Rapid Response for June 2026 Venezuela Earthquakes An AI-based building damage assessment
In response to the June 2026 Venezuela earthquakes, the AI for Good Lab ran a building-level damage assessment over five post-disaster satellite scenes. This report summarizes the combined results across the full area of analysis.
Summary
What we did. Following the June 2026 Venezuela earthquakes, the Microsoft AI for Good Research Lab ran an AI-based building damage assessment on five post-disaster optical satellite scenes covering the La Guaira / Vargas coast, including Catia La Mar, La Guaira, and Caraballeda.
What we found. Across the full ~210 km² area of analysis we assessed 72,162 distinct buildings. 3,004 (4.2%) were obscured by clouds and could not be assessed; of the 69,158 cloud-free buildings, the model flagged 8,410 as damaged to some extent — about 12.0% of the buildings it could see. Damage is concentrated in Catia La Mar and the immediate La Guaira coastline.
Why it matters. Satellite-based assessments deliver a fast, wall-to-wall picture of where buildings were affected when ground access is limited. The results below are intended to help responders localize damage and prioritize recovery, relief, and rebuilding efforts.
Download. Combined results (footprints) · AOI
1. Area of analysis
The shaded boundary below is the combined area of analysis — the union of the five satellite scenes. Drag to pan and click the map to enable scroll-zoom; zoom in over the satellite basemap to inspect any area.
2. Headline results
The assessment covers five overlapping scenes; because the same building can appear in more than one scene, the totals below count each building once, so coverage is not double-counted.
| Measure | Buildings | Share |
|---|---|---|
| Buildings analyzed | 72,162 | 100% |
| Cloud-covered (not assessable) | 3,004 | 4.2% |
| Cloud-free (assessed) | 69,158 | 95.8% |
| Damaged — of all buildings | 8,410 | 11.7% |
| Damaged — of cloud-free buildings | 8,332 | 12.0% |
3. How the assessment was made
We create building-level damage assessments by training and then running an AI model on the post-disaster imagery. The model classifies every pixel as “building”, “damaged”, “cloud”, or “other” (background). These predictions are joined to Overture Maps building footprints, which represent the state on the ground pre-event, and distributed as a vector GeoPackage with the following per-footprint attributes:
| Attribute | Meaning |
|---|---|
id | The Overture Maps unique ID for each footprint |
damage_pct_0m | Percentage area of the footprint classified as “damage” |
damage_pct_10m | Percentage classified “damage” in a 10 m buffer around the footprint (denominator = footprint area, capped at 1) |
damage_pct_20m | Percentage classified “damage” in a 20 m buffer around the footprint (denominator = footprint area, capped at 1) |
damaged | 1 if damage_pct_0m > 0, else 0 |
unknown_pct | 1 if the building was covered by clouds, else 0 |
Because the five scenes overlap, a single building can appear in more than one response; how those overlapping observations are reconciled into the combined totals is described in §4. Building-level results for each individual response are available on OCHA's HDX platform, or via the direct links in §5.
4. Merging methodology
The five responses overlap along the coast, so the same building can be observed by up to three scenes. To report each building once, footprints are compared in a common projection (UTM 19N) and matched across responses by spatial overlap: any two footprints from different scenes whose intersection-over-union exceeds 0.5 are treated as the same building and linked into a single cluster. Each cluster keeps one representative, preferring a cloud-free observation; where the Catia La Mar scenes overlap, the next-day Planet read is preferred over the Vantor read. This collapses 97,439 raw footprints to 72,162 unique buildings (25,277, or 25.9%, removed as duplicates).
Agreement between responses. 25,094 buildings (34.8%) were seen by more than one scene, which lets
us check how consistently the responses call damage. We record a per-building uncertainty — the
spread (max − min) of damage_pct_0m across the scenes that saw it. The
responses agree exactly for 20,707 of these buildings; 4,387 (17.5%) disagree to some
degree, and for 3,893 (15.5%) one scene flags damage while another calls the building intact. The
single largest source of disagreement is the Catia La Mar overlap, where the Vantor (Jun 25) read flags
~2,600 buildings as damaged that the Planet (Jun 26) read calls intact — the reason that overlap is resolved
in Planet's favour. Every building in the merged dataset carries num_observations,
min/max_damage_pct_0m, and uncertainty, so these conflicts stay visible for
downstream review.
Merged dataset. ALL_AOIS_building_predictions_deduplicated.gpkg · combined area of analysis: valid_area_mask_union.geojson
uncertainty
(|max − min| of damage_pct_0m; 0 = the responses agree exactly, 1 = full
disagreement). Most overlapping buildings agree; a long tail disagrees completely.5. Individual responses
The combined assessment is made up of the five per-scene responses below. The table links each response's building-level dataset (Data), area-of-interest mask (AOI), interactive visualizer (Viz), and HDX dataset; the cards that follow give the per-response detail.
| Response | Sensor | Date | Footprints | Cloud | Damaged (of clear) | Links |
|---|---|---|---|---|---|---|
| Catia La Mar | Vantor | Jun 25, 2026 | 29,027 | 1,734 | 9,134 (33.5%) | Data · AOI · HDX |
| La Guaira (East) | BlackSky | Jun 25, 2026 | 5,411 | 0 | 112 (2.1%) | Data · AOI · Viz · HDX |
| Caraballeda | Vantor | Jun 26, 2026 | 10,392 | 3,239 | 587 (8.2%) | Data · AOI · Viz · HDX |
| Catia La Mar (East) | Planet SkySat | Jun 26, 2026 | 24,732 | 0 | 1,209 (4.8%) | Data · AOI · Viz · HDX |
| La Guaira & surrounding | Vantor | Jun 26, 2026 | 26,142 | 2,842 | 729 (3.1%) | Data · AOI · Viz · HDX |
Catia La Mar · Vantor · June 25, 2026
41 km² of Vantor imagery. Out of 29,027 footprints in the study area, 1,734 were obscured by clouds. Of the remaining 27,293 non-cloudy footprints, the model predicted 9,134 (33.5%) damaged to some extent.
La Guaira (East) · BlackSky · June 25, 2026
25 km² of BlackSky imagery over the eastern portion of La Guaira. Out of 5,411 footprints, 0 were cloud-obscured. The model predicted 112 (2.1%) damaged. We independently labeled 200 footprints, all of which could be reliably assessed; the model showed an estimated recall of 100% and precision of 57%. Extrapolated to the 3,072 buildings > 50 m², we estimate ~62 damaged buildings (2.0%), 95% CI [4, 120].
Caraballeda · Vantor · June 26, 2026
15 km² of Vantor imagery. Out of 10,392 footprints, 3,239 were cloud-obscured. Of the remaining 7,153 non-cloudy footprints, the model predicted 587 (8.2%) damaged. We independently labeled 200 footprints, 135 of which were uncloudy and assessable; the model showed an estimated recall of 92% and precision of 100%. Extrapolated to the 6,019 buildings > 50 m², we estimate ~535 damaged buildings (8.9%), 95% CI [248, 822].
Catia La Mar (East) · Planet SkySat · June 26, 2026
109 km² of Planet SkySat imagery over Catia La Mar (East). Out of 24,732 footprints, 0 were cloud-obscured. The model predicted 1,209 (4.8%) damaged to some extent.
La Guaira & surrounding · Vantor · June 26, 2026
73 km² of Vantor imagery over La Guaira and surrounding areas. Out of 26,142 footprints, 2,842 were cloud-obscured. Of the remaining 23,300 non-cloudy footprints, the model predicted 729 (3.1%) damaged. We independently labeled 300 footprints, 251 of which were uncloudy and assessable; the model showed an estimated recall of 88% and precision of 88%. Extrapolated to the 14,730 buildings > 50 m², we estimate ~469 damaged buildings (3.2%), 95% CI [151, 787].
6. Interactive visualizers
The interactive visualizer for each response is embedded below. Click a header to expand it; opening one collapses the others. Use “Open in new tab” for a full-screen view.
7. Validation against crowdsourced reports
As an independent check, we compared the model's damage predictions against citizen-submitted reports collected at terremotovenezuela.com. When scraped (afternoon 6/29/2026), the map held 925 reports, each labelled with a severity category (parcial, severo, total). 221 of these fall within our area of analysis. Because the reports are hand-placed pins, their locations are approximate — some points marked as severe or total damage do not line up with the destroyed building visible in the imagery (the median offset to the nearest footprint is 2.1 m, with a long tail).
We use these reports to estimate the model's recall as a function of buffer distance. For each report we look within a radius of X metres:
- If there are no building footprints within the radius, the report is dropped.
- If there are footprints but none are flagged damaged, it counts as a false negative.
- If there is at least one damaged footprint, it counts as a true positive.
Recall is then true positives ÷ (true positives + false negatives) — the share of reports near a mapped building that have a damaged prediction nearby. Conditioning on a footprint being present separates the damage classifier from footprint detection.
| Reported severity | 20 m | 50 m | 100 m |
|---|---|---|---|
| total (n=155) | 67% | 83% | 96% |
| severo (n=50) | 30% | 69% | 84% |
| parcial (n=16) | 40% | 63% | 94% |
| all (n=221) | 57% | 79% | 93% |
Read this as: among reports labelled total damage that sit within reach of a mapped building, 67% have a predicted damaged footprint within 20 m, 83% within 50 m, and 96% within 100 m. The 20 m figure excludes 14 total-damage reports that have no mapped building within 20 m; measured against all 155 total-damage reports it is 61%. By 50 m almost every report has a footprint nearby, so the distinction disappears.
8. Limitations
The assessment is automated and based on single-date satellite imagery. Damage labels for buildings seen by more
than one response can differ between scenes; the merged dataset captures this through its uncertainty
field (§4), and contested buildings should be reviewed before use.
This analysis is based on satellite imagery and automated assessments, which include false positives and false negatives and should not substitute for on-the-ground validation. This report was created to help; its purpose is to inform damage localization, recovery, relief, and rebuilding efforts. This report and data should not be shared or distributed without prior authorization.