Flood water depth estimation by DEM differencing
Combining SAR-derived flood extents with high-resolution elevation models lets analysts estimate water depth and inundated volume, but DEM vertical error sets a hard floor on what the method can honestly claim.
Sensors
- Sentinel-1 A/B (C-band SAR): Provides flood-extent polygons at 10–20 m ground resolution with 6-day repeat at mid-latitudes (12-day with one satellite). Cloud-penetrating, day/night capable. Backscatter contrast between open water and dry land typically exceeds 5 dB, enabling reliable binary inundation masks that feed the differencing workflow.
- Copernicus DEM GLO-30: Global 30 m horizontal posting, derived from TanDEM-X interferometric SAR. Absolute vertical accuracy (LE90) quoted at approximately 4 m over flat terrain and better over open land; in practice, relative accuracy within a small catchment is often 1–2 m. Freely available under open licence.
- TanDEM-X (bistatic X-band InSAR): The source data for GLO-30. The full-resolution 12 m DEM has a relative height error specification of 2 m (LE90) and absolute vertical accuracy of 10 m (LE90). Commercial 12 m tiles can be licensed for higher-precision work where the 30 m product is insufficient.
- ALOS-2 PALSAR-2 (L-band SAR): L-band penetrates partial vegetation canopy, improving flood detection under crops and sparse forest where C-band may underestimate inundated area. Spotlight mode reaches 3 m resolution. Revisit is approximately 14 days, which limits its role to post-event snapshots rather than rapid onset mapping.
- Airborne or UAV lidar (locally derived DSM/DTM): Where available, lidar-derived bare-earth models reach vertical accuracy of 0.1–0.25 m RMSE, reducing depth uncertainty by an order of magnitude compared with TanDEM-X. Coverage is limited and pre-event acquisition is rarely guaranteed, but national mapping agencies in flood-prone countries increasingly hold lidar archives.
The arithmetic is simple. The error budget is not.
Flood water depth at any pixel is, in principle, the local water surface elevation minus the bare-earth terrain elevation beneath it. Estimate the water surface elevation from the flood perimeter intersected with the DEM (assuming a locally flat water surface within a hydraulic reach), subtract the terrain, and you have a depth raster. Volume follows by integration.
The problem is that every term in that subtraction carries uncertainty. The DEM has a vertical error. The flood-extent polygon has an edge uncertainty of at least one pixel width (10–20 m for Sentinel-1). The assumption of a flat water surface breaks down where hydraulic gradients are steep or where multiple disconnected water bodies are merged into one polygon. In flat deltaic terrain, where depths of 0.3–0.8 m are operationally significant for evacuation decisions, those errors can exceed the quantity being measured. Stating that plainly is not a weakness in the method; it is the starting point for using it correctly.
What each DEM tier can and cannot resolve
The Copernicus DEM GLO-30 is the practical default for global flood response. Its 30 m posting and approximately 1–4 m relative vertical accuracy (depending on terrain slope and land cover) mean that depth estimates below roughly 1.5–2 m should be treated as order-of-magnitude indicators rather than precise measurements. Above that threshold, the method produces useful relative depth patterns even if absolute values carry uncertainty bands that must be communicated to end users.
The TanDEM-X 12 m commercial product improves horizontal detail but not dramatically on vertical accuracy in flat terrain, where the limiting factor is the radar's sensitivity to vegetation and surface roughness rather than the interferometric baseline. Lidar changes the picture entirely. A national lidar DTM with 0.15 m vertical RMSE reduces depth uncertainty to a level where the SAR-derived flood mask becomes the dominant error source, not the elevation model. Countries investing in pre-event lidar acquisition, as the Netherlands and parts of the United Kingdom have done, gain a qualitatively different analytical capability when flooding occurs.
SRTM, the predecessor to TanDEM-X products, records the first-return surface rather than bare earth, inflating apparent terrain elevation under vegetation. Using SRTM for depth estimation in forested or crop-covered floodplains introduces systematic underestimation of depth. GLO-30 is preferable in almost every case where SRTM might otherwise be reached for.
Propagating error honestly into a depth product
A defensible depth raster should carry a per-pixel uncertainty estimate, not just a single value. The standard approach combines DEM vertical error (taken from the product's published accuracy specification or a locally validated figure), flood-mask edge uncertainty (one half-pixel at the perimeter, zero in the interior), and water-surface slope uncertainty (estimated from hydraulic reach length and assumed gradient bounds). These can be propagated in quadrature to produce a depth uncertainty layer alongside the depth estimate itself.
In practice, most rapid-response products omit this layer because it complicates visualisation. That is a mistake. A depth map showing 0.6 m in a zone where the combined uncertainty is plus or minus 1.2 m is not a depth map; it is a flood-extent map with spurious precision attached. Delivering the uncertainty layer, even as a simple low/medium/high confidence classification, is the difference between a product that informs decisions and one that misleads them.
Volume calculations and hydraulic model validation
Integrating a depth raster over the inundated area gives an estimated flood volume. This figure is useful for reservoir operators assessing downstream storage, for logistics planners estimating pumping requirements, and for post-event loss modellers correlating inundation depth with damage functions. Published damage-depth curves for residential buildings, for instance, show non-linear relationships where the difference between 0.5 m and 1.0 m of inundation translates to a substantial change in expected structural damage, which makes even approximate depth information valuable for insurance and reconstruction planning.
Hydraulic modellers use satellite-derived depth rasters differently: as boundary conditions or validation datasets for 1D and 2D models such as HEC-RAS or LISFLOOD-FP. The satellite product constrains the model's initial or boundary water surface elevation, and the model then fills in the physics of flow routing that remote sensing cannot observe. This hybrid approach, satellite extent plus DEM depth plus hydraulic model, consistently outperforms either satellite or model alone in studies published in journals such as Remote Sensing and Hydrology and Earth System Sciences.
Operational limits that matter in the field
Several failure modes are worth naming directly. First, urban areas: buildings occlude SAR backscatter and create layover artefacts, so flood extent under dense building stock is systematically underestimated, and depth estimates in those zones are unreliable regardless of DEM quality. Second, vegetated floodplains: C-band SAR underestimates inundation under tall crops or reed beds; L-band (ALOS-2) partially mitigates this but does not eliminate it. Third, rapidly receding floods: by the time a Sentinel-1 acquisition is processed and delivered, water may have retreated from shallow margins, causing the depth estimate to reflect a historical state rather than the current one. Revisit frequency is not a minor operational detail.
Finally, the method assumes that the DEM represents the terrain at the time of the flood. Post-event sedimentation, levee breaches, or channel avulsion can alter the terrain enough to invalidate the pre-event DEM. In geomorphically active systems, this is not a theoretical concern.
Satellize applies this workflow, including per-pixel uncertainty classification, on Copernicus open data and licensed commercial DEMs. The same analytical pipeline that underpins the Tonga crop-estimation programme handles raster algebra and uncertainty propagation at scale; flood depth is a different application of the same infrastructure.
Typical figures
| Flood mask spatial resolution | 10–20 m (Sentinel-1 IW mode); 3–10 m (ALOS-2 Spotlight/StripMap) |
| DEM horizontal posting (global) | 30 m (Copernicus GLO-30); 12 m (TanDEM-X commercial) |
| DEM vertical accuracy (LE90, flat terrain) | ~1–4 m relative (GLO-30); ~2 m relative (TanDEM-X 12 m); 0.1–0.25 m (lidar DTM) |
| Minimum reliably detectable depth | ~1.5–2 m with GLO-30; ~0.5–1 m with TanDEM-X 12 m; ~0.3 m with quality lidar |
| Sentinel-1 revisit (mid-latitudes) | 6 days (two satellites); 12 days (one satellite post-Sentinel-1B loss) |
| Processing latency (rapid response) | 3–12 hours from SAR acquisition to flood mask; depth raster adds 1–3 hours |
| Global DEM coverage | Copernicus GLO-30: 90% of land surface; TanDEM-X: near-global between 84°N and 84°S |
| Depth uncertainty output | Per-pixel RMSE or low/medium/high confidence class, depending on delivery format |
| Delivery formats | GeoTIFF depth raster, GeoJSON flood polygon, uncertainty layer, summary CSV of volume by administrative unit |
Analytics Satellize can run
| Flood depth raster with uncertainty layer | DEM differencing: water surface elevation from perimeter-DEM intersection minus bare-earth terrain; uncertainty propagated from published DEM LE90 and mask edge error | GeoTIFF depth grid (metres) plus co-registered confidence classification layer |
| Inundated volume estimate by catchment or administrative unit | Pixel-wise depth integration over flood mask area; aggregated by polygon overlay | CSV table of volume (cubic metres) and area (square kilometres) per zone, with uncertainty bounds |
| Hydraulic model boundary conditions | Water surface elevation profile extracted along channel centreline from DEM-perimeter intersection; formatted for HEC-RAS or LISFLOOD-FP input | Structured text file of cross-section water surface elevations and flood polygon for model domain boundary |
| Building-level depth exposure classification | Depth raster intersected with building footprint layer (OpenStreetMap or national cadastre); depth at centroid or maximum depth within footprint extracted | GeoJSON of building footprints attributed with estimated depth class (0–0.5 m, 0.5–1 m, 1–2 m, >2 m) and confidence flag |
| Multi-date depth change raster | Differencing of depth rasters from sequential SAR acquisitions over the flood event; identifies recession rate and persistent deep-water zones | Animated GeoTIFF time series and summary statistics of recession by zone |
| DEM suitability assessment for target area | Comparison of available DEM vertical accuracy specifications against expected flood depth range; flags zones where uncertainty exceeds depth and recommends alternative data sources | Written assessment report with coverage map and data-source recommendation |
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.