Seagrass bed distribution mapping in shallow coastal waters
Satellite water-column correction methods can map submerged seagrass meadows to depths of roughly five to ten metres, giving coastal managers areal estimates that field surveys alone cannot match in scale or frequency.
Sensors
- Sentinel-2 MSI: 10 m resolution in visible bands (B2 blue, B3 green, B4 red) with a 5-day revisit at the equator under the two-satellite constellation. Free and globally archived from 2015, making it the workhorse for change detection over time. The coastal aerosol band (B1, 60 m) aids atmospheric correction but is too coarse for substrate classification.
- Planet SuperDove: 3 m resolution across eight spectral bands including a dedicated coastal blue band, with near-daily revisit. The finer pixel footprint resolves patch edges and small meadow fragments that Sentinel-2 blurs into mixed pixels. Requires a commercial licence; archive depth varies by site.
- WorldView-2 / WorldView-3: 0.5 m panchromatic, 2 m multispectral across eight bands including coastal blue and yellow bands particularly useful for water-column penetration. Best spatial detail available commercially, but narrow swath (16.4 km) and tasking cost mean it is used for validation sites or high-value patches rather than regional surveys.
- NASA PACE OCI: Launched March 2024. Hyperspectral ocean colour instrument covering 340 to 890 nm at roughly 5 nm spectral resolution and approximately 1 km spatial resolution. Designed for open-ocean colour but its fine spectral sampling improves water-leaving radiance decomposition in optically complex coastal waters. Spatial resolution limits it to regional-scale seagrass distribution rather than patch mapping.
Why the water column is the problem, not the water
Light entering the sea is selectively absorbed and scattered before it reaches the bottom, and again on the way back up. Blue wavelengths penetrate furthest; red is gone within a metre or two in most coastal water. What a satellite sensor records is a mixture of surface reflectance, water-column attenuation and bottom reflectance. Disentangling these three contributions is the entire analytical challenge of shallow-water substrate mapping.
The Lyzenga log-ratio method addresses this by taking the natural logarithm of two visible bands (typically blue and green) and forming a ratio that, under certain assumptions about water optical properties, is depth-invariant. Pixels of the same substrate type cluster together in log-ratio space regardless of depth, allowing classification of seagrass, sand, coral and algae. The Sagawa depth-invariant index extends this logic and has been applied to WorldView and Sentinel-2 imagery in Indo-Pacific settings. Neither method is magic: both assume spatially uniform water optical properties, which turbid or heterogeneous coastal waters violate.
What the spectral bands actually see
Seagrass absorbs strongly in the red and near-infrared due to chlorophyll and accessory pigments, but those wavelengths rarely survive a return trip through more than one or two metres of water. In practice, seagrass discrimination at depth relies on the contrast between its relatively low blue-green reflectance compared with bright carbonate sand, and its different spectral shape compared with macroalgae. WorldView-2 and SuperDove both carry a coastal blue band centred near 425 to 450 nm, which penetrates slightly deeper than the standard blue band and improves depth-invariant index performance.
Hyperspectral data from airborne sensors (CASI, HyMap) has long been the research standard for seagrass mapping because narrow contiguous bands allow explicit modelling of water absorption features. PACE OCI brings something close to that spectral resolution from orbit for the first time, though at 1 km pixels it cannot map individual meadow patches. Its value lies in calibrating the atmospheric and water-column correction applied to finer-resolution multispectral sensors.
Depth limits are honest limits, not marketing caveats
The practical depth ceiling for satellite seagrass detection is approximately five metres in turbid estuarine water and up to ten metres in clear oceanic conditions such as the Coral Sea or Caribbean. Beyond those depths, the bottom signal is buried in water-column noise and atmospheric correction residuals. No post-processing method recovers it. This is a physical constraint, not a software limitation.
Turbidity is the other hard wall. Suspended sediment concentrations above roughly 5 to 10 mg per litre typically destroy bottom detectability even at one metre depth. Seasonal river plumes, storm resuspension and tidal fronts all create windows of poor visibility that can span weeks. Time-series analysis helps: compositing multiple cloud-free, low-turbidity images over a season improves the probability of capturing a usable observation for any given pixel. Sentinel-2's five-day revisit makes this feasible without commercial cost.
From classified image to areal estimate
A seagrass map is only as good as its validation. Accuracy assessment requires field survey points or high-resolution aerial photography to assign ground truth to substrate classes. Published studies using Sentinel-2 and WorldView-2 in the Indo-Pacific report overall classification accuracies of 70 to 85 percent for four-class substrate schemes (seagrass, coral, sand, algae), with seagrass producer accuracy often lower than user accuracy because sparse or patchy meadows generate mixed pixels. These figures should be treated as indicative; site-specific accuracy depends on water clarity, tidal state at image acquisition and the density of the meadow itself.
Areal estimates derived from classified maps carry uncertainty from both classification error and pixel-edge effects. Reporting a confidence interval around the estimated hectarage, rather than a single number, is the honest approach. Change detection between two epochs adds a further source of uncertainty: apparent loss or gain can reflect differences in tidal height or water clarity at acquisition rather than genuine ecological change. Normalising acquisition conditions, or using tidal modelling to select images at comparable water levels, is standard practice in published seagrass monitoring programmes.
Where satellite mapping fits in a monitoring programme
Satellite mapping does not replace in-water survey. It does something different: it provides synoptic spatial coverage that a dive team cannot achieve in any practical timeframe. A team of six divers working transects might characterise ten hectares per day in good conditions. A single Sentinel-2 scene covers 290 by 290 kilometres. The appropriate division of labour is to use satellite classification to identify where meadows exist and to quantify broad areal change, then direct field effort to validate boundaries, assess shoot density and collect the biological data the sensor cannot see.
Satellize runs depth-invariant index workflows on Sentinel-2 and, where clients hold licences, on Planet SuperDove, producing classified substrate layers and change-detection reports. The Tonga crop-estimation programme demonstrated the organisation's approach to small-island-state analytics under data-scarce conditions, a context directly relevant to the Pacific seagrass nations where meadow monitoring is most urgently needed and least resourced. For a new site, the first step is a scene-quality audit: identifying which archived acquisitions meet the turbidity and tidal criteria for usable bottom signal before any classification is attempted.
Typical figures
| Typical spatial resolution | 10 m (Sentinel-2 visible bands); 3 m (Planet SuperDove); 2 m multispectral / 0.5 m pan (WorldView-2/3); ~1 km (PACE OCI) |
| Revisit frequency | 5 days (Sentinel-2, two-satellite); near-daily (Planet SuperDove); tasked on demand (WorldView-2/3) |
| Effective depth limit | ~5 m in turbid water; up to ~10 m in clear oceanic conditions |
| Turbidity threshold | Bottom signal typically lost above ~5–10 mg/L suspended sediment |
| Key spectral bands | Coastal blue (~440–450 nm), blue (~490 nm), green (~560 nm); red and NIR used only in very shallow (<2 m) clear water |
| Minimum mappable patch size | ~1 ha reliably at 10 m resolution; ~0.1 ha at 3 m resolution (dependent on meadow density) |
| Archive depth | Sentinel-2: from 2015; Landsat (for coarser change context): from 1972; Planet SuperDove: varies by site, generally post-2017 |
| Delivery formats | GeoTIFF classified raster, vector polygon layer (GeoPackage or Shapefile), accuracy assessment report, change-detection summary |
| Classification accuracy (published range) | 70–85% overall for 4-class substrate scheme in Indo-Pacific studies; site-specific validation required |
Analytics Satellize can run
| Depth-invariant substrate classification | Lyzenga log-ratio or Sagawa depth-invariant index applied to atmospherically corrected blue and green bands | GeoTIFF and vector polygon layer of seagrass, sand, coral and algae classes with per-class area statistics |
| Seagrass areal change report | Multi-epoch classification comparison with tidal-state normalisation; change pixels flagged only where acquisition conditions are comparable | PDF report with before/after maps, net area change in hectares and confidence intervals |
| Scene-quality audit and acquisition calendar | Automated turbidity proxy (NDWI variants, Rrs 490/560 ratio) and cloud-mask screening across archive | Ranked list of usable historical acquisitions and recommended future acquisition windows by tidal phase |
| Meadow boundary vector layer | Object-based image analysis (OBIA) on classified raster to delineate contiguous meadow polygons and compute patch perimeter and fragmentation metrics | GeoPackage polygon layer with attribute table including area, perimeter and centroid coordinates |
| Accuracy assessment with field-point integration | Stratified random sampling of classified map against client-supplied or open-access validation points; confusion matrix computation | Accuracy report with overall accuracy, kappa coefficient and per-class producer and user accuracy |
| Seasonal composite seagrass probability surface | Pixel-wise median compositing of depth-invariant index values across low-turbidity acquisitions within a defined season | Continuous probability raster (0–1) indicating likelihood of seagrass presence, suitable for threshold-based mapping or probabilistic reporting |
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.