Multispectral coral reef habitat classification

Multispectral satellites exploit spectral contrasts between coral, algae, seagrass and sand in the blue-green window to map reef habitats, but water-column attenuation and turbidity impose hard limits that no algorithm fully escapes.

Sensors

WorldView-2 / WorldView-3 (Maxar): WorldView-3 delivers 30 cm panchromatic and 1.24 m multispectral resolution with eight VNIR bands including a dedicated coastal-blue band near 427 nm, essential for shallow-water penetration. Revisit is roughly 1 to 4.5 days depending on latitude and tasking priority. The resolution is sufficient to resolve individual coral heads and patch-reef boundaries at the sub-metre scale, though archive coverage of remote reefs is sparse.
Sentinel-2 MSI (ESA / Copernicus): 10 m resolution in the visible bands (Band 2 at 490 nm, Band 3 at 560 nm, Band 4 at 665 nm) with a 5-day revisit at the equator from the two-satellite constellation. Free and open, with global coverage and an archive back to 2015. Adequate for reef-scale habitat mapping and change detection, but cannot resolve individual coral colonies or fine-scale zonation patterns.
Planet SuperDove: 3 m resolution across eight spectral bands including a coastal-blue band at 444 nm. Near-daily revisit globally makes it attractive for monitoring bleaching events as they develop. Radiometric consistency across the fleet has improved substantially since the introduction of SuperDove, though absolute calibration for water-column work still requires careful cross-calibration against in-situ measurements.
Landsat 8 / 9 OLI (USGS / NASA): 30 m resolution with a 16-day revisit per satellite (8 days combined). Band 1 (coastal aerosol, 435 to 451 nm) was specifically designed for water-column and coastal applications. The long archive stretching back to 1984 (Landsat 5 onwards) makes it the primary source for multi-decadal reef change studies, though 30 m pixels conflate habitat types across most patch-reef structures.

Why the blue-green window is the only window that matters

Water absorbs red and near-infrared light within the first few metres of the water column. By 5 m depth, virtually no red light reaches the seafloor and returns to a satellite sensor. The blue band (roughly 430 to 490 nm) and the green band (490 to 580 nm) are the only spectral regions where solar irradiance penetrates shallow coastal water and reflects back in detectable quantities. This is not a design choice; it is physics.

Different benthic substrates reflect these wavelengths differently. Coral carbonate skeletons scatter broadly across the blue-green range. Live coral with zooxanthellae absorbs more blue light than bare skeleton. Seagrass absorbs blue and reflects green. Algal turf has a distinct green peak. Sand reflects strongly and fairly uniformly across both bands. These contrasts are real but modest, typically separated by 2 to 8 per cent reflectance, which means sensor calibration and atmospheric correction must be precise before any classification attempt is credible.

Removing the water column before you can see the bottom

The confounding problem is depth. A coral patch at 2 m looks brighter than the same coral at 8 m, simply because less water attenuates the signal. Without correcting for this, a classifier will conflate depth variation with substrate variation and produce maps that are partly bathymetric artefacts.

The standard approach is the Lyzenga log-ratio depth-invariant index, published in the early 1980s and still widely used. It exploits the fact that two bands attenuate at different rates with depth. By taking the log ratio of blue and green reflectance, it is possible to produce a depth-invariant bottom index that separates substrate type from water-column depth, provided the water is optically homogeneous. The method requires areas of known uniform substrate (typically deep sand patches) to calibrate the attenuation coefficients. More recent physics-based approaches such as the Maritorena model and semi-analytical inversion methods retrieve water optical properties and bottom reflectance simultaneously, but they demand more spectral bands and more careful parameterisation. Neither approach is immune to turbidity.

Resolution sets the ceiling on what you can classify

At 30 m (Landsat), a single pixel on a complex reef might contain coral, rubble, algae and sand in unknown proportions. Classification accuracy on structurally complex reefs at this resolution is typically reported in the literature at 60 to 75 per cent overall accuracy, with substantial confusion between algae and seagrass. At 10 m (Sentinel-2), major habitat zones become separable: fore-reef, reef flat, lagoonal seagrass beds, and sand aprons can be reliably distinguished on reefs with clear water and simple zonation patterns.

At 30 cm to 1.24 m (WorldView-2/3), individual coral colonies, rubble patches and algal mats become resolvable. Published studies using WorldView-2 have reported overall classification accuracies above 80 per cent for four to six habitat classes in clear, shallow water. The trade-off is cost and coverage: tasking a commercial satellite over a remote reef is expensive, and the archive of historically tasked imagery over most reefs outside the Caribbean and Great Barrier Reef is thin.

Ground truth is not optional

Every classifier, whether a maximum-likelihood algorithm, a random forest or a convolutional neural network, requires field-collected ground truth to train and validate against. Satellite pixels over a reef carry no inherent label. Someone has to put a diver or a drop camera over representative patches of each habitat class, record GPS positions, and match those observations to image pixels. The minimum credible dataset for a four-class map is typically 50 to 100 field points per class, distributed across the range of depths and conditions present in the image.

This is the practical bottleneck for most reef mapping projects. Remote reefs in the Pacific or Indian Ocean may not have been surveyed in years. Opportunistic photographs from recreational divers, while increasingly aggregated through citizen science platforms, are rarely geolocated precisely enough to serve as classifier training data. Without contemporary ground truth, accuracy assessments are guesses.

Where the method fails: turbidity, glint and depth

Surface sun glint is the specular reflection of direct sunlight off the water surface. It saturates pixels in the blue and green bands and cannot be corrected after the fact if it is severe. Glint correction algorithms (the Hedley method is the most widely cited) work adequately under moderate glint conditions but require glint-free deep-water pixels as a reference. Imagery acquired at low sun angles or with wind-roughened surfaces reduces glint but introduces other radiometric complications.

Turbidity is a harder problem. Suspended sediment and phytoplankton scatter and absorb light in the blue-green window in ways that mimic benthic reflectance signals. In waters with Secchi depth below roughly 5 m, bottom reflectance at depths greater than 3 to 4 m becomes indistinguishable from the water column signal at the signal-to-noise ratios available from current sensors. Seasonal turbidity events, river plumes and storm resuspension can render an entire reef unmappable for weeks at a time. The practical depth limit for reliable classification in clear oceanic water is approximately 15 m for WorldView-class sensors and closer to 8 to 10 m for Sentinel-2, where the signal-to-noise ratio in the blue band is lower.

Satellize applies these methods on open-access Sentinel-2 and Landsat archives, adding commercial WorldView tasking on client licence where sub-metre resolution is required. The Tonga crop-estimation programme demonstrated the same depth-invariant correction logic in a coastal context, which transfers directly to reef habitat work across the Pacific island region.

What a finished map can and cannot tell a reef manager

A well-executed multispectral habitat map distinguishes broad classes: live coral-dominated substrate, algae-dominated substrate, seagrass, rubble and sand. It can track the spatial extent of each class over time, identify zones where coral cover has declined and algae has expanded (a common bleaching or eutrophication signature), and provide area statistics for marine protected area planning and reporting.

What it cannot do is measure coral species composition, percentage live coral cover at the colony scale, or bleaching intensity within a mixed pixel. Those require in-water survey or, at the frontier of current research, hyperspectral airborne or spaceborne data. The map is a spatial framework, not a substitute for ecological survey. Used honestly, it tells a reef manager where to send divers, not what those divers will find.

Typical figures

Best available spatial resolution	30 cm (WorldView-3 panchromatic); 1.24 m (WorldView-3 multispectral); 3 m (Planet SuperDove); 10 m (Sentinel-2 visible bands); 30 m (Landsat 8/9 OLI)
Effective depth limit (clear oceanic water)	~15 m for WorldView-class sensors; ~8 to 10 m for Sentinel-2; shallower in turbid or coastal-influenced water
Revisit frequency	1 to 4.5 days (WorldView-3, tasked); near-daily (Planet SuperDove); 5 days (Sentinel-2, two-satellite); 8 days combined (Landsat 8+9)
Key spectral bands for reef mapping	Coastal blue (427 to 444 nm), blue (450 to 490 nm), green (490 to 580 nm); red bands largely attenuated below 3 to 5 m depth
Typical classification accuracy (published literature)	60 to 75% overall at 30 m; up to 80 to 85% at 1 to 3 m resolution in clear water with adequate ground truth; lower in turbid or deep conditions
Minimum habitat patch detectable	Approximately 3 to 5 pixels across for reliable classification; ~1 m patches with WorldView-3, ~30 m patches with Landsat
Archive depth	Sentinel-2: 2015 to present; Landsat: 1984 to present (Landsat 5 onwards); WorldView-2: 2009 to present (sparse for remote reefs); Planet: 2016 to present
Primary failure modes	Surface sun glint, turbidity (Secchi depth below ~5 m), depth beyond sensor penetration limit, absence of contemporary field ground truth
Deliverable formats	GeoTIFF habitat classification raster, vector polygon habitat map (GeoPackage / Shapefile), area-statistics report, change-detection layer between two epochs

Analytics Satellize can run

Benthic habitat classification map	Depth-invariant index (Lyzenga log-ratio) followed by supervised classification (random forest or maximum-likelihood) trained on field ground-truth points	GeoTIFF raster and vector polygon layer with four to six habitat classes (coral, algae, seagrass, rubble, sand, deep/unclassifiable), with per-class area statistics
Multi-epoch habitat change detection	Post-classification comparison of two or more atmospherically corrected and depth-corrected image dates; change matrix quantifying transitions between habitat classes	Change-detection GeoTIFF and tabular report showing net gain or loss per class between selected dates, drawn from Sentinel-2 or Landsat archive
Bleaching extent indicator	Anomaly detection on the coral-class reflectance signal in the blue-green bands; bleached coral carbonate reflects more strongly than healthy zooxanthellate coral, producing a detectable brightness anomaly at 3 to 10 m resolution	Spatial alert layer flagging pixels with statistically significant brightening relative to a baseline period, updated per available cloud-free acquisition
Depth-invariant bottom index mosaic	Lyzenga or Maritorena semi-analytical model applied to atmospherically corrected, glint-corrected imagery to produce a substrate index independent of water-column depth variation	Single-band GeoTIFF index mosaic suitable for input to client classification workflows or visual interpretation
Habitat area statistics for MPA reporting	Zonal statistics extracted from classified habitat map clipped to marine protected area boundary polygons supplied by client	Tabular report of habitat class areas (hectares) within each MPA zone, formatted for standard reporting frameworks
Ground-truth sampling design	Stratified random sampling of classified habitat strata to optimise field survey effort; sample size calculated to achieve target confidence intervals per class	GPS waypoint file and field survey protocol document specifying target sample counts per stratum and recommended observation method

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.