hotosm/venezuela_eq_2026

2026 Venezuela earthquake: AI building and damage assessment

Download the data from HDX: Venezuela - M 7.5 Earthquake - Damage Assessment Area (Validated)

AI-derived building footprints and building-level damage for the 24 June 2026 Venezuela earthquake, organized by area. Damage comes from up to three independent AI sources: HOTOSM fAIr (primary), Microsoft AI for Good Lab, and an OSU/CUNY Sentinel-1 radar product.

Interactive map: view it in map.

Building Damage Assessment Working Flow

fAIr detects building footprints on the pre-disaster image; the damage AI compares pre and post imagery. fAIr, Microsoft and OSU are fused on an H3 grid (more sources agreeing means higher confidence), then MapSwipe volunteers vote on each flagged area and the majority verdict decides the outcome: confirmed damaged areas are published (HF, HDX, uMap); everything the crowd does not confirm is uncertain and feeds back to the fAIr damage model for retraining. Per-building outputs are advisory; the H3 neighbourhood layers are the robust unit, and coverage is limited to where pre and post imagery overlap.

Areas

fAIr is the primary damage source for every area (all rows). Microsoft AI for Good Lab and the OSU/CUNY Sentinel-1 radar product are included where available. Building and damage counts are estimated; imagery and per-area details are in each area's README.

Area	Folder	Buildings	Microsoft	OSU	Human validated?	Publication
Caracas	caracas/	157,252	no	assessed, none detected	yes	uMap / HF
La Guaira	la_guaira/	18,430	yes	yes	yes	HDX / uMap
Catia La Mar	catia_la_mar/	24,495	yes	yes	yes	HF
Caraballeda	caraballeda/	5,366	yes	yes	yes	HF
Naiguata	naiguata/	2,438	no	yes	yes	HF
Morón	moron/	12,369	no	yes	yes	HF
El Junquito	el_junquito/	62,620	no	yes	yes	HF
Carayaca	carayaca/	2,356	no	yes	yes	HF

The dataset-level aoi.geojson indexes all areas assessed so far; each area folder also has its own aoi.geojson. Validation counts and MapSwipe project ids for every area are tracked in one table, in validated/.

Notes

• Caracas: inland; damage is largely interior or soft-story and not visible from nadir, so the layer is a screening aid for human review.
• La Guaira: the validated product, cross-checked by a MapSwipe campaign and published on HDX.
• Catia La Mar, Caraballeda: coastal.
• Naiguata: coastal, on very-high-resolution Vantor imagery.
• Morón: Carabobo coast, 120 km west of the eastern cluster; Vantor post, ESRI z18 pre.
• El Junquito: mountain corridor between Caracas and the La Guaira coast; Vantor WV03 post, ESRI z19 pre.
• Carayaca: Vargas coast, ridge-top town ~15 km west of La Guaira; Vantor WV03 post, ESRI z18 pre. No agent review pass; all damage classes flagged for human adjudication.

Damage sources and where to find them

Inside each area's damage_assessment/, every source is provided as polygons, points and H3 resolution-11 cells, each with a source.json for provenance and license:

• fair/ (primary): graded minor / major / destroyed.
• microsoft/: Microsoft AI for Good Lab, binary damaged, CC-BY 4.0.
• osu/: OSU/CUNY Sentinel-1 radar, binary damaged, experimental.
• combined/: fAIr and Microsoft merged for validation (H3 resolution-11). Each cell is tagged with which sources flag it (sources, n_sources) and carries a blank validated field for downstream review. OSU is excluded from the combined layer (radar, not confirmable on optical imagery).

Human validation: validated/ holds the merged, all-area MapSwipe results. All eight areas are covered. This is the single copy of validation results.

Every area has MapSwipe results. Download the underlying data from validated/.

Multi-source agreement across all areas and all three sources (fAIr, Microsoft, OSU) on one H3 grid: damage_analysis_multi_source/.

Each area coloured by the highest agreement tier any of its cells reach. Download the data and see the per-area breakdown in damage_analysis_multi_source/.

Locked column schemas for every GeoJSON in this dataset: SCHEMAS.md.

fAIr models

• Buildings: dinov3s-buildings (DINOv3 ViT-S/16 encoder, UperNet decoder), trained on hotosm/vhr-building-segmentation. Inference: full-resolution continuous sliding window (256 px window, 192 px stride), confidence threshold 0.4371.
• Damage: dinov3-damage-assessment (frozen DINOv3 ViT-L/16, UperNet decoder, ordinal 4-class head), trained on xBD / xView2.

Limitations

Every damage layer in this dataset is a screening aid for human review, not a verified damage census. Confidence is the mean probability of the assigned class over the footprint pixels; low values flag buildings for review, not certainty that a building is undamaged. Per-area README "Limitations" sections cover only what's specific to that area's imagery, terrain or damage signature.

License

This section applies dataset-wide; per-area READMEs defer to this section.

• fAIr damage and building layers (<city>/damage_assessment/fair/, <city>/buildings/, per-area and root aoi.geojson): the damage predictions come from a model trained on xBD / xView2 (CC BY-NC-SA 4.0), so all fAIr predictions in this dataset inherit non-commercial share-alike terms. Building footprints are shared under the same terms for consistency.
• Microsoft AI for Good Lab (<city>/damage_assessment/microsoft/): CC-BY 4.0, distributed via HDX.
• OSU/CUNY Sentinel-1 radar (<city>/damage_assessment/osu/): experimental, citation required. Cite Corey Scher (CUNY Graduate Center) and Jamon Van Den Hoek (Oregon State University). Uses Overture Maps building footprints (ODbL) and Copernicus Sentinel-1 data via ASF.
• Combined layers (<city>/damage_assessment/combined/): inherit the most restrictive term across their contributing sources, currently CC BY-NC-SA 4.0.
• MapSwipe validation (validated/): CC-BY-SA 4.0 (MapSwipe project data).
• Imagery:
  • Caracas, La Guaira, Naiguata, Catia La Mar, Caraballeda (post-event): Vantor Inc. 2026.
  • Morón (post-event): Vantor Inc. 2026 (LG05).
  • El Junquito (post-event): Vantor Inc. 2026 (WV03 scenes 10400100B979DD00, B040001100075610).
  • Carayaca (post-event): Vantor Inc. 2026 (WV03 scene 10400100B979DD00).
  • Morón, Caracas, El Junquito, Carayaca (pre-event): ESRI World Imagery, subject to ESRI Terms of Use.

See each source folder's source.json for the exact provenance.

Changelog

• v4.0 (2026-07-13) - MapSwipe validation: added El Junquito (projects 3211 and 3210, 2,844 areas, 63 confirmed) and Carayaca (project 3231, 46 areas, 5 confirmed). All eight areas are now human validated. Totals across the dataset: 7,774 areas reviewed, 706 confirmed. Merged by scripts/append_mapswipe_validation.py.
• v3.9 (2026-07-10) - Carayaca: added multi-source agreement map (carayaca/viz/carayaca_agreement_map.png); merged Carayaca into the dataset-level damage_analysis_multi_source/venezuela_multi_source_damage_h3_res11.geojson (14,656 → 14,763 cells, 8 areas covered).
• v3.8 (2026-07-10) - Carayaca: added per-area README.md and the six standard viz PNGs (viz/). Damage layer unchanged; a viz-only cloud + AOI-edge filter keeps cloud-adjacent false positives out of damage_gallery.png / damage_crops.png while damage_polygons.geojson still carries every row for downstream review.
• v3.7 (2026-07-10) - Carayaca: added OSU/CUNY Sentinel-1 layer (143 damaged structures, mean damage_probability 0.90). Prior "no OSU" was wrong.
• v3.6 (2026-07-10) - added Carayaca (2,356 buildings; Vantor WV03 post, ESRI z18 pre; fAIr only). No agent review pass; 243 damaged buildings (188 minor, 30 major, 25 destroyed) carry review=True for downstream human adjudication.
• v3.5 (2026-07-07) - MapSwipe validation: added Morón (project 3195, 34 confirmed of 199 areas). Total confirmed 638 across 6 cities.
• v3.4 (2026-07-07) - MapSwipe validation v2: added Catia La Mar (206 confirmed) and Naiguata (4 confirmed), expanded La Guaira (215 confirmed, up from 36), rewrote validated/venezuela_validated_*.geojson with sources derived per cell from the fAIr+Microsoft combined H3 layer. Total confirmed 604 across 5 cities; rebuilt validated/venezuela_validation_overview.png in the AOI-map style with a validated/not-validated legend.
• v3.3 (2026-07-03) - El Junquito: added OSU/CUNY Sentinel-1 layer (1,418 damaged structures). The area DOES have OSU coverage; prior "no" was wrong.
• v3.2 (2026-07-03) - El Junquito rebuilt on Bing Maps aerial z19 pre-imagery (0.29 m) instead of ESRI z18 (0.59 m); buildings 30,919 -> 62,620; severe (major+destroyed) 2,752 -> 4,094 post agent review.
• v3.1 (2026-07-02) - added El Junquito (30,919 buildings; Vantor WV03 post, ESRI z18 pre; fAIr only).
• v3 (2026-07-02) - locked column schemas (see SCHEMAS.md); moved every fAIr layer into each area's damage_assessment/fair/ subfolder; renamed combined files to <city>_fair_microsoft_combined_*; standardised H3 grids to resolution 11; added city and disaster columns to every row; dropped tm_project_id, damage_model and pre_cov columns from published fAIr layers; consolidated licence terms into this README.
• v2 (2026-07-02) - refreshed counts across all areas after re-running the pipeline on land-clipped AOIs with the current pre-imagery composites; added Morón as the sixth area.
• v1 (2026-06-28) - initial release: Caracas, La Guaira, Catia La Mar, Caraballeda, Naiguatá.