Satellite-derived bathymetry for optically shallow rivers and lakes
In clear, shallow water, sunlight penetrates to the bed and returns through the water column carrying depth information in its spectral ratios. Sentinel-2 and WorldView imagery can resolve that signal to roughly 10 m depth, provided turbidity stays below about 5 NTU.
Sensors
- Sentinel-2 MSI: 10 m resolution in visible bands (B2 blue, B3 green, B4 red), 5-day revisit at the equator with both satellites. The blue and green bands penetrate to roughly 10 m in clear water. Free and open archive from 2015. Cloud cover is the dominant operational constraint.
- WorldView-2 / WorldView-3: Panchromatic at 0.31–0.46 m, 8-band multispectral at 1.24–1.85 m. The coastal-blue and yellow bands improve shallow-water column discrimination relative to Sentinel-2. Tasked on demand; cost limits systematic coverage of large catchments.
- Planet SuperDove: 3 m multispectral with 8 bands including coastal blue, daily revisit over most land. Spatial resolution suits narrow channels. Radiometric calibration is sufficient for log-ratio bathymetry but less well characterised than Sentinel-2 for physics-based inversion.
- Landsat 8/9 OLI: 30 m resolution, 16-day revisit per satellite (8-day combined). The coastal-aerosol band (Band 1) aids water-column correction. Primary value is the archive depth back to 1984, enabling long-term change in bed morphology rather than high-resolution single-epoch mapping.
What the water column does to light
Sunlight entering a water body is absorbed and scattered as it travels to the bed and back. The rate of attenuation is wavelength-dependent: red light is absorbed within the first metre or two, green penetrates to perhaps 5–7 m in clean water, and blue can reach 10 m or beyond before the signal is indistinguishable from deep-water radiance. A multispectral sensor above the water surface therefore receives a mixture of surface reflectance, water-column scattering and bed reflectance, all entangled with atmospheric path radiance.
Separating depth from this mixture requires either a calibrated physical model of the water's optical properties or an empirical relationship anchored to known-depth control points. Both approaches have been formalised in the published literature. The Lyzenga method uses the natural logarithm of band reflectance and a linear regression against in-situ depths. The Stumpf log-ratio method divides the log of one band by the log of another, cancelling much of the bed-albedo variation and leaving a signal that is more nearly proportional to depth alone. Neither method is magic: both assume the water column is optically homogeneous, the bed is reasonably uniform in reflectance, and the surface is calm enough not to introduce sun-glint noise.
The 5 NTU wall and the 10-metre ceiling
Turbidity is the practical enemy of satellite-derived bathymetry. When suspended sediment or phytoplankton raise turbidity above roughly 5 NTU, the water-column signal begins to dominate over bed reflectance and depth estimates become unreliable. Published studies of optically shallow systems consistently place this threshold in the 3–8 NTU range depending on sediment type and particle size. Rivers in flood, glacially fed lakes, and any system downstream of active construction are typically excluded.
The depth ceiling is set by the blue-band extinction coefficient of the water. In exceptionally clear oligotrophic lakes, blue-band penetration may reach 12–15 m, but most rivers and productive lakes cap out at 8–10 m before the bed signal is lost in noise. Beyond that depth the method returns a flat, saturated value rather than a meaningful number. This is not a processing failure; it is physics. Any deliverable from this method should carry an explicit validity mask showing where turbidity and depth conditions fall within the retrievable envelope.
Sun-glint from a ruffled surface adds a further complication. Sentinel-2 scenes acquired at high solar zenith angles or over wind-exposed lakes require glint correction before band ratios are computed. Several published correction schemes exist, including the Hochberg et al. approach, but residual glint artefacts can still bias shallow-water retrievals by 0.5–1.5 m in affected pixels.
Calibration: what you need on the ground
Empirical methods require control points: a set of locations where both the satellite reflectance and the true depth are known simultaneously. For rivers, this typically means an echo-sounder transect or a high-quality LiDAR survey conducted within a short time of the satellite overpass. The number of control points needed depends on the complexity of the bed and the range of depths: published studies suggest a minimum of 30–50 well-distributed points for a stable Stumpf regression, with accuracy improving up to a few hundred points before diminishing returns set in.
Physics-based inversion (solving the radiative transfer equation directly) can in principle operate without in-situ depth data, but it requires independent measurements of the water's absorption and backscattering coefficients. These are rarely available for ungauged rivers. In practice, most operational deployments blend the two approaches: a physics-based model provides the functional form and the in-situ points tune the coefficients. Without any ground truth, treat the output as relative bathymetry (useful for mapping shoals and channels) rather than absolute depth.
Sentinel-2 in practice: resolution, revisit and honest accuracy
Sentinel-2 is the workhorse for this application at basin scale. Its 10 m visible bands resolve channels wider than roughly 30 m with acceptable edge effects, and its five-day revisit gives a reasonable chance of acquiring a cloud-free, low-turbidity scene within any given month. Published accuracy assessments for Stumpf-method bathymetry from Sentinel-2 over clear lakes and coastal lagoons typically report root-mean-square errors of 0.3–0.8 m for depths under 5 m, degrading to 1–2 m at 8–10 m. These figures are from favourable conditions. Expect worse performance in rivers with variable bed albedo, patchy aquatic vegetation, or seasonal turbidity fluctuations.
WorldView imagery is justified when the channel is narrow (under 30 m), when sub-metre bed-feature mapping is required, or when a single high-value survey needs the best possible spatial detail. The cost and the non-systematic acquisition schedule make it unsuitable for routine monitoring across a large catchment. Planet SuperDove fills the middle ground: daily revisit at 3 m resolution is genuinely useful for tracking rapid geomorphic change after flood events, though its radiometric stability is still being characterised across the published literature.
From depth map to operational product
A single-epoch bathymetric map has limited operational value on its own. The more useful products are change products: comparing two cloud-free acquisitions separated by months or years reveals sediment deposition, channel migration, shoal formation and dredging effects. Landsat's archive back to 1984 makes it possible to reconstruct decadal bed-level change in rivers where no historical survey data exists, at the cost of 30 m spatial resolution.
For hydraulic modelling, the deliverable is typically a gridded depth raster registered to a coordinate system compatible with the model grid, accompanied by a confidence layer that flags pixels outside the valid turbidity and depth range. Satellize runs this processing chain on open Sentinel-2 imagery and can add commercial WorldView tasking under client licence. The workflow draws on the same radiative-transfer calibration approach used in the Tonga crop-estimation programme, where spectral signal decomposition under variable atmospheric and surface conditions is a shared methodological challenge.
One underappreciated application is navigation safety on ungauged rivers. In remote basins where no hydrographic survey has ever been conducted, a satellite-derived bathymetric map at even 1 m vertical accuracy can distinguish navigable from impassable channels, informing logistics planning for aid operations, resource extraction and infrastructure surveys.
Typical figures
| Spatial resolution (Sentinel-2) | 10 m in blue, green, red bands |
| Spatial resolution (WorldView-2/3) | 1.24–1.85 m multispectral; 0.31–0.46 m panchromatic |
| Spatial resolution (Planet SuperDove) | 3 m multispectral |
| Revisit (Sentinel-2, both satellites) | 5 days at equator; cloud-free frequency is site-dependent |
| Maximum retrievable depth | ~10 m in clear water; signal saturates beyond this |
| Turbidity limit | Approximately 3–8 NTU; method unreliable above ~5 NTU |
| Typical vertical accuracy (Stumpf method, Sentinel-2) | 0.3–0.8 m RMSE under 5 m depth; 1–2 m at 8–10 m depth |
| Archive depth (Sentinel-2) | From 2015; Landsat extends comparable analysis to 1984 at 30 m |
| Spectral bands used | Coastal blue, blue, green (primary); red for shallow limit |
| Deliverable formats | GeoTIFF depth raster, confidence/validity mask, vector channel polygons |
Analytics Satellize can run
| Single-epoch depth raster with validity mask | Stumpf log-ratio or Lyzenga multi-band linear regression calibrated against in-situ control points | GeoTIFF depth grid with co-registered turbidity-exclusion and depth-saturation mask; delivered per scene |
| Multi-epoch bathymetric change map | Differencing of co-registered depth rasters from two or more cloud-free acquisitions; change significance tested against per-pixel RMSE estimates | GIS layer showing net deposition and erosion, with magnitude in metres and confidence classification |
| Navigable-channel delineation | Depth raster thresholded at client-specified draught; combined with surface-width retrieval for channel geometry | Vector polygon of navigable reach under specified vessel draught, updated on each cloud-free acquisition |
| Hydraulic model bathymetric input | Depth raster resampled and formatted to match model grid; uncertainty layer propagated from regression RMSE | NetCDF or GeoTIFF bathymetric grid with uncertainty field, ready for ingestion into HEC-RAS or similar |
| Decadal bed-level trend (Landsat archive) | Stumpf log-ratio applied to Landsat OLI archive scenes meeting turbidity and cloud thresholds; linear trend fitted per pixel | Report and raster showing estimated bed-level trend (m per decade) with pixel-level confidence, covering available archive period |
| Post-flood shoal and channel change assessment | Pre- and post-event depth rasters from Planet SuperDove or Sentinel-2; differencing with geomorphic feature classification | Briefing report with annotated change map, suitable for navigation authority or emergency logistics planning |
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.