Urban flash flood runoff modelling from impervious surface mapping
Satellite-derived impervious surface maps at sub-10 m resolution feed directly into runoff models, replacing coarse land-use datasets that routinely underestimate sealed area in rapidly urbanising cities.
Sensors
- WorldView-3 (Maxar): 0.31 m panchromatic, 1.24 m multispectral in 8 VNIR bands plus 8 SWIR bands at 3.7 m. The SWIR bands distinguish wet concrete from dry asphalt and separate thin roof coatings that look identical in VNIR alone. Revisit roughly 1 day at mid-latitudes with off-nadir tasking, though cloud cover frequently extends effective wait times in humid climates.
- Pleiades Neo (Airbus): 0.30 m panchromatic, 0.50 m multispectral in 6 bands including a red-edge channel. The red-edge helps isolate sparse urban vegetation from impervious surfaces at sub-metre scale. Daily revisit capacity over any point, but tasking must be planned; archive coverage of informal settlements is patchy before 2021.
- Sentinel-2 MSI (ESA/Copernicus): 10 m in four VNIR bands, 20 m in six red-edge and SWIR bands, 60 m for three atmospheric bands. Free and open, with a 5-day revisit at the equator (2–3 days at mid-latitudes with both satellites). At 10 m, individual buildings are sub-pixel; spectral mixture analysis can still recover impervious fraction per pixel but cannot resolve fine-grained roof typology. The best choice for city-wide or regional baseline mapping.
- Planet SkySat: 0.50 m panchromatic, 0.72 m multispectral in 4 bands (BGRN). Useful for rapid pre- or post-event mapping when WorldView or Pleiades tasking is unavailable. SWIR is absent, which limits material discrimination; classification accuracy for impervious fraction typically falls 5–10 percentage points compared with WorldView-3 SWIR-assisted methods on comparable urban scenes.
Why land-use maps get the hydrology wrong
Most urban flood models still ingest impervious fractions from national land-use/land-cover datasets compiled at 25–100 m resolution, often updated on decade-long cycles. In a city adding 50,000 m² of paved surface per year, that lag is not a rounding error; it is a systematic underestimate of runoff volume that propagates directly into peak-discharge predictions.
The SCS Curve Number method, still the workhorse of municipal drainage design, is acutely sensitive to impervious fraction. A 10-percentage-point error in sealed-area estimate for a catchment with CN around 85 can shift predicted runoff depth by 15–25 mm for a 50 mm storm event. That translates to meaningful errors in peak flow timing and volume, particularly in small, fast-responding urban catchments where the time of concentration is under 30 minutes.
What a floating roof gives away
Impervious surface mapping from satellite imagery rests on two complementary approaches. Spectral mixture analysis (SMA) decomposes each pixel into fractional contributions from endmembers: high-albedo impervious surfaces such as concrete and light roofing, low-albedo impervious surfaces such as dark asphalt, vegetation, soil, and water. At Sentinel-2's 10 m, SMA can estimate impervious fraction with root-mean-square errors of roughly 8–12% in well-calibrated urban scenes, based on published comparisons with airborne LiDAR reference data.
At sub-metre resolution from WorldView-3 or Pleiades Neo, pixel-level classification becomes feasible. Machine-learning classifiers, particularly random forests and convolutional neural networks trained on labelled roof and road samples, routinely achieve overall accuracies of 90–95% for broad impervious/pervious separation in published studies. The harder problem is material-level discrimination: distinguishing a metal roof from a concrete terrace, or a gravel car park from bare soil. WorldView-3's SWIR bands earn their cost here; the shortwave infrared reflectance of roofing materials diverges in ways that VNIR alone cannot resolve.
One honest caveat: shadow. In dense mid-rise districts with narrow streets, shadows from adjacent buildings can obscure 20–30% of ground-level impervious area. Fusing nadir and off-nadir acquisitions, or combining optical classification with a digital surface model, partially compensates, but no purely optical approach eliminates the problem.
From pixels to peak discharge
The classified impervious fraction feeds into runoff models in two main ways. For CN-based approaches, the satellite-derived fraction replaces the tabulated CN lookup for each sub-catchment, producing a spatially distributed CN surface rather than a single lumped value. For Green-Ampt infiltration models, the impervious fraction determines the areal proportion of the catchment where infiltration capacity is effectively zero, with the remainder assigned hydraulic conductivity values from soil texture data.
Routing the resulting runoff through a digital elevation model to produce inundation depth and timing is a separate step, and the accuracy of that step depends heavily on DEM quality, which satellite optical imagery cannot supply on its own. This page does not cover DEM-based depth estimation; that is addressed in the sibling page on flood water depth estimation by DEM differencing. What satellite impervious mapping can supply is a materially better boundary condition than any land-use polygon dataset compiled more than two or three years ago.
Choosing the right sensor for the catchment size
Sensor choice should follow catchment geometry, not procurement habit. For a city-wide hazard assessment covering 200–2,000 km², Sentinel-2 is the practical starting point. It is free, consistently archived since 2015, and its 10 m bands support SMA-derived impervious fraction mapping at a scale appropriate to sub-catchments of 0.5 km² and above. Accuracy degrades in informal settlements with highly heterogeneous roof materials and narrow alleys, where mixed pixels dominate.
For a priority district covering 5–50 km², WorldView-3 or Pleiades Neo imagery at 0.3–0.5 m resolution resolves individual buildings and allows per-parcel impervious fraction estimates. The cost is real: commercial tasking runs to tens of dollars per km² depending on contract structure, and cloud-free acquisition over a specific district on a specific date is not guaranteed. SkySat offers a middle path when speed matters more than material discrimination.
A practical workflow for emergency preparedness combines both: Sentinel-2 for the full urban extent to identify high-impervious sub-catchments, then targeted very-high-resolution tasking over the highest-risk zones for model calibration and detailed drainage planning.
Honest limits of the method
Impervious surface classification from optical imagery is a solved problem in favourable conditions and a genuinely difficult one in others. Persistent cloud cover in tropical cities, where flash flood risk is highest, can prevent cloud-free acquisition for weeks at a time. A single cloud-free Sentinel-2 scene over a monsoon-belt city may require compositing imagery from multiple passes across several months, during which urban development continues.
Classification accuracy also degrades in informal settlements, where roof materials vary at sub-metre scales, standing water on flat roofs mimics pervious surfaces spectrally, and building density creates deep shadow. Published accuracy figures from well-mapped North American or European cities do not transfer directly to these environments without local training data.
Finally, the CN and Green-Ampt models themselves carry substantial uncertainty independent of the impervious input. Satellite mapping improves one input; it does not remove model structural uncertainty or the need for local calibration against gauged flow data where it exists.
Satellize's analytics team has worked through analogous input-uncertainty problems in the Tonga crop-estimation programme, where field-truth data is sparse and spectral conditions are challenging. The lessons transfer: honest uncertainty quantification around the impervious fraction estimate should accompany every model run, not just the headline runoff number.
Satellite inputs versus traditional datasets: a direct comparison
National land-use datasets typically classify parcels as residential, commercial or industrial and assign a tabulated impervious fraction to each class. In practice, a 'residential' parcel in a dense Asian city may be 85% impervious; the same class in a low-density suburb may be 35%. The table cannot know the difference. Satellite-derived impervious fraction is measured, not assumed.
The practical gain is largest in cities that have grown quickly since the last cadastral or land-use survey. Satellite archive depth helps here: Sentinel-2 data runs continuously from 2015, and the Landsat archive extends to 1972 at 30 m resolution, providing a temporal record of urban expansion that no municipal dataset matches. For a flood model calibrated against a historical event, matching the impervious surface map to the date of the event rather than the date of the last survey is a meaningful improvement in model fidelity.
Typical figures
| Best spatial resolution (impervious mapping) | 0.30–0.50 m (WorldView-3, Pleiades Neo); 0.72 m (SkySat); 10 m (Sentinel-2 VNIR) |
| Impervious fraction RMSE (published ranges) | 8–12% at 10 m (Sentinel-2 SMA); 5–8% at sub-metre with SWIR-assisted ML |
| Revisit (tasked VHR) | 1 day or better with off-nadir; cloud-free acquisition not guaranteed in humid tropics |
| Revisit (Sentinel-2) | 5 days at equator; 2–3 days at mid-latitudes (both satellites combined) |
| Spectral bands used | VNIR (all sensors); SWIR at 3.7 m (WorldView-3); red-edge at 0.50 m (Pleiades Neo) |
| Minimum resolvable feature (impervious) | Individual buildings at VHR; sub-catchment fraction (≥0.5 km²) at Sentinel-2 |
| Archive depth | Sentinel-2 from 2015; Landsat 30 m from 1972; commercial VHR from approx. 2008 (limited) |
| Typical processing latency (post-acquisition) | 2–5 days for classified impervious layer; same-day possible with pre-trained models on cloud-free data |
| Delivery formats | GeoTIFF impervious fraction raster; sub-catchment CN surface as vector GIS layer; uncertainty band raster |
| Cloud sensitivity | Optical-only; persistent cloud cover requires multi-date compositing or acceptance of temporal mismatch |
Analytics Satellize can run
| City-wide impervious fraction map | Spectral mixture analysis on Sentinel-2 10 m composites, calibrated against available VHR reference patches | GeoTIFF raster (10 m) with per-pixel impervious fraction 0–1 and associated uncertainty layer |
| Sub-catchment CN surface | Aggregation of satellite impervious fraction to drainage sub-catchments using SCS Curve Number lookup, replacing tabulated land-use values | Vector GIS layer (sub-catchment polygons) with CN value, impervious fraction, and comparison against national land-use dataset value |
| Per-parcel roof and surface classification | Random forest or CNN classifier applied to WorldView-3 or Pleiades Neo imagery at 0.3–0.5 m, using SWIR bands for material discrimination where available | Vector parcel layer with material class (concrete, metal, asphalt, gravel, vegetation, soil) and confidence score |
| Runoff volume and timing estimate for design storm | CN-based or Green-Ampt infiltration model driven by satellite impervious surface inputs and a specified design rainfall hyetograph | Tabular runoff hydrograph per sub-catchment; summary report with sensitivity analysis showing runoff response to ±10% impervious fraction uncertainty |
| Urban expansion impervious change series | Multi-date Sentinel-2 SMA or Landsat-based impervious mapping across archive depth (2015–present for Sentinel-2) | Annual or biennial impervious fraction rasters with change-detection layer highlighting newly sealed areas; suitable for updating legacy drainage models |
| High-risk sub-catchment prioritisation | Overlay of satellite-derived CN surface with slope-derived time-of-concentration estimates and population density to rank sub-catchments by flash flood exposure | Ranked GIS layer with composite risk score; input-ready format for municipal drainage investment planning |
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.