Learning from urban form to predict building heights

Area of Science:

• Urban planning and remote sensing
• Geospatial data analysis
• Machine learning applications in urban studies

Background:

• Sustainable urban planning requires high-resolution building stock data for effective policy implementation.
• Existing 3D building models are costly to create and maintain, posing challenges for many cities.
• There is a need for cost-effective methods to estimate building characteristics where detailed data is unavailable.

Purpose of the Study:

• To develop and validate a machine learning model for predicting building heights.
• To utilize open-access geospatial data on urban form for height prediction.
• To assess the model's performance in regions without existing 3D data and evaluate the impact of citizen-contributed data.

Main Methods:

• A machine learning approach was employed to predict building heights.
• The model was trained using open-access geospatial data, including building footprints and street networks.
• Model performance was evaluated using data from four European countries and tested in Brandenburg, Germany.

Main Results:

• Urban fabric morphology was found to be highly predictive of building height.
• The model achieved an average prediction error well below typical floor height (approx. 2.5m) in Brandenburg.
• Even limited citizen-collected local height data significantly improved prediction accuracy.

Conclusions:

• The developed method enables reliable and cost-effective prediction of building heights using readily available geospatial data.
• This approach facilitates the estimation of missing urban infrastructure data, supporting climate change mitigation strategies.
• Open government data and volunteered geographic information are valuable resources for scalable scientific applications in urban studies.