Flood extent mapping with synthetic aperture radar
SAR backscatter maps flooded land within hours of acquisition, day or night, through cloud cover. Open water returns a specular, near-zero signal; the complication is flooded forest, which can appear bright and fool simple thresholds.
Sensors
- Sentinel-1 A/B (C-band, 5.405 GHz): Primary workhorse for operational flood mapping. Interferometric Wide Swath mode covers 250 km at 5 × 20 m ground range resolution (single-look complex) or 10 m detected product. Revisit 6 days per satellite at mid-latitudes with one satellite operational; 6-day repeat with Sentinel-1C now in commissioning. VV and VH polarisations available; VH cross-pol is critical for separating flooded vegetation from open water.
- ALOS-2 PALSAR-2 (L-band, 1.236 GHz): JAXA's L-band system penetrates forest canopy far more effectively than C-band, making it the preferred sensor for tropical and boreal flooded forests. ScanSAR wide mode delivers 350 km swath at 100 m resolution; fine-beam modes reach 3–10 m. Revisit is 14 days, which limits near-real-time use but the archive depth since 2014 supports change detection.
- RADARSAT-2 (C-band, 5.405 GHz): Commercial Canadian system with programmable polarimetry (quad-pol available) and a 24-day exact repeat, though off-nadir tasking allows more frequent revisit. ScanSAR Wide mode gives 500 km swath at 100 m; Fine mode reaches 8 m. Quad-pol data supports decomposition methods that better separate double-bounce returns in flooded forest.
- Capella Space (X-band, ~9.6 GHz): Commercial constellation offering sub-metre spotlight imagery with tasking latency under a few hours. X-band does not penetrate vegetation, so it is most useful for urban and periurban flood mapping where open-surface water detection is the primary goal. Revisit depends on constellation size and tasking priority.
Why radar, and what the physics actually promises
A microwave pulse hitting calm open water reflects away from the sensor at the specular angle. Very little energy returns to the antenna. On a SAR image, inundated land therefore appears dark, contrasting sharply with rougher surrounding surfaces that scatter energy back. The effect is reliable across C-band and L-band frequencies, and it works at night and through cloud cover, which is precisely when optical sensors fail during active weather events.
The promise is real but bounded. Minimum detectable flood width depends on resolution: Sentinel-1 IW mode at 10 m can detect inundation patches of roughly a few hundred square metres in practice, though narrow drainage channels are often missed. Wind roughens water surfaces, raising backscatter and eroding the contrast. Urban streets flooded to shallow depth may not register at all if the water surface is broken by debris or vegetation. These are not edge cases; they are routine conditions during major flood events.
The flooded-forest problem and how analysts handle it
Beneath a flooded forest canopy, the geometry changes. The water surface acts as a mirror that redirects the radar pulse upward into the trunk, which reflects it directly back to the sensor. This double-bounce mechanism produces a bright return, the opposite of what open water produces. A simple low-backscatter threshold will therefore miss inundated forest entirely, sometimes misclassifying it as dry land.
Two mitigation strategies dominate the published literature. The first is cross-polarisation ratio analysis: VH backscatter (cross-pol) is less affected by double-bounce than VV, so the VH/VV ratio shifts in a characteristic way over flooded vegetation. The second is change detection against a pre-flood reference image acquired under similar conditions. Areas that were dark in the dry-season reference and are now bright have likely flooded beneath a canopy. Neither method is perfect. Change detection requires a cloud-free, seasonally matched reference, and the ratio approach still struggles with dense mangrove and closed-canopy tropical forest. L-band from ALOS-2 PALSAR-2 penetrates deeper into canopy and produces stronger double-bounce signals, which paradoxically makes flooded forest easier to identify as a distinct class rather than confusing it with dry land.
Operational pipelines: from acquisition to emergency map
The Copernicus Emergency Management Service (CEMS) is the most widely used operational system for SAR-based flood mapping. When activated, it draws primarily on Sentinel-1 and produces delineation maps within hours of data acquisition. The service publishes grading maps showing inundated area, damage proxy products and reference maps, all as open GIS deliverables. Activation is free to eligible users including national civil protection authorities, and the public record of past activations is searchable.
JAXA runs its own disaster response protocol, the JAXA Satellite Monitoring for Environmental Studies group, which tasks ALOS-2 for major flood events and releases processed imagery through the Sentinel Asia portal. Response time from event to available imagery is typically 12 to 48 hours depending on the satellite's orbital geometry at the time of the event. For events in tropical Asia, where dense vegetation covers most floodplains, ALOS-2 data often provides the more complete inundation picture.
Beyond these agency pipelines, automated processing chains built on Sentinel-1 data are now well documented. The Otsu thresholding method, applied to VV backscatter histograms, is a common starting point. More sophisticated approaches use random forest classifiers trained on labelled SAR scenes, incorporating auxiliary layers such as slope, land cover and distance to river network to reduce false positives in shadow-affected terrain.
Terrain shadow: the error source analysts rarely advertise
SAR images contain geometric artefacts that mimic flood signatures. In hilly terrain, slopes facing away from the sensor are in radar shadow and return near-zero backscatter, identical to open water. Layover, where elevated features are projected forward in range, creates additional confusion. Any flood map produced without a terrain correction and shadow mask applied is unreliable in anything other than flat floodplains.
The standard mitigation is to mask shadow and layover zones using a digital elevation model before applying any water detection algorithm. The Copernicus DEM at 30 m resolution is freely available and sufficient for most operational work. In steep terrain, residual errors remain even after masking, and analysts should report the fraction of the area of interest affected by shadow as part of the product metadata. Buyers of flood maps should ask for this figure.
Archive depth and the return-period question
Sentinel-1 data is available from April 2014 (Sentinel-1A launch). ALOS-2 PALSAR-2 extends back to 2014 as well, with the original ALOS PALSAR archive reaching back to 2006. This decade-plus record is long enough to begin characterising flood frequency, identifying which areas flood in a one-in-ten-year event versus a one-in-fifty-year event, provided the orbital geometry produced usable acquisitions during past events.
The practical constraint is that Sentinel-1 was not always acquiring over every region on every pass in its early years; the acquisition plan changed as the constellation evolved. Analysts should verify actual coverage dates against the Copernicus archive before committing to a frequency analysis. Satellize runs systematic archive queries as part of flood-risk assessments, including the kind of coverage-gap audit that determines whether a historical analysis is actually feasible before work begins.
Capella Space and other commercial X-band operators do not yet have archives long enough for return-period analysis. Their value is in the present: fast tasking, high resolution, and the ability to image a specific site repeatedly over a single event.
What a flood extent map can and cannot support
A SAR-derived flood extent polygon is a snapshot of surface inundation at the moment of acquisition. It supports emergency response routing, damage assessment, insurance loss estimation and model validation. It does not directly measure water depth, flow velocity or total volume. Depth estimation requires combining the extent map with a terrain model and assumptions about the flood surface slope, which introduces significant uncertainty in low-gradient terrain where the water surface is nearly flat over large distances.
Validation studies comparing SAR flood maps against gauge records and airborne surveys typically report overall accuracy between 85 and 95 percent for open-water classes on flat agricultural land, falling to 70 to 80 percent in vegetated or urban areas. Those figures come from published benchmark studies on European and Asian flood events using Sentinel-1 data. Buyers should treat accuracy claims outside that range with scepticism unless the vendor can point to a specific validation dataset.
Typical figures
| Spatial resolution (Sentinel-1 IW mode) | 10 m detected product; 5 × 20 m single-look complex |
| Spatial resolution (ALOS-2 ScanSAR wide) | 100 m; fine-beam modes 3–10 m |
| Swath width | 250 km (Sentinel-1 IW); 350 km (ALOS-2 ScanSAR wide) |
| Revisit (Sentinel-1, one satellite) | 6 days at mid-latitudes; more frequent at high latitudes due to orbit convergence |
| Frequency bands used | C-band 5.405 GHz (Sentinel-1, RADARSAT-2); L-band 1.236 GHz (ALOS-2); X-band ~9.6 GHz (Capella) |
| Typical map latency (CEMS activation) | Hours to ~24 h from SAR acquisition to published delineation product |
| Minimum detectable inundated patch | Approximately a few hundred square metres at 10 m resolution under calm wind conditions; larger under wind roughening |
| Archive depth | Sentinel-1 from April 2014; ALOS PALSAR from 2006; ALOS-2 from 2014 |
| Cloud penetration | Full; SAR is unaffected by cloud, rain (except very heavy precipitation at X-band) and darkness |
| Delivery formats (CEMS standard) | GeoTIFF, shapefile, GeoPackage; open access via Copernicus portal |
Analytics Satellize can run
| Near-real-time flood extent polygon | Otsu thresholding on VV backscatter histogram, with terrain shadow mask applied using Copernicus DEM 30 m | GeoPackage layer and PDF situation map, delivered within hours of Sentinel-1 acquisition |
| Flooded-vegetation classification | VH/VV cross-polarisation ratio change detection against seasonally matched pre-flood reference image | Raster layer distinguishing open water, flooded vegetation and dry land; included in GIS delivery |
| Flood frequency map | Pixel-level inundation frequency count across Sentinel-1 archive acquisitions, normalised by number of valid (non-shadow) observations | Raster showing fraction of observed dates with inundation, exportable as GeoTIFF with metadata on coverage gaps |
| Urban damage proxy | Coherence change detection between pre- and post-event Sentinel-1 interferometric pairs; coherence loss indicates structural change or inundation | Alert layer with affected building footprints flagged, delivered as GeoJSON feed |
| Event timeline reconstruction | Multi-date SAR stack showing inundation onset, peak extent and recession, one layer per available acquisition | Animated GIF and time-stamped GeoPackage for situation reporting |
| Archive coverage audit | Query of Copernicus and JAXA acquisition histories against area of interest and event dates | Coverage report confirming which historical flood events have usable SAR data, as a prerequisite for return-period analysis |
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.