Google Open Buildings Mapped 1.8 Billion Homes the World Had Never Counted

Roughly half of humanity lives in places that were, until recently, effectively invisible on digital maps. Across much of Africa and South Asia, villages grew into towns and towns into cities faster than any mapping agency could record them. The consequences are practical and severe. You cannot plan an electricity grid, target a vaccination campaign, or deliver disaster aid to homes that officially do not exist. Google Research solved a large part of this problem with an AI system called Open Buildings, which has mapped 1.8 billion building footprints across the Global South and released every one of them for free.

The scale is hard to grasp. The dataset covers an inference area of 58 million square kilometers spanning Africa, South Asia, Southeast Asia, Latin America, and the Caribbean. According to Google, that is about 40 percent of the globe and roughly 54 percent of the world's population. For many of these regions, it represents the first comprehensive census of built structures ever assembled.

An AI Lab in Accra

The story of Open Buildings begins in Ghana, not California. The project started inside Google's AI Research Lab in Accra, the company's first such lab on the continent. Researchers there faced a question that commercial mapping efforts had largely ignored. Detecting buildings in satellite imagery is easy in wealthy countries where roofs are large, regular, and well lit. It is far harder across the Global South, where structures can be small, made of local materials, crowded together, and hidden under tree cover.

The team trained a deep learning model to detect and outline building footprints in high resolution satellite images, then ran it across entire continents. The model had to cope with dusty roads that look like rivers, rectangular farm plots that look like foundations, and dense informal settlements where one roof touches the next. After validation and cleanup, the results were released in 2021 as an open dataset, first covering Africa, and later expanded across the Global South.

Openness was the point. Google could have kept the data as a proprietary layer in Google Maps. Instead, the full dataset of 1.8 billion footprints is downloadable by anyone, with no fee and no special access, through Google's research site and Google Earth Engine.

Why Aid Workers Needed This So Badly

A building map sounds mundane until you see what depends on it. When an earthquake or flood strikes, responders need to know how many structures stood in the affected area to estimate how many people need rescue and shelter. When a health ministry plans a measles or polio campaign, it needs to know where every household is to route vaccinators and estimate coverage. When a utility extends the power grid or a telecom builds towers, settlement maps decide where the money goes.

Open Buildings has become a quiet piece of that machinery. According to Google Research, the dataset has been widely used by UN agencies, NGOs, and researchers for planning electrification, crisis response, vaccination campaigns, and more. The UN Humanitarian Data Exchange, the central clearinghouse that aid organizations rely on during emergencies, curated 20 country level datasets drawn from Open Buildings specifically to speed up humanitarian response.

The work has since moved beyond static footprints. In 2024, Google released the Open Buildings 2.5D Temporal dataset, which adds estimated building heights and tracks how structures change across multiple years of imagery. That temporal dimension matters because growth itself is information. A district adding hundreds of new structures per month needs schools and clinics before official statistics catch up, and a district losing structures may be experiencing conflict or disaster.

From Maps to Damage Assessments

The most consequential application has come during disasters. In partnership with the United Nations, Google Research combined Open Buildings with AI models that assess building damage from satellite imagery taken after a catastrophe. Responders can compare before and after views over a populated area and produce a damage assessment in hours instead of the days or weeks that manual analysis used to require.

Reporting published by ReliefWeb in 2026 described how this partnership gave UN disaster response wider coverage and faster damage assessments, with the Open Buildings model serving as the base layer that knows where the buildings were in the first place. You cannot measure destroyed homes accurately if you never counted the homes to begin with.

This work also complements a broader movement in AI assisted humanitarian mapping. The Humanitarian OpenStreetMap Team, which coordinates volunteer mappers during crises, has found that AI generated datasets can roughly double a mapper's output, from around 1,500 to between 2,500 and 3,000 buildings added per day. Open Buildings feeds directly into that ecosystem, letting volunteers validate and enrich AI detections rather than trace every structure by hand.

The Caveats That Come With the Map

No AI dataset this large is perfect, and Open Buildings is open about its limits. Detection quality varies by region, imagery freshness, and settlement type, and very small or heavily obscured structures can still be missed. Researchers who compared AI generated buildings in OpenStreetMap have noted measurable differences between AI mapped and human mapped structures, including subtle variations in shape and placement.

That is precisely why the data is released openly rather than baked invisibly into a commercial product. When a government statistician in Nairobi or an NGO analyst in Dhaka can inspect, correct, and redistribute the data, errors get caught. The dataset becomes a shared public asset, closer to infrastructure than to a product.

There is a larger lesson in Open Buildings for the AI for good conversation. The flashiest AI stories tend to involve models that talk, diagnose, or generate. But some of the highest impact applications are unglamorous foundations. Counting buildings sounds like bookkeeping. Done across 58 million square kilometers and given away freely, it becomes the substrate on which vaccination campaigns, electrification plans, and disaster responses are built. Visibility, it turns out, is a form of aid.