Research pick: Sites of the underground - "Application of deep learning algorithms in the design of urban subway public art space"

Underground metro (subway) stations are no longer merely points of departure and arrival. As cities grow denser and transit networks expand, these spaces have the potential to function as some of the most widely shared public interiors in urban life. They are places where millions pass daily, cutting across age, income, and neighbourhood. They offer a rare platform for collective cultural experience. Stations can, suggests research in the International Journal of Environment and Sustainable Development, anchor local identity, narrate a city’s history, and shape how residents and visitors alike perceive the character of the urban environment.

The research addresses a practical question confronting transport authorities and urban designers: how can large-scale public art projects fit into this infrastructure as it changes? Traditional artist-led design processes, though highly creative, can be time-intensive. By contrast, deep learning has allowed computers to generate high-quality images at speed. The missing link is that the computer-generated images may not understand the cultural meaning that the images need to convey. There is also a need to take into account how well a design might be installed in a real site.

The researchers hope to bridge this gap and have developed a multi-stage framework that integrates cultural analysis, visual cognition modelling, and spatial feasibility testing into a single pipeline.

Their approach is based on a semantic labelling system. The system can organise cultural concepts, such as local history, regional traditions, and environmental identity, into a knowledge graph. This graph can map relationships between ideas, enabling the computer to understand individual symbols and how they fit with broader narratives.

The framework then uses Contrastive Language-Image Pretraining, CLIP, is a deep neural network trained on vast datasets of containing pairings of images and text. An additional layer simulates human perception through a visual attention prediction network, considering composition, spatial layout, and pedestrian flow. By predicting where passengers are likely to focus while moving through a station, the system can position key symbolic elements in high-attention zones. The researchers suggest this could improve not only the aesthetic impact of the art installation but also the way in which pedestrians navigate the subway stations.

Wang. Q. (2026) ‘Application of deep learning algorithms in the design of urban subway public art space’, Int. J. Environment and Sustainable Development, Vol. 25, No. 5, pp.44–72.