Research map / Label-efficient and robust learning / Recognition and long-tail
Large-Scale School Mapping Using Weakly Supervised Deep Learning for Universal School Connectivity
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds a weakly supervised pipeline that classifies 300 m satellite tiles for schools using ViT, SwinV2, and ConvNext ensembles, then localizes schools via CAMs to produce nationwide school maps for 10 African countries.
Problem
Governments lack complete geolocated school records in low- and middle-income countries, which hinders estimating costs of universal school connectivity; precise school localization usually requires costly bounding-box or pixel annotations.
Approach
Train image-level school classifiers on satellite tiles sampled from built-up areas, then use explainable-AI class activation maps to convert the most important pixel into lat-lon coordinates, so only classification-level labels are needed; ensemble ViT, SwinV2, and ConvNext models per country (VSC-Ensemble) and add an interactive validation tool.
Key results
- VSC-Ensemble reaches AUPRC of 0.984 for the regional model and between 0.966 and 0.998 for the local models across the 10 countries.
- Transformer models (ViT, SwinV2) generally outperform ConvNext CNNs on most countries' image classification.
- GradCAM and GradCAM++/GradCAMElementWise give the highest ROAD confidence drop depending on country, with GradCAM best where ViT-Huge is the best model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Ensemble the best ViT, SwinV2, and ConvNext variants using mean soft voting | VSC-Ensemble · Benin test | AUPRC: 0.984 -> 0.998 | Helped |
| Ensemble the best ViT, SwinV2, and ConvNext variants using mean soft voting | VSC-Ensemble · Botswana test | AUPRC: 0.989 -> 0.997 | Helped |
| Ensemble the best ViT, SwinV2, and ConvNext variants using mean soft voting | VSC-Ensemble · Ghana test | AUPRC: 0.931 -> 0.991 | Helped |
| Ensemble the best ViT, SwinV2, and ConvNext variants using mean soft voting | VSC-Ensemble · Kenya test | AUPRC: 0.916 -> 0.966 | Helped |
| Ensemble the best ViT, SwinV2, and ConvNext variants using mean soft voting | VSC-Ensemble · Malawi test | AUPRC: 0.969 -> 0.983 | Helped |
| Ensemble the best ViT, SwinV2, and ConvNext variants using mean soft voting | VSC-Ensemble · Namibia test | AUPRC: 0.955 -> 0.980 | Helped |
| Ensemble the best ViT, SwinV2, and ConvNext variants using mean soft voting | VSC-Ensemble · Rwanda test | AUPRC: 0.983 -> 0.998 | Helped |
| Ensemble the best ViT, SwinV2, and ConvNext variants using mean soft voting | VSC-Ensemble · Senegal test | AUPRC: 0.980 -> 0.993 | Helped |
| Ensemble the best ViT, SwinV2, and ConvNext variants using mean soft voting | VSC-Ensemble · South Sudan test | AUPRC: 0.971 -> 0.995 | Helped |
| Ensemble the best ViT, SwinV2, and ConvNext variants using mean soft voting | VSC-Ensemble · Zimbabwe test | AUPRC: 0.967 -> 0.996 | Helped |
| Train on combined transnational data instead of country-specific data | Regional model · African country test sets | AUPRC | No clear effect |
| Use transformer-based models rather than CNNs | School image classifier · 10 African country test sets | AUPRC | Mixed |
Surprising findings
- A single regional model trained on pooled data from 10 countries performs about as well as the best country-specific models, so local data alone already yields strong performance.
Limitations
- Government school coordinates are noisy, sometimes hundreds of meters from actual buildings, and labels required manual correction.
- Performance varies by country; countries with few school samples generalize poorly to others.
Experimental setup
Datasets: government school registers via GigaMaps OpenStreetMap Overture Maps GHSL Microsoft Building Footprints Google Open Buildings Maxar satellite imagery ImageNet
Metrics: AUPRC F2 score precision recall average confidence drop (ROAD)
Baselines: ViT-Base SwinV2-Tiny ConvNext-Small GradCAM++ LayerCAM HiResCAM
Builds on: ViT Swin Transformer V2 ConvNeXt GradCAM
Closest papers by meaning
- RoMA: Scaling up Mamba-based Foundation Models for Remote Sensing (NeurIPS 2025 · similarity 0.82)
- Combining Deep Learning and Street View Imagery to Map Smallholder Crop Types (AAAI 2024 · similarity 0.81)
- Learning without Exact Guidance: Updating Large-scale High-resolution Land Cover Maps from Low-resolution Historical Labels (CVPR 2024 · similarity 0.81)
- Multi-Scale VMamba: Hierarchy in Hierarchy Visual State Space Model (NeurIPS 2024 · similarity 0.80)
- OpenStreetView-5M: The Many Roads to Global Visual Geolocation (CVPR 2024 · similarity 0.80)
- Spatial-Logic-Aware Weakly Supervised Learning for Flood Mapping on Earth Imagery (AAAI 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.