evidoria

← Back to browse

Good practice Imported

Giga School Mapping — AI Satellite Analysis Finds 23,100 Missing Schools Across Africa and Asia

Kenya · Nairobi · See the Kenya profile · See the Nairobi profile

Evidence: Descriptive / self-reported Top 43% 53/100 · Ask Evidence Copilot about this practice

UNICEF and ITU's Giga initiative trained CNN classifiers on satellite imagery to find schools missing from national databases, adding 23,100 unmapped schools across 8 countries including Kenya and Rwanda — accuracy fell to ~57% in the poorest, most remote areas.

23100 schools
Previously unmapped schools added to national registries (2020–2021)
8 countries
Countries covered (2020–2021)
71 million tiles (60cm resolution)
Satellite map tiles scanned
18 million supertiles
Supertiles processed (25 hours)
0.85 F1 / 0.94 AUC
Global model detection performance
57 %
Model accuracy in poorest, most remote areas
Giga School Mapping — AI Satellite Analysis Finds 23,100 Missing Schools Across Africa and Asia

Details

Maturity
Scaling
Promoter
UNICEF Office of Innovation & ITU (Giga initiative), with Development Seed
Period
2020–2021 (core mapping campaign); ongoing
Keywords
education infrastructure planning, satellite imagery AI, school connectivity mapping, GIS, education management information systems

Context

Most low- and middle-income countries do not have a complete, accurate record of where their schools physically are. UNICEF's Office of Innovation and the International Telecommunication Union (ITU) built the Giga initiative's school-mapping tool to close that gap.

Objectives

Convolutional neural network classifiers (based on Xception and MobileNetV2 architectures), developed with technical partner Development Seed, were designed to scan high-resolution satellite imagery and detect school buildings missing from national education management information systems, as a first step toward planning electricity and internet connectivity for schools governments did not know they had.

Activities

Between 2020 and 2021, the system scanned 71 million map tiles at 60cm resolution across Kenya, Rwanda, Sierra Leone, Ghana, Niger, Kazakhstan, Uzbekistan and Honduras, processing 18 million supertiles in 25 hours at up to 2,000 supertiles per second.

Results

The system added 23,100 previously unmapped schools to national registries. Country-specific model variants reached F1 scores of 0.94 in Rwanda, 0.90 in Kenya and 0.91 in Sierra Leone, with a global model at 0.85 and an overall AUC of 0.94. The Giga team's own code is open-sourced, and an independent review by IRCAI (a UNESCO-affiliated international research centre on AI) confirmed the technical results.

Conclusions

IRCAI's review also reported a significant equity limitation: model accuracy falls to roughly 57% in areas characterised by poverty and remoteness, meaning the tool is least reliable exactly where unmapped, unconnected schools are most likely to be found. No independent evaluation has yet shown that the newly mapped schools translated into actual connectivity investment or budget reallocation.

Implementation
Indicative cost
Medium (€50k–€500k)
Time to results
Medium (1–3 years)
Staffing & skills
UNICEF Office of Innovation, International Telecommunication Union (ITU), Development Seed (technical partner, model development), IRCAI (independent technical review)

Conditions for success

  • Access to high-resolution (60cm) satellite imagery at scale
  • Country-specific model variants trained per national context
  • Open-sourced code for transparency and reuse
  • Integration pathway into national education management information systems

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