Global population dataset uses buildings to reveal people's daily movement patterns

Researchers at Oak Ridge National Laboratory (ORNL) have introduced LandScan Mosaic, an upgraded global population distribution dataset that leverages building information to estimate daily movement patterns. This new dataset builds upon ORNL's established LandScan Global, which is renowned for its accuracy in representing population distribution.

LandScan Mosaic's innovative approach focuses on buildings, utilizing machine learning techniques to estimate missing characteristics like height, floor count, function, and use type, even in areas with limited building data. By incorporating building and land use information along with standardized occupancy distributions, the model creates a high-resolution picture of ambient population—showcasing the average number of people present at any given location throughout a 24-hour period.

This captures daily movement between various activity spaces, such as homes, workplaces, and schools. One of the dataset's standout features is its provision of explicit uncertainty measures, allowing users to gauge the confidence in each population estimate. This is particularly valuable for decision-makers who must act on incomplete information, as it offers transparency about the reliability of the data.

ORNL's Daniel Adams, lead author of the Scientific Reports paper detailing the methodology, highlighted that this is the first globally accessible population dataset to include such uncertainty measures. ORNL's principal investigator, Marie Urban, emphasized the team's multidisciplinary approach, combining expertise from computer science, data science, geography, and other fields to achieve this breakthrough.

The team is also developing a historical version called LandScan Mosaic Timeseries (LSM-TS), which provides a 50-year look-back from 1975 to 2024, using a backcasting method to produce historical population distributions. This extended dataset enables researchers to analyze long-term population trends and the long-term impact of events, offering the same fine-grained building-level modeling for historical data as it does for modern data.