Soil organic carbon mapping in cropland using bare-soil composites

Topsoil organic carbon can be estimated from satellite optical imagery when fields are bare, exploiting the reflectance contrast between carbon-rich and carbon-depleted soils. Sentinel-2 and Landsat archives make cloud-free bare-soil composites feasible at regional scale, though the method is strictly surface-limited without ground-truth depth profiles.

Sensors

Sentinel-2 MSI: 10 m resolution in visible bands, 20 m in SWIR (bands 11 and 12 at 1610 nm and 2190 nm); 5-day revisit at mid-latitudes with both satellites. The archive from 2015 onward provides enough seasonal observations to isolate multiple bare-soil windows per field per year, which is the minimum needed for a statistically stable composite.
Landsat 8/9 OLI: 30 m resolution across visible, NIR and SWIR bands; 16-day revisit per satellite, 8-day combined. The archive extends to 1984 (Landsat 5 TM), enabling multi-decadal trend analysis of soil brightness change. OLI's SWIR-2 band at 2200 nm is particularly sensitive to clay mineralogy and organic matter absorption features.
PRISMA (ASI): Hyperspectral imager with 239 contiguous bands from 400 to 2500 nm at roughly 30 m spatial resolution and approximately 29-day revisit. Contiguous spectral coverage allows direct modelling of specific organic matter absorption features near 2100 nm and 2300 nm, reducing the ambiguity that multispectral indices carry when clay content varies.
EnMAP (DLR/GFZ): Hyperspectral at 30 m resolution, 228 spectral bands from 420 to 2450 nm, launched 2022. Designed explicitly for soil and vegetation applications; published pre-launch studies demonstrated retrieval of soil organic carbon with root mean square errors in the range of 2 to 5 g/kg under controlled bare-soil conditions, though operational accuracy depends heavily on local calibration data.

Why bare soil is the only honest window

Organic carbon in topsoil absorbs solar radiation across the visible and short-wave infrared more strongly than mineral-dominant soils do. The physical mechanism is straightforward: humic substances have broad, featureless absorption that darkens the reflectance spectrum from roughly 400 nm through to 2500 nm. A field with 4% organic carbon by weight will reflect measurably less red and SWIR radiation than an adjacent field at 1%, all else being equal.

The critical qualifier is 'all else being equal'. Moisture content, clay mineralogy, iron oxides and surface roughness all shift soil reflectance independently of organic matter. Green vegetation, crop residue and even sparse weed cover mask the soil signal almost entirely. This is why the bare-soil composite approach exists: by stacking every cloud-free, low-vegetation observation across several growing seasons and selecting only those pixels where a spectral vegetation index falls below a strict threshold (commonly NDVI below 0.15 to 0.25), analysts build a synthetic bare-soil image that represents the soil itself rather than whatever happened to be growing on it last Tuesday.

Building the composite: patience over precision

The construction of a reliable bare-soil composite is primarily a data-management problem. For a temperate cereal-growing region, a single Sentinel-2 tile might yield only 10 to 20 usable bare-soil observations per field across an entire year once cloud, cloud shadow and vegetation masks are applied. Across three or four years of archive, that number rises to 40 to 80 per pixel, which is generally sufficient to compute a stable median or geometric mean reflectance in each spectral band.

The standard workflow filters scenes by cloud cover, applies a scene-classification layer or external cloud mask, computes a vegetation index per pixel, retains only bare-soil observations, then aggregates to a single composite image. The BSDB (Bare Soil Detection and Building) approach published by Demattê and colleagues, and the Copernicus Global Land Service work on soil reflectance, both follow this general logic. Sentinel-2's five-day revisit makes it the workhorse; Landsat fills in where Sentinel-2 has data gaps or where the longer archive matters for trend detection.

One practical complication: fields under permanent crops, orchards or perennial pasture never go bare. The method simply cannot retrieve carbon estimates for those land covers without physical sampling. A good composite product is explicit about which pixels have sufficient bare-soil observations and which are masked as unreliable.

From reflectance to carbon: the calibration problem

Spectral reflectance is a proxy, not a direct measurement. Converting it to organic carbon content in grams per kilogram requires a calibration model trained on co-located field samples with laboratory-measured carbon values. Partial least squares regression, random forests and Gaussian process regression are all established in the published literature for this task; none is universally superior, and all are sensitive to the representativeness of the training set.

Transferability is the persistent difficulty. A model calibrated on soils in northern France may perform poorly when applied to soils in sub-Saharan Africa because the relationship between reflectance and organic matter is modulated by clay type, iron content and carbonate presence. Regional re-calibration with local samples is not optional; it is the difference between a plausible map and a misleading one. Published studies using Sentinel-2 composites report validation R² values ranging from roughly 0.5 to 0.75 for topsoil organic carbon, with errors typically in the range of 3 to 8 g/kg depending on soil variability and sample density. Hyperspectral sensors narrow that error range somewhat by resolving specific absorption features, but they do not eliminate the calibration requirement.

The depth limitation deserves emphasis. Satellite sensors observe the top few millimetres of the soil surface under dry conditions. Topsoil samples used for calibration are typically taken to 20 or 30 cm depth. The two are correlated but not identical. Carbon stocks at plough depth, let alone subsoil carbon, require physical cores. Satellite data can guide where to sample efficiently; it cannot replace the sampler.

What the map can and cannot support

A well-constructed bare-soil composite product at Sentinel-2 resolution gives a spatially continuous picture of relative organic carbon variation across a landscape at roughly 20 m posting. That is genuinely useful for prioritising soil health interventions, identifying fields that have lost carbon over time, and stratifying a region for more targeted physical sampling campaigns. It can detect change between composite periods of three to five years if the calibration is consistent and the sample network is maintained.

It cannot verify carbon credits at the individual field level without additional ground-truth data. The uncertainty at a single 20 m pixel is too large for most carbon market protocols, which typically require errors below 10 to 20% of the measured value. The appropriate use is as a spatial layer that informs where to concentrate expensive soil sampling, not as a substitute for it. National soil monitoring programmes in France (RMQS), Germany and the UK have all explored satellite-derived SOC maps as a spatial interpolation aid rather than a standalone measurement.

Cloud cover remains an operational constraint in humid tropical regions. In areas where fields are rarely bare and skies are rarely clear simultaneously, composite quality degrades sharply. Analysts should report the number of valid bare-soil observations per pixel alongside any SOC estimate.

Hyperspectral sensors: more information, same caveats

PRISMA and EnMAP represent a step forward in spectral information content. Rather than inferring organic matter from four or five broad bands, hyperspectral sensors resolve hundreds of narrow contiguous channels, allowing direct identification of absorption features associated with humic acids and aliphatic C-H bonds near 2100 and 2300 nm. Laboratory spectroscopy at these wavelengths has long been used in soil science; satellite hyperspectral instruments bring that capability to field scale.

The practical gain is real but bounded. Hyperspectral retrieval reduces the confounding effect of clay mineralogy because clay absorption features can be modelled and partially separated from organic matter features. Published EnMAP pre-launch simulations and early PRISMA results suggest improvements of 15 to 30% in prediction error compared to multispectral approaches on the same soils. The revisit rate of both sensors (roughly 29 days for PRISMA, variable tasking for EnMAP) is lower than Sentinel-2, which makes composite construction slower. For national-scale annual monitoring, Sentinel-2 composites with good calibration data remain more practical. For high-value research sites or targeted surveys, hyperspectral adds meaningful precision.

Satellize processes both multispectral composites and PRISMA scenes through its analytics pipeline; the Tonga crop-estimation programme demonstrated the archive-processing workflow that underpins composite construction at small-island scale, where observation frequency is a genuine constraint.

Delivering a product that survives contact with a soil scientist

The output of a satellite SOC mapping exercise should be a georeferenced raster of estimated topsoil organic carbon in g/kg or percentage, accompanied by a per-pixel uncertainty layer, a data-quality mask indicating the number of bare-soil observations used, and a calibration report documenting the training sample locations, laboratory methods, model type and cross-validation statistics. Without those accompanying layers, the carbon map is an assertion rather than a measurement.

Change detection products comparing two composite periods require that the calibration approach and composite construction method are held constant between epochs. Apparent trends caused by changes in sensor, atmospheric correction version or sample depth convention are a known source of error in the literature. Honest reporting distinguishes between statistically significant change and noise.

Typical figures

Typical spatial resolution	20 m (Sentinel-2 SWIR bands); 30 m (Landsat OLI, PRISMA, EnMAP)
Revisit for composite building	5 days (Sentinel-2 dual satellite); 8 days (Landsat 8+9 combined); ~29 days (PRISMA); variable tasking (EnMAP)
Minimum bare-soil observations for stable composite	Typically 20 to 40 per pixel; fewer observations increase uncertainty
Spectral bands most relevant to SOC	Visible (red ~665 nm), SWIR-1 (~1610 nm), SWIR-2 (~2190 nm); full 400–2500 nm for hyperspectral
Typical retrieval accuracy (multispectral)	Validation R² 0.5 to 0.75; RMSE 3 to 8 g/kg depending on soil variability and calibration sample density
Depth sensitivity	Surface millimetres under dry conditions; calibration samples typically 0 to 30 cm
Archive depth	Sentinel-2 from 2015; Landsat from 1984 (Landsat 5 TM)
Cloud cover constraint	Composite quality degrades in humid tropics; per-pixel observation count must be reported
Minimum mapping unit	Practically ~1 ha for reliable field-level estimates at 20 m resolution
Delivery formats	GeoTIFF (SOC estimate + uncertainty + QA mask), GeoPackage, cloud-optimised GeoTIFF for web delivery

Analytics Satellize can run

Bare-soil composite raster	Time-series filtering using NDVI threshold (typically <0.15 to 0.25) and cloud masking over Sentinel-2 or Landsat archive; median or geometric-mean aggregation per band	Cloud-free bare-soil mosaic GeoTIFF per season or multi-year epoch, with per-pixel observation-count layer
Topsoil organic carbon map	Partial least squares regression or random forest model trained on co-located field samples and laboratory measurements, applied to composite reflectance; cross-validated against held-out samples	Georeferenced SOC raster in g/kg with per-pixel prediction uncertainty layer, calibration report
SOC change detection layer	Epoch-to-epoch differencing of calibration-consistent composite SOC estimates; significance testing against per-pixel uncertainty to flag statistically meaningful change	Change raster (g/kg delta) with significance mask; summary statistics by administrative unit or farm boundary
Sampling stratification layer	Spectral clustering of bare-soil composite bands to identify spectrally distinct soil zones; used to design spatially balanced physical sampling campaigns that minimise laboratory cost	Stratification GIS layer with recommended sample locations and stratum areas
Hyperspectral SOC retrieval (PRISMA/EnMAP)	Continuum-removed spectral feature analysis at C-H and humic absorption wavelengths (~2100–2300 nm); PLSR or Gaussian process regression on full spectral vector	Higher-precision SOC raster for targeted survey areas, with comparison report against multispectral baseline
Data-quality and coverage report	Per-pixel tallying of cloud-free bare-soil observations across the archive; flagging of fields with insufficient observations for reliable retrieval	Coverage report PDF and QA raster identifying pixels with fewer than the minimum observation threshold

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.