Using mobile phone data to map human population distribution Pierre Deville, Vincent D. Blondel Université catholique de Louvain, Belgium Andrew J. Tatem University of Southampton, UK Marius Gilbert Université Libre de Bruxelles, Belgium Samuel Martin Université de Lorraine CNRS, France Catherine Linard Biological Control & Spatial Ecology Université Libre de Bruxelles Andrea Gaughan, Forrest Stevens University of Louisville, USA Vienna, September 23-26, 2014 GIScience 2014 - Eighth International Conference on Geographic Information Science
Gridded population datasets: the options Dataset Coverage Spatial resolution Year(s) represented GPW v3 Global 2.5 arcminutes (~5 km) 1990, 1995, 2000, 2005, 2010, 2015 GRUMP Global 30 arcseconds (~1 km) 1990, 1995, 2000 UNEP Africa, Asia, South America 2.5 arcminutes (~5 km) 2000 LandScan Global 30 arcseconds (~1 km) 1998-2012 WorldPop Africa, Asia, South America 3 arcseconds (~100 m) 2010, 2015, 2020
Population Counts Population Density (pph) Intro to gridded population data Introduction Methods Results Conclusions Census data linked to GIS administrative boundaries Ancillary data e.g. Settlements, roads Spatial modelling rules to disaggregate census counts Estimates of number of people in each grid cell
Population distribution in Africa in 2010 Introduction Methods Results Conclusions A. C. How can we better inform on temporal changes? B. Linard et al (2012) PLoS ONE
Mobile phone usage data X User makes a call from location X Y User travels to Y and makes a call Call routed through nearest tower Network operator records time and tower of call for billing Penetration rates (2013) Global: 96% Developed countries: 128% Developing countries: 89%
Objectives Use MP data to map the spatio-temporal distribution of human population over large spatial scales Develop a method that: Is easy to implement Minimizes the impact of phone usage heterogeneities Preserves users privacy Mobile phone towers in France ~17,000 towers May-October 2007 > 1 billion calls 17 millions users
MP call density by admin. unit Number of calls aggregated by tower Coverage area approximated using Voronoi polygons Density of calls estimated for each administrative unit Census data 2007 Population Density (people/km²) < 10 10-50 51-100 101-500 501-1,000 1,001-5,000 5,001-10,000 > 10,000
Calibration c c Population density in admin. unit c Phone call density in admin. unit c α and β fitted by a linear regression α = scale ratio β = super linear effect of population density Adjustment of population estimates using national population
Training data Training data ~1,000 communes (ADM-5) 2 sampling procedures: random and spatial 1000 bootstraps Spatiallystratified random sampling
Coefficient estimates Introduction Methods Results Conclusions
Census-derived population density MP method WorldPop method
Relative error MP RMSE: 517 COR: 0.85 WorldPop RMSE: 539 COR: 0.9
Seasonal movements Relative difference in pop. density between holidays and working periods Asterix Park CDG Airport Disneyland Versailles Brest Rennes Nantes
Weekly movements Relative difference in pop. density between weekends and weekdays Asterix Park CDG Airport Disneyland Versailles Brest Rennes Nantes
Extrapolation ability Stability of β within and between countries Random sampling Spatial sampling
Extrapolation ability Stability of β within and between countries Sensitivity analysis of pop. estimates to α and β +15%
Conclusions Mobile phone method: WorldPop method: Dynamic Static Higher accuracy in urban areas Higher accuracy in rural areas Very simple aggregated data Many input data required Easy to implement More complex implementation Combine both methods?
Conclusions MP call activities can be used to produce spatially and temporarily explicit estimations of population densities across countries and their changes over multiple timescales, while preserving the anonymity of individual users. Limitations: Density of calls vs. density of users Daily-aggregated data Variations in phone usage behaviours not taken into account Partnerships between governments and phone companies could enable fast and cheap production of accurate maps of population distribution for every country in the world for every month
Thank you! E-mail: linard.catherine@gmail.com Reference Deville P., Linard C., et al. (2014) Dynamic population mapping using mobile phone data. Proc Natl Acad Sci:201408439. www.worldpop.org.uk