Monitoring access to piped water and sanitation infrastructure in Africa at disaggregated scales using satellite imagery and self-supervised learning

Imagine trying to fix a leaky roof in a massive, sprawling city, but you don't have a map, and you can't afford to send a team of inspectors to check every single house. That is essentially the challenge facing many African nations when it comes to tracking who has access to clean water and proper sewage systems.

This paper presents a clever new solution: using satellite eyes and artificial intelligence to "see" where the plumbing is, without ever needing to visit the site.

Here is a breakdown of how they did it, using simple analogies:

1. The Problem: The "Blind Spot"

Governments and the UN want to know if everyone has clean water (SDG 6). Usually, they rely on surveys where people are asked, "Do you have running water?"

The Issue: Surveys are like sending a few scouts into a giant forest. They can tell you about the trees they walk past, but they miss the deep, remote corners. It's also expensive, slow, and often outdated. By the time the data is collected, the situation might have changed.

2. The Solution: The "Satellite Detective"

The researchers built a digital detective using satellite images (photos taken from space) and AI.

The Analogy: Imagine you are trying to guess if a house has a high-speed internet connection just by looking at a photo of the neighborhood from a drone. You might not see the internet cable itself, but you can see the fiber-optic box on the pole, the modern router on the roof, or the fact that the house is in a dense, paved neighborhood with streetlights.
The Method: The AI looked at satellite photos of neighborhoods across Africa. It learned that areas with piped water and sewage often look a certain way: they have denser roads, more buildings, and specific patterns of development. Even though the pipes are underground and invisible, the "fingerprint" of the neighborhood gives them away.

3. The "Self-Taught" Student (Self-Supervised Learning)

Usually, to teach an AI, you need a teacher to show it thousands of pictures and say, "This one has water, this one doesn't." But they didn't have enough labeled pictures.

The Analogy: Instead of a teacher, they gave the AI a massive library of unlabeled photos (like a student reading a library of books without a test key). The AI taught itself to recognize patterns, shapes, and textures in the landscape. It learned, "Okay, this type of texture usually means a city, and cities usually have pipes."
The Result: They used a specific AI model called DINO (which is like a super-smart student that learns by comparing different views of the same image). This model became an expert at spotting the subtle clues that indicate infrastructure.

4. The Results: A Crystal Clear Map

When they tested this AI:

Accuracy: It was incredibly good. It correctly identified water access about 96% of the time and sewage access about 97% of the time.
The Big Picture: They combined these tiny neighborhood predictions with population data (knowing how many people live in each spot). This allowed them to create a "heat map" of the entire continent.
Validation: When they compared their AI-generated numbers with the official UN statistics, the numbers matched up almost perfectly (like two clocks ticking in sync).

5. Why This Matters: The "Flashlight" for Policy Makers

Think of this technology as a flashlight for policymakers.

Before: They were walking in the dark, guessing where the problems were based on old, incomplete maps.
Now: They can shine a light on specific, remote villages that are being overlooked. They can see exactly which neighborhoods are underserved, even if no one has ever surveyed them before.
The Benefit: This saves money and time. Instead of sending expensive survey teams everywhere, governments can use this map to send help exactly where it's needed most.

The Catch (Limitations)

The authors are honest about one potential flaw: The AI is very good at spotting "cities" vs. "countryside." Since cities usually have water and villages often don't, the AI might be guessing based on "urban vibes" rather than seeing the actual pipes.

The Fix: They acknowledge this and suggest that in the future, they will need to teach the AI to look for even finer details to make sure it's not just guessing based on how "city-like" an area looks.

In a Nutshell

This paper shows that we don't always need to be on the ground to solve big problems. By teaching AI to read the "visual language" of satellite photos, we can create a real-time, high-definition map of who has clean water and who doesn't, helping to ensure that no community is left behind in the race for a better future.

1. Problem Statement

Access to clean water and sanitation (UN Sustainable Development Goal 6) is critical for health and development, yet significant data gaps persist, particularly in Africa.

Data Limitations: Traditional monitoring relies on national surveys (e.g., censuses) and field data, which are costly, logistically challenging, and often lack the spatial resolution needed to identify regional inequalities or reach remote communities.
Reporting Gaps: National-level statistics often obscure sub-national disparities. Furthermore, many developing countries lack the resources to sustain continuous, data-driven monitoring.
Research Gap: While satellite remote sensing and deep learning have been applied to water quality and hydrology, their application to water supply and sanitation infrastructure (specifically piped water and sewage systems) remains underexplored. Existing methods often rely on labeled datasets that are scarce in developing regions.

2. Methodology

The authors propose a remote sensing framework that integrates survey data, multispectral satellite imagery, and self-supervised learning (SSL) to estimate infrastructure access at a disaggregated scale.

Data Sources

Ground Truth: Afrobarometer survey data (Rounds 7–9, 2019–2023) covering 40 African countries with ~147,000 georeferenced interviews. Binary labels were created for household access to piped water and sewage.
Satellite Imagery: Sentinel-2 multispectral imagery (10m resolution) retrieved via Google Earth Engine. Patches of $256 \times 256$ pixels were extracted around survey locations, filtered for cloud cover (<10%).
Population Data: Meta's High-Resolution Population Density Maps (HRSL) (~30m resolution) to convert patch-level predictions into population-weighted national estimates.
Validation Data: Official statistics from the WHO/UNICEF Joint Monitoring Programme (JMP) for SDG Indicators 6.1.1 (water) and 6.2.1 (sanitation).

Model Architecture & Training

The study evaluates Vision Transformer (ViT) based foundation models using Self-Supervised Learning (SSL) to learn transferable representations without extensive labeled data.

Models Evaluated: DINO, DINOv2, DINOv3, Galileo, and Prithvi-EO-2.0.
Pretraining Strategy:
- DINO and DINOv2: Pretrained in-house on 186,701 unlabeled Sentinel-2 patches using a student-teacher self-distillation framework.
- Others: Used publicly available pretrained weights.
Classification: Instead of fine-tuning the entire network, the study extracts feature embeddings from the "teacher" encoder and uses a lightweight $k$ -Nearest Neighbor ( $k$ -NN) classifier to predict access (binary classification).
Validation Strategy:
1. Held-out Test: A spatial split (70% train, 15% val, 15% test) of survey locations to prevent spatial data leakage.
2. External Validation: The best model was applied to 50 African countries. Predictions were aggregated using population weights and compared against JMP national statistics.

3. Key Contributions

Novel Framework: First study to integrate Afrobarometer survey data with Sentinel-2 imagery and SSL foundation models specifically for mapping piped water and sewage infrastructure across Africa.
Self-Supervised Optimization: Demonstrated that models pretrained directly on Sentinel-2 data (DINO/DINOv2) outperform general-purpose or other Earth Observation foundation models for this specific infrastructure task.
Scalable Disaggregation: Successfully generated spatially disaggregated estimates that can be aggregated to national levels, bridging the gap between local satellite observations and global SDG reporting.
Open Science: The code, pretrained weights, and inference results are publicly available, promoting reproducibility.

4. Results

Held-out Test Performance

The models achieved high accuracy in classifying access at the survey location level:

Best Model: DINOv2 (pretrained on Sentinel-2) achieved the highest performance.
- Piped Water: 84.19% Accuracy, 83.87% Recall, 83.94% F1-score.
- Sewage: 87.17% Accuracy, 84.29% Recall, 85.00% F1-score.
Observation: Sewage access was predicted slightly more accurately than piped water, likely because sewage infrastructure is more strongly correlated with dense urban built-environment patterns visible in satellite imagery.

External Validation (National Scale)

When aggregated to national levels and compared against JMP statistics:

Piped Water: Strong correlation ( $R^2 = 0.95$ ). The model tends to slightly underestimate coverage but captures national variations effectively.
Sewage/Sanitation: Moderate correlation ( $R^2 = 0.77$ ). The lower score is attributed to definition mismatches (JMP includes septic systems/latrines, while the model focuses on sewage infrastructure).
Generalization: The model performed well even in countries without survey data in the training set.
- For piped water in non-survey countries, the Mean Absolute Error (MAE) was 9.5 percentage points.
- For sanitation, the MAE was 10.7 percentage points.
- This indicates the model learned transferable "urban signatures" rather than overfitting to specific survey regions.

5. Significance and Policy Implications

Cost-Effective Monitoring: Provides a scalable, low-cost alternative to expensive field surveys, enabling continuous monitoring of SDG 6 progress.
Targeted Intervention: The spatially disaggregated nature of the data allows policymakers to identify underserved regions (both urban and rural) that are often hidden in national averages, facilitating targeted infrastructure investment.
Data-Driven Decision Making: Offers a mechanism to validate and complement official statistics, particularly in data-scarce environments.
Broader Applicability: The methodology can be adapted for other infrastructure-related SDGs (e.g., energy, transport).

6. Limitations

Urban-Rural Bias: The model may inadvertently rely on "urban signatures" (road density, building footprints) rather than the infrastructure itself, as piped water and sewage are highly correlated with urbanization.
Data Availability: Reliance on cloud-free satellite imagery can introduce geographic bias in regions with frequent cloud cover.
Definition Mismatch: The study's definition of "sewage" is narrower than the JMP's broad definition of "safely managed sanitation," leading to validation discrepancies.
Decentralized Systems: Decentralized water/sewage systems may not be detectable via optical satellite imagery.

Conclusion

This study successfully demonstrates that self-supervised learning on satellite imagery can effectively estimate access to piped water and sanitation infrastructure across Africa. By achieving high accuracy and strong alignment with official national statistics, the proposed framework offers a powerful tool for accelerating progress toward SDG 6, particularly in regions where traditional data collection is limited.