Introduction
The rapid and reliable identification of infrastructure damage is a major concern for security and territorial monitoring actors. In response, Earth observation provides unique capabilities for regular monitoring. Sarimpact was developed in this context as part of the SECURE-SAT project funded by the University of Strasbourg and the Carnot Institute. The tool aims to strengthen rapid decision-support capabilities in crisis management and surveillance, with later applications in infrastructure monitoring.
1. Context
Considering SAR imagery merely as a substitute for high-resolution optical images would overlook one of its main strengths. A SAR sensor emits a wave and records the echo reflected back by the surface. It provides its own illumination, operates day and night, is unaffected by cloud cover, and measures quantities including the signal phase, which is sensitive to changes in the urban environment. Although it is generally considered more complex to interpret than optical imagery, its use is becoming increasingly widespread. As a result, a growing number of applications in the security domain are emerging (E. Colin 2025), and radar capability is becoming more integrated into the spatial component of public policy, notably in defense (IRIS 2026).
In open-source research, proposed solutions often rely on platforms such as Google Earth Engine (Bellingcat 2023), which distribute geospatial data. These methods may fail to exploit the depth originally offered by radar sensors, because so-called “complex” data are not always distributed there. This data type is therefore less accessible and requires expert capabilities to process. However, in the field of urban damage assessment, this depth provides real informational detail, which is necessary to improve result reliability.
For practical reasons, we make use of data from European Sentinel-1 satellites, available as open data (Copernicus 2026). The ground resolution of the products used is on the order of several meters. However, the phase of the signal, sensitive to a fraction of the wavelength, enables the SERTIT processing chains to measure or detect centimeter-scale changes in the surface (Applisat 2019, SERTIT 2025). This level of precision allows us to address urban built-up applications rigorously.
This article focuses on two test cases that illustrate the current capabilities and limitations of the tool. First, the double explosion at the Port of Beirut on 4 August 2020 is of interest because it is a sudden, large-scale event in a dense urban setting. The contrast between the stability of the urban fabric before the event and the damage caused by the explosion is therefore strong. Irpin, on the northwestern outskirts of Kyiv, was the subject of the second case study. The damage signature there is more diffuse because the city experienced a prolonged period of fighting and strikes at the beginning of the Russian invasion of Ukraine, from February to March 2022.
2. Method
Sarimpact’s methodology is based on comparing several pre-event radar acquisitions with a single post-event acquisition. For each AOI, the tool assembles a stack of so-called “complex” Sentinel-1 images. A minimum of three pre-event acquisitions and one post-event acquisition are finely co-registered. These images contain two types of information: amplitude, that is, the power of the radar echo returned to the sensor by ground targets, and phase, which indicates their distance from the satellite.
Phase 1: generation of the index stack: the N pre-event Sentinel-1 images and the post-event image are co-registered. Coherence is computed for image pairs, and calibrated γ0\gamma_0γ0 amplitudes for each acquisition are combined into eight indicators, including the coherence-drop z-score, which is the main detection signal. The stack is then orthorectified (Copernicus DEM, 10 m pixel) and clipped to the area of interest to produce the ImpactStack raster.
Phase 2: analysis: the z-score is denoised using an NLM filter, thresholded at 5σ (exceedance in VV or VH), restricted to built-up areas (GHSL layer), then vectorized into polygons (minimum area 200 m²); each polygon receives a probability index divided into three classes: possible, probable, and highly probable. Detections from ascending (ASC) and descending (DSC) orbits are merged into a single layer. The two phases are run separately for each orbit, up to the final ASC/DSC intersection step.
The main signal is coherence, that is, the similarity of phase between two radar passes over the same area. The pre-event images establish the usual behavior of each pixel: its average coherence and natural variability. The coherence measured between the last image before the event and the first image after the event is then compared with this reference: an abnormal drop, expressed in number of standard deviations (z-score), indicates a probable surface change. This normalization prevents naturally unstable areas, such as vegetation or water bodies, from being flagged and thereby reduces false alarms. Changes in amplitude, after speckle noise filtering, are retained as a complementary attribute.
Figure 2b. Principle for a pixel showing coherence loss, indicating the destruction state of the built environment around the Port of Beirut.
a) Chronology of the Beirut ASC test case (Sentinel-1): three pre-event images (5, 17, and 29 July 2020), one post-event image (10 August); port explosion on 4 August (red line).
All coherence pairs are formed against the most recent pre-event image (29 July): 5 July → 29 July and 17 July → 29 July in blue (pre-pre pairs), 29 July → 10 August in orange (co-event pair, the only one spanning the event).
(b) Principle on a pixel of destroyed built-up area at the Port of Beirut: pre-event coherences of 0.85 and 0.77 around a mean of 0.81, and co-event coherence of 0.10, much lower.
The drop is measured in standard deviations: z=(pre mean−co-event coherence)/σprez = (\text{pre mean} – \text{co-event coherence}) / \sigma_{\text{pre}}z=(pre mean−co-event coherence)/σpre, here z≈12.9z \approx 12.9z≈12.9. This normalization makes pixels comparable: the same absolute drop matters more for a pixel that is usually stable than for one that naturally fluctuates.
The z-score map, orthorectified on a 10 m grid, is then denoised by a filter and thresholded: only pixels whose coherence drop exceeds five times the usual variability are retained, and only within built-up zones identified by the European GHSL dataset (Pesaresi et al. 2024). Small clusters are removed. The retained pixels are grouped into polygons, to which a damage probability index is assigned.
The tool then runs this processing on both Sentinel-1 polarizations (VV and VH) and repeats it on both satellite viewing geometries, ascending and descending orbits, before combining the detections: depending on façade orientation relative to the sensor, each pass reveals damage that the other does not.
(a) Coherence-drop z-score map. (b) Thresholding: a pixel is retained if z>5z > 5z>5, with exceedance in VV or VH; the GHSL built-up filter applied at this stage is not shown. (c) Clusters with area < 200 m² (2 pixels at 10 m) are removed. (d) Each cluster becomes a polygon described by three statistics: mean z level, heterogeneity, and size. These are merged into a single score, yielding a probability index divided into three classes: possible, probable, and highly probable.
3. Results and validation
Sarimpact results were compared with independent references established by visual interpretation of very high-resolution satellite images: the Copernicus rapid mapping service analysis for Beirut (CEMS 2020) and the UNOSAT team analysis for Irpin (UNOSAT 2022).
A reference building is counted as detected when an area classified as damaged by Sarimpact intersects that building.
a) Case study 1 – Beirut:
Following the double explosion at the Port of Beirut on 4 August 2020, the Sarimpact method detected all 25 destroyed buildings and 89% of the buildings classified as damaged by CEMS (CEMS 2020) (Figure 4). We observed that integrating both ascending and descending acquisitions into the method improves detection. Thus, detection of buildings classified as “possibly damaged” increased from 77% to 90%.
CEMS annotated each building individually, totaling 12,237 buildings (633 damaged, 11,604 with “no visible damage”). Before the event, the city is stable and the coherence between two satellite passes remains high (Figure 5a, Figure 5b). On the image pair spanning the explosion, coherence collapses around the port (Figure 5c). It is this change, and its intensity, that serve as the detection marker (Figure 5d).
(a) Pre-event mean radar amplitude. (b) Pre-event mean coherence. (c) Co-event coherence; coherence collapses around the port. (d) z-score of coherence drop with detection contours from the fusion of ascending and descending orbits in red.
We therefore miss almost no damage: 100% of destroyed buildings are found, 89% of damaged buildings, and 80% of possibly damaged buildings. However, we also detect 23% of buildings classified as “no visible damage” by CEMS. Yet these 23% may include buildings whose damage is not visible in optical imagery.
To verify these buildings, we used open sources (OSINT): we searched for press articles or organization reports documenting the event and mentioning damage to buildings not detected by optical imagery. For example, in Beirut, the Grand Serail Governmental Building was affected by the blast (CNN 2020). The building was classified as having “no visible damage” by CEMS but was indeed detected by Sarimpact.
In cartographic representation, the probably affected sectors are symbolized by a gradient. The polygon edges are more transparent than the center, so as not to give the impression that these areas are completely destroyed, but rather that they contain buildings that are probably damaged (Figure 6).
b) Case study 2 – Irpin
During the fighting phase in Irpin, 1,060 buildings were damaged, including 115 fully destroyed, according to the United Nations Satellite Centre analysis (UNOSAT 2022).
UNOSAT mapped the impacted buildings, but not those with “no visible damage.” We therefore used the global Overture Maps database to map all buildings in the area of interest. This database is recent, whereas the damage in Irpin dates from 2022. Some destroyed buildings are therefore no longer present in Overture. We implemented a validation methodology:
- A building with an UNOSAT point within 10 meters is considered damaged; the point is sometimes offset relative to the Overture outline, or a point may represent a group of buildings.
- An UNOSAT point with no building within 10 m on the map (154 cases, buildings destroyed and then removed from maps or never mapped) is replaced by a 15 m radius disk.
- Everything else is assumed to be unaffected.
In Irpin, the results show that the tool detects 81% of destroyed buildings, 75% of severe damage, and 63% of moderate damage. The two orbits complement each other well: 62% in ascending orbit and 38% in descending orbit, for a total of 73% overall (Figure 7).
We find that 20% of buildings assumed to be undamaged overlap with a detection, comparable to Beirut’s 23% of false detections.
4. Discussion
The results show that the tool flags areas where the reference identifies no damage: in Beirut, nearly one quarter of buildings with no visible damage overlap with a detection (23%), and in Irpin, 20% of the supposedly intact built environment.
However, this apparent over-detection should be interpreted with caution. The reference data are based on vertical optical views, recognized through visual interpretation. Radar measures something different and is more sensitive to fine displacements on the order of a few centimeters. Thus, a building whose upper floor has collapsed inward, or whose structure has been breached under an intact roof, may not be recognized in optical photo-interpretation, whereas the coherence loss induced by its change of state may be detected by radar.
Between two satellite passes, tillage, vegetation growth, or snow cause coherence to drop even when no damage has occurred. The tool removes most of these false alarms by cross-referencing detections with a built-up layer (GHSL data). In Irpin, for example, more than one third of the pixels exceeding the detection threshold fall outside built-up areas.
This urban filter is not without bias. It may be incomplete (diffuse urban area, new construction, and, for events treated retrospectively, older destroyed buildings that are no longer present). Thus, in Irpin, 3.5% of the damaged buildings recorded had been removed from the GHSL urban layer and were therefore mechanically missed by the detection.
5. Conclusion and outlook
Sarimpact shows very promising results. However, moving toward operational use in crisis contexts requires more: the method must remain reliable across the diversity of image acquisition conditions, observed environments, and other factors affecting data quality.
Sarimpact is therefore entering a second phase, focused on improving result robustness and reducing over-detection.
Radar and high-resolution optical approaches are complementary. Their combination, potentially strengthened by artificial intelligence—for example with the open-source change detection chain Picanteo from CNES (Bellet et al. 2026)—is a natural direction for future development.
Article by Cléa Péculier (remote sensing engineer), Maxime Azzoni (remote sensing engineer, SAR specialist), and Hugo Regad (SECURE-SAT project manager).
Glossary
AOI: Area of interest
CEMS: Copernicus Emergency Management Service
GHSL: Global Human Settlement Layer
SAR: Synthetic-aperture radar
UNOSAT: United Nations Satellite Centre
**
z-score: deviation from the norm, measured in standard deviations. Here, it measures how much coherence dropped relative to its usual pre-event variability.
coherence: similarity between the phase of two radar images over the same area, between 0 and 1. High if the scene remained stable between the two passes, low if it changed.
phase: position of the radar wave in its cycle at the time of reception. Sensitive to changes on the order of a centimeter. Coherence measures its stability.
orbit: the satellite’s trajectory around Earth. The same area is observed during an ascending pass (northbound) and a descending pass (southbound), from two different viewing angles.
polarization: orientation of the emitted and received radar wave (vertical or horizontal). Sentinel-1 provides two combinations (VV and VH), which respond differently to scene changes.
Sources and references
Applisat 2019. Collective, L’imagerie satellitaire radar: comment ça marche et exemples d’application, Applisat, 2019, link
Bellingcat 2023. Ollie Ballinger, A New Tool Allows Researchers to Track Damage in Gaza, Bellingcat, 2023/11/23, link
CEMS 2020. CEMS, Explosion in Beirut, Lebanon, 05/08/2020, link
CNN 2020. CNN, Here’s what the damage looks like inside Lebanese prime minister’s headquarters, 04/08/2020, link
Copernicus 2026. Copernicus Data Space Ecosystem, Copernicus Browser, 2026, link
E. Colin 2025. Elise Colin, Le radar en observation terrestre: nouvelles perspectives, IPGP seminar, 2025/12/05, link
Pesaresi et al. 2024. Pesaresi, M., Schiavina, M., Politis, P., Freire, S., Krasnodębska, K., Uhl, J. H., Kemper, T. (2024). Advances on the Global Human Settlement Layer by joint assessment of Earth Observation and population survey data. International Journal of Digital Earth, 17(1), 18/03/2024, link
V. Bellet et al. 2026. Valentine Bellet, Dimitri Lallement, David Youssefi, Dawa Derksen, and Erwin Bergsma, La 3D par satellite, comment ça marche ?, CNES Campus de la donnée, 2026/01/16, link
IRIS 2026. Sandra Mezzadri, Building a European geospatial intelligence capability: mapping initiatives, identifying gaps, and outlining strategic options, IRIS (Ares Group), 2026/01, link
SERTIT 2025. Maxime Azzoni, Camille Caffarel, Séisme au Myanmar (2025) vu par le satellite Sentinel-1, SERTIT, 09/04/2025, link
UNOSAT 2022. Irpin Building Damage Assessment Overview Map, UNOSAT, 2022/09/15, link