Soil carbon proxy monitoring for agricultural offset programmes

Bare-soil reflectance in the visible-to-SWIR range correlates with organic carbon content, but only in the top few millimetres. Satellite data cannot replace field sampling; it can, however, cut the number of cores needed by stratifying fields into homogeneous carbon-change zones.

Sensors

Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands; 20 m in SWIR bands 11 and 12 (1610 nm and 2190 nm), which are the most diagnostic for bare-soil carbon proxies. Five-day revisit at the equator with two satellites. Cloud cover over temperate farmland means usable bare-soil observations may arrive only a handful of times per year, so multi-year compositing is standard.
Landsat 8/9 OLI-2: 30 m resolution across all bands including SWIR-1 (1560 nm) and SWIR-2 (2200 nm). Sixteen-day single-satellite revisit, reduced to eight days with both platforms flying. The 40-plus-year Landsat archive is the primary tool for establishing pre-project soil-reflectance baselines, which offset registries increasingly require.
DESIS (ISS hyperspectral): Hyperspectral imager on the International Space Station, 2.5 nm spectral sampling from 400 to 1000 nm, 30 m ground sampling. Does not reach SWIR, so its value is in finer discrimination of visible-range soil colour indices rather than full SWIR-based carbon proxies. Non-sun-synchronous orbit means acquisition timing is irregular and cannot be scheduled around tillage windows reliably.
Sentinel-1 SAR: C-band synthetic aperture radar at 5.405 GHz, 10 m resolution in Interferometric Wide Swath mode. Cloud-independent. Backscatter intensity and coherence are sensitive to soil moisture and surface roughness, both of which confound optical carbon proxies. Used here as a correction layer rather than a primary carbon signal, to flag observations where soil moisture is too high for reliable reflectance interpretation.

What a satellite can and cannot see in a field

Electromagnetic radiation at visible and shortwave-infrared wavelengths penetrates agricultural soil to a depth of roughly one to two millimetres under dry conditions. Soil organic carbon (SOC) affects reflectance across this skin because humic compounds absorb strongly in the visible range, making carbon-rich soils darker, and show characteristic absorption features near 2100 nm in the SWIR. The relationship is real and has been confirmed in laboratory spectroscopy and field studies for decades.

The practical ceiling is that offset programmes care about SOC to 30 cm depth, sometimes deeper. A satellite observing the top two millimetres is not measuring the carbon pool that generates credits. It is reading a surface signal that correlates with bulk SOC under a specific set of conditions: the soil must be bare, recently tilled or at least free of standing residue, and it must be dry enough that moisture does not swamp the carbon signal. When those conditions are not met, the spectral proxy is unreliable. This is not a limitation that better sensors will fix; it is physics.

The moisture and roughness problem

Soil moisture is the dominant confound. Water absorption features in the SWIR overlap substantially with organic matter features, and a wet soil surface can look spectrally similar to a dry carbon-rich one. Published studies using Sentinel-2 data report that SOC prediction errors roughly double when soil moisture exceeds 15 to 20 percent volumetric water content. Tillage roughness adds a second layer of noise: freshly ploughed furrows create within-pixel shadow fractions that darken the surface independently of carbon content.

Sentinel-1 C-band backscatter provides a partial remedy. Sigma-nought in VV polarisation is sensitive to surface moisture and roughness, and can be used to screen out or weight-down observations acquired within a day or two of rainfall or immediately after primary tillage. The combination of optical and SAR data for bare-soil compositing is now standard in published MRV methodologies, though it adds processing complexity and does not eliminate the problem entirely.

Building a bare-soil composite: the tillage-window approach

The practical workflow starts with identifying the period in each agricultural year when fields are most likely to be bare: typically post-harvest and pre-sowing windows that vary by crop system and climate zone. For temperate cereal systems in Europe, this often means September to November and again in February to April. A multi-year stack of Sentinel-2 or Landsat scenes is filtered to remove cloud-affected pixels, then further filtered using a bare-soil index such as the Bare Soil Index (BSI) or the Modified Soil Adjusted Vegetation Index (MSAVI) to exclude pixels with any significant canopy cover.

From the remaining cloud-free, vegetation-free observations, a percentile composite is built. The 25th percentile of SWIR reflectance over several years tends to approximate the driest observed state for each pixel, reducing moisture bias. The resulting composite is not a snapshot; it represents the central tendency of bare-soil surface conditions over the composite period. Spatial resolution of 10 to 30 m means a single pixel may contain soil from multiple management zones within a large field, which introduces within-pixel heterogeneity that field sampling must account for.

Archive depth matters here. Landsat data from 1984 onward is freely available through USGS, giving nearly four decades of bare-soil observations for baseline construction. Sentinel-2 adds high-resolution coverage from 2015. Together they allow a project developer to show how surface reflectance, and by proxy surface SOC conditions, have evolved over time before the project start date.

Stratification: where satellite data earns its keep

Even if satellite-derived SOC proxies are too uncertain to replace field measurements, they can do something almost as valuable: tell you where to sample. A project area of 10,000 hectares cannot be sampled exhaustively. Statistical sampling designs require some prior understanding of spatial variability to allocate cores efficiently.

Clustering pixels by their multi-year bare-soil reflectance profile, combined with ancillary layers such as slope, drainage class and historical tillage intensity, produces strata within which SOC is expected to be more homogeneous than across the project area as a whole. Sampling effort can then be allocated proportionally to stratum area and variance. Published work using this approach has demonstrated reductions in required sample density of 30 to 50 percent for equivalent statistical precision, though the exact figure depends heavily on the landscape and the index used. The satellite data does not measure carbon; it organises the space in which carbon is measured.

This stratification function is also the most defensible use of satellite data in front of a verification body. It does not require the spectral-SOC relationship to be quantitatively precise, only that it captures enough spatial structure to be a useful stratification variable.

What registries and verifiers currently accept

No major carbon registry, as of publicly available methodology documents, accepts satellite-derived SOC estimates as a primary measurement of carbon stock change. Verra's VM0042 and the Gold Standard's soil carbon methodology both require physical soil sampling as the basis for credit issuance. Satellite data appears in these frameworks as a supplementary tool for spatial stratification, change detection to identify anomalous areas requiring additional sampling, and documentation of land management practices such as tillage reduction.

The honest implication for project developers is that satellite data reduces cost and improves sampling design but does not replace the laboratory. A credible MRV plan will specify the satellite-derived stratification method, the sampling design it informs, and the uncertainty bounds on the resulting stock-change estimate. Satellize produces stratification layers and multi-year bare-soil composites as GIS deliverables that slot directly into this kind of plan; the Tonga crop-estimation programme used a structurally similar compositing approach for a different agricultural variable, confirming that the processing pipeline is operational on open-constellation data.

One emerging development worth watching is the incorporation of hyperspectral data from missions such as PRISMA (ASI) and the forthcoming CHIME mission (ESA, planned for the late 2020s). CHIME will cover 400 to 2500 nm at roughly 20 m resolution with a 10-day revisit, which would substantially improve the SWIR coverage that current Sentinel-2 and Landsat composites provide. Whether it will close the gap between surface proxy and full-profile SOC is an open research question.

Typical figures

Optical spatial resolution	10 m (Sentinel-2 visible/NIR), 20 m (Sentinel-2 SWIR), 30 m (Landsat 8/9)
SAR spatial resolution	10 m (Sentinel-1 IW mode, ground range)
Revisit (optical, dual-satellite)	5 days (Sentinel-2A+B); 8 days (Landsat 8+9 combined)
Effective bare-soil observation frequency	Typically 2 to 8 usable scenes per year per field in temperate climates, after cloud and vegetation filtering
Key spectral bands for SOC proxy	SWIR-1 (~1560–1610 nm), SWIR-2 (~2100–2200 nm), Red (~665 nm), Blue (~490 nm)
Sensing depth (optical)	~1–2 mm under dry bare-soil conditions; effectively zero through any crop canopy or wet surface
Archive depth	Landsat from 1984 (USGS); Sentinel-2 from 2015 (Copernicus Dataspace)
Minimum stratification unit	~0.1 ha practical minimum at 10 m resolution; field-scale stratification most reliable above ~1 ha
Composite latency	Multi-year composites produced retrospectively; near-real-time bare-soil flagging possible within 3–5 days of acquisition
Delivery formats	GeoTIFF stratification rasters, sampling-zone shapefiles, per-field time-series CSV, PDF uncertainty summary

Analytics Satellize can run

Multi-year bare-soil composite	Percentile compositing of cloud-masked, vegetation-filtered Sentinel-2 and Landsat scenes over user-defined tillage windows; BSI and MSAVI masking	GeoTIFF raster at 10–30 m resolution, one layer per SWIR band and per composite period, with pixel-count quality layer
Soil moisture screening mask	Sentinel-1 VV backscatter thresholding and temporal filtering to flag high-moisture acquisition dates; applied before optical compositing	Binary GeoTIFF mask per scene date, integrated into composite pipeline; flagged-date log in CSV
SOC-proxy stratification map	Unsupervised clustering (k-means or hierarchical) on multi-band bare-soil composite, optionally combined with slope and drainage ancillary layers	Polygon shapefile of strata with recommended sampling density per stratum, formatted for import into standard MRV sampling-design tools
Baseline reflectance trajectory	Per-pixel linear trend analysis on annual bare-soil composites from Landsat archive (1984 onward) and Sentinel-2 (2015 onward), with Mann-Kendall significance test	Raster of trend slope and p-value; per-field summary table in CSV; narrative PDF for verifier submission
Anomaly detection for reversal monitoring	Comparison of current-year bare-soil composite against multi-year baseline; z-score flagging of pixels with significant reflectance increase (potential SOC loss or management change)	Alert shapefile of flagged parcels with confidence score; quarterly update cycle
Land management practice documentation	Temporal NDVI and BSI time-series to classify tillage intensity (conventional vs. reduced vs. no-till) based on timing and frequency of bare-soil exposures, following published classification schemes	Per-field practice classification table with supporting time-series plots; suitable for registry additionality documentation

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.