evidoria

← Back to browse

Good practice Imported

AI Roof Classification Helps Dominica Rebuild Housing Resilience After Hurricane Maria

Dominica · Roseau · See the Dominica profile · See the Roseau profile

Evidence: Observational / pre–post Top 12% 80/100 · Ask Evidence Copilot about this practice

World Bank/GFDRR fused CNNs with post-disaster drone and satellite imagery to classify roof type and material for 8,345 Dominica and Saint Lucia buildings, reaching F1 scores of 0.92–0.93, to help target hurricane-vulnerable housing for resilience investment.

8,345 buildings
Buildings classified (pilot) (2018–2023 pilot)
0.93 F1 score
F1 score, roof type (satellite + LiDAR) (pilot)
0.92 F1 score
F1 score, roof material (satellite + LiDAR) (pilot)
0.86 F1 score
F1 score, roof type (drone imagery alone) (pilot)
0.90 F1 score
F1 score, roof material (drone imagery alone) (pilot)
AI Roof Classification Helps Dominica Rebuild Housing Resilience After Hurricane Maria

Details

Maturity
Pilot
Promoter
World Bank Group – Global Facility for Disaster Reduction and Recovery (GFDRR), Digital Earth Partnership, with the Government of the Commonwealth of Dominica and the Government of Saint Lucia
Period
2018–2023
Keywords
disaster risk management, geospatial AI, housing, digital public infrastructure

Context

Hurricane Maria struck Dominica in 2017 and damaged an estimated 90% of the island's housing stock, exposing how little systematic, up-to-date data governments had on which roof types and materials were most vulnerable to storm damage.

Objectives

As part of the World Bank's GFDRR-funded 'Digital Earth Project for Resilient Housing and Infrastructure in the Caribbean,' undertaken with the governments of Dominica and Saint Lucia, the project aimed to automatically classify roof type and material to help target hurricane-vulnerable housing for resilience investment.

Activities

Researchers trained convolutional neural networks (ResNet50, InceptionV3, EfficientNet-B0 feature extractors feeding Logistic Regression, Random Forest and SVM classifiers) on RGB aerial imagery fused with airborne LiDAR-derived surface models, classifying roof type and material across 8,345 annotated buildings sampled from 80 tiles on the two islands. The project also trained local government staff to operate drones and manage large geospatial datasets.

Results

Combining satellite-resolution imagery with LiDAR achieved F1 scores of 0.93 for roof type and 0.92 for roof material; lower-cost drone imagery alone still reached F1 scores of 0.86 and 0.90 respectively.

Conclusions

This remains a technical pilot covering sampled tiles rather than the full national housing stock, and no public report yet documents the data being used in an actual reconstruction, building-code or insurance decision.

Implementation
Indicative cost
Medium (€50k–€500k) — Not publicly disclosed; combines satellite imagery, airborne LiDAR capture and drone imagery acquisition plus local staff training.
Time to results
Long (> 3 years) — Project period 2018–2023, following Hurricane Maria's 2017 damage to an estimated 90% of Dominica's housing stock.
Staffing & skills
World Bank Group / GFDRR Digital Earth Partnership research team, Governments of Dominica and Saint Lucia as partner administrations, Local government staff trained to operate drones and manage geospatial datasets

Conditions for success

  • Fusing higher-resolution satellite/LiDAR imagery with lower-cost drone imagery to balance accuracy and affordability
  • Training local government staff on drone operation and geospatial data management so agencies can keep the housing intelligence current themselves
  • Openly publishing methodology, training data references and code (arXiv preprint and public GitHub repository) for scrutiny and reuse

Common failure modes

  • Pilot covers only sampled tiles (8,345 buildings across 80 tiles) rather than full national coverage
  • No public report yet documents the classified data being used in an actual reconstruction, building-code or insurance decision
  • Long-term institutional ownership after the project ends is not documented

Where it fits

Governance type
national government with multilateral technical partner
Scale
two-island pilot (Dominica, Saint Lucia)
Income level
upper-middle income

Replication kit

Reusable artefacts from this practice — as published by their sources.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful