Sensors
- Sentinel-2 MSI: 10 m resolution in the visible bands (443, 490, 560, 665 nm) with a 5-day revisit at the equator from the two-satellite constellation. The coastal aerosol band (Band 1, 443 nm) and the three red-edge bands are useful for water-column correction and shallow-water bottom discrimination. Free and globally archived from 2015.
- WorldView-2 / WorldView-3: Eight multispectral bands at 1.85 m (WV-2) or 1.24 m (WV-3) ground sample distance, including a dedicated coastal band (400–450 nm) and yellow band (585–625 nm) that improve discrimination between seagrass species and bare sediment in optically shallow water. Tasked commercially; revisit is typically 1–4 days depending on latitude and tasking priority.
- Planet SuperDove: Eight bands including a coastal blue band, 3–5 m resolution, near-daily revisit globally. Radiometric consistency across the constellation has improved with the SuperDove generation, making it useful for change detection between seasons. Depth penetration is limited to the same physics as Sentinel-2 but at finer spatial grain.
- ICESat-2 ATLAS: Photon-counting lidar operating at 532 nm (green), capable of detecting the seabed in waters shallower than roughly 30–40 m under clear conditions, with depth precision of approximately 10 cm vertically. Along-track spacing is 0.7 m at the surface; cross-track coverage is sparse (six beams, ~3.3 km apart), so ATLAS functions as a calibration transect rather than an area-mapping instrument. Repeat cycle is 91 days.
Why seagrass carbon is systematically undercounted
Seagrass meadows sequester carbon primarily in below-ground biomass and the organic-rich sediments they stabilise. Published estimates place sediment carbon stocks in the range of 83 to 141 Mg C per hectare, comparable to many mangrove systems. Yet fewer than a third of the countries with significant seagrass coastlines include them in national greenhouse gas inventories submitted to the UNFCCC. The gap is not political indifference. It is a measurement problem.
Field surveys in shallow coastal water are expensive and slow. Acoustic methods work for bathymetry but do not discriminate vegetation type. Aerial photography gives good spatial detail but poor temporal coverage at national scale. Satellite remote sensing is the only method that can produce consistent, repeatable, archive-backed evidence across an entire exclusive economic zone, and it has matured enough in the last decade to support credible MRV.
What the physics allows, and where it stops
Light penetrating the water column is attenuated and scattered before it reaches the seabed, reflects off the bottom, and must travel back up through the same water before reaching the sensor. The practical depth limit for passive optical mapping depends on water clarity. In the clearest tropical waters (low suspended sediment, low chlorophyll), Sentinel-2 and WorldView imagery can resolve bottom features to roughly 15–20 m. In turbid estuarine conditions that limit drops to 3–5 m, or less. This is not a solvable engineering problem with current passive sensors. It is physics.
Several published water-column correction methods exist to separate the bottom-reflectance signal from the water-column contribution. The Lyzenga log-ratio method uses pairs of bands with different attenuation coefficients to produce a depth-invariant bottom index. The Lee radiative transfer model (also called the semi-analytical approach) inverts a physics-based model of water optical properties to retrieve both depth and bottom albedo simultaneously. Both methods require either in-situ depth measurements or ICESat-2 transects for calibration. Neither works reliably when turbidity is high enough to obscure the bottom entirely, and neither can distinguish seagrass from dark algae or dark sediment without additional spectral information.
Species discrimination: what spectral libraries actually show
Published spectral libraries for seagrass genera document measurable differences in leaf reflectance. Posidonia oceanica, dominant in the Mediterranean, has a relatively high near-infrared reflectance compared with bare sand, though the NIR signal is strongly attenuated in even a few metres of water. Thalassia testudinum (turtle grass, Caribbean and Gulf of Mexico) and Zostera marina (temperate Atlantic and Pacific) show distinct green-peak reflectance signatures when measured in situ, but these differences compress significantly after water-column attenuation and atmospheric correction.
WorldView-2 and WorldView-3, with their eight bands including coastal blue and yellow channels, offer the best current passive-sensor basis for species-level discrimination in clear, shallow water. Sentinel-2 can separate seagrass from bare sand and from coral reliably in good conditions, but species-level separation is uncertain without ground truth. Planet SuperDove's eight-band configuration is intermediate. Honest practice is to report a single seagrass class with a confidence interval, rather than species maps, unless site-specific spectral calibration has been done.
From extent map to carbon stock: the chain of inference
A satellite-derived extent map is not a carbon stock estimate. Converting area to stock requires allometric relationships between canopy cover, shoot density, below-ground biomass, and sediment organic carbon. These relationships are species- and site-specific. Published values for sediment carbon density vary by an order of magnitude across meadow types and geographic regions. The IPCC Wetlands Supplement (2013) provides Tier 1 default values, but Tier 2 and Tier 3 approaches require field sampling co-located with the remote-sensing footprint.
The practical workflow is: produce a classified extent map with uncertainty bounds; stratify by depth zone and canopy density class using the depth-invariant indices; apply published allometric coefficients per stratum with propagated uncertainty; report a stock estimate with a confidence range that honestly reflects both the classification error and the allometric variance. Change detection between epochs, using consistent preprocessing, is often more defensible for MRV than an absolute stock number, because systematic errors in the water-column correction partially cancel when the same method is applied to the same site at different times.
ICESat-2 as the calibration backbone
ICESat-2 ATLAS is the only spaceborne instrument currently providing direct bathymetric depth measurements at the precision needed to calibrate water-column correction models. Its 532 nm photon-counting lidar penetrates clear water to depths well beyond the passive-sensor limit, and its along-track depth precision of roughly 10 cm allows accurate construction of depth profiles. Several published studies have used ATLAS transects to validate and calibrate Sentinel-2 and WorldView-derived depth maps in coral reef and seagrass environments.
The limitation is coverage. Six beams separated by 3.3 km cross-track, on a 91-day repeat, means that most coastal sites will have at most a handful of transects in the archive. ATLAS is a calibration instrument, not a mapping instrument. The workflow is to extract depth points from available ATLAS passes, use them to fit the water-column correction model for the site, then apply that model to the denser passive-sensor archive. Where no ATLAS transect exists, boat-based echo-sounding or airborne lidar remains necessary for calibration.
Putting it into practice
A credible seagrass MRV product requires at minimum: a multi-date Sentinel-2 archive for the site (free from the Copernicus Data Space); at least one high-resolution image (WorldView or Planet) for classification validation; one calibration depth source, preferably ICESat-2 supplemented by field points; a published water-column correction method applied consistently across dates; and a reported classification accuracy assessed against independent field or very-high-resolution reference data. Accuracy assessments in the published literature for seagrass mapping in clear-water tropical settings typically report overall accuracies of 70–90%, with lower performance in mixed or turbid conditions.
Satellize applies this workflow on open and commercial constellations for coastal-state clients building blue-carbon inventories. The same radiometric preprocessing pipeline used in the Kingdom of Tonga crop-estimation programme underpins the coastal-water atmospheric correction step here. For sites with persistent cloud or high seasonal turbidity, the realistic output is a best-available-conditions composite rather than a continuous time series. Buyers should be sceptical of any vendor claiming annual, wall-to-wall seagrass maps in monsoon-affected or high-sediment-load coastlines. The physics does not support it. What satellite data does support is a repeatable, auditable, archive-backed baseline that field campaigns alone cannot provide at national scale.
Typical figures
| Spatial resolution (passive optical) | 1.24–1.85 m (WorldView-2/3); 3–5 m (Planet SuperDove); 10 m (Sentinel-2 visible bands) |
| Revisit cadence | Near-daily (Planet); 5 days at equator (Sentinel-2 dual satellite); 1–4 days tasked (WorldView); 91-day repeat (ICESat-2) |
| Effective depth penetration (passive) | Up to ~15–20 m in clearest tropical water; 3–5 m or less in turbid or estuarine conditions |
| ICESat-2 bathymetric depth precision | ~10 cm vertical; penetration to ~30–40 m in clear water; along-track point spacing ~0.7 m |
| Key spectral bands for seagrass | Coastal blue (400–450 nm), blue (450–510 nm), green (530–590 nm), yellow (585–625 nm, WV-2/3 only), red-edge (Sentinel-2) |
| Minimum mappable patch size | ~100 m² at WorldView resolution; ~900 m² at Planet; ~1 ha at Sentinel-2 (subject to contrast with substrate) |
| Sentinel-2 archive depth | 2015 to present (global); earlier coverage from Landsat 8 (2013–) at 30 m |
| Typical classification accuracy (clear water) | 70–90% overall accuracy reported in peer-reviewed literature; lower in turbid or mixed-substrate conditions |
| Delivery formats | GeoTIFF extent map, GeoPackage per-stratum stock table, PDF MRV evidence report, uncertainty-bounded carbon stock raster |
Analytics Satellize can run
| Seagrass extent map with canopy-density classes | Supervised classification (random forest or support vector machine) on water-column-corrected, atmospherically corrected multispectral imagery, with accuracy assessment against field or very-high-resolution reference data | GeoTIFF classified raster and vector polygon layer with per-class area statistics and confusion matrix |
| Depth-invariant bottom index | Lyzenga log-ratio or Lee semi-analytical radiative transfer model, calibrated with ICESat-2 ATLAS transects or field depth points | Depth-invariant index raster used as input to classification and stock stratification |
| Bathymetric depth map | Empirical or physics-based inversion of Sentinel-2 or WorldView multispectral data, validated against ICESat-2 along-track depth profiles | GeoTIFF depth raster with RMSE uncertainty estimate per depth zone |
| Blue-carbon stock estimate by stratum | Area-weighted application of published allometric coefficients (IPCC Wetlands Supplement Tier 1, or site-specific Tier 2 values) to classified extent and density strata | Tabular stock report with confidence intervals, suitable for national inventory submission or voluntary carbon registry |
| Change detection between baseline and monitoring epochs | Consistent water-column-corrected classification applied to multi-date image pairs; change matrix quantifying gain, loss and stable area | Change GeoTIFF, area-change table with dates, and narrative MRV evidence summary |
| Turbidity and cloud-cover quality flag per scene | Automated scene-quality screening using water-leaving radiance thresholds and cloud/shadow masks (Sen2Cor or equivalent) | Per-scene quality flag layer appended to image archive metadata; used to exclude unreliable acquisitions from stock calculations |
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.