Poverty is usually measured by knocking on doors and asking families about their income, their belongings and their living conditions. This works, but it is slow, expensive, and it cannot cover all households, especially for fast-growing cities in low-income countries. The new study from Tanzania asks a simple but powerful question i.e., what if the water-related things we can already see from the sky such as a tank on a roof, a toilet without a roof, a pond of dirty water, a house in flooded area, a house near wastewater stabilization/treatment ponds could tell us weather who is poor and who is not, without ever needing to knock on a single door? Researchers from Global Water Partnership Tanzania, working under the DEEP Challenge Fund, tested exactly this idea in selected wards in Dar es Salaam. They combined high-resolution aerial photos, street-level images, and an AI vision model to build a fast, low-cost way of mapping poverty across the city. The results are striking, and they point to a future where poverty maps can be updated faster in a large area.
Water is not just something people need to survive. It is also a mirror of how well-off a household is. Richer families can afford large overhead tanks that store water for weeks. Poorer families often cannot, so they buy small amounts of water every day, sometimes at high prices from private vendors. Richer families can afford a proper roofed toilet. Poorer families often cannot, and use unroofed latrines instead. Richer families tend to live on higher, safer ground. Poorer families are often pushed into cheaper, flood-prone land near rivers and swamps. Because these patterns are so consistent, the research team believed that water-related features, visible in ordinary aerial and street photographs, could act as reliable stand-ins for poverty data that is otherwise very hard and expensive to collect.
The research focused on an urban catchment within the Wami-Ruvu basin, one of the most water-stressed basins in Tanzania. This catchment includes most of Dar es Salaam city plus a small part of the neighbouring Coast region, covering 47 wards and 167sub-wards. It was chosen because it mixes wealthy neighbourhoods with some of the city's poorest and most flood-exposed settlements, giving the researchers a realistic picture of urban inequality.
Working with a panel of water, statistics and policy experts, the team identified a set of "non-traditional" indicators that are not part of normal poverty surveys but that can be spotted in photos. These fell into three broad groups including
More than 400high-resolution aerial images, taken at resolutions as fine as 7 centimetres, along with open street-level photos from the Mapillary platform, were used hotspot these features across tens of thousands of households.
To see whether these water-related clues really did line up with poverty, the team compared them against an established benchmark i.e., the wealth index from Tanzania's 2022 Demographic and Health Survey. Statistical tools were used to test whether wealthy and poor areas cluster together geographically, and whether the presence or absence of each water-related feature made a real difference to a household's wealth score.
From the indicators that lined up well with the survey data, the team built a new measure called the Spatial Relative Wealth Index, or SRWI for short. Instead of relying on door to door interviews, this index is built entirely from what can be seen in photographs, scored and combined using the existing statistical technique called principal component analysis.
Wealthy households tend to cluster near other wealthy households, and poor households tend to cluster near other poor households. This pattern was extremely strong and clearly not down to chance, which means poverty in the city is concentrated in identifiable hotspots rather than being spread out evenly.
Neighbourhoods with many overhead water tanks tended to be wealthier, while neighbourhoods with many unroofed toilets tended to be poorer. These two features turned out to be some of the clearest visual dividing lines between richer and poorer areas.
Poorer households were far more likely to live on low-lying, flood-prone areas. While some richer households were also found in these areas, the poorest neighbourhoods showed the strongest overlap between flood risk and deprivation.
When compared with the official national wealth index, the SRWI showed a strong and statistically meaningful match. A visual check against aerial photos also confirmed that areas the index labelled as poor did, in fact, look poorer and more densely packed, with worse sanitation and visible flood exposure.
The study also measured inequality directly. Roughly one-fifth of households were found to control only around a tenth of the area's total measured wealth, while the richest tenth controlled close to 40 percent. This gap matters for water access too i.e., households that can afford storage tanks are effectively able to hoard a larger share of available water, leaving less for everyone else.
The most forward-looking part of the study was training an AI vision model, built on GPT-4o, to automatically recognise these water-related features in aerial and street images, instead of having a person label every single photo by hand.
The team fine-tuned the AI model using a set of labelled sample images, then tested it on hundreds of new images it had never seen before, checking its answers against those of trained human experts. The AI performed remarkably well, reaching an accuracy level described in the study as close to that of a human expert, and it was able to explain roughly 88 percent of the variation in the true poverty scores.
The AI model could identify water tanks, toilets, ponds and flood zones from aerial images almost as reliably as a trained human but far faster.
In practical terms, this means that once trained, the AI can scan huge volumes of aerial imagery in a fraction of the time it would take a team of human analysts, opening the door to poverty maps that can be refreshed frequently rather than once every few years.
The findings point to a situation the researchers describe as "private water, public grime" i.e., water storage is concentrated among wealthier households, while poorer households are left to buy small amounts of water at high, unregulated prices from private vendors. During dry periods, prices for vended water can roughly double, placing a heavy burden on families who already have the least.
The study suggests two practical directions for policymakers. First, better regulation of private water vending prices could ease the burden on poorer households. Second, investing in shared or community water storage in deprived neighbourhoods could reduce dependence on expensive vendors altogether.
More broadly, because poverty was found to cluster in specific pockets rather than spreading evenly across administrative boundaries, the study argues that aid and infrastructure investment should be targeted at these small, specific hotspots rather than spread thinly across entire districts or wards.
This study shows that everyday water-related features, the kind visible in an aerial photograph, can serve as a fast, affordable and surprisingly accurate stand-in for traditional poverty surveys. By pairing these visual clues with an AI model trained to recognise them automatically, researchers were able to build detailed, updated poverty maps for the major African city without a single door to door interview.
The approach will not replace household surveys altogether, since surveys still capture details about people's lives that no photograph can show. But as a complement, it offers something surveys struggle to provide i.e., speed, low cost, and the ability to be updated again and again as a city keeps changing. The same method, the researchers note, could be adapted to track other kinds of poverty indicators beyond water, offering the mechanism that other cities and countries facing similar data gaps could follow.
...
