Flood insurance loss estimation using SAR-derived inundation duration
Flood damage scales with how long water sits on a property, not just how far it spreads. Multi-date Sentinel-1 SAR time series can reconstruct inundation recession curves at parcel level, giving insurers and reinsurers an independent, physics-based input to actuarial loss models.
Sensors
- Sentinel-1 A/B SAR (C-band, 5.405 GHz): 10 m ground range resolution in Interferometric Wide Swath mode, 250 km swath, 6-day repeat per satellite. The workhorse for multi-date inundation time series over large areas; free and open archive from 2014.
- ICEYE SAR constellation (X-band): Sub-metre spotlight mode and roughly 3 m stripmap resolution, with revisit intervals as short as a few hours over a target area when multiple satellites are tasked. X-band penetrates light vegetation better than C-band and is less affected by urban multipath in some geometries, though it shares the dense-canopy limitation.
- Capella Space SAR (X-band): Spotlight imagery at approximately 0.5 m resolution. Useful for parcel-level damage confirmation at individual high-value properties, but coverage per pass is narrow and tasking cost limits its role to sample validation rather than area-wide recession mapping.
- Sentinel-2 MSI (optical, 10–20 m): 10 m visible and near-infrared bands, 20 m shortwave infrared. SWIR bands sharply distinguish open water from moist soil during cloud-free windows, providing cross-validation of SAR-derived recession curves where sky conditions allow. Cloud cover during active flood events is the principal constraint.
Why duration is the number actuaries actually need
Flood extent maps answer one question: was this parcel underwater? That is necessary but not sufficient for loss estimation. Structural damage to masonry and timber-frame buildings is a function of water depth and the time that depth is maintained. Drywall and insulation that drains within 24 hours may be salvageable; the same materials submerged for five days typically require full replacement. Agricultural crop loss follows a similar but faster curve: most cereal crops suffer irreversible yield loss after 48 to 72 hours of root-zone inundation, depending on growth stage.
Insurers have long known this. The problem is that traditional loss adjustment relies on field visits, which are slow, expensive, and impossible to scale across thousands of simultaneous claims after a major event. A satellite-derived recession curve, one that shows the date water first appeared and the date it cleared each parcel, gives the claims function an independent, spatially consistent input before a single adjuster reaches the site.
How a Sentinel-1 time series reconstructs a recession curve
Sentinel-1 operates in C-band (5.405 GHz) and images through cloud, day or night. Open water returns very low backscatter in SAR imagery because the flat surface specularly reflects the radar pulse away from the sensor. Flooded land therefore appears distinctively dark against the rougher backscatter of dry vegetation and soil. By stacking multiple passes over a flood event and comparing each to a pre-event baseline, analysts can assign a binary wet/dry classification to each 10 m pixel on each acquisition date.
With a single Sentinel-1 satellite the repeat cycle is 6 days, which means a 10-day flood event might be captured by only two or three passes during the inundation period. The ESA two-satellite configuration (Sentinel-1A and the now-degraded Sentinel-1B) halved that to 3 days over Europe; with Sentinel-1B's current reduced availability, coverage frequency outside Europe has reverted toward 6-day intervals. That sampling rate is coarse for fast-receding flash floods but adequate for the slow drainage typical of lowland river and coastal plain events. Recession curves are fitted by interpolating between classified acquisition dates, using a physically plausible drainage model constrained by local terrain slope derived from a digital elevation model.
The output is a per-parcel inundation duration raster: each pixel carries an estimated number of days submerged, with a confidence interval that reflects the temporal sampling gap. That raster can be intersected with cadastral parcel boundaries and building footprints to produce a claim-level duration estimate.
Where C-band SAR fails, and what to do about it
C-band has two well-documented blind spots. First, dense vegetation canopy (closed-canopy forest, mature maize, tall sugarcane) attenuates the signal before it reaches standing water beneath. The double-bounce return that sometimes reveals flooded woodland is inconsistent and geometry-dependent. If a parcel is under a mature tree canopy, the SAR-derived duration estimate is unreliable and should be flagged as such in any deliverable.
Second, urban canyons trap and scatter the radar signal in ways that mask water returns. In dense city blocks the backscatter from building walls dominates, and a flooded street at the base of a five-storey terrace may be invisible to Sentinel-1. Commercial X-band systems at sub-metre resolution improve the situation somewhat by resolving individual structures, but they do not eliminate the problem.
The practical mitigation is sensor fusion. Sentinel-2 SWIR imagery fills temporal gaps on cloud-free days. ICEYE or Capella tasking can provide higher-resolution validation passes at critical moments. For heavily vegetated parcels, field-survey triggers based on satellite-flagged uncertainty are more honest than a fabricated duration estimate. Honest uncertainty quantification is not a weakness in an insurance context; it is what separates a defensible loss input from a liability.
Connecting recession curves to loss models
The standard actuarial approach to flood loss uses a depth-damage function: a curve that maps inundation depth at a property to expected damage as a percentage of replacement value. Duration modifies that function. Published research in flood risk assessment, including work underpinning European flood directive implementation, treats duration as a multiplier on the depth-damage relationship, particularly for contents and agricultural assets.
A SAR-derived duration layer plugs directly into that framework. Parcel-level duration estimates, expressed in days with associated uncertainty bounds, can be joined to policy databases by cadastral identifier. The resulting loss distribution gives the claims team a ranked list of expected severity before field visits are scheduled, allowing adjusters to prioritise the highest-exposure properties. For reinsurers assessing aggregate exposure across a portfolio, the spatial consistency of satellite-derived data is particularly valuable: it removes the adjuster-to-adjuster variability that inflates uncertainty in post-event loss development.
Crop loss is a related but distinct calculation. Inundation duration during specific phenological windows drives yield loss more than total submersion time. That calculation is covered in the sibling page on crop loss estimation from flood recession timing; this page focuses on structural and contents losses.
Archive depth and event reconstruction
Sentinel-1's open archive extends to 2014, which means historical events can be reconstructed for portfolio back-testing. An insurer pricing a flood endorsement for a property in a frequently inundated river floodplain can request a retrospective duration analysis across all flood events in the archive, not just the most recent one. That historical depth is something no field-survey programme can replicate at scale.
Latency for a current event depends on Sentinel-1's acquisition schedule over the affected area and the time to process and classify each pass. From satellite acquisition to a classified inundation layer, processing pipelines using ESA's Sentinel Application Platform or equivalent can typically deliver results within 24 to 48 hours of data download. The 6-day repeat remains the binding constraint on how finely the recession curve can be resolved, not the processing speed.
Satellize runs Sentinel-1 flood analytics on open-archive data and can add commercial SAR tasking for clients requiring higher revisit. The workflow draws on the same analytical infrastructure used in the Tonga crop-estimation programme, adapted for structural loss rather than agricultural yield. Teams wanting to trial the method against a historical event in their portfolio can request a bounded retrospective run as a first step.
Typical figures
| Primary sensor spatial resolution | 10 m (Sentinel-1 IW mode); 3 m stripmap / sub-metre spotlight (ICEYE, Capella) |
| Sentinel-1 revisit (single satellite) | 6 days at equator; 1–3 days at high latitudes due to orbit convergence |
| Temporal resolution of recession curve | One classification per SAR pass; interpolated to daily with DEM-constrained drainage model |
| Minimum detectable open-water patch | Approximately 0.01 ha in open terrain at 10 m resolution; larger under vegetation or in urban areas |
| Radar frequency | C-band 5.405 GHz (Sentinel-1); X-band ~9.6 GHz (ICEYE, Capella) |
| Swath width | 250 km (Sentinel-1 IW); 15–30 km typical (commercial X-band stripmap) |
| Cloud penetration | SAR: all-weather, day/night. Sentinel-2 optical cross-validation: cloud-free acquisitions only |
| Archive depth | Sentinel-1: from 2014; Sentinel-2: from 2015; ICEYE commercial archive from ~2018 |
| Processing latency (current event) | 24–48 hours from satellite acquisition to classified inundation layer, depending on data download scheduling |
| Delivery formats | GeoTIFF raster (duration in days per pixel), GeoPackage or Shapefile (parcel-level summary), CSV for actuarial model ingestion |
Analytics Satellize can run
| Per-parcel inundation duration raster | Multi-date Sentinel-1 backscatter change detection with pre-event baseline; binary wet/dry classification per pass; temporal interpolation using DEM-constrained recession model | GeoTIFF layer: estimated inundation days per 10 m pixel, with uncertainty band, clipped to event area of interest |
| Parcel-level duration summary table | Zonal statistics applied to duration raster using cadastral or policy parcel boundaries | CSV or GeoPackage: parcel ID, mean duration, max duration, area flooded (ha), confidence flag (open/vegetated/urban) |
| Recession curve time series | Percentage of parcel area classified as inundated at each SAR acquisition date, fitted with exponential or piecewise-linear recession function | Per-parcel JSON time series and PDF chart, suitable for claims file attachment |
| Uncertainty and coverage quality flag | Classification of each parcel by dominant land cover (open, vegetated canopy, dense urban) using Sentinel-2 NDVI and NDBI; SAR reliability score applied per class | Attribute column in parcel summary: HIGH / MODERATE / LOW reliability, with explanatory notes for underwriters |
| Historical event reconstruction for portfolio back-testing | Retrospective Sentinel-1 archive processing over user-defined event dates and bounding box; same classification pipeline as current-event product | Duration rasters and parcel summaries for each historical flood event, formatted for ingestion into catastrophe model back-test workflows |
| Sensor-fused inundation extent cross-validation | Sentinel-2 SWIR band ratio (MNDWI) on cloud-free acquisitions during event window, compared against SAR classification to identify systematic commission or omission errors | Discrepancy map and summary statistics report, flagging parcels where SAR and optical classifications diverge |
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.