Human Movement
Anonymised, aggregated flows of people across the planet — where they go, when, and the activity they are doing.
What this dataset is
A planet-wide picture of how people move — where they go, when, and what they are doing — assembled from many anonymised mobility panels, fused together, and calibrated against national census and transport statistics. The dataset never contains individual trajectories. Every value you can read is an aggregate over many people, many devices, and many days.
Three products sit inside the dataset:
- Presence — the number of distinct people present in each 250-metre hexagon, each hour of each day, broken down by inferred activity (home, work, study, transit, leisure, errands, other).
- Flows — origin-to-destination matrices between hexagons, aggregated to hour-of-day and day-of-week, for trips above a minimum count threshold.
- Activity profiles — for each hexagon, the daily and weekly rhythm of activities, useful for typology and clustering.
What you get
- A global GeoParquet table for presence, partitioned by country and month, ready to query with DuckDB, BigQuery, or Spark.
- A Zarr store of the gridded presence layer for fast spatial-temporal slicing in xarray.
- Flows as country-level GeoPackages, with a topology that loads directly into routing and visualisation tools.
- A taxonomy lookup defining the seven activity classes and the inference rules used to assign them.
Typical uses
- Measuring the recovery of city centres, transit corridors and tourism after disruption.
- Estimating exposure populations for air-quality, noise and heat-stress assessments.
- Calibrating transport, retail and emergency-response models.
- Tracking displacement, return movements and humanitarian access.
- Detecting and quantifying changes in commuting, remote work and the geography of leisure.
How we handle privacy
This dataset only exists because we believe it can be released safely. Five things make that true:
- Aggregation thresholds: every cell, every hour, every flow is suppressed if it would be derived from fewer than a fixed minimum number of distinct devices.
- Differential privacy: a small amount of calibrated noise is added to every released value, with an audited ε budget.
- Coarsening near sensitive places: cells covering schools, places of worship, hospitals, refuges, and other sensitive land uses are reported only at coarser spatial resolution.
- No trajectories, ever: the underlying panel data never leaves our pipeline. We do not redistribute device-level traces, and there is no API path that would let you reconstruct them.
- Independent review: the methodology is reviewed annually by an external privacy panel, whose report is published with the next vintage.
If you find an aggregation that you believe could be re-identifying, please contact us. We will investigate and, if necessary, suppress and reissue.
Notes on bias
Mobility panels do not sample the population evenly — they over-represent younger, more urban, more connected people, and under-represent the elderly, rural communities, and households with no smartphone. We correct for the most important biases using national statistics, but residual bias remains, especially in low-income and remote regions. The accompanying bias report documents what we can and cannot say about each country.
