AI that spots inequality could monitor living conditions in cities

By | April 18, 2019
London from above

Some measures of inequality can be inferred from images

TangMan Photography/Getty

Social and economic inequality has no easy fix, but now a system that automatically detects signs of inequality from street images could be used to help.

Esra Suel and colleagues at Imperial College London trained artificial intelligence to detect inequalities in four UK cities, using a combination of government statistics and public images taken from Google Street View.

The AI was trained on 525,860 images from 156,581 postcodes across London, along with income, health, crime, housing, and living environment statistics about the areas.


A fifth of the data was withheld to test how closely the algorithm’s estimation matched real distributions of inequality in London.

The AI was most successful at spotting differences in quality of the living environment and mean income, scoring 0.86 for both on a statistical test of how closely its predictions matched with the real data, where a score of 1 is a complete overlap.

It was least successful at predicting differences in crime rate and self-reported health, scoring 0.57 and 0.66, respectively.

Reducing inequality

The team then used the AI to perform the same estimation in Birmingham, Manchester and Leeds, after being fine-tuned with some additional images collected in those cities. It scored 0.68, 0.71, and 0.66 respectively, compared to an overall correlation of 0.77 in London.

Some features of a living environment, such as pollution and signs of disrepair, are directly linked to visual elements that the algorithm could recognise, but others are less so, says Suel. “What is usually perceived as unsafe is not necessarily correlated with actual crime rates,” she adds.

Street imagery could be a helpful tool in monitoring the success of policies to reduce inequality, because they are updated more frequently than some government surveys or census data.

The team now plans to train the algorithm to detect inequalities in cities in developing countries, where statistical data is not as widely available.

Journal reference: Scientific Reports, DOI: 10.1038/s41598-019-42036-w

More on these topics:

New Scientist – Health