SAR coherence-loss damage detection in denied or clouded areas
When buildings collapse or ground is disturbed, the radar signal that bounced predictably from a surface stops bouncing the same way. Interferometric coherence loss between repeat SAR passes makes that change legible even under cloud, smoke, or access denial.
Sensors
- Sentinel-1 IW (C-band, ESA): 5.5 GHz C-band, 5 x 20 m ground range resolution in Interferometric Wide swath mode, 250 km swath, six-day exact repeat at the equator (three days with both satellites). Free archive to 2014. Coherence is computed between two passes on the same relative orbit; C-band is sensitive to vegetation moisture and canopy motion, which is the primary ambiguity in non-urban scenes.
- ICEYE X-band SAR (commercial): 9.65 GHz X-band, sub-1 m resolution in Spot mode, daily revisit over a target with tasking. X-band coherence decorrelates faster over vegetation than C-band but is less sensitive to soil moisture, making it better suited to urban damage where structural scatterers dominate. Latency from acquisition to delivery can be under two hours.
- Capella Space X-band SAR (commercial): X-band, spotlight mode at 0.5 m resolution, revisit configurable via tasking. Useful for small-footprint damage confirmation where Sentinel-1 resolution is insufficient to distinguish a collapsed building from an intact neighbour in the same pixel.
- Sentinel-1 archive (historical baseline): The Copernicus archive holds consistent repeat-pass Sentinel-1 acquisitions back to 2014 for most of the globe. Pre-event coherence stacks over months establish a seasonal baseline, which is essential for separating genuine structural change from normal agricultural or vegetation cycles.
Why a radar signal stops agreeing with itself
Interferometric SAR coherence measures how similar the phase of a radar return is between two passes over the same area. When nothing on the ground has moved, the signal from each scatterer, a wall corner, a roof edge, a road surface, arrives at almost the same phase angle on both dates. Coherence is high, close to 1.0. When a building collapses, the scatterers that produced that stable return are gone or rearranged. The two signals no longer agree. Coherence drops, in heavily damaged urban areas often below 0.3, against a pre-event urban baseline that typically sits above 0.7.
The physics is wavelength-dependent. C-band (Sentinel-1, roughly 5.6 cm) is sensitive to centimetre-scale surface motion. X-band (ICEYE, Capella, roughly 3.1 cm) is more sensitive still, which cuts both ways: finer sensitivity to small structural shifts, but faster decorrelation over vegetation from wind or moisture. Knowing which band you are working with matters when you are trying to explain a coherence anomaly to a decision-maker who wants certainty.
What the 2023 Türkiye earthquake and UNOSAT conflict work established
The February 2023 Kahramanmaraş earthquakes provided a well-documented validation case. Multiple published analyses, including work drawing on Sentinel-1 and ICEYE acquisitions, showed that coherence-difference maps correctly identified the highest-damage districts in Antakya and Kahramanmaraş city within 24 to 48 hours of the event, before optical surveys were complete and while cloud persisted over parts of the affected area. The method did not require ground truth to generate a damage proxy map; it required only a pre-event coherence baseline and a post-event acquisition on the same geometry.
UNOSAT, the UN satellite centre, has applied coherence-loss analysis in conflict settings including Syria and Ukraine, where smoke, cloud, and access denial make optical assessment intermittent at best. Their published methodology uses coherence difference thresholding combined with pre-event building footprint data to estimate the fraction of structures affected in a given grid cell. The output is explicitly labelled a damage proxy, not a confirmed damage assessment, which is the honest framing any analyst should carry forward.
The ambiguity problem is real and must not be minimised
Coherence loss is not unique to structural damage. Three other processes produce similar signatures and must be actively excluded before a map is handed to a commander or a humanitarian coordinator.
Vegetation change is the largest confound. A field that is ploughed, flooded, or simply rained on between two passes will decorrelate at C-band. Seasonal agricultural cycles produce coherence patterns that look, on a single coherence-difference image, indistinguishable from widespread ground disturbance. The fix is a multi-temporal baseline: if coherence drops in a field every spring, a single-event drop in spring proves nothing. Urban pixels are less ambiguous because buildings are stable scatterers by design.
Soil moisture variation matters in bare or sparsely vegetated terrain. A heavy rainfall event between two passes changes the dielectric constant of the soil surface, altering the phase of the return. This is a known issue in arid conflict zones where the ground between settlements is the primary background.
Temporal baseline length compounds all of these. Sentinel-1's six-day repeat is long enough for vegetation to change meaningfully in growing season. ICEYE's ability to acquire on consecutive days can reduce this source of ambiguity significantly, at the cost of commercial tasking budget. The honest answer is that coherence maps require corroboration: building footprint data, optical imagery when available, or local knowledge of land cover.
Operational workflow from acquisition to damage proxy
A standard coherence-loss workflow begins with co-registration of the pre- and post-event SAR images to sub-pixel accuracy, then computation of the complex coherence magnitude over a spatial window, typically 5 x 20 looks for Sentinel-1 IW, which trades spatial resolution for coherence estimation stability. The coherence difference image is then thresholded or classified against the pre-event baseline.
In conflict settings the pre-event baseline is often a multi-temporal average of six to twelve coherence pairs from the preceding months, which smooths out transient decorrelation from weather. The post-event image is differenced against this average. Pixels falling more than two standard deviations below the baseline mean are flagged as anomalous. Building footprint layers, where they exist, are used to mask non-urban pixels and reduce false positives from agriculture.
Latency depends on sensor choice. Sentinel-1 data is typically available in the Copernicus Data Space within one to three hours of acquisition; processing to a coherence product adds another hour or two with automated pipelines. ICEYE can deliver a tasked acquisition with shorter total latency when the satellite geometry is favourable. For a fast-moving conflict situation, the difference between a six-day revisit and a one-day revisit is operationally significant.
What the method cannot tell you
Coherence loss identifies that something changed on the surface. It does not distinguish a collapsed residential building from a deliberately demolished one, a bombed factory from one that was evacuated and stripped of equipment, or a flooded basement from a structurally intact structure with a wet roof. Casualty estimation from coherence maps alone is not defensible; population data, building typology, and time-of-day assumptions all have to be layered in separately and carry their own uncertainties.
Spatial resolution is a genuine limit. At Sentinel-1 IW resolution of roughly 5 x 20 m, a single pixel may contain several small buildings. Partial damage to one structure in a pixel may not produce enough coherence loss to cross a detection threshold. ICEYE spotlight at sub-1 m resolution reduces this problem but does not eliminate it, because the coherence estimation window still averages over multiple pixels to produce a stable estimate.
Satellize processes coherence-loss products over open Sentinel-1 archives and commercial SAR on client licence, with output formatted for GIS ingestion or direct overlay on command dashboards. The Tonga crop-estimation programme demonstrated the same multi-temporal baseline logic in an agricultural context; the conflict-damage application is more demanding because the stakes of a false positive or false negative are higher.
Typical figures
| Spatial resolution (Sentinel-1 IW) | 5 x 20 m (single-look); coherence estimation window typically 5 x 20 looks, effective output ~100 m |
| Spatial resolution (ICEYE / Capella spotlight) | Sub-1 m single-look; coherence estimation window degrades effective resolution to ~5–10 m |
| Revisit (Sentinel-1, both satellites) | 6-day exact repeat at equator; 3-day with Sentinel-1A and 1B combined on overlapping orbits at mid-latitudes |
| Revisit (ICEYE commercial tasking) | Daily or sub-daily with tasking; exact geometry depends on orbital slot and target latitude |
| Radar frequency | C-band 5.405 GHz (Sentinel-1); X-band ~9.65 GHz (ICEYE, Capella) |
| Minimum detectable damage proxy | Coherence drop of ~0.3–0.4 below pre-event baseline; individual building collapse may not be detectable at Sentinel-1 resolution without commercial SAR |
| Archive depth (Sentinel-1) | Global coverage from 2014; pre-event baseline construction requires 6–12 prior coherence pairs on same relative orbit |
| Latency (end-to-end, automated pipeline) | 2–4 hours post-acquisition for Sentinel-1; potentially under 2 hours for ICEYE with pre-positioned processing |
| Cloud and smoke penetration | Full: SAR is unaffected by cloud, smoke, haze, or darkness |
| Delivery formats | GeoTIFF coherence-difference raster, GeoJSON damage-proxy polygons, PDF assessment report |
Analytics Satellize can run
| Damage proxy map (DPM) | Interferometric coherence difference thresholding against multi-temporal pre-event baseline; published UNOSAT methodology | GeoTIFF raster with per-pixel coherence-change score; classified into low/medium/high anomaly bands |
| Urban damage fraction by grid cell | DPM intersected with building footprint layer (OpenStreetMap or commercial); fraction of footprint area flagged as anomalous | GeoJSON grid with damage-fraction attribute, suitable for humanitarian prioritisation or battle-damage triage |
| Temporal change series | Coherence computed for every available Sentinel-1 pair over a rolling 90-day window; anomaly flagged when single-pair coherence falls below seasonal mean minus 2 SD | Time-series chart per area of interest, CSV export, alert feed triggered on threshold breach |
| Ambiguity screening report | Coherence anomalies cross-checked against precipitation records (ERA5 reanalysis), NDVI change (Sentinel-2 when cloud-free), and land-cover classification to flag likely false positives | Annotated PDF with confidence rating per flagged zone; distinguishes high-confidence structural change from probable vegetation or moisture artefact |
| Before/after coherence mosaic | Multi-look coherence images for pre- and post-event epochs mosaicked and colour-coded for rapid visual interpretation | Web-tile layer or static GeoTIFF suitable for overlay in QGIS, ArcGIS, or command GIS dashboard |
| Rapid-onset alert (commercial SAR) | ICEYE or Capella tasking triggered by event notification; coherence computed against nearest prior acquisition on same geometry | Alert GeoJSON with damage-proxy polygons delivered within agreed SLA of acquisition; flagged for analyst review before dissemination |
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.