Building characteristics, such as storeys and types, play an important role across a multitude of domain. However, geospatial data on the building stock is often fragmented and incomplete.

Here, we propose a novel method to predict a set of building characteristics within diverse contexts, addressing the existing data gaps. Our method exploits the geospatial connectivity between street-level contexts and building stocks, employing Graph Neural Networks (GNNs) for modelling spatial patterns to infer multiple building characteristics.

For demonstration, we provide a sample_data.csv in Boston for running this notebook. This include:

• Building data in Boston
• Street-level urban objects, obtained from OSM. Here is an example to retrieve the spatial distribution of points of interests via Overpass API:

[out:json];
{{geocodeArea:Boston}}->.searchArea;
// gather results
(
  // query part for: amenity=*”
  node["amenity"](area.searchArea);
  way["amenity"](area.searchArea);
  relation["amenity"](area.searchArea);
);
// print results
out body;
>;
out skel qt;

We are using graph neural networks (GNNs) to model the spatial connectivity between buildings and their surrounding urban objects at street scale. Therefore, such spatial instrcutrue can be exploited to infer building characterstics.

To aggregate street-level contexts, we map urban objects around a building of interest within three scales - 50m, 200m and 500m.

To infer multiple building characteristics across cities, we use our Graph-based method gnn_building_prediction.ipynb to train and evaluate the predictions.

1. Compare predicted building storeys and ground truth in Boston

2. Compare predicted building type and ground truth in Boston

The predicted building characteristics have potential to support a variety of use cases. For instance, the number of storeys can be used for generating a 3D city model. We take Boston as an example to demonstrate the adoption of height-related information in 3D GIS.

To do so, a simple way is to extrude footprints to a height estimated from the number of storeys to a 3D city model and output a CityJSON dataset:

• Boston buiding footprints. The prepared data is reprojected to USA Massachusetts Mainland CRS (EPSG:2249)
• Extrude footprints to 3D buildings (The code mostly courtesy of cityjson-software)
• The generated 3D city model for Boston

A paper about the work was published in Computers, Environment and Urban Systems.

If you use this work in a scientific context, please cite this article.

Lei B, Liu P, Milojevic-Dupont N, Biljecki F (2024): Predicting building characteristics at urban scale using graph neural networks and street-level context. Computers, Environment and Urban Systems 111: 102129. doi: 10.1016/j.compenvurbsys.2024.102129

@article{lei2024predicting,
  title={Predicting building characteristics at urban scale using graph neural networks and street-level context},
  author={Lei, Binyu and Liu, Pengyuan and Milojevic-Dupont, Nikola and Biljecki, Filip},
  doi={10.1016/j.compenvurbsys.2024.102129},
  journal={Computers, Environment and Urban Systems},
  volume={111},
  pages={102129},
  year={2024},
  publisher={Elsevier}
}

It is licensed under the terms of the MIT license.