Permafrost terrain carbon-pool mapping from SAR and optical fusion

Permafrost soils hold roughly twice the carbon currently in the atmosphere, yet mapping those stocks remotely is genuinely hard. SAR backscatter and optical vegetation indices together constrain the problem, within honest limits.

Sensors

ALOS-2 PALSAR-2: L-band (1.27 GHz) SAR; 3–10 m resolution in spotlight/stripmap modes, 14-day repeat. L-band penetrates the vegetation canopy and is sensitive to soil moisture and above-ground biomass in tundra and boreal settings, making it the primary SAR choice for carbon-pool proxies.
Sentinel-1 A/B C-band SAR: C-band (5.4 GHz), 10 m GRD resolution, 6-day repeat at mid-latitudes, somewhat longer at high latitudes depending on orbit geometry. Less soil-penetrating than L-band but useful for surface soil moisture and freeze–thaw state discrimination.
Sentinel-2 MSI: 10–20 m optical, 13 spectral bands including red-edge (705 nm, 740 nm) and SWIR. 5-day revisit with both satellites. Provides NDVI, EVI and plant-community classification that stratify the landscape into carbon-relevant vegetation types.
Landsat 8/9 OLI: 30 m optical, 16-day repeat, archive back to 1972 (TM/ETM+/OLI). Longer archive enables multi-decadal vegetation change detection; OLI SWIR bands constrain surface water fraction and organic-matter proxies.

Why permafrost carbon is hard to count from orbit

Permafrost soils are estimated to contain somewhere between 1,460 and 1,600 Gt of organic carbon, much of it accumulated over millennia in frozen peat and mineral horizons. That figure comes from synthesis of field cores, not from satellites. The satellite problem is that carbon itself has no direct spectral signature at the surface. What sensors can observe are proxies: vegetation type and density, surface soil moisture, standing water fraction and, indirectly, microtopographic roughness.

The depth to the permafrost table, which controls how much organic material can thaw and decompose in a given season, remains largely unobservable from orbit. Field measurement, ground-penetrating radar and modelled soil temperature are the only routes to that variable. Any satellite-derived carbon map carries this as a fundamental, irreducible uncertainty. Buyers should understand that what remote sensing delivers is a spatial stratification of carbon-pool likelihood, not a direct assay of soil carbon content.

What L-band backscatter actually measures in tundra

L-band radar (around 1.27 GHz on PALSAR-2) penetrates the low tundra canopy and interacts with the top 20–30 cm of soil, depending on moisture content. Wetter soils increase the dielectric constant, which raises backscatter. Dry, frozen soils produce markedly lower returns. This freeze–thaw sensitivity is well documented in the published literature and forms the basis of JAXA's global freeze–thaw products derived from PALSAR data.

Above-ground biomass in low-stature tundra communities is also correlated with L-band backscatter, though the relationship saturates at relatively modest biomass levels, typically below 50–80 t/ha, which is not a severe constraint in tundra but becomes one in boreal forest edges. HV cross-polarisation is more sensitive to volume scattering from vegetation structure than HH, so dual-polarisation acquisitions are preferred. C-band Sentinel-1 adds freeze–thaw discrimination and surface soil moisture but loses the canopy-penetration advantage.

Optical indices and the vegetation-community stratification problem

Carbon stocks vary enormously between tundra vegetation communities. Sedge-dominated wet meadows, Sphagnum bogs, dwarf-shrub heaths and lichen-covered fell-fields each carry different organic-matter accumulation histories. Sentinel-2's red-edge bands (705 nm and 740 nm) improve discrimination of these communities compared with broadband NDVI alone, because chlorophyll absorption features shift with canopy structure and leaf area index in ways that red-edge reflectance captures more precisely.

Landsat's 50-year archive matters here for a different reason: it documents where shrubification has occurred, a documented response to Arctic warming in which dwarf shrubs expand into former sedge or lichen terrain. Shrub expansion changes both the carbon accumulation rate and the SAR backscatter character of a pixel. Fusion workflows that use Landsat change vectors to stratify current Sentinel-2 classifications are more defensible than single-epoch optical approaches.

Cloud cover is a genuine operational problem at high latitudes. In the Siberian and Alaskan Arctic, summer cloud fractions can exceed 70 % in any given overpass. Median compositing over the June–September window across multiple years is standard practice, but it means the optical component of any map represents a multi-year average rather than a single-season snapshot.

Fusing the two data streams: where the method gains and where it does not

The fusion logic is straightforward in principle. Optical classification stratifies the landscape into vegetation communities, each assigned a prior carbon-density range from field literature. SAR backscatter then adjusts the soil-moisture and biomass estimates within each stratum, tightening the uncertainty bounds. The combined product is a raster of carbon-pool estimates with associated uncertainty, typically expressed as a range in kg C per square metre.

Published studies using PALSAR and Landsat over Alaskan tundra report spatial uncertainties on the order of 30–50 % at the pixel level, improving to perhaps 15–25 % when aggregated to watershed or landscape units. Those are honest figures. The method does not resolve peat depth, which can vary from centimetres to several metres across distances of tens of metres in polygonal terrain. Peat depth is the single largest source of error in any remote-sensing carbon estimate for permafrost landscapes.

Practical delivery: what a government or research buyer receives

A working carbon-pool map for a permafrost region typically involves three processing stages. First, a multi-year optical composite classifies vegetation communities at 10–30 m resolution. Second, time-series SAR backscatter, processed to normalised gamma-naught values, provides per-pixel soil moisture and biomass adjustment. Third, a lookup table derived from published field-data syntheses, such as the Northern Circumpolar Soil Carbon Database maintained by the International Permafrost Association, converts the combined signal into carbon-density estimates with uncertainty bounds.

The deliverable is a GIS-ready raster, typically in GeoTIFF with associated uncertainty layers, accompanied by a methods report that states explicitly what the map cannot tell the user. For climate modellers using this as a boundary condition, the vegetation-community classification layer is often more useful than the carbon estimate itself, because land-surface models apply their own carbon-density parameters. Satellize has run similar multi-source fusion workflows for agricultural applications, including crop estimation in the Kingdom of Tonga, and applies the same fusion architecture to permafrost terrain on client request.

Update frequency is limited by the short Arctic optical window. Annual updates during the June–September season are realistic; sub-annual updates add little given cloud constraints and the slow pace of permafrost change at decadal timescales. Rapid-change scenarios, such as post-fire soil exposure or sudden thermokarst expansion, are better tracked through SAR time series alone, which is covered separately in the sibling page on thermokarst lake dynamics.

Typical figures

Primary SAR spatial resolution	3–10 m (PALSAR-2 stripmap/spotlight); 10 m (Sentinel-1 GRD)
Primary optical spatial resolution	10 m (Sentinel-2 MSI); 30 m (Landsat 8/9 OLI)
SAR revisit (PALSAR-2)	14 days; polar latitudes may receive more frequent coverage due to orbit convergence
SAR revisit (Sentinel-1)	6 days (both satellites operational); single-satellite 12 days
Optical revisit (Sentinel-2)	5 days (both satellites); usable cloud-free acquisitions typically 2–4 per summer season at high latitudes
SAR frequency bands used	L-band 1.27 GHz (PALSAR-2); C-band 5.4 GHz (Sentinel-1)
Optical spectral bands of interest	Red-edge 705 nm, 740 nm; NIR 842 nm; SWIR 1610 nm, 2190 nm (Sentinel-2); OLI equivalent bands
Carbon-pool estimate uncertainty (pixel level)	Typically 30–50 % at 10–30 m pixel; 15–25 % aggregated to watershed scale (published range)
Archive depth	Landsat from 1972; Sentinel-1 from 2014; PALSAR from 2006 (ALOS-1); PALSAR-2 from 2014
Delivery formats	GeoTIFF (carbon-density and uncertainty rasters), GeoPackage (vegetation classification), PDF methods report

Analytics Satellize can run

Vegetation-community classification map	Random-forest or support-vector classification on Sentinel-2 red-edge and SWIR composites, with Landsat change vectors as ancillary features	GeoTIFF raster, 10 m resolution, with class legend and accuracy matrix
Surface soil moisture anomaly layer	Change-detection on normalised Sentinel-1 C-band gamma-naught time series; freeze–thaw state flag derived from PALSAR-2 seasonal contrast	Seasonal GeoTIFF stack with per-pixel anomaly scores
Above-ground biomass estimate	Empirical regression of PALSAR-2 HV backscatter against published tundra biomass field data; saturation flagged above 50 t/ha	GeoTIFF with biomass estimate and saturation mask
Soil organic carbon density map	Lookup-table fusion of vegetation-community classification and SAR-derived moisture/biomass, calibrated against Northern Circumpolar Soil Carbon Database field values	GeoTIFF carbon-density raster (kg C m⁻²) with paired uncertainty layer
Multi-decadal shrubification change map	Landsat NDVI and SWIR time-series trend analysis (1985–present) using annual summer composites; breakpoint detection for vegetation-type transitions	Change-vector GeoTIFF and summary report of affected area by decade
Carbon-pool boundary condition package for land-surface models	Aggregation of pixel-level estimates to user-defined spatial units (watershed, grid cell); formatted to CF-convention NetCDF with uncertainty metadata	NetCDF file and accompanying data-quality statement

Sources

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.