Generative AI for Spatial Planning Tasks
In collaboration with the Province of Utrecht (PU) feasibility studies were performed and generative AI (genAI) systems were built to examine the spatial reasoning capabilities of genAI and their potential value for the Spatial Planning department of the PU.
The DSP was sole responsible for developing the genAI systems in collaboration with the PU.
Objectives
The project consisted of 3 feasibility studies, all with the same aim of examining spatial reasoning capabilities of generative AI models, specifically Large Language Models (LLMs).
- Feasibility study 1: post-training LLM's on maps [fail]
- Feasibility study 2: fine-tuning LLM's on 12 years of archival spatial planning data [fail]
- Feasibility study 3: building an agentic reasoning workflow for performing a spatial planning task [success]
The main limitations for the project were the necessary time, costs, and expertise for successfully finishing each feasibility study itself. As such, the first 2 feasibility studies failed due to not meeting these limitations. Important to note is that this does not mean the approaches themselves were deemed not to be feasible, merely that the resources available to conduct the feasibility tests were inadequate to provide a conclusive answer.
Feasibility study 1: post-training LLM's on maps
The first feasibility study looked at connecting LLM's with visual reasoning modules with stronger textual reasoning LLM's. The LLM's were to be provided maps and metadata for post-training, and using these generate coordinate vectors that were proper outlines of elements within the map.
This first study was unable to answer the question due to the technical complexity of post-training LLM's -- plus the little available pre-prepped data with metadata proper for spatial planning use -- costing more time than available within the project.
Feasibility: fail
Feasibility study 2: fine-tuning LLM's on 12 years of archival spatial planning data
The second feasibility study looked at fine-tuning existing LLM's -- being less technically complex than post-training an LLM as attempted in the first study -- using a large archive of spatial plans from ruimtelijkeplannen.nl (a repository of Dutch spatial plans ranging from 2012-2024).
This second study was unable to answer the question due to the contamination present in the archive, leading to an impossible amount of data to clean given the time allocated for the project.
Feasibility: fail
Feasibility study 3: building an agentic reasoning workflow for performing a spatial planning task
The third feasibility study used a smaller scope of a single task of spatial planners within the PU, namely that of creating 'kanskaarten' (opportunity maps) that indicate where certain objects can or cannot be placed within the confines of the province looking at legal constraints and geographic realities.
The third study was a success, showcasing an agentic workflow that utilised legal reasoning to provide lists of limitations (lijst van belemmeringen) including 3 categories of limitations (strong, complex, weak), which was subsequently used by a geographic reasoning agent to create polygonal maps of the province where limited areas were accurately categorised and 'cut out' of the map.
Feasibility: success
Conclusion
Within this project we showcased the feasibility of using genAI within tasks of the spatial planning department of the PU. Code was properly handed over, including documentation and report, to technical specialists of the PU that are able to continue the work provided by the DSP. The PU was very happy with the results and the possibilities of the feasibility study, leading to follow-ups internally.
Contact
Contact Fabian (fabian.kok@hu.nl) for more information about this project.