Coastal saltmarsh carbon habitat mapping and change detection
Saltmarsh blue-carbon accounting demands community-level vegetation mapping, tidal-mask correction and creek-network delineation that generic land-cover products cannot provide. Resolution, timing and spectral depth all matter.
Sensors
- WorldView-3 (Maxar): 0.31 m panchromatic, 1.24 m multispectral across 8 bands (coastal blue through near-infrared), plus 8 SWIR bands at 3.7 m. At this resolution, individual creek channels down to roughly 1–2 m width are resolvable, and the fine-scale patchwork of Spartina, Salicornia and bare mud can be separated by object-based classification. Revisit is roughly 1–4.5 days at mid-latitudes depending on tasking priority.
- Planet SuperDove (PlanetScope): 3 m multispectral across 8 bands including red-edge and two near-infrared bands. Daily revisit globally. The red-edge band is particularly useful for separating actively photosynthesising saltmarsh canopy from senescent or stressed vegetation, and for estimating above-ground biomass via red-edge NDVI variants. At 3 m, narrow creek networks are partially resolved but not fully delineated.
- Sentinel-2 MSI (ESA Copernicus): 10 m (visible and NIR), 20 m (red-edge and SWIR), 60 m (coastal aerosol). Free, 5-day revisit at the equator and better at higher latitudes with two satellites. Adequate for mapping saltmarsh extent and broad community zones in larger marshes, but 10 m pixels conflate community boundaries in small or highly fragmented sites. The red-edge bands (B5, B6, B7) add meaningful separation of vegetation types beyond simple NDVI. Archive runs from 2015.
- DESIS (DLR hyperspectral, ISS): 235 spectral bands across 400–1000 nm at roughly 2.55 nm spectral sampling, 30 m spatial resolution. Hyperspectral data allows spectral unmixing to estimate fractional cover of multiple saltmarsh species within a single pixel, which is not achievable with broadband multispectral sensors. Coverage is non-systematic given ISS orbital constraints; targeted acquisitions must be planned. Revisit is irregular and cannot be relied upon for time-series monitoring.
- Landsat 8/9 OLI (USGS/NASA): 30 m multispectral, 16-day revisit per satellite (8-day combined). The archive runs to 1972 across the broader Landsat series, making it indispensable for long-term change detection: marsh loss, invasion by Phragmites, or recovery after restoration. At 30 m, community-level mapping is unreliable in narrow or fragmented marshes but the temporal depth has no equivalent in commercial constellations.
Why saltmarsh is harder to map than it looks
Saltmarsh sits at the intersection of several remote-sensing problems at once. The canopy is low, typically 20–80 cm for Spartina alterniflora, shorter still for Salicornia, so there is no canopy height signal to exploit from lidar or radar in the way forest programmes rely on. The spectral differences between community types are real but subtle: Spartina, Phragmites and Salicornia overlap substantially in broadband near-infrared reflectance, and separating them requires either hyperspectral data or a combination of red-edge bands with high spatial resolution and careful phenological timing.
Phenology is the key variable most programmes underweight. Salicornia is highly seasonal, turning red-purple in autumn and nearly invisible against bare mud in winter. Phragmites reaches peak greenness later than Spartina. Acquiring imagery at a single date risks misclassifying community types that happen to be at similar phenological stages. A minimum of two acquisition windows, one in early growing season and one at peak biomass, is standard practice in published field-validated studies.
The tidal inundation masking problem
Tidal state at acquisition time is not a nuisance: it is a fundamental confound. A pixel that is bare mud at low tide and submerged at high tide will produce radically different reflectance values, and both states can be misclassified as non-vegetation. The practical fix requires pairing imagery acquisition with tidal predictions from a reliable gauge or hydrodynamic model, then restricting classification to images acquired within a defined tidal window, typically within two hours of mean low water for intertidal zones.
This constraint interacts badly with cloud cover. In temperate marshes, the probability of acquiring a cloud-free image within a narrow tidal window on a given day is low. Planet's daily revisit rate makes it the most practical sensor for assembling a tidal-window-filtered composite, but even at 3 m resolution the spectral library must be built from images acquired under the same tidal conditions as the classification target. Sentinel-2's 5-day revisit can satisfy the tidal constraint at sites with predictable clear-sky periods; in persistently cloudy regions such as the British Isles or Pacific Northwest, multi-year compositing is often necessary.
Creek networks, carbon burial and the resolution floor
Tidal creek density is a meaningful predictor of carbon burial rates. Marshes with dense, well-connected creek networks export more organic material seaward but also promote deeper anoxic soil development, which is the primary carbon storage mechanism. Mapping creek networks from satellite imagery requires spatial resolution fine enough to resolve channel widths that are often 1–5 m. At Sentinel-2's 10 m, most secondary creeks are invisible. At Planet's 3 m, primary creeks are detectable but secondary networks require careful edge-enhancement or fusion with higher-resolution panchromatic data. WorldView-3 at sub-metre resolution is the only operational satellite sensor that reliably resolves the creek hierarchy needed for carbon-burial modelling.
Creek extraction is typically done through a combination of NDWI thresholding on NIR and SWIR bands, morphological thinning and manual quality control. Automated creek mapping from satellite data remains error-prone at the margins, particularly where creeks are narrow, turbid or partially vegetated. Any carbon-burial model that incorporates creek density should carry an explicit uncertainty term derived from the resolution limit of the input imagery.
Above-ground biomass: what satellites can and cannot say
Above-ground biomass (AGB) in saltmarsh can be estimated from vegetation indices. Red-edge NDVI and the red-edge chlorophyll index (CIre) correlate reasonably well with AGB in Spartina-dominated marshes, with published R² values in the range of 0.6–0.8 in field-validated studies using Sentinel-2 and similar sensors. The relationship degrades at high biomass densities due to canopy saturation, and it varies substantially between species: a Phragmites stand and a Spartina stand at the same NDVI may carry quite different carbon densities.
The more important limitation is that AGB is typically only 10–30% of total saltmarsh carbon stock. The dominant store is below-ground: root biomass and, critically, centuries of accumulated organic matter in anoxic soils that can extend several metres deep. No operational satellite sensor measures below-ground carbon directly. Soil carbon must be estimated from field sampling combined with spatial covariates such as vegetation community type, elevation (from lidar or Structure from Motion surveys) and inundation frequency. Satellite data defines the spatial framework; it does not replace soil cores.
Change detection and the permanence question
Carbon registries require evidence that a saltmarsh carbon project has not lost stock between verification periods. Change detection at the extent level, tracking marsh loss to erosion, sea-level rise or land reclamation, is tractable with Sentinel-2 or Landsat time series. The Landsat archive back to the 1980s is particularly valuable for establishing pre-project baselines and demonstrating additionality.
Community-level change is harder. Detecting the spread of invasive Phragmites into a Spartina marsh, or the die-back of Salicornia after a salinity shift, requires the red-edge sensitivity of SuperDove or Sentinel-2 combined with consistent phenological timing across years. A change that is real in the field can be invisible in imagery acquired at the wrong tidal state or growth stage. Satellize structures multi-year time-series packages for carbon auditors around these constraints, drawing on the same open-constellation approach used in its Tonga crop-estimation programme. The output is a documented, reproducible change record rather than a single snapshot classification.
One honest caveat: saltmarsh extent change can be mapped with high confidence at scales above roughly 0.1 ha using Sentinel-2. Below that threshold, or in marshes with complex creek-edge dynamics, uncertainty rises sharply and higher-resolution commercial tasking is needed.
Building a defensible MRV package
A satellite-based MRV submission for a saltmarsh blue-carbon project needs to document four things clearly: the spatial extent of each vegetation community at baseline and each verification date; the biomass-to-carbon conversion factors applied to each community type; the tidal and atmospheric conditions at image acquisition; and the uncertainty propagated through each step.
The community classification should be validated against field plots distributed across the spectral variability of the site, not just across geographic area. Accuracy assessments reported only as overall accuracy obscure the class-level confusion that matters most for carbon accounting, where misclassifying a low-carbon Phragmites stand as high-carbon Spartina directly inflates the credit count. Separating error in extent mapping from error in carbon-density assignment, and reporting both, is the standard that credible registries are moving toward.
Typical figures
| Spatial resolution (community mapping) | 1.24 m (WorldView-3 MS) to 30 m (Sentinel-2 SWIR); 3 m (Planet SuperDove) recommended minimum for community-level work |
| Creek network resolution floor | Sub-metre panchromatic (WorldView-3 0.31 m) required to resolve secondary creek networks; 3 m resolves primary channels only |
| Revisit (cloud-free, tidal-window constrained) | Planet SuperDove: daily globally; Sentinel-2: 5-day repeat; WorldView-3: 1–4.5 days (tasked); DESIS: irregular (ISS orbit) |
| Key spectral bands | Red-edge (705–740 nm), NIR, SWIR for vegetation separation; coastal blue and SWIR for water/mud masking; 400–1000 nm hyperspectral (DESIS) for unmixing |
| Minimum mappable unit (extent) | Approximately 0.1 ha at Sentinel-2 10 m; approximately 0.01 ha at WorldView-3 1.24 m |
| Archive depth | Landsat: 1972 to present; Sentinel-2: 2015 to present; Planet SuperDove: 2017 to present; WorldView-3: 2014 to present |
| Tidal window constraint | Imagery should be acquired within approximately 2 hours of mean low water for intertidal classification; tidal gauge or model data required |
| Above-ground biomass estimation accuracy | Published R² 0.6–0.8 for red-edge index methods in Spartina; degrades at high canopy density and varies by species |
| Below-ground carbon | Not directly measurable by satellite; requires field soil cores with satellite-derived community type and elevation as spatial covariates |
| Delivery formats | GeoTIFF community-classification rasters, GeoPackage or Shapefile extent polygons, tabular biomass estimates with uncertainty bounds, PDF audit-ready change reports |
Analytics Satellize can run
| Saltmarsh extent map with community classification | Object-based image analysis (OBIA) or supervised pixel classification using red-edge and NIR bands; phenologically timed dual-date composites; tidal-window filtering | GeoTIFF and polygon layer with per-class accuracy matrix; suitable for registry submission as baseline or verification extent evidence |
| Tidal-mask corrected vegetation composite | Time-series filtering of Planet SuperDove or Sentinel-2 stack using tidal predictions; cloud and shadow masking; median compositing within tidal window | Cloud-free, tide-consistent reflectance composite with acquisition metadata; input layer for all subsequent classification and biomass steps |
| Above-ground biomass estimate by community zone | Red-edge chlorophyll index (CIre) or red-edge NDVI regression against field-calibrated biomass data; species-specific conversion factors applied per classified community | Raster biomass map (tonnes dry matter per hectare) with per-pixel uncertainty; tabular summary by project zone for carbon accounting |
| Creek network delineation | NDWI thresholding on WorldView-3 NIR/SWIR bands; morphological skeleton extraction; manual quality control at network junctions | Vector creek network layer with channel-width attributes; input for carbon-burial spatial covariate modelling |
| Multi-year change detection report | Bi-temporal or time-series classification comparison using Landsat and Sentinel-2 archive; change magnitude and direction mapped at pixel level; class-transition matrix computed | Annotated change map and transition matrix PDF; machine-readable GeoPackage of changed polygons with date attribution; reproducible processing log |
| Hyperspectral community unmixing (DESIS) | Linear spectral unmixing using endmembers derived from field spectra or image-derived pure pixels; fractional cover estimated per dominant species per 30 m pixel | Fractional cover rasters per species; uncertainty map based on residual error; advisory note on coverage gaps from ISS orbital constraints |
| MRV uncertainty propagation summary | Monte Carlo propagation of classification accuracy, biomass regression error and community-carbon density variance; class-level rather than overall accuracy reported | Structured uncertainty table in registry-compatible format; separates extent error from carbon-density assignment error |
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.