Critical infrastructure flood exposure from LiDAR-fused elevation models

Coarse global DEMs misplace flood risk for hospitals, substations and water-treatment plants by metres. Fusing airborne LiDAR bare-earth models with SAR-derived flood extents narrows that uncertainty to operationally useful bounds.

Sensors

Airborne LiDAR (national survey datasets): Point densities of 1–20 pts/m² yield bare-earth DEMs at 0.25–1 m spatial resolution with vertical RMSE typically 5–15 cm over open ground. Accuracy degrades under dense canopy. UK Environment Agency, USGS 3DEP and several European national datasets are publicly accessible.
Sentinel-1 SAR (C-band, ESA): 6–12 day repeat at mid-latitudes in Interferometric Wide Swath mode; 10 m ground range resolution. Penetrates cloud and rain, making it the primary source for near-real-time flood extent mapping. Backscatter thresholding and change detection against pre-event composites are the standard published methods.
ICEYE SAR (X-band commercial): Sub-metre spotlight mode and hourly revisit possible through constellation tasking. Useful for capturing rapidly evolving inundation on a specific asset when Sentinel-1's 6-day repeat is too slow. Stripmap mode delivers 3 m resolution over wide areas.
TanDEM-X (X-band, DLR/Airbus): Global DEM at 12 m posting with absolute vertical accuracy quoted at 10 m (90th percentile) and relative accuracy of about 2 m over flat terrain. A degraded but globally consistent fallback where LiDAR is absent. The 2 m relative figure is the honest planning floor for infrastructure risk work.
CoastalDEM (regression-corrected SRTM derivative): Published by Climate Central; reduces SRTM's systematic positive bias in coastal urban areas from several metres to roughly 0.4–1.1 m RMSE in validation studies. A meaningful improvement over raw SRTM for coastal assets, but still insufficient for centimetre-scale threshold decisions without independent ground truth.

Why a 2-metre DEM is the wrong tool for a 30-centimetre problem

The global DEMs most agencies reach for first, SRTM and its derivatives, carry vertical errors that are not random noise. In urban areas, SRTM systematically captures rooftop and canopy heights rather than bare earth, inflating ground elevations by 2–8 m in dense cities. That bias is not uniformly distributed: it clusters around the very structures, hospitals, substations, treatment works, that planners most need to assess correctly.

A water-treatment plant sitting 0.4 m above the modelled 1-in-100-year flood surface is either safe or catastrophically exposed depending on whether that 0.4 m figure is real or an artefact of the DEM. Getting it wrong in one direction means unnecessary evacuation of irreplaceable staff. Getting it wrong in the other means a city loses clean water mid-disaster. The vertical accuracy of the elevation model is not a technical footnote; it is the decision.

What airborne LiDAR actually delivers, and where it falls short

Airborne LiDAR surveys flown at 1–4 pts/m² produce bare-earth DEMs with vertical RMSE of 5–15 cm over open, hard surfaces after ground-filtering. That is the accuracy class that makes centimetre-threshold decisions meaningful. Several national programmes now publish these datasets openly: the UK Environment Agency's 1 m composite LiDAR covers most of England; USGS 3DEP targets 1 m resolution nationally; the Netherlands AHN4 product reaches 0.5 m. Coverage is patchy across lower-income countries, which is precisely where disaster resilience investment is most urgent.

The honest limits matter. Vertical accuracy degrades to 20–50 cm under dense tree canopy because ground-return filtering is imperfect. Surveys age: a LiDAR dataset flown before major construction or land reclamation will misrepresent current ground levels. And the bare-earth model says nothing about the flood water; it must be combined with a separate inundation surface derived from SAR or hydraulic modelling to produce a risk output.

Fusing LiDAR elevations with SAR flood extents

The operational workflow has two inputs. First, a flood extent polygon, typically derived from Sentinel-1 backscatter thresholding or change detection against a pre-event composite, giving a binary wet/dry boundary at roughly 10 m resolution. Second, the LiDAR bare-earth DEM clipped to the area of interest. The fusion step intersects the SAR-derived water edge with the DEM to extract the water surface elevation at that boundary, then propagates that elevation surface inland using a connected-component or hydraulic fill algorithm. The result is an inundation depth raster at the spatial resolution of the LiDAR, typically 1 m.

Each infrastructure asset is then queried against that depth raster. A hospital's lowest finished floor elevation, taken from the LiDAR or from building survey data, is compared to the modelled water surface. The output is not a binary flooded/not-flooded flag but a probability surface that carries the vertical uncertainty of both inputs explicitly: the LiDAR RMSE and the SAR edge localisation error, which is typically half a pixel, or about 5 m for Sentinel-1. Planners who see a single deterministic map without those uncertainty bounds are looking at false precision.

ICEYE's hourly revisit capability adds a temporal dimension that Sentinel-1 alone cannot provide during fast-rising events. Repeated X-band acquisitions over a 12-hour flood peak allow the inundation boundary to be tracked in near-real time, updating the depth surface as water levels change rather than relying on a single snapshot.

When LiDAR is absent: TanDEM-X and CoastalDEM as honest substitutes

For the majority of the world's critical infrastructure, particularly across South and Southeast Asia, sub-Saharan Africa and Pacific island states, airborne LiDAR surveys either do not exist or are not publicly accessible. The fallback hierarchy is TanDEM-X at 12 m posting, then CoastalDEM for coastal assets, then SRTM as a last resort.

TanDEM-X's relative vertical accuracy of approximately 2 m over flat terrain means that any infrastructure asset within 2 m of the modelled flood threshold must be treated as genuinely ambiguous. That is not a failure of the analysis; it is the correct characterisation of the risk. Planners who receive a TanDEM-X-based inundation map should be told explicitly: assets flagged as marginally safe may not be. CoastalDEM's regression correction reduces SRTM's urban bias meaningfully, but published validation puts its RMSE at 0.4–1.1 m, still an order of magnitude coarser than LiDAR. For a substation 0.5 m above the modelled flood line, that uncertainty swallows the margin entirely.

The practical recommendation for governments without national LiDAR coverage is to commission targeted airborne surveys over the specific asset clusters that matter most. A survey corridor covering 50 km² around a city's critical infrastructure nodes costs a fraction of a post-disaster restoration bill and produces a dataset that remains useful for 10–20 years of infrastructure planning.

Producing an inundation probability surface planners can actually use

The deliverable that matters is not a map but a table: each critical infrastructure asset, its modelled flood depth under a specified return period, and the vertical uncertainty band around that depth figure. A hospital at +0.8 m ± 0.12 m (LiDAR-derived) is a different risk conversation from one at +0.8 m ± 2.1 m (TanDEM-X-derived). Both can be expressed honestly in the same GIS layer using confidence interval attributes.

Satellize structures this analysis as an asset-level exposure register, cross-referenced to the SAR-derived flood extent from the event or scenario in question. For the analyst, the key output is a ranked list of assets by inundation probability, with the DEM source and its associated uncertainty documented for each record. That provenance matters when decisions are contested.

Return-period scenarios, whether the 1-in-20, 1-in-100 or 1-in-500 year flood, are typically sourced from national hydrological agencies or global datasets such as the JRC Global Flood Database. The elevation model determines whether an asset sits above or below the modelled water surface for each scenario; the SAR data validates the model against observed events and grounds the uncertainty estimate in real acquisition geometry rather than theoretical sensor specs.

Limits that responsible analysis must state

SAR-derived flood extents have known failure modes. Dense urban canyons produce double-bounce returns that can mask open water. Slow-moving or shallow water under vegetation may not register as flooded. Sentinel-1's 10 m resolution means that narrow drainage channels and small courtyards are below the detection threshold. These gaps affect the accuracy of the water surface elevation extracted at the flood boundary, which in turn propagates through the entire depth calculation.

LiDAR surveys capture a moment in time. Ground subsidence, particularly in cities built on deltaic sediments or where groundwater extraction is heavy, can shift actual ground elevations by centimetres to decimetres per decade. Jakarta, Ho Chi Minh City and parts of coastal Bangladesh are documented examples. A LiDAR dataset more than five years old in a subsiding area should be treated with the same scepticism as a coarser DEM until its vintage is confirmed against recent ground control.

None of this makes the analysis unusable. It makes the uncertainty explicit, which is what separates a useful risk product from a confident-looking map that misleads the people who act on it.

Typical figures

LiDAR bare-earth vertical RMSE (open ground)	5–15 cm (national survey datasets at 1–4 pts/m²)
TanDEM-X vertical accuracy (flat terrain, relative)	~2 m; absolute 90th-percentile spec 10 m
CoastalDEM vertical RMSE (coastal urban)	0.4–1.1 m (published validation range)
Sentinel-1 spatial resolution (IW mode)	10 m ground range; flood edge localisation ±5 m
Sentinel-1 revisit (mid-latitudes)	6 days single satellite; 3–4 days with Sentinel-1A and -1B combined
ICEYE SAR resolution (Stripmap / Spotlight)	3 m / sub-1 m; tasked revisit potentially hourly over a target
Inundation depth raster output resolution	Matches LiDAR input, typically 0.5–1 m; 12 m where TanDEM-X is used
Minimum detectable water depth (SAR-LiDAR fusion)	No reliable lower bound below ~10 cm; shallow sheet flow is frequently missed
LiDAR archive depth (national datasets)	UK EA from ~2000; USGS 3DEP from ~2012; coverage and vintage vary by region
Delivery format	GeoTIFF depth raster, GeoPackage asset register with uncertainty attributes, PDF scenario summary

Analytics Satellize can run

Asset-level inundation probability register	LiDAR DEM intersection with SAR flood extent polygon; uncertainty propagation from both input RMSEs	GeoPackage table: each asset, modelled depth, vertical uncertainty band, DEM source, return-period scenario
Flood depth raster at LiDAR resolution	Water surface elevation extracted at SAR flood boundary; propagated inland by connected-component fill over bare-earth DEM	GeoTIFF at 0.5–1 m resolution with per-pixel depth and confidence interval bands
Degraded-DEM risk flags (TanDEM-X or CoastalDEM zones)	Automatic detection of assets within vertical uncertainty band of flood threshold; flagged as ambiguous rather than safe or exposed	Separate GIS layer of ambiguous assets with recommended LiDAR survey priority score
Temporal inundation tracking during event	Multi-pass ICEYE or Sentinel-1 SAR change detection; flood boundary updated per acquisition; depth surface recalculated each pass	Time-stamped GeoTIFF series and alert feed showing which assets cross inundation threshold as event evolves
Return-period scenario comparison	JRC Global Flood Database or national hydrological model water surfaces intersected with LiDAR DEM for 1-in-20, 1-in-100, 1-in-500 year events	Three-scenario PDF report with ranked asset exposure tables and map panels per scenario
Subsidence-adjusted exposure update	Published InSAR subsidence rates (e.g. from Sentinel-1 time series) applied as vertical correction to aged LiDAR datasets in known subsiding areas	Corrected DEM layer with subsidence adjustment documented; revised asset register

Sources

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.