Rapid post-disaster building damage mapping from satellite & drone imagery

After an earthquake, flood, storm or conflict, responders need to know within hours which buildings and districts are hit hardest. Analysts traditionally compare before/after imagery by eye, which takes days for large areas. Deep-learning change detection pre-grades every building and lets human analysts validate, correct and publish the map. The result is a faster, more consistent first picture; it does not replace field verification.

Problem & value

• Today: the EU Copernicus Emergency Management Service (CEMS) Rapid Mapping, UNOSAT and the International Charter "Space and Major Disasters" task satellites, then analysts digitise damage manually and publish grading maps, typically within hours to days of activation [1][2][3]. Large events (the 2023 Türkiye-Syria earthquakes) overwhelm manual capacity.
• AI value: triage of very large areas, consistent grading, rapid re-runs as new imagery arrives, and a prioritised list of where analysts should look first. Insurers use the same maps to estimate loss and route adjusters (portfolio exposure vs. damaged buildings).
• Benchmarks, not guarantees: the xBD dataset (xView2 challenge) provides pre/post imagery with building polygons and four damage classes across multiple disaster types [4]. Models localise buildings well; the intermediate classes (minor/major) remain the hardest and least reliable.

Variants

• Civil protection / humanitarian: CEMS-style grading maps, UNOSAT products, Red Cross / HOT.
• Insurance & reinsurance: event footprint intersected with portfolio, parametric triggers, adjuster dispatch (see PRED-010, AGENT-014, VIS-023).
• Drone-based local mapping: sub-10 cm orthomosaics and oblique views for facade and roof damage by fire brigades or municipalities (see VIS-003).
• Flood-specific: SAR flood extent plus building exposure, rather than structural grading (PRED-027).
• Conflict damage monitoring: human-rights and humanitarian documentation (e.g. UNOSAT); see the defence and ISR boundary under Risks.
• Private use: none realistic; consumers rely on public maps.

How it works

1. event trigger (shake map / flood alert / CEMS activation)
2. imagery tasking & acquisition (pre-event archive + post-event, optical and SAR)
3. co-registration & orthorectification (sub-pixel alignment is critical)
4. building footprints (open footprints or segmentation model)
5. pre/post change model (Siamese / two-stream segmentation or ViT, per-building damage class)
6. aggregation per building and per block, uncertainty score
7. analyst validation (human-in-the-loop, corrections, local knowledge)
8. grading map + stats published; field/drone/ground reports feed back

• Damage model: a Siamese encoder takes pre and post tiles; a head predicts building mask and a damage class. Typical winners of xView2 used U-Net/ResNet-style ensembles; newer work uses foundation-model encoders (e.g. geospatial foundation models) and SAR/optical fusion.
• Human-in-the-loop: analysts review low-confidence and mid-grade buildings; publication of any official grading map stays an analyst responsibility.
• VLMs can help with oblique drone photos and generate textual situation summaries but are not reliable for per-building structural grading from nadir satellite imagery.

Data preparation

Imagery is available quickly after a disaster, but usable training and reference data is not.

• Data and location: pre-event satellite, aerial and drone imagery in Copernicus, commercial providers or agency archives; building footprints and cadastre data in GIS systems and open data portals; insured-object lists in claims or policy systems; field reports on paper, phones and file shares.
• Typical gaps: mismatched imagery dates, clouds and viewing angles, missing or outdated building footprints, few labelled examples from European building types, inconsistent damage scales, personal data in addresses and ownership records, and access restrictions on sensitive sites.
• Preparation steps: assemble pre- and post-event pairs, co-register and tile them, align footprints, adopt a documented damage scale, harvest labels from past assessments and field reports, build an evaluation set from historical events, and separate or pseudonymise address data. Prepare the pipeline before an event, not during it.
• Tools: GDAL and QGIS, open model repositories, a labelling tool, a geospatial database.
• Share of effort: high, typically 50-60 percent.

Helpful building blocks: PREP-005, PREP-019, PREP-020, PREP-022 and PREP-025.

Implementation paths

Which path for whom

• Public (civil protection, municipalities, relief agencies): Path B on sovereign EU cloud, because procurement rules and sovereignty favour EU-hosted services (for example Copernicus-based pipelines), and open model APIs can be combined with existing GIS systems.
• Enterprise (insurers, reinsurers, loss adjusters): Path B, because the damage maps must feed claims and portfolio systems, and cloud APIs scale for peak events without owning GPU capacity.
• Developer (geospatial vendors, research teams): Path C, because open-weight segmentation models and own pipelines on HW3 give control and are the base for custom training in Path D.

Path A — Buy

• Public services (free): Copernicus EMS Rapid Mapping (activation by authorised users, EU) [1], UNOSAT [2], International Charter [3]. A municipality or NGO should request these rather than build.
• Commercial analytics: Maxar (Vantor) with Open Data Program imagery for major events [5], Planet, Airbus Defence and Space (EU), ICEYE (Finland, EU, SAR), Esri ArcGIS (damage assessment tools), and Sentinel Hub (Sinergise, Slovenia, EU; now part of Planet).
• Insurance: specialised cat-modelling and geospatial providers (e.g. Verisk, Nearmap) sell post-event damage analytics; check EU coverage and data handling.
• Choose A when: you need results for an event rather than owning a model.

Path B — Build on cloud APIs

• Imagery access: Sentinel Hub / Copernicus Data Space Ecosystem APIs (EU) [6], STAC catalogues, Planet and Maxar APIs, ICEYE tasking API.
• Compute: managed GPU inference on EU cloud regions; Google Earth Engine or Microsoft Planetary Computer for archive processing.
• Orchestration: event-triggered pipeline (Airflow / Argo), results to GIS (ArcGIS, QGIS Server, GeoServer) and a web map.
• Choose B when: you are an insurer, utility or agency with recurring needs and a GIS team.

Path C — Self-host (open source)

• Models: xView2 first-place solution and other open xBD baselines; segmentation frameworks (PyTorch, segmentation-models-pytorch, TorchGeo); geospatial foundation models (IBM/NASA Prithvi, Clay) as fine-tuning bases.
• Stack: GDAL, rasterio, STAC, QGIS, PostGIS; OpenDroneMap for drone orthomosaics.
• Hardware: HW3 (one 48 GB GPU) processes a city-scale event in hours; HW2 suffices for tiles of a few hundred km2 when images are tiled.

Path D — Custom training

Often needed: xBD is dominated by US and a few global events, with limited European building types, roof styles (e.g. masonry vs. timber) and sensor characteristics. Fine-tune with a few thousand locally labelled buildings from past CEMS products (published grading vectors are a useful label source), and evaluate by event, not random tiles, to avoid leakage. Pre-train on xBD, then adapt to your sensor and region.

State of the art (2026)

• xBD / xView2 remains the standard benchmark; later datasets (e.g. BRIGHT, a multimodal optical-SAR building damage benchmark) broaden the sensors and events covered [4][7].
• Optical + SAR fusion is the main practical advance: SAR works through cloud and at night, which matters for storms and floods; but SAR alone is weak for grading individual building damage.
• Foundation models (Prithvi, Clay and others) reduce labelled-data needs for new regions.
• Operational reality: CEMS and UNOSAT use AI as pre-screening with analyst validation.
• Limits and failure modes: nadir imagery sees roofs, not facades, so pancake or soft-storey collapse and interior damage are missed; off-nadir angle and shadows differ between pre and post images; smoke, cloud and debris confuse models; mid-class grades are inconsistent even between human analysts; models trained on one disaster type transfer poorly to another (earthquake vs. flood vs. wildfire). Reported scores on benchmarks should not be expected on a new event.

Shared resources & combinations

• Same building blocks (remote-sensing, segmentation, 3d-vision): VIS-021 general change detection, VIS-038 EO compliance, VIS-061 InSAR ground motion, VIS-049 EO verification (where present).
• Combinations: drone close-ups via VIS-003 and VIS-023; flood forecasting PRED-027 as trigger; portfolio impact PRED-010, claims handling AGENT-014 and fraud screening ANOM-003.
• Shared infrastructure: one STAC catalogue plus GPU inference service serves all EO use cases.

Effort, cost & KPIs

• Path A: hours to activate free services; commercial analytics per km2 or subscription.
• Path B: 3-6 months for an event-ready pipeline with GIS integration.
• Path C/D: 6-12 months incl. labelling, event-based validation and exercises; imagery licensing (VHR) is usually the main recurring cost.
• KPIs: time from event to first grading map, per-class F1 on held-out events, building-level agreement with analyst/field reference, share of buildings needing manual review, area covered per hour.

People & change

Involve management and the head of civil protection, claims or GIS early; they decide mandates and budget. Analysts and field teams are the end users, and local responders should help define the damage scale and validate maps. IT provides GIS integration and compute, data owners (surveying offices, imagery providers) grant access, and legal or compliance clarify liability for wrong assessments, imagery licences and procurement.

Data protection: building-level maps can be linked to residents, so involve the DPO; a DPIA is likely where maps are combined with addresses or claims data, and an AVV is needed with every cloud or imagery processor. Works council: depends. Analysis of buildings does not monitor staff, but case-handling systems that log user performance in claims or crisis teams would trigger §87(1) no. 6 BetrVG; inform the works council or the Personalrat under §90(1) no. 3 as the work process changes.

AI literacy: support analysts with training on confidence, resolution limits and typical misses such as roof-only damage (AI Act Art. 4). Pilot on a past event, compare with ground truth, then run a live exercise before any real deployment. Typical concerns: wrong maps misdirecting help or denying claims; address them by treating output as triage and prioritisation, with a human verifying before any decision about aid or payment.

See GUIDE-005, GUIDE-009, GUIDE-008 and GUIDE-004.

Risks, regulation & compliance

• EU AI Act: classification depends on use. Pure damage mapping of buildings is generally not an Annex III category, but systems used as safety components in critical infrastructure management or to dispatch/prioritise emergency first-response services (Annex III, point 5(d)) can be high-risk; assess per deployment. Insurance use for loss estimation of buildings is not the life/health risk pricing listed in Annex III, but check claims decisions affecting individuals.
• Art. 5: no prohibited practice applies to building damage mapping. Do not extend to identifying or tracking individuals from imagery; conflict-zone and military use belongs under VIS-051 and is outside the AI Act scope for military purposes (Art. 2(3)).
• GDPR: very-high-resolution imagery and drone footage can reveal people and, via addresses, homeowners; damage grades linked to a property become personal data. Use legal bases for emergency tasks (Art. 6(1)(e)), minimise, blur people and plates where published, and limit address-level release.
• Other rules: drone operation rules (EU 2019/947), national civil protection laws, imagery licence terms (commercial imagery often restricts redistribution), Copernicus data policy.
• Risks: wrong grades misdirect rescue or aid; false "undamaged" labels delay help; publishing address-level damage can enable looting or stigmatise areas. Mitigate with uncertainty maps, clear "preliminary" labelling, analyst sign-off and ground-truth feedback.

Getting started

1. Register as an authorised CEMS user (or find your national contact point) and review sample grading products from past activations.
2. Download xBD and run an open baseline on a past event in your region to learn failure modes.
3. Build a labelled test set of 500-1000 buildings from published CEMS vectors for your own events.
4. Set up a STAC-based imagery pipeline (Copernicus Data Space) and pre-event archive for key regions.
5. Run a tabletop exercise: trigger, imagery, model, analyst review, publication, within a target time.

Sources

1. Copernicus Emergency Management Service, Rapid Mapping — https://emergency.copernicus.eu/
2. UNOSAT (UNITAR) — https://unosat.org/
3. International Charter Space and Major Disasters — https://disasterscharter.org/
4. Gupta et al., xBD: A Dataset for Assessing Building Damage from Satellite Imagery (2019) — https://arxiv.org/abs/1911.09296
5. Maxar (Vantor) Open Data Program — https://registry.opendata.aws/maxar-open-data/
6. Copernicus Data Space Ecosystem — https://dataspace.copernicus.eu/
7. Chen et al., BRIGHT: multimodal optical-SAR building damage dataset — https://arxiv.org/abs/2501.06019
8. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj