Causally-informed geospatial artificial intelligence for scalable, low-carbon urban planning

Cities are essential for climate change mitigation, as they concentrate large shares of population, economic activity, and emissions. At the same time, unprecedented amounts of open, geospatial data on urban form – characterized by a city’s buildings, streets, land use, points of interest, and sociodemographics – enable the use of artifi- cial intelligence (geospatial AI) to enhance urban decision-making. However, despite geospatial AI’s great potential, this thesis identifies four gaps that prevent its uptake for low-carbon urban planning: It is unclear (1) how open but scattered urban form data can be utilized across many cities. In addition, it has hardly been demonstrated how geospatial AI can develop planning advice that (2) is neighborhood-specific, (3) considers cause-effect relationships between observed variables, and (4) is scalable across many cities. Here, this thesis seeks to close these gaps by answering the overarching research question: How can open urban form data and geospatial AI enhance low-carbon urban planning at scale? It assesses the individual components of this question via three interconnected studies. At first, this thesis investigates the usability of open but scattered data by har- monizing 206 million urban form data points, specifically building footprints, across Europe. It demonstrates that data is, in principle, available for climate-relevant anal- ysis. Yet, a lack of data infrastructure and critical gaps in data coverage, attributes, and temporal dimensions constrain urban planning utility at scale. Then, it investigates the potential of geospatial AI for urban planning using the example of urban form and inner-city travel data. To evaluate how geospatial AI can reveal neighborhood-specific insights, it applies explainable machine learning to a sample of 3.5 million car commutes and high-resolution urban form data in Berlin. Assessing the non-linear relationship between urban form variables and travel per neighborhood reveals the importance of subcenters, inner-city compact development, and releasing peripheral low-income communities from car dependence. To assess the cause-effect relationships between observed variables and to scale the analysis, this thesis extends the modeling approach with a causal structure and applies it to 10 million mobility data points of six cities across three continents. Using causal graph discovery reveals significant causal dependencies between individual urban form variables, which had been neglected in previous work. Considering the causal struc- ture, the explainable machine learning model shows that high access matters more across cities than high density or street connectivity. At the same time, effect magni- tudes and locations vary depending on a city’s centrality and size. Finally, it identifies city-specific suburban neighborhoods that benefit more from densification than higher access, highlighting the role of subcenter development. This thesis underscores the need for harmonized, openly available data and location- specific, causally informed, scalable urban planning approaches. The findings provide a foundation for developing targeted, evidence-based strategies and tools for creating low-carbon urban futures, adaptable to diverse city contexts worldwide.