Conflict-related building destruction detection with SAR
SAR coherence loss and backscatter change between pre- and post-event image pairs can locate collapsed urban structures in conflict zones where no ground access exists. The method has been applied publicly in Syria, Ukraine, and Gaza by UNOSAT and academic groups.
Sensors
- Sentinel-1 (ESA): C-band SAR at 5.6 cm wavelength. Interferometric Wide Swath mode gives 5 x 20 m ground resolution and 6-day repeat at the equator (12-day for a single satellite). Free, open archive from 2014. Cloud-independent. The standard workhorse for UNOSAT conflict-damage assessments because of its consistent acquisition geometry and long archive depth.
- ICEYE SAR constellation: X-band (3.1 cm) commercial smallsats. Spotlight mode delivers sub-1 m resolution; Strip mode around 3 m. Revisit can be tasked to less than 24 hours over a specific city. X-band is more sensitive to fine structural detail and surface roughness changes than C-band, but also more sensitive to vegetation and moisture variation, which can introduce false positives.
- Capella Space: X-band commercial SAR. Spotlight imagery at approximately 0.5 m resolution, suitable for individual building-level damage assessment rather than neighbourhood-scale statistics. Useful for corroborating or refining coarser coherence-change maps in high-priority areas.
- COSMO-SkyMed (ASI): X-band constellation of four satellites. Spotlight mode at approximately 1 m resolution with tasking revisit of 12 hours or better. Has been used in academic conflict-damage studies for its combination of high resolution and multi-temporal availability. Requires a data-access agreement with the Italian Space Agency or licensed distributors.
What a building facade gives the radar, and what rubble gives instead
An intact multi-storey building presents a dihedral corner reflector to a side-looking SAR: the vertical facade and the horizontal ground plane bounce energy directly back to the sensor. This double-bounce return is strong, stable, and coherent across repeat passes. Interferometric coherence, the correlation between the phase of two SAR acquisitions taken from nearly identical geometry, stays high for solid structures because the scatterers do not move between passes.
Collapse destroys both signals simultaneously. The facade disappears, removing the double-bounce geometry. The rubble field that replaces it is a chaotic mix of concrete chunks, rebar, glass, and dust at scales comparable to the radar wavelength, producing volume scattering that is bright but incoherent. Coherence drops sharply, often below 0.3 on a 0 to 1 scale, while backscatter intensity shifts in ways that depend on rubble geometry. Detecting both changes together, coherence loss and backscatter change, reduces false alarms compared with using either signal alone. This dual-criterion approach underpins the UNOSAT methodology applied in Syria from 2013 onward and subsequently in Ukraine and Gaza.
Temporal baseline: choosing the right image pair
The pre-event image should be acquired close enough to the conflict period that it reflects the building stock as it stood before hostilities, but not so close that incidental construction or demolition contaminates the baseline. For Sentinel-1, analysts typically use a stack of several pre-event acquisitions to establish a stable coherence baseline and filter out seasonal decorrelation from vegetation or soil moisture. Using a single pre-event scene is riskier.
The temporal baseline between the two images in the coherence pair matters. Very short baselines (6 days for Sentinel-1) maximise coherence for stable structures, making collapse stand out clearly. Longer baselines allow natural decorrelation to accumulate and can obscure the damage signal. In practice, UNOSAT and academic groups have used 6- to 12-day Sentinel-1 pairs for conflict mapping, accepting that some ambiguity remains in areas with dense vegetation or active construction.
Seasonal change is a genuine confound. Agricultural fields, orchards, and parks decorrelate naturally between growing seasons. A city surrounded by farmland can produce coherence-loss patches that have nothing to do with damage. Masking urban extent using a pre-existing building footprint layer, or restricting analysis to pixels classified as built-up in the pre-event period, is standard practice to reduce this noise.
The ambiguity problem: damage, demolition, or new construction?
SAR coherence loss is a physical signal, not a label. It records that something changed, not why. A building demolished by an airstrike, a building demolished by a municipal contractor clearing a site, and a building under active construction all produce similar coherence signatures. In active conflict zones, this ambiguity is not merely academic: authorities have been documented clearing rubble, demolishing damaged structures, and constructing new buildings in areas under their control, sometimes within weeks of fighting.
Distinguishing conflict damage from deliberate post-conflict demolition requires temporal sequencing. If coherence loss appears in an image pair that straddles a documented strike event, the attribution is stronger. If the loss appears months later, alternative explanations must be considered. Cross-referencing with optical imagery, when cloud cover and access permit, and with open-source geolocation of strike evidence, is standard practice in UNOSAT assessments. Honest conflict-damage maps carry a damage-likelihood score rather than a binary destroyed/intact classification.
Access constraints make satellite the only viable method, and also the hardest to validate
The reason SAR conflict-damage mapping exists as a discipline is straightforward: ground surveys are impossible or dangerous in active conflict zones. No insurance adjuster, no UN field team, and no journalist with a clipboard can systematically assess building stock in Mariupol or northern Gaza while fighting continues. Satellite is not a second-best option here. It is the only option for systematic, city-scale assessment.
That same access constraint makes validation difficult. Ground-truth data, meaning confirmed building states from field surveys or high-resolution optical imagery with clear sight lines, is sparse and often delayed by months. Published studies on Syria and Ukraine have used post-conflict field surveys, crowdsourced damage reports, and very high resolution commercial optical imagery to validate SAR-derived damage maps, typically finding overall accuracies in the 70 to 85 percent range at the building level, with higher accuracy at neighbourhood scale. Errors tend to cluster in areas of mixed damage severity, where partially damaged structures produce ambiguous backscatter signatures.
The honest position is that SAR conflict-damage maps are best used as a spatial prioritisation tool, directing limited field resources and humanitarian response to the highest-probability damage areas, rather than as a definitive count of destroyed structures.
Operational workflow and what a delivered product actually contains
A standard conflict-damage assessment begins with identifying a consistent set of pre-event Sentinel-1 acquisitions over the target city, selecting the appropriate beam mode and orbit direction to maximise coverage and geometric consistency, and computing mean coherence across the pre-event stack. Post-event coherence is then computed for each new acquisition pair, and the change relative to the baseline is thresholded and classified.
Backscatter change analysis runs in parallel: pixels showing both coherence loss and a significant backscatter shift are assigned higher damage confidence. The output is typically a raster of damage likelihood and a vector layer of affected building footprints, derived by intersecting the raster with a pre-existing footprint dataset such as OpenStreetMap or a national cadastre. Deliverables from UNOSAT assessments are published as GeoTIFF and shapefile packages with accompanying PDF reports; the same formats are standard for operational use.
Satellize can run this workflow on Sentinel-1 open data, with commercial tasking from ICEYE or Capella added for priority areas requiring sub-metre resolution or sub-24-hour revisit. The analytics layer draws on the same published coherence-change methodology used in UNOSAT's publicly documented Syria and Ukraine programmes.
Resolution floors, coverage ceilings, and what the method cannot see
Sentinel-1 at 5 x 20 m cannot resolve individual buildings in dense urban fabric. It works at the level of city blocks or clusters of structures. A single collapsed building in a row of intact ones will not register unless its footprint occupies several pixels. ICEYE and Capella at sub-1 m can resolve individual structures, but their swath widths are narrow, typically 5 to 25 km in spotlight mode, making city-wide coverage expensive and time-consuming to assemble.
Partial damage is systematically underdetected. A building that has lost its upper floors but retains its ground-floor structure may still produce a double-bounce return. Fire damage without structural collapse produces almost no SAR signal change. Underground damage, basements, tunnels, and buried infrastructure, is invisible to all SAR frequencies in common use.
Cloud cover does not affect SAR, which is the central advantage over optical methods in conflict zones where access is already constrained. Atmospheric conditions, including smoke from fires and dust from explosions, are similarly transparent to C- and X-band radar.
Typical figures
| Spatial resolution (Sentinel-1 IW mode) | 5 x 20 m ground range |
| Spatial resolution (ICEYE / Capella spotlight) | 0.5 to 1 m |
| Revisit (Sentinel-1, single satellite) | 12 days; 6 days with two satellites in constellation |
| Revisit (commercial SAR, tasked) | Under 24 hours over a specific target |
| Radar frequency | C-band (5.4 GHz, Sentinel-1); X-band (9.6 GHz, ICEYE, Capella, COSMO-SkyMed) |
| Archive depth (Sentinel-1) | From 2014; consistent IW acquisitions over most conflict zones from 2015 |
| Minimum detectable damage unit | Neighbourhood block at C-band; individual large building at X-band spotlight |
| Cloud and smoke penetration | Full: SAR is unaffected by cloud, smoke, or dust |
| Typical damage-map accuracy (building level) | 70 to 85 percent, per published Syria and Ukraine validation studies |
| Delivery formats | GeoTIFF (coherence and backscatter change rasters), GeoPackage or shapefile (building footprint damage classification), PDF assessment report |
Analytics Satellize can run
| Coherence-change damage likelihood raster | Interferometric coherence differencing between pre-event stack mean and post-event pair, dual-polarisation (VV+VH) where available | GeoTIFF classified into four damage likelihood bands (low, moderate, high, very high), updated per new Sentinel-1 acquisition cycle |
| Building footprint damage classification layer | Intersection of coherence-change raster with OpenStreetMap or supplied cadastral footprints; per-footprint damage score aggregated from pixel statistics | GeoPackage with per-building damage score and confidence flag, suitable for loading into QGIS or ArcGIS |
| Temporal damage progression series | Multi-date coherence-change stack producing a timeline of damage accumulation across the conflict period | Animated GIF and CSV time series of affected area by district, with acquisition dates and baseline parameters documented |
| High-confidence damage confirmation layer | Dual-criterion filter requiring both coherence loss below 0.3 and backscatter change exceeding 3 dB, reducing false positives from vegetation or moisture | Separate GeoTIFF and vector layer of highest-confidence collapsed structures, recommended for humanitarian resource allocation |
| Commercial SAR spotlight overlay for priority blocks | ICEYE or Capella sub-1 m imagery tasked over blocks flagged as high-damage by Sentinel-1 screening; visual and automated building-state classification | Annotated image chips with structure-level damage labels, delivered as georeferenced TIFF with accompanying spreadsheet |
| Ambiguity flag layer (damage vs. demolition vs. construction) | Temporal sequencing of coherence-loss events cross-referenced against documented strike dates and optical imagery availability windows | Vector layer with per-polygon temporal attribution notes and confidence rating, formatted for inclusion in humanitarian situation reports |
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.