AI-Driven Development of New Open Buildings Dataset Unveiled

Home

News

April 10, 2025

HarryGonzález

164

# ai # Maps # research

AI-Driven Development of New Open Buildings Dataset Unveiled

In 2021, the Google Research Africa team kicked off Open Buildings, an open-source dataset that maps building footprints across the Global South using AI and high-resolution satellite imagery. Their goal was pretty straightforward: to plug a big hole in the data about population and density in developing countries. Now, with the third version out, their dataset boasts polygons for a whopping 1.8 billion buildings spread over 58 million km² in Africa, South and Southeast Asia, Latin America, and the Caribbean. A bunch of folks, from governments to the UN, and even researchers and nonprofits, have been using Open Buildings to get a handle on population size and distribution. This has helped them plan better for things like vaccination drives and disaster response. Plus, it's even beefed up Google Maps by adding millions of buildings that weren't mapped before. The team, based in Ghana but with members scattered across places like Tel Aviv and Zurich, has been on a mission to make the project even more useful. "We're always in hackathon mode, trying out new ideas and tackling challenges," says Google Research program manager Abdoulaye Diack. "One thing we couldn't do with the original dataset was show how areas change over time—it was static. And that's something our partners really wanted." Commercial satellite image providers usually focus on areas that bring in the bucks, leaving about 40% of the world, mostly the Global South, without regular high-res coverage. Some remote spots and informal settlements don't get any coverage at all. Meanwhile, open-source imagery from the European Space Agency's Sentinel-2 satellite, which snaps a global pic every five days, was thought to be too low-res for building detection. But the team figured it might not be as big a deal as they thought, so they gave it a shot. First, they fed a single low-res frame from Sentinel-2 into their model and asked it to draw building polygons. "It was tough, but we saw potential," Abdoulaye says. "So we told the model to just give us the building masks—binary pixel data tied to specific spots. It did okay, and we thought, 'Hey, we can do this.'" After a year of tweaking the model, they rolled out the Open Buildings 2.5D Temporal Dataset last month. It covers 2016 to 2023 and gives an annual snapshot of building presence and counts across much of the Global South, plus building heights. This shows how cities change due to development, disasters, and other factors. Users can pick a region, flip through the years, and watch the world grow and shrink in a colorful display of shapes. "By 2050, about 2.5 billion more people might move to cities, mostly in the Global South. This dataset could be a game-changer for governments and organizations dealing with that growth," says Google Research product manager Olivia Graham. "If a city's planning where to put essential services like healthcare and education, or where to build infrastructure like water and energy supplies, this dataset shows which areas are growing fast." On September 28, 2018, a massive 7.4 magnitude earthquake off Indonesia's coast triggered a tsunami, impacting around 1.5 million people on Sulawesi. The dataset shows how the built area pulled back from the coast after the disaster. You can check it out in our interactive Earth Engine app. You can also see the construction of New Cairo, Egypt, in the Open Buildings 2.5D Temporal Dataset demo. So, how did the team manage to get their model to read Sentinel-2's fuzzy satellite images and confidently detect buildings? They started by sharpening things up. "We used a teacher-student model setup to both 'super-resolve' the low-res images and pull out the building footprints," says Google Research software engineer Krishna Sapkota. "The teacher model learns to spot buildings in high-res images and gives labels to the student model. The student model, which actually creates the dataset, learns from the teacher's output. It can then take low-res images from Sentinel-2 and guess what a higher-res version would look like." The teacher model gives high-res training labels to the student model, which then figures out building presence from low-res imagery. To get the detail needed for building footprints, the model uses up to 32 frames of Sentinel-2 images of a spot for any prediction. Each frame is a bit different from the others, thanks to the tiny time gap between captures, which helps boost resolution—kind of like how Pixel phones use multiple shots for sharper images. Unlike the original dataset, which gave precise polygonal outlines of buildings, the new temporal dataset uses raster data for building footprints. It also predicts building heights, crucial for estimating population density, with an error of just 1.5 meters, or less than one story. Before its official release, the temporal dataset was shared with trusted partners like Ugandan nonprofit Sunbird AI. "About 73% of Ugandans don't have electricity, and Sunbird AI used our original database to help the government figure out where to put microgrids or solar panels," Olivia says. "With the new dataset, they're looking at Jinja and Fort Portal, creating visuals that help city councils see where growth is happening fast and adjust their plans. It shows how both datasets can be part of a bigger toolkit to understand a population and how it's changing." The same curiosity that led to the temporal dataset is pushing the team to keep improving it. "I live in Ghana and see the impact our work is having and can have," Abdoulaye says. "Many places here struggle with resources, which leads to data gaps with big consequences. Being part of a team working to fix that and make a difference is a real honor."

Meta Enhances AI Security with Advanced Llama Tools Meta has released new Llama security tools to bolster AI development and protect against emerging threats.These upgraded Llama AI model security tools are paired with Meta’s new resources to empower c

NotebookLM Unveils Curated Notebooks from Top Publications and Experts Google is enhancing its AI-driven research and note-taking tool, NotebookLM, to serve as a comprehensive knowledge hub. On Monday, the company introduced a curated collection of notebooks from promine

Alibaba Unveils Wan2.1-VACE: Open-Source AI Video Solution Alibaba has introduced Wan2.1-VACE, an open-source AI model poised to transform video creation and editing processes.VACE is a key component of Alibaba’s Wan2.1 video AI model family, with the company

Comments (27)

0/200

Submit

JustinHarris

July 30, 2025 at 9:41:20 PM EDT

Super cool to see AI mapping buildings in the Global South! 😮 Makes me wonder how this could help urban planning or disaster response.

KennethJohnson

July 22, 2025 at 2:33:07 AM EDT

This dataset sounds like a game-changer for urban planning! 😮 I wonder how accurate those AI-mapped footprints are compared to real-world buildings.

CarlHill

April 18, 2025 at 3:10:36 PM EDT

新しいオープンビルディングデータセットのAI駆動開発はすごくクール！AIと衛星画像でグローバルサウスの建物のフットプリントをマッピングできるなんて驚きだ。でも、データが少し圧倒的すぎる。とはいえ、人口密度を理解するのに素晴らしいツールだよ！🌍📊

LarryMartin

April 17, 2025 at 12:08:18 PM EDT

새로운 오픈 빌딩 데이터셋의 AI 기반 개발은 정말 멋져요! AI와 위성 이미지를 통해 글로벌 사우스의 건물 발자국을 매핑할 수 있다는 게 놀랍네요. 하지만 데이터가 조금 압도적이에요. 그래도 인구 밀도를 이해하는 데 좋은 도구예요! 🌍📊

StevenGreen

April 16, 2025 at 12:37:39 AM EDT

El Desarrollo Impulsado por IA del Nuevo Conjunto de Datos de Edificios Abiertos es súper genial. Es increíble cómo la IA y las imágenes satelitales pueden mapear las huellas de los edificios en el Sur Global. Pero, los datos pueden ser un poco abrumadores. Aún así, es una gran herramienta para entender la densidad de población. 🌍📊

AlbertScott

April 14, 2025 at 7:26:14 AM EDT

The AI-Driven Development of New Open Buildings Dataset is super cool! It's amazing how AI and satellite imagery can map building footprints in the Global South. But, the data can be a bit overwhelming. Still, it's a great tool for understanding population density! 🌍📊