This study presents a workflow for learning compact geometric embeddings of urban form directly on vector data of building groups. It leverages a spectral model based on a Non-Uniform Fourier Transform and a Multi-Layer Percepton to encode building groups representations in an unsupervised way. The learned embeddings are evaluated on real-world urban data, highlighting strengths and limitations and demonstrating the potential of the approach.
All data used to are derived from OpenStreetMap and retrived via the osmnx Python library. The code to gather the data and contruct the building groups is provided in the notebook 1_BuildingBlocks.ipynb.
Most dependencies are specified within the requirements.txt file. Additional dependencies are contained in the \dependencies directory and derive from previous work by [1] and [2]. Please refer to the notebook 0_SetUp.ipynb for additional information and requirements instructions.
All code required to reproduce the experiment is provided. The notebooks should be executed in the following order:
0_SetUp.ipynb: setup and requirements instructions.1_BuildingBlocks.ipynb: retrieval of OpenStreetMap data and construction of building blocks and clusters.2_BaselineGroupModel.ipynb: computation of geometric embeddings for building clusters and related analysis.
It is recommended to execute the code in Google Colab and to enable GPU, especially for2_BaselineGroupModel.ipynb.
The folder land_cover_task contains additional experiments; see the README inside for details.
[1] Lo Presti L., Mooney P., "Road network simplification to support air quality analysis", GISRUK conference proceedings, 2025. https://doi.org/10.5281/zenodo.15089554.
[2] Mai, G., Jiang, C., Sun, W. et al., "Towards general-purpose representation learning of polygonal geometries". Geoinformatica 27, 289-340 (2023). https://doi.org/10.1007/s10707-022-00481-2.