Freshwater lake aquatic macrophyte mapping and change detection
Satellite multispectral imagery can map emergent, floating and submerged macrophyte communities in inland lakes at 10 m resolution, but submerged vegetation becomes undetectable once turbidity exceeds roughly 5 NTU. This page explains what the physics allows, where it fails, and how to build a defensible change-detection record.
Sensors
- Sentinel-2 MSI: 10 m resolution in Blue, Green, Red and NIR bands; 20 m in four red-edge and SWIR bands. Five-day revisit at the equator (two-satellite constellation). The red-edge bands at 705 nm and 740 nm are the primary discriminators between emergent and floating-leaf communities. Free and open archive from 2015.
- Landsat 8/9 OLI: 30 m multispectral resolution with a 16-day revisit per satellite (8-day combined). Coarser than Sentinel-2 for small lake patches but provides an archive back to 1984 (Landsat 5 TM) for long-term change baselines. Coastal/Aerosol band at 443 nm aids water-column depth estimation in optically shallow pixels.
- Planet SuperDove: 3 m resolution, 8 spectral bands including red-edge at 705 nm, near-daily revisit globally. Commercially tasked. Resolves individual macrophyte patches smaller than 10 m, which Sentinel-2 cannot. Cloud cover remains the same limiting factor; no archive depth comparable to Landsat.
- Copernicus Marine Service (CMEMS) inland water products: Provides atmospherically corrected water-leaving reflectance and turbidity/chlorophyll products derived from Sentinel-2 and Sentinel-3 for selected inland water bodies. Turbidity estimates from these products are the practical gate for deciding whether submerged-vegetation retrieval is worth attempting on any given acquisition.
What the water column does to your signal
Mapping macrophytes from space is a water-optics problem before it is a classification problem. Photons that reach a submerged canopy must travel down through the water column, reflect off the plant tissue, and travel back up through the same column before reaching the sensor. Every extra 10 cm of depth attenuates the red and NIR wavelengths that carry vegetation information. In optically deep or turbid water, the bottom signal disappears entirely into the water-leaving radiance of the water itself.
The practical threshold is roughly 5 NTU of turbidity. Below that, submerged vegetation in water shallower than about 2 to 3 m can produce a detectable spectral contrast in Sentinel-2 blue and green bands, particularly when the water column is corrected using published shallow-water radiative-transfer approaches such as the Lee et al. semi-analytical model. Above 5 NTU, the signal is ambiguous: you may be detecting suspended sediment, phytoplankton or macrophyte canopy, and no spectral unmixing will resolve the ambiguity with confidence. This is not a processing failure; it is physics.
How red-edge bands separate the community types
Emergent macrophytes, the reed beds and bulrush stands that grow above the waterline, behave spectrally like terrestrial vegetation. They show strong NIR reflectance and a pronounced red-edge inflection between 680 nm and 740 nm. Floating-leaf communities such as water lilies also show NIR reflectance but with a different red-edge slope because their lower leaf surfaces are in contact with water rather than air, suppressing internal scattering slightly. Submerged vegetation, where detectable, appears as a muted green signal in the blue and green bands with little or no NIR return, because water absorbs NIR almost completely before it reaches the canopy.
Sentinel-2 Band 5 (705 nm) and Band 6 (740 nm) are the operative discriminators. A simple red-edge chlorophyll index, (B7/B5) minus 1, separates emergent stands from open water and from floating-leaf zones with reasonable reliability in clear-water lakes. Landsat OLI lacks dedicated red-edge bands, so community discrimination there relies on NIR/Red ratios and tasselled-cap wetness, which conflate floating-leaf and emergent classes more often. Planet SuperDove's red-edge band at 705 nm gives it discrimination power comparable to Sentinel-2 but at 3 m, resolving narrow fringing reed beds that are sub-pixel in Sentinel-2 imagery.
Atmospheric correction is not optional here
Inland water bodies are among the most demanding targets for atmospheric correction. The water-leaving radiance in the NIR is near zero for deep, clear water, which means even small aerosol residuals create large relative errors in the bands most important for macrophyte mapping. The standard Sen2Cor processor, designed for land surfaces, tends to over-correct over dark water pixels. The ACOLITE processor, developed specifically for aquatic applications and published by the Royal Belgian Institute of Natural Sciences, performs better for inland water targets and is widely used in the published literature.
Sun-glint is a separate problem. Specular reflection from the water surface can saturate pixels in NIR bands, mimicking emergent vegetation. Glint correction requires either a near-infrared dark-pixel subtraction or a physical glint model using wind-speed estimates. Acquisitions taken at low solar zenith angles in summer, when solar elevation is high, are most affected. Scheduling imagery for early morning or late afternoon passes, where the geometry is less favourable for glint, reduces the problem but does not eliminate it entirely.
Building a change-detection record that holds up
A single-date macrophyte map is of limited conservation value. What managers need is a time series that shows whether reed beds are advancing into open water, whether submerged meadows are contracting under eutrophication pressure, or whether an invasive species such as water hyacinth is colonising new areas. Sentinel-2's archive from 2015 and Landsat's archive from 1984 make multi-decadal baselines possible, but only if the classification is applied consistently across dates.
Phenology complicates this. Emergent macrophytes reach peak biomass in mid-summer; floating-leaf communities may be absent in early spring. A change-detection workflow must compare like-season imagery, not arbitrary acquisition dates. The standard approach is to build seasonal composites, typically a peak-growing-season median from June to August in the northern hemisphere, and then classify each annual composite using a consistent spectral decision tree or supervised classifier trained on stable reference polygons. Change between annual composites is then attributable to real vegetation dynamics rather than phenological noise.
Cloud cover is the operational enemy. Many temperate and tropical lakes have fewer than 10 cloud-free Sentinel-2 acquisitions per growing season at useful solar angles. In persistently cloudy regions, annual composites may be based on only two or three scenes, which introduces sampling uncertainty into the change signal. Honest reporting requires stating the number of cloud-free observations that contributed to each annual composite.
What the method cannot do
Species-level discrimination is generally beyond what 10 m multispectral imagery can deliver. Sentinel-2 can separate community functional types (emergent, floating-leaf, submerged) but cannot reliably distinguish common reed from reedmace, or pondweed from hornwort. Hyperspectral imagery from airborne sensors or from future spaceborne systems can narrow that gap, but it is outside the scope of current open-constellation workflows.
Small lakes below roughly 1 to 2 hectares in area are problematic at 10 m resolution because mixed pixels dominate the shoreline. Planet SuperDove at 3 m extends the minimum mappable lake size considerably, but at commercial tasking cost. Depth estimation for submerged vegetation requires field-collected bathymetric data or a calibrated radiative-transfer inversion; satellite imagery alone cannot provide it reliably.
Satellize applies these workflows on open Sentinel and Landsat archives and adds commercial Planet tasking where patch-scale precision matters, as it does in the Tonga crop-estimation programme, where small field polygons set the resolution requirement. The same logic applies to small lake systems.
From pixel to conservation decision
The output that matters to a lake manager is not a classified raster. It is an area estimate with uncertainty bounds, a trend line across years, and an alert when a defined threshold, say a 15% reduction in submerged meadow extent, is crossed in a single season. That requires the classification to be accompanied by a per-class accuracy assessment using independent validation points, an area-adjusted estimate following the Olofsson et al. good-practice framework published in Remote Sensing of Environment, and a documented decision rule for what triggers a management response.
Delivering that chain from raw imagery to decision-ready output is where the analytical work sits. The satellite physics sets the ceiling on what is knowable. The processing chain, the validation, and the honest communication of uncertainty determine whether the output is actually used.
Typical figures
| Spatial resolution (primary sensor) | 10 m (Sentinel-2 visible and NIR); 20 m (Sentinel-2 red-edge and SWIR); 30 m (Landsat OLI); 3 m (Planet SuperDove) |
| Revisit frequency | 5 days (Sentinel-2, two-satellite); 8 days combined (Landsat 8+9); near-daily (Planet SuperDove, commercial) |
| Key spectral bands | Blue 490 nm (water-column penetration); Green 560 nm; Red 665 nm; Red-edge 705 nm and 740 nm (community discrimination); NIR 842 nm (emergent biomass); SWIR 1610 nm (water/land boundary) |
| Turbidity limit for submerged vegetation | Approximately 5 NTU; above this threshold submerged canopy signal is not reliably separable from water-column scattering |
| Minimum mappable lake area | Approximately 1 to 2 hectares at 10 m (Sentinel-2); approximately 0.1 hectares at 3 m (Planet SuperDove) |
| Minimum mappable patch width | Approximately 20 to 30 m for reliable classification at 10 m resolution (two to three pixel minimum) |
| Archive depth | Sentinel-2 from 2015; Landsat from 1984 (Landsat 5 TM); Planet SuperDove from approximately 2021 |
| Atmospheric correction (aquatic) | ACOLITE or equivalent aquatic-optimised processor; Sen2Cor not recommended over dark inland water |
| Typical processing latency | 2 to 5 days from acquisition for standard open-archive products; same-day possible with direct-access pipelines |
| Delivery formats | GeoTIFF classified raster, vector polygon GIS layer (GeoPackage or Shapefile), area-summary CSV with uncertainty bounds, PDF monitoring report |
Analytics Satellize can run
| Macrophyte community-type map | Supervised classification (random forest or support vector machine) on Sentinel-2 red-edge and NIR bands with ACOLITE-corrected surface reflectance; training labels from field survey or high-resolution imagery | Annual GeoTIFF and vector polygon layer with emergent, floating-leaf, submerged and open-water classes; per-class area table with confidence intervals |
| Multi-year change-detection time series | Peak-season median compositing across Sentinel-2 and Landsat archives; consistent classifier applied to each annual composite; area-adjusted accuracy assessment following Olofsson et al. good-practice protocol | Time-series chart of class area by year (1984 to present where Landsat used, 2015 to present for Sentinel-2); change-magnitude GeoTIFF highlighting gain and loss polygons |
| Turbidity screening and data-quality flag | Turbidity retrieval from Copernicus Marine Service inland water products or ACOLITE-derived suspended particulate matter index; pixel-level flag masking acquisitions above 5 NTU threshold | Per-acquisition quality report stating usable pixel fraction and turbidity distribution; flagged raster included in all classified outputs |
| Submerged vegetation extent (clear-water lakes only) | Semi-analytical radiative-transfer inversion (Lee et al. model class) applied to atmospherically corrected blue and green bands in pixels with turbidity below threshold and estimated depth below 3 m | Submerged canopy probability raster with explicit depth and turbidity caveats; suitable for lakes with field-validated bathymetry |
| Seasonal phenology profile | Red-edge chlorophyll index time series extracted from all cloud-free Sentinel-2 acquisitions per season; harmonic regression to characterise growing-season onset, peak and senescence dates | Per-community-type phenology curves; inter-annual comparison table flagging anomalous seasons |
| Threshold-breach alert | Automated comparison of current-season classified area against user-defined baseline and alert threshold; triggered when class area change exceeds agreed percentage in a single season | Email or API alert with supporting map excerpt and area statistics; recommended for monitoring sites under active conservation management plans |
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.