No matter where in the world, local communities are largely shaped by residents — who themselves are, in turn, shaped by the communities in which they live. Community land use, infrastructure and resource allocation policies however are typically the product of models favoured by professional urban planners.
Urban planning deals with the physical layout of human settlements and guides orderly development in urban, suburban, and rural areas. Effective urban planning mitigates the operational and social vulnerabilities of an urban system, striving to improve quality of life while reducing traffic congestion and accidents, waste and pollution, and crime rates and tax burdens.
In the recent paper Reimagining City Configuration: Automated Urban Planning via Adversarial Learning, two University of Central Florida (UCF) PhD students specializing in spatiotemporal data mining, along with advisors from UCF, the Chinese University of Hong Kong and Virginia Tech, propose reducing the workloads of urban planners by introducing deep learning systems to handle some of their responsibilities.
“Traditional urban planning is time-consuming and laborious, and many factors need to be considered when generating the final planning scheme,” Dongjie Wang, a first-year UCF PhD student and first author of the paper, told Synced. “We wonder if AI can be used to automatically generate urban planning solutions.”
While the concept of AI-enabled automated urban planning is appealing, the researchers quickly encountered three challenges: how to quantify a land-use configuration plan, how to develop a machine learning framework that can learn the good and the bad of existing urban communities in terms of land-use configuration policies, and how to evaluate the quality of the system’s generated land-use configurations.
The researchers began by formulating the automated urban planning problem as a learning task on the configuration of land-use given surrounding spatial contexts. They defined land-use configuration as a longitude-latitude-channel tensor with the goal of developing a framework that could automatically generate such tensors for unplanned areas.
The team developed an adversarial learning framework called LUCGAN to generate effective land-use configurations by drawing on urban geography, human mobility, and socioeconomic data. LUCGAN is designed to first learn representations of the contexts of a virgin area and then generate an ideal land-use configuration solution for the area.
They researchers identified residential communities as “central areas” based on their latitude and longitude, then observed related contexts and extracted explicit features of these contexts from value-added space, points-of-interest (POIs) distribution, and traffic conditions. They then mapped the explicit feature vectors to the geographical spatial graph as attributes of corresponding nodes.
The researchers obtained context embeddings by utilizing graph embedding techniques to fuse all explicit features and spatial relations in the context. Based on expert knowledge, they distinguished between “excellent” and “terrible” land use configurations, inputting that information and the context embedding into LUCGAN to teach it how to distribute “excellent” plans.
“Our research is use-oriented,” Wang explains. “Since we’ve already seen preliminary results of automatic planning of urban layouts, we believe this path will work.”
The approach does have its limitations. For example, the semantics and stability of the solution generated by the algorithm still have room for improvement, and the team says it plans to address this in future work.
Wang believes the framework can eventually be developed into an open-source software package that can be used by urban planning professionals. With some corresponding preferences manually set in the algorithm, it can automatically prioritize various influencing factors in the surrounding environment and generate an appropriate planning scheme tailored to community needs. Human urban planners will also be able to adjust the generated results.
“Although AI and automation help improve work efficiency, we believe humans are the ultimate decision-makers,” Wang says. “The algorithms can touch on areas often ignored by people’s inertial thinking while humans can modify the algorithms’ results to better meet real-world policies and regulations. The combination of the two can truly provide reasonable solutions.”
The paper Reimagining City Configuration: Automated Urban Planning via Adversarial Learning has been accepted for oral presentation at the ACM SIGSPATIAL 2020, and is available on arXiv.
Reporter: Yuan Yuan | Editor: Michael Sarazen
This report offers a look at how China has leveraged artificial intelligence technologies in the battle against COVID-19. It is also available on Amazon Kindle. Along with this report, we also introduced a database covering additional 1428 artificial intelligence solutions from 12 pandemic scenarios.
Click here to find more reports from us.
We know you don’t want to miss any story. Subscribe to our popular Synced Global AI Weekly to get weekly AI updates.