Flood extent mapping with synthetic aperture radar
Synthetic aperture radar detects flooded land by the near-total absence of backscatter from smooth open water, day or night, through cloud. The method is fast and operationally proven, but flooded vegetation and wind chop introduce systematic errors that demand careful handling.
Sensors
- Sentinel-1 IW (C-band, 5.6 cm): 10 m ground range resolution in Interferometric Wide swath mode, 250 km swath, 6-day repeat at the equator with two satellites (shorter at higher latitudes). Free and open archive from 2014. The workhorse for operational flood mapping globally.
- ALOS-2 PALSAR-2 (L-band, 23.6 cm): L-band penetrates forest canopy far better than C-band, making it the preferred sensor for detecting inundation beneath dense vegetation. Stripmap mode delivers 3 m resolution; ScanSAR covers 350 km at 100 m. Revisit is 14 days, limiting timeliness for fast-moving events.
- ICEYE X-band SAR constellation: Sub-metre resolution (0.5 m spotlight) with same-day tasking and revisit intervals as short as a few hours through constellation scheduling. X-band is more sensitive to small surface roughness changes, which helps in urban margins but increases misclassification risk on wind-roughened water.
- Capella Space X-band SAR: Spotlight mode at 0.5 m resolution, with rapid tasking turnaround. Complements ICEYE for high-priority urban or infrastructure sites where 10 m Sentinel-1 pixels are too coarse to distinguish flooded streets from dry ones.
Why radar works when clouds do not care
Optical sensors go blind the moment convective cloud builds over a flood. That is precisely when you most need imagery. SAR transmits its own microwave pulse and records what bounces back, so overcast skies and night-time are irrelevant. The physics is straightforward: calm open water acts as a specular reflector, bouncing almost all energy away from the sensor and returning a very low backscatter signal. Dry soil, crops and urban surfaces scatter energy back strongly. The contrast is large enough to detect with simple thresholding on a single acquisition.
C-band (Sentinel-1 at 5.6 cm wavelength) is the operational standard because the archive is deep, the data are free and the 10 m resolution is adequate for field-scale and catchment-scale mapping. L-band (ALOS-2 at 23.6 cm) trades resolution and revisit for the ability to see through forest canopy, which C-band cannot do reliably. X-band commercial systems trade swath width for spatial detail, useful when you need to know whether a specific road or levee is passable.
The two failure modes that cause real operational errors
Double-bounce is the first. When a radar pulse hits the water surface beneath a vertical structure, it reflects upward to the structure and back to the sensor, producing an anomalously high return. Flooded forest and flooded urban areas both generate this signature. The result is that inundated vegetation can appear bright rather than dark, causing automated classifiers to miss the flood entirely in those pixels. L-band is more prone to this in forest because its longer wavelength penetrates canopy and reaches the trunk-water interface.
Wind roughening is the second. Whitecaps and ripples on open water scatter energy back toward the sensor, raising backscatter above the calm-water threshold. A fast-moving storm surge with 15-knot winds can produce water backscatter values indistinguishable from dry grassland in C-band VV polarisation. Analysts working storm events must apply wind-speed corrections or cross-check against auxiliary wind data. Neither failure mode is exotic: both appear routinely in operational flood products, and any honest flood-mapping workflow names them explicitly in its uncertainty layer.
Change detection versus single-image thresholding
A single SAR acquisition can be thresholded for low backscatter, but the result conflates permanent water bodies, sand, smooth tarmac and actual flood. The standard improvement is bi-temporal change detection: subtract a pre-event image (same sensor, same orbit geometry, same polarisation) from the post-event image. Areas that have become significantly darker are candidates for new inundation. This approach is the basis of the Copernicus Emergency Management Service (CEMS) rapid-mapping products and the Dartmouth Flood Observatory's archive.
Polarimetric decomposition adds another layer. Sentinel-1 IW acquires VV and VH polarisations simultaneously. Open water suppresses both, but the ratio behaves differently over flooded vegetation versus open flood. Dual-pol classifiers trained on labelled samples outperform single-band thresholds in mixed land-cover scenes, though they require more processing and a reference dataset. For fast operational response, a threshold on VV change detection is still the most widely deployed method because it runs in minutes.
Sensor choice is a trade-off, not a hierarchy
Sentinel-1 is the default for most national disaster agencies because it is free, consistent and globally archived. Its 6-day revisit (with both satellites active) is acceptable for slow-rising river floods but inadequate for flash floods that peak and recede within 24 hours. ICEYE and Capella can task within hours, but their narrow swaths mean you must know where to point them. Ordering commercial SAR without a prior Sentinel-1 pass to identify the affected area is an expensive way to image the wrong field.
ALOS-2 PALSAR-2 fills the forest-flood gap that no C-band or X-band sensor addresses well. In tropical river systems with dense riparian vegetation, C-band products can underestimate inundated area by 30 to 50 percent compared with L-band, according to published comparisons in the peer-reviewed literature. JAXA activates ALOS-2 emergency observations through the International Charter on Space and Major Disasters, making L-band data accessible without a commercial contract during declared events.
What the output actually tells a response coordinator
A flood extent polygon in a GIS layer answers one question: where is water now? That is necessary but not sufficient. Depth is not directly observable from backscatter alone (though it can be estimated by combining extent with a digital elevation model). Velocity is not observable. Contamination, structural damage and access routes require additional data layers. SAR flood extent is best understood as the spatial trigger that focuses other assessments, not the complete picture.
Latency matters enormously. A flood extent map delivered 72 hours after acquisition is useful for recovery planning but not for evacuation routing. The operational target for emergency response is acquisition-to-delivery under six hours. Sentinel-1 NRT (near-real-time) processing through the Copernicus Data Space Ecosystem can approach this for pre-planned monitoring areas. Commercial X-band providers can in principle deliver faster, but the analytic pipeline, not the satellite, is usually the bottleneck. Satellize runs automated change-detection pipelines on Sentinel-1 acquisitions that flag new inundation extents as GIS layers with embedded uncertainty polygons, feeding directly into client emergency operations centres.
Honest limits of the method
SAR flood mapping has a minimum detectable area that depends on resolution and speckle filtering. At 10 m Sentinel-1 resolution, isolated flooded pixels smaller than roughly 0.1 hectares are unreliable after speckle filtering. Urban flooding in narrow streets is largely invisible at this resolution; X-band sub-metre imagery is needed there, and even then shadow and layover from tall buildings create geometric gaps.
Permanent water bodies must be masked using a pre-existing water mask (the JRC Global Surface Water dataset is the standard reference) or they inflate the flooded-area estimate. Transitional wetlands and rice paddies in flooded state are spectrally ambiguous and require seasonal baseline imagery to classify correctly. Finally, the method produces a binary or probabilistic surface-water mask, not a validated flood impact assessment. Translating extent into affected population or economic loss requires population grids and exposure databases that are external to the SAR product itself.
Typical figures
| Spatial resolution (Sentinel-1 IW) | 10 m ground range, 250 km swath |
| Spatial resolution (ALOS-2 PALSAR-2 Stripmap) | 3 m; ScanSAR 100 m over 350 km swath |
| Spatial resolution (ICEYE / Capella Spotlight) | 0.5 m |
| Revisit (Sentinel-1, two satellites) | 6 days at equator; 1-3 days at mid-latitudes |
| Revisit (ALOS-2) | 14 days |
| Revisit (ICEYE constellation) | Hours to same-day with tasking |
| Frequency bands used | C-band (5.405 GHz, Sentinel-1); L-band (1.2 GHz, ALOS-2); X-band (~9.6 GHz, ICEYE, Capella) |
| Minimum reliably detectable inundation patch (Sentinel-1) | Approximately 0.1 ha after speckle filtering; smaller patches unreliable |
| Archive depth (Sentinel-1) | From April 2014 (Sentinel-1A launch) |
| Typical acquisition-to-product latency (operational) | Under 6 hours for NRT pipelines; 24-72 hours for manual expert mapping |
Analytics Satellize can run
| Binary flood extent mask | Bi-temporal VV backscatter change detection with adaptive thresholding | GeoTIFF and vector polygon layer (GeoPackage or Shapefile) with acquisition timestamp |
| Probabilistic inundation layer | Bayesian mixture-model classification on VV/VH dual-pol Sentinel-1 | Raster layer with per-pixel flood probability 0-1, plus binary mask at chosen confidence threshold |
| Flooded-vegetation detection | Dual-pol decomposition and double-bounce index on L-band or C-band data | Supplementary polygon layer flagging areas where inundation beneath canopy is probable, with confidence flag |
| Flood extent time series | Automated stack processing of all available Sentinel-1 passes over a defined AOI during an event | Animated GIS stack and CSV of inundated-area statistics (hectares) per acquisition date |
| Affected-population estimate | Spatial intersection of flood extent with WorldPop or national census grid | Tabular report of estimated population within flood extent, disaggregated by administrative unit |
| Near-real-time flood alert | Automated Sentinel-1 NRT ingestion with change-detection trigger against pre-event baseline | Push alert (email or API) with preliminary extent polygon within hours of SAR overpass |
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.