Floodplain vegetation roughness mapping for hydraulic model parameterisation
Spatially distributed Manning's n values are the weakest link in most 2-D flood models. Combining TanDEM-X canopy height, GEDI waveform structure and Sentinel-1 SAR backscatter cuts the number of hand-drawn roughness zones without replacing field calibration.
Sensors
- TanDEM-X WorldDEM: Bistatic X-band InSAR digital surface model at 12 m posting (standard) and 6 m (high-resolution product). Phase-coherent canopy surface height above bare-earth DEM gives a first-order canopy-height proxy, though X-band penetrates only the uppermost canopy layer; errors over dense forest can reach 3-5 m relative to ground truth.
- GEDI L2A / L2B: NASA spaceborne full-waveform lidar on the ISS. Footprint diameter approximately 25 m, along-track spacing 60 m, cross-track spacing roughly 600 m between beams. L2A provides relative height metrics (RH25, RH50, RH98); L2B provides plant area index and canopy cover fraction. Sparse sampling means GEDI is best used to calibrate wall-to-wall DSM-derived height, not as a standalone roughness layer. ISS precession limits repeat coverage to approximately 25-day cycles at mid-latitudes.
- Sentinel-1 SAR (C-band): 6-day repeat at mid-latitudes (12-day for a single satellite). IW mode at 10 m ground range resolution. VV and VH polarisations respond differently to vegetation structure: VH backscatter increases with canopy volume and is useful for separating bare floodplain from shrub and forest. Inundation under canopy can suppress backscatter (double-bounce) or enhance it depending on geometry, which must be accounted for when building roughness classes near the channel edge.
- Sentinel-2 MSI: 10 m visible and near-infrared bands, 20 m red-edge and shortwave infrared. NDVI, NDWI and red-edge chlorophyll indices stratify herbaceous, shrub and tree cover and help resolve class boundaries that SAR alone conflates. Cloud contamination is the main operational constraint; a multi-temporal composite over a dry-season window is usually required to get clean floodplain imagery.
Why Manning's n is the number modellers argue about most
A 2-D hydrodynamic model solves the shallow-water equations across a gridded floodplain. Every cell needs a Manning's roughness coefficient. Bare gravel might take n = 0.025; dense riparian forest can sit at 0.10 or higher. The difference is not cosmetic. A misclassified stand of willows in the conveyance zone can shift peak water-surface elevation by tens of centimetres and move the inundation boundary by hundreds of metres laterally.
Traditionally, modellers draw roughness polygons by hand from aerial photography and field visits, then calibrate against gauged flood events. That works on well-instrumented rivers in wealthy countries. On ungauged reaches, or where riparian vegetation has changed since the last survey, the hand-drawn zones carry hidden uncertainty that the model happily propagates into its output. Satellite-derived roughness maps do not eliminate that uncertainty, but they make it spatially explicit and traceable.
What a floating roof gives away: reading vegetation structure from orbit
TanDEM-X measures the phase difference between two radar pulses separated by a known baseline. The result is a digital surface model that records the top of whatever is there, whether bare ground, shrub canopy or forest crown. Subtracting a bare-earth DTM (derived from the same data using filtering, or from an independent airborne survey) yields a canopy height model. At 12 m posting this is sufficient to separate bare floodplain (height < 0.5 m), herbaceous and low shrub (0.5 to 3 m), tall shrub (3 to 8 m) and riparian forest (> 8 m), which maps reasonably well onto the four broad Manning's n bands used in most 2-D model setups.
GEDI fills the structural gap that X-band InSAR leaves open. Because X-band barely penetrates a closed canopy, TanDEM-X height over dense forest underestimates true canopy height. GEDI's full waveform captures energy returned from multiple vertical layers, so its RH98 metric (the height below which 98 % of returned energy falls) is a reliable canopy-top estimate. The practical workflow is to regress GEDI RH98 against TanDEM-X DSM minus DTM across a sample of footprints, then apply the correction spatially. Studies on temperate riparian corridors have shown this reduces canopy-height RMSE from around 4 m to under 2 m.
Sentinel-1 and Sentinel-2 as the wall-to-wall classifier
GEDI and TanDEM-X provide height. Sentinel-1 and Sentinel-2 provide the dense, repeatable spatial coverage that turns point and strip measurements into a continuous map. A multi-temporal Sentinel-1 VH backscatter stack captures seasonal phenology: deciduous riparian forest shows a pronounced winter-to-summer backscatter increase that separates it from evergreen scrub or bare gravel, which remain relatively stable. Sentinel-2 NDVI and red-edge indices add biochemical information that SAR cannot see.
A supervised classification combining these four inputs, trained on GEDI-corrected height strata and validated against field-measured vegetation plots, typically achieves overall accuracies of 80 to 90 % across four to six land-cover classes on temperate floodplains. Accuracy drops in narrow riparian corridors where mixed pixels dominate at 10 m resolution, and in tropical systems where cloud cover limits clean Sentinel-2 composites. Those limits matter for the hydraulic modeller: a 10 m pixel straddling the boundary between bare gravel and dense willow carries real ambiguity that should be flagged rather than silently assigned to one class.
From land-cover class to Manning's n: the translation layer
The classification output is a raster of vegetation classes. Converting that to Manning's n requires a lookup table anchored to published hydraulic literature. Chow's 1959 tables remain the standard reference for base n values by vegetation type; subsequent work by Arcement and Schneider (US Geological Survey) extended this to forested floodplains with explicit correction factors for stem density, surface irregularity and channel obstruction. These corrections can themselves be estimated from GEDI plant area index (L2B) and from Sentinel-1 backscatter as a proxy for stem-area density.
The result is a spatially distributed n raster that can be resampled to any 2-D model grid. For a LISFLOOD-FP or HEC-RAS 2-D setup at 10 to 25 m resolution, the satellite-derived map typically reduces the number of manually defined roughness zones from several dozen to fewer than ten, with the remaining manual zones concentrated at hydraulic structures and channel cross-sections where satellite data cannot resolve sub-grid geometry. That is the honest scope of what remote sensing contributes: it handles the floodplain; the channel itself still needs survey.
Honest limits: what the method cannot see
Cloud is the most immediate constraint. A single-date Sentinel-2 scene over a tropical floodplain is frequently unusable; multi-month compositing is standard practice but introduces phenological mixing. Sentinel-1 is cloud-independent but cannot distinguish species within a structural class, which matters where hydraulic roughness differs between, say, reed beds and sedge marsh at similar canopy heights.
Spatial resolution sets a hard floor. At 10 m, isolated trees and narrow levee-top hedgerows are subpixel features that the classification misses entirely. Airborne lidar at 0.5 to 1 m point density resolves these, but its cost and coverage limits mean it is rarely available for reaches longer than a few kilometres. TanDEM-X at 6 m (the high-resolution product, available on commercial licence) helps at the margins but does not close the gap. Finally, Manning's n is not a fixed physical constant; it varies with flow depth and seasonal vegetation state. A single roughness map, however accurate, is a snapshot. Hydraulic models that run multi-season scenarios should ideally use seasonally updated roughness layers, which is operationally feasible with Sentinel-1 and Sentinel-2 but adds processing overhead.
Satellize applies this multi-sensor classification workflow on open Copernicus data combined with GEDI and commercial TanDEM-X products, producing GIS-ready roughness rasters formatted for direct ingestion into LISFLOOD-FP, HEC-RAS 2-D and similar solvers. The approach is the same one underpinning the spatial data layers in our Tonga crop-estimation programme, where multi-sensor fusion at the field scale required the same kind of honest class-boundary uncertainty accounting.
Putting the map to work: calibration still earns its keep
A satellite-derived roughness map is a prior, not a posterior. It should enter the hydraulic model as a spatially structured starting point for calibration, not as a fixed parameter set. Where a gauged flood event exists, the modeller can adjust n values within physically plausible bounds for each class rather than treating the entire floodplain as a single calibration zone. That structured approach to calibration is more defensible in regulatory submissions and more transferable to adjacent ungauged reaches than a single global n adjustment.
The practical gain is speed and auditability. A river authority managing several hundred kilometres of floodplain can generate a first-pass roughness map for any reach within days rather than months, update it after a major vegetation-clearing event or flood-deposited sediment pulse, and document the remote-sensing basis for each roughness zone in a way that a hand-drawn polygon cannot. That is a modest but real improvement in how flood risk is communicated to the people who live on floodplains.
Typical figures
| TanDEM-X WorldDEM spatial resolution | 12 m standard; 6 m high-resolution product (commercial licence) |
| GEDI footprint diameter / along-track spacing | ~25 m / 60 m; cross-track beam separation ~600 m |
| Sentinel-1 IW mode resolution | 10 m ground range; 6-day repeat at mid-latitudes (single-satellite 12-day) |
| Sentinel-2 MSI resolution | 10 m (VIS/NIR); 20 m (red-edge, SWIR); 5-day repeat at equator |
| Canopy height accuracy (GEDI-corrected TanDEM-X) | RMSE typically 1.5–2.5 m over temperate riparian forest; degrades over dense tropical canopy |
| Land-cover classification accuracy (4–6 classes) | 80–90 % overall accuracy on temperate floodplains; lower in narrow corridors and cloud-affected tropical scenes |
| Minimum resolvable vegetation patch | ~30 m diameter at Sentinel-1/2 resolution; sub-pixel features require airborne lidar |
| TanDEM-X archive depth | Global coverage acquired 2010–2015; high-resolution updates available commercially |
| GEDI operational period | April 2019 to present (ISS-mounted; subject to mission extension decisions) |
| Output delivery format | GeoTIFF roughness raster, GeoPackage class polygons, HEC-RAS / LISFLOOD-FP compatible land-cover tables |
Analytics Satellize can run
| Vegetation-structure canopy height model | GEDI RH98 regression against TanDEM-X DSM minus DTM difference; spatial interpolation of correction surface | GeoTIFF canopy height raster at 10–12 m, with per-pixel height uncertainty band |
| Multi-temporal Sentinel-1 VH backscatter composite | Radiometric terrain correction, speckle filtering (Lee or refined Lee), seasonal median stack | GeoTIFF backscatter stack (dry season / wet season) as input to classifier |
| Floodplain land-cover classification | Random forest or gradient-boosted classifier trained on GEDI height strata and Sentinel-1/2 multi-temporal features; validated against field plots or high-resolution imagery | GeoTIFF and GeoPackage: 4–6 vegetation classes with per-class confidence score |
| Spatially distributed Manning's n raster | Class-to-n lookup anchored to Chow (1959) and Arcement & Schneider USGS tables; GEDI plant area index used to modulate within-class n variation | GeoTIFF roughness raster resampled to client model grid; companion CSV of class n ranges and literature basis |
| Roughness change detection layer | Bi-temporal classification differencing between baseline and post-event Sentinel-1/2 composites to flag vegetation removal, flood-deposited sediment or regrowth | Change polygon GeoPackage with class-transition labels; suitable for triggering roughness-map update in operational model chains |
| Hydraulic model input package | Reprojection, resampling and format conversion to target model coordinate system; manual roughness zones for channel and structures flagged as requiring field input | Model-ready land-use file (HEC-RAS .hdf or LISFLOOD-FP .asc) with documented uncertainty notes per zone |
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.