The Bartlett


Sensing Digitally Gets Us Closer to Reality

Sharon Richardson estimates the local active population from digitised interactions.

Plot of mobile data points in East London

15 December 2017

While urban planning often takes the long view, from five to fifty years and beyond, the increasing digitisation of everyday life is making it possible to gather data at much smaller spatial and temporal scales. One aspect of urban analytics to benefit from this change is the science of where people are. This is at the heart of my doctoral research in behavioural data science, examining how environmental and behavioural data traces can provide insights into human dynamics in urban open spaces.

Traditionally, the gold standard for estimating the population of an area is administrative data – the residential population provided by the most recent census. In reality, that tells us where people are mostly likely to be when they are asleep. Using the residential census makes sense for urban decisions about fixed local infrastructure needs, such as how many school buildings are required to serve a district. But a residential count can result in misleading statistics when applied to dynamic phenomena such as quantifying the population at risk of exposure to air pollution or as a normalisation measure to calculate street crime rates.

In recognition that the residential population does not reflect where people are during the daytime, a recent innovation has been the ambient population count, an estimate of where people are during an average working day. The ambient population is usually calculated by redistributing the residential population according to land use data, identifying where people work, study and play. The 2011 census included a workday population measure for the first time.

However, even the ambient population count is a generalisation. Part of my research is exploring whether or not we can make use of mobile phone data to produce an active population count in near real-time. The image at the beginning of this article shows the movement traces captured by a mobile phone app for three consecutive Sundays in May 2017 at the Queen Elizabeth Olympic Park in Stratford, east London. It represents a very small sample, less than 100 individual devices across the three dates. But there are visible differences in the spatial distribution.  Without providing any labelling, it is not difficult to guess on which of the three dates there was an event taking place at the London Stadium (left of centre).

Landscan grid Queen Elizabeth Olympic Park

But can such a small sample of movements provide an active population estimate? I believe it can, if combined with the robustness and completeness of administrative data sources. The image above is a grid of LandScan pixels covering the Queen Elizabeth Olympic Park. LandScan is an ambient population measure provided by Oak Ridge National Laboratory and updated annually. The highlighted pixel contains no residential properties. It does, however, contain various work facilities and two event venues – the London Stadium and the Copper Box indoor arena. The residential population count for this pixel would be zero. The LandScan ambient population estimate for 2015 was 5,907. Using the count of devices running a mobile phone app as a weighting measure, we can explore how much the population might vary over time. The table below is based on device readings captured during June 2017.  

 AmbientActive (weight)
Friday daytime5,9075,907 (1.00)
Friday normal at 6pm5,9079,865 (1.67)
Friday events at 6pm5,90755,112 (9.33)

Setting Friday daytime as the base line, we can see that the active population increases by 67 per cent during the Friday rush hour, when people are commuting home. This would make sense because the pixel is located between a residential district to the west and a transport hub to the east. On two Friday evenings in June, music concerts were held at the London Stadium located within the pixel. The number of devices detected in the area on these days increased by an order of magnitude, leading to an active population estimate of over 55,000. Again, this fits with expectations because the stadium seating capacity is 55,000. If anything, the estimate is conservative and should probably be a little higher. But it does provide a reading much closer to reality than either the residential population provided by the most recent census or ambient population estimate provided by LandScan.

New urban data sources such as readings from mobile device apps and social media interactions shared online are providing new insights into human dynamics. Whilst there are many valid questions and concerns about using such data, such as demographic bias in adoption and small sample sizes, this study suggests there may be potential in combining the robustness of administrative data sources with the finer spatial and temporal resolution of ‘reality’ data to better understand dynamic urban and social phenomena. Further research is now being undertaken to validate the findings and determine the feasibility of providing an active population estimate at street-level in near real-time.

Sharon Richardson is undertaking a PhD on Digitally sensing and evoking behavioural change in urban spaces at CASA, UCL, funded by an EPSRC grant. This article is based on a presentation on 29th November 2017, the Bartlett Centre for Advanced Spatial Analysis (CASA) hosted Smart Cities and Planning: New Urban Agenda, New Urban Analytics. The conference was part of a two-year research program in Applicable Urban Informatics kindly sponsored by the MacArthur Foundation. Spatial background images courtesy of OpenStreetMap contributors.