Governments and humanitarian organizations need reliable data on building and infrastructure changes over time to manage urbanization, allocate resources, and respond to crises. However, many regions across the Global South need more access to timely and accurate data on buildings, making it difficult to track urban growth and infrastructure development. The absence of this data hinders effective planning and disaster response efforts. Current methods for detecting buildings generally rely on high-resolution satellite imagery, which provides detailed images of building footprints. However, high-resolution images are often captured sporadically, sometimes years apart, making it difficult to track changes in buildings over time, especially in rural or rapidly developing areas.
Google researchers introduced the Open Buildings 2.5D Temporal Dataset to address the issue of rapid urban population growth, particularly in the Global South, where urban areas are projected to expand significantly by 2050. Unlike previous efforts, which relied on high-resolution images, this new dataset utilizes Sentinel-2 satellite imagery captured by the European Space Agency, which offers lower resolution but captures images every five days globally. By using a novel machine learning approach, the dataset can estimate changes in building presence and height over time, covering a large geographic region from 2016 to 2023.
The core idea behind the dataset involves using a combination of student-teacher models based on HRNet architecture. The teacher model is trained on high-resolution satellite imagery, providing ground truth labels. The student model, trained on lower-resolution Sentinel-2 images, aims to recreate the teacher’s predictions without directly seeing the high-resolution images. By leveraging multiple time frames of Sentinel-2 data (up to 32 images for each location), the model enhances resolution and detects building footprints with high accuracy. This approach allows the model to achieve a mean Intersection over Union (IoU) of 78.3%, which is close to the 85.3% accuracy obtained using high-resolution imagery. The dataset also includes tools for estimating building heights and counts, with a mean absolute error of 1.5 meters for height estimates and reliable building count predictions.
In conclusion, Google’s Open Buildings 2.5D Temporal Dataset offers a significant advancement in detecting and monitoring building changes across the Global South using public satellite imagery. By employing a combination of frequent, lower-resolution Sentinel-2 images and machine learning models, it provides an innovative solution for addressing the lack of accurate, up-to-date data on buildings. The proposed method not only improves the ability to track changes in urban areas but also supports better planning and crisis response in regions that are often data-poor.
Check out the Dataset and Blog. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and join our Telegram Channel and LinkedIn Group. If you like our work, you will love our newsletter..
Don’t Forget to join our 50k+ ML SubReddit
Pragati Jhunjhunwala is a consulting intern at MarktechPost. She is currently pursuing her B.Tech from the Indian Institute of Technology(IIT), Kharagpur. She is a tech enthusiast and has a keen interest in the scope of software and data science applications. She is always reading about the developments in different field of AI and ML.
