Episode 3 of our Mapping Urban AI Series
Urban planners in many parts of the world contend with ensuring the livability of cities in the face of rapid urbanization and growth. The dynamism of urban expansion poses a particular challenge when municipalities lack maps and data on the forms within its built environment. For a city, like eThekwini, in South Africa, where 25% of its population lives in 580 different informal settlements, grasping the nature of habitation within the city is a constantly moving target as settlements morph and grow without any record of the changes. For this reason, eThekwini approached the United Nations Innovation Technology Accelerator for Cities (UNITAC), to assess the ways in which they could use data and technology to map settlements within the city, as a means of informing the delivery of basic services to the settlements, as well as other urban interventions.
For a young organization (UNITAC was created less than 2 years ago), it has already made a broad impact across Africa and South America. UNITAC partners with UN Habitat in those areas to support local and national governments in their digital transformation. It aims to ensure that governments use technology and data in an equitable, human-centered manner, and that the implementation strategies serve to reduce the digital divide. To this end, it utilizes a three-pronged approach: offering strategic planning guidance, technological tools, and capacity building to support governments in the Global South in incorporating data and technology.
Sophie Naue, an Urban Innovation Specialist at UNITAC, joined us for the third episode of the “Mapping Urban AI” webinar series, to discuss the use of AI as a tool for urban planning in the context of urban informality. The Mapping Urban AI series, which is organized by Urban AI, CIDOB, and the Global Observatory of Urban AI (GOUAI), hosts representatives from cities around the world to showcase and discuss their projects and policies through the lens of an overarching theme: holistic strategies, health, urban planning, sustainability. The concept for the series draws from both the project “The Atlas of Urban AI,” an interactive online map of AI initiatives that cities throughout the world have undertaken and also from the newly released “Urban AI Guide,” which presents detailed case studies on AI project implementation in local government.
In the episode, Naue focused on the eThekwini project, describing how UNITAC leveraged an object detection algorithm to identify rooftops and derive structure footprints from aerial imagery. The general process of artificial intelligence-based land use detection, such as this, is fairly straightforward. First, people select parts of images and label them as what they are: buildings, water, roads, etc. Then, the algorithm predicts the likelihood that other clusters of pixels within the photograph might represent any one of those land cover types, based on the previously determined classifications assigned to other groupings of pixels. Elements like building standards, which lend regularity to the built environment, can aid classification and lead to more accurate models, since they narrow the scope of what constitutes a building or a road and make pattern recognition easier.
Informal settlements, however, pose a difficulty when algorithms attempt to generalize a rooftop from a collection of pixels, since virtually any material could comprise a roof in an informal settlement, leading to less precise classifications. Similarly, rooftops in informal settlements can overlap or have irregular shapes that make it difficult to identify single dwellings. Since most commercially available building detection tools have been trained on data from conventionally planned cities, model performance degrades in the differently-patterned context of urban informality. Thus, UNITAC trained its algorithm using data on informal settlements in eThekwini specifically, to ensure that the particular characteristics of informal settlements would be detected and identified, rather than serving as barriers to detection (as they might in other systems).
The resulting tool, dubbed “BEAM” (which stands for “Building & Establishment Automated Mapper”) has significantly aided eThekwini’s mapping efforts. Previous attempts at mapping informal settlements in the city required a team of 15 people to perform very time-intensive manual processes, including physically walking around the city. Given the dynamism of informal settlements, any maps produced would soon fall out of date — constituting a never ending game of catch-up. BEAM, on the other hand, can map the entire city within 72 hours, providing a much more timely picture of informality in eThekwini and allowing for more detailed analyses on settlement growth patterns. As a key part of the project, UNITAC emphasized collaboration and skills transfer, to empower the city to partake in the design process and train the city on the tool and its underlying technology. Moving forward, UNITAC hopes to extend the collaborative aspects of the tool to include a higher degree of public participation, to ensure that the tool caters to the needs of the public. It also aims to incorporate other datasets, so the tool can better support decision-making, beyond mapping.
By Sarah Popelka, Head of Education Programs at Urban AI
