evidoria

← Back to browse

Good practice Imported

AFAD's AI-Powered Building-Damage Assessment — Microsoft's AI for Good Lab Mapped 2023 Earthquake Destruction from Satellite Imagery

Türkiye · Kahramanmaraş · See the Türkiye profile

Evidence: Observational / pre–post Top 25% 73/100 · Ask Evidence Copilot about this practice

Within three days of the 2023 Kahramanmaraş earthquakes, Microsoft's AI for Good Lab used deep learning on satellite imagery to flag 3,849 damaged buildings for AFAD, the World Bank, IFRC and WFP — a one-off crisis collaboration, not an institutionalised AFAD capability.

3,849
Buildings flagged as damaged/destroyed (Feb 2023 (3-day assessment))
7.44 %
Visible buildings damaged in Kahramanmaraş (Feb 2023)
>0.99
Independent academic AUC for collapsed-building detection (post-2023 academic study)
AFAD's AI-Powered Building-Damage Assessment — Microsoft's AI for Good Lab Mapped 2023 Earthquake Destruction from Satellite Imagery

Details

Maturity
Pilot
Promoter
Türkiye Disaster and Emergency Management Presidency (AFAD), Ministry of Interior, with Microsoft AI for Good Lab
Period
February 2023 (3-day rapid assessment)
Keywords
disaster response, remote sensing, emergency management, computer vision

Context

Two earthquakes (magnitude 7.8 and 7.7) struck south-central Türkiye on 6 February 2023, killing more than 40,000 people and devastating cities including Kahramanmaraş, Gaziantep, Hatay and Adıyaman. Ground surveys across a disaster zone this size take days to weeks, but search-and-rescue teams and aid agencies needed to know quickly which buildings had collapsed.

Activities

Microsoft's AI for Good Lab worked with Türkiye's Disaster and Emergency Management Presidency (AFAD), under the Ministry of Interior, to run deep-learning models over pre- and post-earthquake satellite imagery from Planet Labs (50cm resolution) and Maxar Technologies (30cm resolution), covering the first three days after the disaster.

Results

The models flagged 3,849 damaged or destroyed buildings across four of the hardest-hit cities, with Kahramanmaraş showing 7.44% of visible buildings damaged. The estimates were shared with AFAD's own response teams as well as the World Bank, the International Federation of Red Cross and Red Crescent Societies, and the World Food Programme to help direct relief. Independent academic teams later applied comparable deep-learning approaches to the same earthquake sequence and, benchmarking against ground-truth surveys, reported area-under-curve accuracy scores above 0.99 for detecting collapsed buildings in some study areas.

Conclusions

This was a rapid, ad hoc collaboration between a private AI lab and a government disaster agency for a single event, not a system AFAD owns, operates or has written into its standing emergency-management doctrine. There is no public evidence Türkiye has since adopted satellite-based AI damage assessment as a routine, institutionalised capability.

Implementation
Indicative cost
Low (< €50k) — No cost figures are disclosed; the assessment relied on existing commercial satellite-imagery subscriptions and Microsoft's own AI for Good Lab resources provided as a pro-bono crisis response rather than a funded government programme.
Time to results
Short (< 1 year) — Time-boxed to the first three days after the 6 February 2023 earthquakes; no evidence of continued or repeated use since.
Staffing & skills
Microsoft AI for Good Lab (model development and imagery analysis), AFAD, Türkiye Disaster and Emergency Management Presidency (response-team recipient)

Conditions for success

  • Rapid access to pre- and post-event high-resolution commercial satellite imagery (Planet Labs, Maxar)
  • Pre-existing relationship between Microsoft AI for Good Lab and a government disaster agency enabling fast collaboration
  • Public methodology report and open-sourced toolkit enabling external review and reuse

Common failure modes

  • One-off crisis collaboration rather than an institutionalised AFAD capability — no public evidence of routine adoption since 2023
  • Detections used for response decisions were not independently ground-truthed by AFAD at the time
  • System was designed, hosted and run by Microsoft with no disclosed AFAD-side oversight or audit framework

Where it fits

Governance type
private AI lab collaborating ad hoc with a national disaster agency
Scale
four cities
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