratification layers with explicit uncertainty bands, not as point estimates of carbon density. The map tells a project developer where the high-variance zones are and how many field cores are needed to close the uncertainty to a given confidence interval. That is a more honest and more useful product than a false-precision carbon map.
For the field team, the deliverable includes a prioritised list of sampling locations generated by a variance-minimising algorithm applied to the proxy strata. This is the same logic Satellize applied in its crop-estimation work for the Kingdom of Tonga: proxy layers guide sampling, and sampling validates the proxy.
Honest limits, stated plainly
Several constraints are non-negotiable. C-band SAR does not penetrate closed forest canopy to the soil surface in any meaningful way; Sentinel-1 alone cannot produce a soil carbon proxy in dense tropical forest. L-band penetration improves under dry conditions but degrades rapidly when the litter layer is saturated, which is common in the forests where soil carbon stocks are largest. SMAP's 36 km resolution is a blunt normalisation tool. Sentinel-2 optical indices apply only to exposed soil, which is a small fraction of a healthy forest.
Deep peat soils, which can hold thousands of tonnes of carbon per hectare, are a separate problem addressed by peatland-specific methods and are outside the scope of this approach. Mineral forest soils with thin organic horizons are the target domain. Even there, the proxy relationship between surface backscatter and total soil organic carbon stock (which integrates to a metre or more depth) is weak. Published root-mean-square errors in cross-validated models are typically in the range of 15–40% of mean carbon density. That is a useful constraint on uncertainty; it is not a measurement.
Typical figures
| SAR spatial resolution (ALOS-2 PALSAR-2) | 3–10 m single-look complex; 25 m in mosaic products |
| SAR spatial resolution (Sentinel-1) | 10 m (IW mode GRD) |
| Optical resolution (Sentinel-2 MSI) | 10 m (visible/NIR), 20 m (red-edge, SWIR) |
| SAR revisit | 6 days (Sentinel-1 combined A+B); 14 days (ALOS-2 PALSAR-2) |
| Soil moisture ancillary (SMAP) | 36 km effective resolution; 2–3 day revisit |
| Frequency bands used | C-band 5.405 GHz (Sentinel-1); L-band 1.27 GHz (PALSAR-2); L-band passive (SMAP); 443–2190 nm optical (Sentinel-2) |
| Minimum detectable organic horizon signal | Approximately 1–3 dB HV backscatter difference across realistic litter depth ranges; detectable only after moisture normalisation |
| Cloud-cover constraint (optical) | Severe in humid tropics; multi-year compositing required to obtain sufficient gap-pixel coverage |
| Archive depth | Sentinel-1 from 2014; Sentinel-2 from 2015; ALOS PALSAR (predecessor) from 2006; ALOS-2 from 2014 |
| Typical deliverable format | GeoTIFF stratification raster with uncertainty bands; CSV of prioritised field sampling locations; PDF methodology report |
Analytics Satellize can run
| L-band litter-moisture proxy layer | ALOS-2 PALSAR-2 HV backscatter normalised against SMAP surface moisture; published two-component scattering model separating canopy and floor contributions | GeoTIFF raster at 25 m, with per-pixel uncertainty estimate, covering the project area |
| Forest gap soil organic matter index composite | Multi-year Sentinel-2 bare-soil composite using SWIR/red ratio and Bare Soil Index on gap pixels only; cloud masking via Scene Classification Layer | GeoTIFF composite with gap-pixel density layer indicating data confidence per grid cell |
| Soil carbon proxy stratification map | Unsupervised clustering of SAR and optical proxy layers into 4–8 strata; variance partitioning to assign uncertainty ranges to each stratum | Polygon GIS layer (GeoPackage or Shapefile) with stratum labels and associated carbon density uncertainty ranges |
| Field sampling allocation plan | Stratified random sampling design using Neyman allocation applied to proxy strata; minimises total field effort for a target confidence interval | CSV of GPS coordinates for recommended field core locations, with stratum membership and sampling priority rank |
| REDD+ baseline uncertainty quantification | Monte Carlo propagation of proxy-layer uncertainty through a published soil carbon stock model; output expressed as confidence interval around mean carbon density per stratum | PDF technical report with uncertainty tables suitable for submission to a REDD+ verification body |
| Temporal change flag layer | Sentinel-1 coherence and backscatter change detection between baseline and monitoring epochs; flags areas where surface conditions have changed sufficiently to invalidate the baseline proxy map | GeoTIFF change flag raster with date of detected change; triggers re-sampling recommendation for affected strata |
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.