Three questions
How good are the assessments at flagging damaged buildings?
Not on their own: no single assessment supports acting on an individual flag.
Do they rank neighbourhoods well enough to guide triage?
Yes, inside a known disaster zone.
How should they be used together?
Count agreement between them rather than choosing one.
Assessments flagging the building
●none
●one
●two or three
●four or more
●Copernicus EMS recorded damage
●flag with no Copernicus EMS record
Copernicus EMS recorded damage per cell (~0.7 km²)
lessmore
▬Microsoft
▬IMPACT
▬OSU
▬UH
▬LIST
▬Copernicus EMS mapped areas (dashed)
▮where all overlap (~61 km²)

Every assessment ranks neighbourhoods fairly well…and combining them ranks best

How well each assessment's ranking of the neighbourhood cells matches the Copernicus EMS ordering (longer is better).The same comparison, now with the agreement count added: cells ranked by how many buildings have three or more assessments in agreement.
Inside the known disaster zone, each one on its own is a usable guide to where first.Beyond ranking, can these assessments build a shortlist of priority sites? Asked for the 20 worst-hit cells, counting where five of the six agree gets 15 right, against 13 for the best single assessment.

The more the assessments agree, the more confidence we can have in each building flagged

We can fine-tune the number of assessments required to agree before a building counts as flagged. As we raise that number we flag fewer buildings, but become more confident in each flag: the share with Copernicus-recorded damage underneath climbs.
The threshold is a choice responders can make to fit the stage of the response: wide coverage in the first hours, higher confidence once teams are being sent building by building.
In short Requiring agreement trades coverage for confidence: from 37,346 flags, 5% with Copernicus-recorded damage, to 211 at 79%.

How much of what an assessment flags has recorded damage underneath? (recorded in our mixed-reference dataset)

The share of each product's flagged buildings with reference-recorded damage underneath. The dot counts only buildings Copernicus EMS graded damaged or destroyed, the lower bound. The bar extends to the upper bound, which also credits Copernicus's possibly-damaged points and flags in cells the MapSwipe crowd judged damaged.
0%50%100% of flagged buildings
24 Jun · M7.5

How satellite damage assessment can effectively support relief operations

LESSONS FROM THE 2026 VENEZUELA EARTHQUAKE · UN OCHA CENTRE FOR HUMANITARIAN DATA, DATA SCIENCE TEAM

On the night of 24 June 2026, two earthquakes, forty seconds apart, struck the coast of Venezuela. Within days, at least eight organisations had turned satellite imagery into assessments of building damage.

Advances in AI and satellite imaging have made assessments like these far faster and cheaper to produce, and more organisations than ever are producing them. The abundance is new, and the humanitarian system is still learning how to make use of it operationally. This story evaluates the assessments produced after the Venezuela earthquake and asks what responders can, and cannot, do with them.

Scroll ↓
The response begins

In the immediate aftermath of an earthquake, responders need to know where the damage is. Satellite-based damage assessments are often used to attempt to quickly provide an answer.

After the 24 June earthquake in Venezuela, those assessments began to arrive within days. OCHA's Data Science team at the Centre for Humanitarian Data tracked them in support of the UN Assessment and Analysis Cell.

Search-and-rescue teams at a collapsed building, Venezuela, 30 June 2026. © UNOCHA/Enrique Luzardo
Part one

Timeline of satellite assessments

Day 1 · Copernicus EMS
The Copernicus Emergency Management Service (EMS), the European Union's emergency mapping service, was activated the night of the earthquake. Upon activation, analysts begin to define areas to prioritise for mapping, the dashed outlines here. They manually map damage from very-high-resolution imagery as it becomes available. The final building-level products were released on 6 July.
Day 2 · the first model
Two days after the earthquake, the first modelled assessment arrived: a building-level classification based on high-resolution optical imagery from Microsoft AI for Good. It flags 8,410 buildings in the coastal strip area.
Day 4 · radar
IMPACT Initiatives delivers an analysis based on synthetic aperture radar (SAR). Radar works through cloud and darkness, covers a far larger area, and measures change rather than appearance. It flags 81,446 buildings.
Day 5 · a third view
A team from Oregon State University, working with NASA, delivers another radar-based analysis, with different methods and a different area of interest: 58,870 buildings.
Day 7 · two more
The UNEP/OCHA Joint Environment Unit (UN Environment Programme and OCHA) delivers another radar-based analysis, which also turns detected damage into an estimate of debris tonnage (96,046 buildings). Later that week, another assessment arrives from a collaboration between the World Food Programme (WFP), the Luxembourg Institute of Science and Technology (LIST) and the European Organization for Nuclear Research (CERN) (163,225 buildings).
Additional assessments
Meanwhile, the UN Assessment and Analysis Cell becomes aware of a further analysis by an academic group, UH SAIL at the University of Houston: an AI model that classifies buildings from optical imagery across eight coastal areas. 76,378 buildings flagged. Two further products arrived in the same window but are not evaluated here. HOT's fAIr published no analysed extent. DISHA (UN Global Pulse) covers less than half of the shared comparison area; including it would have halved the ground on which all six assessments are compared.
The information coordination issue

After a disaster, responders can end up with several damage assessments, in PDFs, shapefiles and dashboards, each using its own method and area, with counts ranging from 8,410 to 163,225 damaged buildings, a factor of nineteen, and no clear sense of which one to act on.

In those first days a coordinator has to decide where search-and-rescue teams go and who needs help first. Answers this far apart cannot all be right, and there is no time to find out which is best to use.

Within days the OCHA Data Science team stood up an ad hoc viewer so responders could at least see which assessment had covered which ground.

Part two

Making sense of these different assessments

Three questions
This evaluation aims to answer three questions an emergency or search-and-rescue coordinator would ask:
  1. How good are the assessments at flagging damaged buildings?
  2. Do they rank neighbourhoods well enough to guide triage?
  3. How should they be used together?
Where they can be compared
The outlines show where each assessment actually looked: areas spanning three orders of magnitude, and no two alike. The dashed white line is where the Copernicus EMS analysts worked. The small red strip is the one place all of them overlap: 61 km² and 57,241 shared buildings, where the assessments can be compared fairly, building by building. Even here, on identical buildings, the six assessments count anywhere from 5,443 to 25,882 damaged.
The reference dataset

Inside that shared strip, we measured every assessment against a reference dataset built from three sources that became more fully available in the weeks after the earthquake: 1. manually mapped damaged buildings (Copernicus EMS), 2. crowd-sourced volunteer review (MapSwipe, 400+ volunteers), and 3. reports from field teams on the ground (ChatMap).

Copernicus EMS-recorded potential damage (teal dots on the map: 2,064 buildings) comprises the main part of the reference.

For every assessment we then asked: of the buildings it flags, how many had damage recorded in our constructed reference dataset? And of the buildings where the reference dataset recorded damage, how many does it find?

Note: this reference is not a complete, ground-verified count of damaged buildings. Copernicus experts mapped what they could see from above in very-high-resolution optical imagery, and optical analysis is known to miss certain kinds of building damage. The true extent of damage is therefore greater than the Copernicus component of the reference indicates.

Flagging damaged buildings
A team following one assessment's flags on the ground would have to visit 8 to 15 flagged buildings to find one that Copernicus EMS recorded as damaged, depending on the assessment.
Ranking neighbourhoods
Zoom out, from buildings to neighbourhood cells of about 0.7 km² each. The colour shows how much damage Copernicus EMS recorded in each cell.
What the ranking can and cannot do
At this scale, every one of the six assessments ranks the cells fairly well on its own. That matters: in the first 72 hours, search-and-rescue teams need above all to narrow down the worst-hit areas quickly. One caveat carries through: this holds inside an area already known to be badly hit. Ranking orders that zone. It cannot say where that zone is, and it cannot point to a building. Across the wider ground each assessment covered, the rankings weaken.
Using the assessments together
In real time, as products arrive during a crisis, there is no definitive answer on which one to use or how accurate it will be. So rather than choosing, we count: for every building, how many of the six assessments flag it.
Combining the assessments
This agreement analysis indicates that there is promise in a cross-product composite that could give responders more meaningful, actionable building-level analysis. As we add more product-agreement requirements to the output, our confidence in each building flag grows and noise is lowered. On the map: the buildings where four or more assessments agree (3,123 of them). It needs no training data and no waiting; it works as soon as the second assessment arrives, and it improves with every one that follows, using more of the available information instead of betting on one source.
Combining also ranks best
Ranking neighbourhoods by how many assessments agree beats the ranking from any single assessment, a promising way to guide interpretation and to decide where to allocate resources first.
Using them together
Combining builds confidence both in the individual buildings flagged and in the ranking of where to go first, and responders do not have to choose one product when, in real time, the accuracy of each is unknown.
UN relief chief Tom Fletcher (centre) at the site of a collapsed building in La Guaira, Venezuela, 7 July 2026. © UNOCHA

Conclusion

1 · Identifying building-level damage

Compared with our mixed reference dataset, every product flags far more buildings than the reference records as damaged, so acting on a single flagged building from any one product could be inaccurate.

2 · Ranking neighbourhoods for triage

In the first 72 hours, search-and-rescue teams have a pressing need to quickly narrow down the worst-hit areas, and this is where the products are strongest. Within a subset area already known to be heavily damaged, they rank neighbourhoods for triage fairly well (three of the six assessments agree closely with the Copernicus EMS ordering). Evaluated against the full geographic scope each product delivered, their ranking ability declines markedly.

3 · Using the products together

Combining products by counting agreement improves performance and lets stakeholders set the threshold that defines the balance of coverage against confidence for their operational needs at different stages of a response. This simple method shows promise and substantial improvement at both the building level and the neighbourhood-ranking level. And because, in real time, there is still no definitive answer on when to use which product or how accurate it will be, compositing is a way to cut back that uncertainty and use more of the available information for better estimates.

What this response demonstrated

Advances in AI and satellite imaging have made damage assessments cheaper and faster to produce, so future disasters will likely bring more of them, from more organisations. This crisis and the emergency response demonstrated the need for two things:

  1. A shared, live picture of which analyses have been produced and where they looked.
  2. An interpretive layer to help key stakeholders make sense of and evaluate new maps and analyses as they become available.

Recommendations

For responders and response coordinators:

  • Acting on a single flagged building from any one product could be inaccurate.
  • Trust the neighbourhood ranking within an already-identified damage zone, where it is most useful in the first days.
  • Rather than choosing one product, evaluate agreement between them. Requiring more products to agree sharply raises the precision of the damage classifications, at the cost of finding fewer of the damaged buildings. The number required to agree can therefore be tuned to strike a balance between finding more of the damage and being confident in each flag.
  • Keep one shared, live picture of what has been analysed where.

For providers: publish the analysed extent and per-building acquisition metadata, and keep a channel open for feedback during the activation while reprocessing can still help.

For evaluators: repeat this evaluation in other crises, and make it routine. That needs two things: a minimum specification for any contributed assessment, meaning its analysed extent, its metadata and a common building layer to score against, and a shared platform providers deliver into, so comparison starts the day the first assessment arrives rather than weeks later.

Limitations

This is one earthquake, in one dense coastal setting, and the shared comparison area is only 61 km², with most of the reference damage in a single area. The references have gaps of their own: the crowd review saw only places the products had flagged, and judged them by area rather than building by building, and the field reports are positives-only with no defined survey area.

The full method, every sensitivity analysis and the per-product results are in the technical brief.