Satellite-derived bathymetry in shallow coastal waters
Multispectral satellites can estimate water depth in clear, optically shallow coastal zones by exploiting how blue and green light attenuate with depth. The method works to roughly 15–25 metres in clean water and fails entirely when turbidity or algae obscure the bottom.
Sensors
- Sentinel-2 MSI: 10 m resolution in the visible bands (B2 blue, B3 green, B4 red); 5-day revisit at the equator with both satellites. Free archive from 2015. The coastal aerosol band (B1, 60 m) aids atmospheric correction but is too coarse for fine-scale depth mapping. The workhorse for open-access bathymetric retrieval.
- Landsat 8/9 OLI: 30 m resolution across all visible bands including a dedicated coastal/aerosol band (band 1, 443 nm). 16-day revisit per satellite, 8-day combined. Archive extends to 1984 for Landsat 5/7, giving decades of coastal change context. Coarser than Sentinel-2 but the long record is unmatched.
- Planet SuperDove: 3 m resolution, 8 spectral bands including two blue bands (450 nm and 490 nm) that improve atmospheric and water-column separation. Near-daily revisit globally. Commercial licence required. Particularly useful for small reef structures and port approaches where 10 m pixels blur critical features.
- WorldView-2/3: WorldView-2 carries a dedicated coastal band at 427 nm and a yellow band, giving 8 multispectral bands at 1.85 m (multispectral) and 0.46 m (panchromatic). WorldView-3 adds shortwave infrared. Sub-2 m multispectral resolution resolves individual coral heads and channel edges. Tasked commercially; expensive per km².
What a floating roof gives away
Sunlight entering the water column is absorbed and scattered at rates that vary by wavelength. Red light is absorbed within the first few metres. Blue light (roughly 400–500 nm) penetrates furthest in clear ocean water, followed by green (500–570 nm). A satellite sensor looking down sees a mixture of surface-reflected light, water-column backscatter, and bottom reflectance. In optically shallow water, where the seabed is close enough to contribute signal, that bottom reflectance carries depth information.
The physics was formalised by Lyzenga in the early 1980s and refined repeatedly since. The core insight is that the ratio of log-transformed radiances in two bands, typically blue and green, is approximately linear with depth over a sandy or coral substrate. Deeper water absorbs more blue relative to green, shifting the ratio in a predictable direction. Invert that relationship with a calibration against known depths and you have a bathymetric map.
Empirical versus semi-analytical: two routes to the same seabed
Empirical methods, such as the Stumpf ratio transform published in 2003, require a set of in-situ depth soundings or lidar control points to anchor the regression. The band-ratio image is regressed against those known depths, and the resulting coefficients are applied across the scene. Simple, fast, and reasonably accurate in the calibration zone, but the coefficients are scene-specific. A different water clarity, substrate type, or sun angle in an adjacent tile can shift results by several metres.
Semi-analytical approaches, such as the Lee et al. quasi-analytical algorithm and its derivatives, attempt to separate the contributions of the water column's inherent optical properties (absorption and backscattering coefficients) from bottom reflectance. They require assumptions or measurements about water constituents but are more transferable across scenes and conditions. The Hydrolight radiative transfer model underpins much of this work. In practice, operational programmes often blend both: semi-analytical physics to normalise across tiles, empirical calibration to pin absolute depth.
Both approaches share a hard constraint. The bottom must be visible. Once the water column is optically deep, meaning bottom reflectance contributes negligibly to the top-of-atmosphere signal, no amount of spectral algebra recovers depth. In clear tropical water this ceiling sits around 20–25 m for Sentinel-2. In turbid estuaries or algae-bloomed coastal zones it can fall below 2–3 m. The retrieval does not degrade gracefully; it simply stops working.
Atmospheric correction is where retrievals quietly fail
Water-leaving radiance is a small fraction of the total signal at the top of the atmosphere. Over dark water, the atmosphere can account for 80–90% of what the sensor records. Errors in atmospheric correction, whether from aerosol optical depth misestimation or adjacency effects from bright land nearby, translate directly into depth errors. The ACOLITE processor developed by RBINS and the 6SV-based corrections in the Sen2Cor and LaSRC chains are commonly used, but none is foolproof.
Sunglint adds another complication. Specular reflection of direct sunlight off the water surface can saturate pixels or introduce a spatially variable bias. Deglinting algorithms, typically regressing near-infrared radiance against visible bands over deep water, help but can over-correct in shallow areas where the NIR signal contains genuine bottom information. Imagery acquired at low sun angles or with significant wave action is often unusable for bathymetry regardless of the algorithm applied.
What the method is actually good for
Reef mapping is the clearest application. The Great Barrier Reef Marine Park Authority and the Allen Coral Atlas project have both used Sentinel-2 and Planet imagery to map benthic habitat and approximate depth across reef structures that are impractical to survey by boat at scale. Depth accuracy of roughly plus or minus 1–2 m at the 10 m pixel scale is achievable over clean carbonate sand and coral in well-calibrated scenes, based on published comparisons against airborne lidar.
Port approach charting is more demanding. Navigational safety requires accuracy that satellite-derived bathymetry cannot yet guarantee without dense lidar or sonar control. The value is in change detection: comparing a new satellite-derived depth surface against a charted baseline to flag areas of possible sedimentation or scour for targeted re-survey, rather than replacing hydrographic survey entirely.
Coastal change monitoring over years or decades is where the archive depth of Landsat and Sentinel-2 pays off. Tracking shoreline migration, sand bar movement, or the progressive bleaching and flattening of reef structures is feasible at decadal scale. Satellize supports this kind of time-series analysis through its open-constellation analytics, including work in Pacific island contexts not unlike the bathymetric conditions encountered in its Tonga crop-estimation programme area.
Honest limits: what breaks the retrieval
Turbid water is the primary failure mode, and it is common. River plumes, resuspended sediment after storms, and seasonal phytoplankton blooms can all render the water column optically deep at wavelengths where the bottom would otherwise be visible. There is no spectral workaround. SAR can detect surface roughness and some near-surface features, but it cannot see the seabed in the way optical sensors can. Spaceborne lidar, specifically ICESat-2's photon-counting system, can penetrate the water column and retrieve depth profiles along its ground tracks, but its spatial coverage is linear and sparse compared to an image.
Substrate matters too. Dark seagrass beds absorb more light than bright sand, producing an apparent depth bias if the calibration was built on sandy bottom. Mixed substrates within a single pixel, common at 10–30 m resolution, produce depth estimates that are averages of uncertain composition. Very high resolution imagery from WorldView-2/3 reduces this problem but does not eliminate it.
Finally, tidal stage affects the retrieval. A scene acquired at high tide shows different apparent depths than one at low tide. Tidal correction using modelled or measured water levels is necessary for any quantitative comparison across dates or between sensors.
Typical figures
| Typical spatial resolution | 10 m (Sentinel-2 visible bands), 30 m (Landsat 8/9 OLI), 3 m (Planet SuperDove), 1.85 m multispectral (WorldView-2/3) |
| Maximum detectable depth (clear water) | Approximately 15–25 m depending on water clarity, substrate brightness and sensor SNR; can fall below 5 m in turbid or algae-laden conditions |
| Typical depth accuracy (calibrated, clear water) | ±1–2 m RMSE over sandy/coral substrate in well-calibrated scenes; degrades significantly with substrate variability or poor atmospheric correction |
| Key spectral bands | Blue (~440–490 nm) and green (~540–570 nm) for depth ratio; coastal/aerosol band (~400–450 nm) for atmospheric correction; NIR for sunglint correction and water masking |
| Revisit cadence | 5 days (Sentinel-2A+B combined at equator); 8 days (Landsat 8+9 combined); near-daily (Planet SuperDove); on-demand tasking (WorldView-2/3) |
| Archive depth | Sentinel-2 from 2015; Landsat from 1984 (Landsat 5); Planet from approximately 2016 |
| Coverage per acquisition | 290 km swath (Sentinel-2); 185 km (Landsat); ~3.7 km strip width typical (WorldView); variable (Planet tasked or routine) |
| Delivery formats | GeoTIFF depth raster, vector contour lines, GIS-ready point cloud (for ICESat-2 integration), change-detection map in GeoPackage or shapefile |
| Atmospheric correction requirement | Mandatory; common processors include ACOLITE (RBINS), Sen2Cor, LaSRC; errors of even 1–2% in surface reflectance translate to metre-scale depth errors |
Analytics Satellize can run
| Calibrated depth raster | Stumpf (2003) log-ratio transform or Lyzenga linear transform, calibrated against available sonar or ICESat-2 control points | GeoTIFF depth grid at native sensor resolution with uncertainty band per pixel, delivered per scene or as a cloud-free composite |
| Benthic depth change map | Multi-date depth raster differencing with tidal normalisation; anomalies flagged against a baseline survey | Change polygon layer (GeoPackage) with magnitude and direction of depth change, suitable for directing hydrographic re-survey |
| Optically shallow water extent mask | NIR-based water masking combined with blue/green ratio threshold to delineate zones where bottom is detectable; outputs a confidence classification | Binary and confidence-graded raster mask, updated per new cloud-free acquisition |
| Turbidity flag and retrieval validity layer | Green-band reflectance threshold and NDWI-based turbidity proxy; pixels exceeding turbidity limits are masked as unreliable | Per-scene validity raster included with every depth product; summary statistics on retrievable area fraction |
| Reef structure and sand-flat classification | Substrate classification using bottom-of-water-column reflectance spectra (from semi-analytical inversion) compared against spectral libraries for coral, sand, and seagrass | Benthic habitat map (vector polygons) with depth layer overlaid, formatted for marine spatial planning GIS workflows |
| Decadal coastal bathymetric trend report | Landsat and Sentinel-2 time-series stacking with annual compositing; linear trend fitted per pixel across the archive | PDF report with trend maps, area statistics, and flagged zones of significant accretion or erosion; underlying rasters in GeoTIFF |
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.