Impervious surface mapping for urban flood and runoff modelling

Sealed surfaces drive flash-flood risk, yet most cities lack current maps of their true impervious fraction. Spectral unmixing of Sentinel-2 and Landsat imagery, checked against SAR backscatter, produces impervious-fraction grids that feed directly into hydrological models.

Sensors

Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands, 20 m in red-edge and SWIR. Five-day revisit at the equator (two-satellite constellation). SWIR Band 11 (1610 nm) is the primary discriminator between impervious surfaces and bare soil; both look bright in VNIR but diverge sharply in SWIR. Free archive from 2015.
Landsat 8/9 OLI: 30 m multispectral resolution including SWIR1 (1565–1651 nm) and SWIR2 (2107–2294 nm), which anchor the V-I-S spectral unmixing model. Sixteen-day single-satellite revisit, eight days with both satellites combined. Archive extends to 1972 across the Landsat series, enabling long-run impervious growth analysis.
Sentinel-1 SAR (C-band): C-band synthetic aperture radar at 5.405 GHz, IW mode at 10 m range resolution. Smooth wet impervious surfaces produce specular reflection that sharply reduces backscatter relative to vegetated or rough soil surfaces. Useful as a post-rainfall discriminator when optical imagery is cloud-obscured, but SAR alone cannot reliably distinguish dry concrete from calm water.
Maxar WorldView-3: 31 cm panchromatic, 1.24 m multispectral, and 3.7 m SWIR (eight bands). The SWIR capability at sub-5 m resolution is uncommon and allows rooftop material classification at individual-building scale. On-demand tasking; costly per square kilometre. Best reserved for validation strips or high-value urban cores rather than city-wide mapping.

Why the V-I-S model is still the right starting point

Ridd's vegetation-impervious-soil (V-I-S) triangle, published in 1995, remains the conceptual backbone of urban land-cover remote sensing. The premise is simple: most urban pixels are mixtures of three end-members, and the spectral signature of each end-member is well characterised. Impervious surfaces cluster toward high reflectance in visible bands and moderate-to-low reflectance in SWIR. Vegetation sits at high near-infrared and low visible. Bare soil occupies the middle ground.

Spectral unmixing extracts the fractional contribution of each end-member per pixel rather than forcing a hard classification. A 10 m Sentinel-2 pixel in a suburban street might return 55 % impervious, 30 % vegetation and 15 % soil. That fraction, not a binary label, is what hydrological models actually need. SWMM and similar storm-water models accept impervious-fraction grids directly; feeding them a hard binary map introduces systematic error in runoff volume estimates.

The dry-soil problem: where VNIR bands mislead you

Dry concrete and dry compacted soil are nearly indistinguishable in the visible and near-infrared. Both are bright, both are spectrally flat, and both suppress the NDVI signal. A classifier trained only on VNIR bands will routinely misclassify unpaved car parks, construction sites, and bare agricultural fields as impervious surface, inflating runoff estimates for the wrong reasons.

SWIR bands resolve most of this ambiguity. Concrete and asphalt have relatively flat SWIR reflectance compared with soil minerals, which show pronounced absorption features tied to clay and iron-oxide content. Sentinel-2 Band 11 at 1610 nm and Band 12 at 2190 nm are the workhorses here. Landsat OLI's SWIR1 and SWIR2 perform the same function at 30 m. The honest caveat: some light-coloured roofing membranes and pale limestone pavements still confuse the classifier. Field-collected spectral libraries or WorldView-3 SWIR strips are the practical fix for high-stakes mapping.

What SAR adds after the rain stops

Smooth wet surfaces behave like mirrors to C-band radar. After rainfall, roads, car parks and flat rooftops return very low backscatter in Sentinel-1 imagery because incident energy scatters away from the sensor rather than back toward it. Vegetated surfaces, by contrast, remain rough at the 5.6 cm wavelength and return moderate-to-high backscatter. This contrast is exploitable.

The practical workflow is to run a change-detection pass on Sentinel-1 VV polarisation before and after a rain event, then intersect low-backscatter zones with candidate impervious areas from the optical unmixing. Agreement between the two methods increases confidence; disagreement flags pixels worth inspecting. The limitation is real: calm water bodies, smooth sand, and freshly rolled soil also go dark in SAR. The method works best as a consistency check, not as a standalone classifier. Cloud cover during rainfall events, which is precisely when you want optical data, is the operational reason SAR earns its place in the stack.

From fraction grid to flood model input

The output of spectral unmixing is a continuous raster, typically 10 m or 30 m cells, where each cell holds a value between 0 and 1 representing impervious fraction. SWMM's subcatchment parameterisation accepts this directly as the percent impervious parameter. Hydrological models that use the curve number method, such as HEC-HMS, translate impervious fraction to CN values through published look-up tables from the USDA Technical Release 55.

Accuracy matters more than it sounds. A 10-percentage-point error in impervious fraction across a 50-hectare catchment can shift peak discharge estimates by 15 to 25 percent in moderate-intensity events, based on published sensitivity analyses using SWMM. City drainage engineers sizing culverts and detention basins need the fraction grid to be current, since a suburb that was 40 % impervious a decade ago may now be 65 %. Sentinel-2's free five-day revisit makes annual updates affordable. The 10 m cell size is sufficient for catchments above roughly two hectares; smaller urban micro-catchments may need WorldView-3 input.

Honest limits of the method at city scale

Spectral unmixing is not infallible. Persistent cloud cover in tropical and monsoonal cities can reduce usable Sentinel-2 acquisitions to a handful per year, forcing analysts to composite imagery across months and accept that the fraction map represents an average condition rather than a single date. Sentinel-1 SAR partly compensates but cannot replace optical data for end-member separation.

Vertical complexity defeats 2D mapping. A multi-storey car park reads as impervious at roof level but conceals sealed surfaces on lower decks that also contribute to runoff through internal drainage. Green roofs present the inverse problem: a vegetated rooftop may be classified as pervious when its drainage layer is effectively impervious. These structures require building-level data from planning records or high-resolution 3D point clouds to model correctly. Satellite imagery alone will undercount sealed area in dense, vertically complex urban cores. Knowing where the method is weakest is how you decide where to spend money on validation.

Putting it to work: from commission to model-ready grid

A practical city-scale commission typically runs in three stages. First, a cloud-screened Sentinel-2 seasonal composite is built, end-members are selected from spectrally pure training pixels, and the unmixing is run to produce the initial fraction grid. Second, Sentinel-1 post-rain passes are used to flag high-uncertainty pixels for review. Third, a sample of WorldView-3 SWIR tiles over representative land-cover types validates the fraction estimates before delivery.

Satellize runs this workflow on open-constellation data with commercial tasking added for validation strips. The deliverable is a GeoTIFF impervious-fraction grid, a confidence layer, and a methodology note that drainage engineers can hand directly to their hydrological modellers. The Tonga crop-estimation programme established the same three-stage composite-unmix-validate pattern in a different spectral domain; the infrastructure transfers. If your city's drainage model is currently parameterised from a land-cover survey that is more than five years old, the gap between that map and current reality is probably already influencing your flood-risk conclusions.

Typical figures

Spatial resolution (optical)	10 m (Sentinel-2 VNIR), 20 m (Sentinel-2 SWIR), 30 m (Landsat OLI)
Spatial resolution (SAR)	10 m ground range (Sentinel-1 IW mode)
Revisit interval	5 days (Sentinel-2 two-satellite), 8 days (Landsat 8+9 combined), 6–12 days (Sentinel-1)
Key spectral bands for impervious discrimination	SWIR1 ~1610 nm, SWIR2 ~2190 nm (Sentinel-2 Bands 11/12; Landsat OLI Bands 6/7)
SAR frequency	C-band 5.405 GHz, VV and VH polarisation (Sentinel-1)
Minimum mapping unit	~100 m² at 10 m resolution; sub-pixel fractions recoverable via unmixing
Archive depth	Sentinel-2 from 2015; Landsat from 1972 (OLI from 2013); Sentinel-1 from 2014
Typical mapping latency	2–6 weeks from commission to validated fraction grid, depending on cloud frequency
Delivery formats	GeoTIFF (fraction grid + confidence layer), GeoPackage, compatible with SWMM, HEC-HMS, QGIS, ArcGIS
Accuracy (published benchmarks)	Overall accuracy 85–92 % for binary impervious/pervious; RMSE on fraction estimates typically 0.08–0.15 in published V-I-S studies

Analytics Satellize can run

City-wide impervious fraction grid	Spectral unmixing of V-I-S end-members using Sentinel-2 SWIR composite; linear or Monte Carlo unmixing	GeoTIFF raster (10 m or 30 m cells, 0–1 fraction values) with per-pixel confidence layer
Dry-soil / impervious disambiguation layer	SWIR-based spectral indices (e.g. NDISI, MNDISI) applied to Sentinel-2 Band 11/12 and Landsat SWIR1/SWIR2	Classified GIS layer flagging high-uncertainty pixels for field or WorldView-3 validation
Post-rainfall SAR consistency check	Sentinel-1 VV backscatter change detection before and after precipitation events; intersection with optical impervious candidates	Agreement/disagreement raster delivered as GeoPackage layer with confidence attribution
Annual impervious-change time series	Multi-year Sentinel-2 and Landsat composites processed through consistent unmixing pipeline; change vector analysis	Annual fraction grids (GeoTIFF stack) and summary statistics table per administrative zone or catchment
SWMM-ready subcatchment parameterisation	Zonal statistics of impervious fraction grid aggregated to user-supplied catchment polygons; look-up conversion to curve numbers	CSV or SWMM .inp snippet with percent-impervious values per subcatchment, ready for import
High-resolution rooftop material classification	Supervised classification using WorldView-3 eight-band SWIR data; spectral library matching for membrane, tile, metal, green-roof types	Building-level GIS polygon layer with material class and estimated drainage coefficient

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.