Spatiotemporal characterization of urban activity and environment with imagery and deep learning

Abstract

BACKGROUND AND AIM: There are limited data on human activity and the environment needed to inform policies and target infrastructures to improve the health and wellbeing of residents in cities in sub-Saharan Africa, the world’s fastest urbanising region. METHODS: We collected a bespoke dataset of 2.10 million images in Accra, Ghana, captured at five-minute internals over ~15 months at 145 representative locations. We retrained a convolutional neural network using a manually labelled subset of images to identify people (including street vendors) and 18 objects – categorised into large vehicles, small vehicles, two wheelers, objects from the market, refuse and animals – that collectively represent important features of human activity and the environment in the city. RESULTS:We identified 23.5 million of these objects in our dataset. Of these, 9.66 million (41%) were humans, followed by cars (4.19 million; 18%). We found strong correlation among the number of people, large vehicles and market-related objects, which were typically captured in the business and commercial core and high-density residential areas; moderate correlation between these three categories and small vehicles; weak correlation with two wheelers; and inverse correlation with refuse and animals which were more common in the peripheral areas of the city. The frequency of objects changed throughout the day with the extent of variation dependent on the type of object and location. There were noticeable reductions in the number of people, vehicles and market related activity in commercial and business areas during the Covid-19 lockdown, but smaller reductions observed in high-density residential areas. CONCLUSIONS:Contextual adaptation of computer vision tools can reduce the global gap in data on cities to advance sustainable and healthy urban development. Our data and approach have the potential to be applied to a range of urban environmental topics, including estimating road-traffic volume/flows and identifying sources of air and noise pollution. KEYWORDS: Big data, imagery, deep learning, built Environment; covid-19; traffic-related