GEO / TorchSDG4
TorchSDG4

High-resolution mapping of
SDG4 sub-indicators

Deep learning models for estimating SDG4 sub-indicators at high spatial resolution from satellite imagery — enabling education outcome estimation in regions where EMIS data is sparse, delayed, or unreliable.

Coming Soon

About

What TorchSDG4 does

TorchSDG4 trains deep learning models on freely available satellite imagery to estimate SDG Indicator 4 sub-indicators — including school access, enrollment rates, and learning outcomes — at sub-kilometer resolution globally. Models and weights will be released open-source.

🛰️
Satellite-native inputs

Models trained on Sentinel-2, Landsat 8/9, and Sentinel-1 SAR composites — no field survey data required at inference time.

📐
Sub-kilometer resolution

Estimates produced at 100m–1km spatial resolution, suitable for school catchment area analysis and within-country inequality mapping.

🔓
Fully open

All model weights, training code, and inference pipelines released on GitHub under MIT license. Designed to be run in Google Earth Engine or locally.

SDG4 Targets

Indicators in scope

The following SDG4 sub-indicators are planned for the first model release.

${[ ["4.1.1", "Minimum proficiency in reading & mathematics (Grades 2/3 and end of primary)"], ["4.1.2", "Completion rate (primary, lower secondary, upper secondary)"], ["4.2.2", "Participation rate in organized learning one year before primary"], ["4.5.1", "Parity indices for all education indicators (sex, location, wealth)"], ["4.a.1", "Schools with access to electricity, internet, water, and sanitation"], ].map(([code, desc]) => `
${code}
${desc}
`).join('')}