Prediction building age from urban form at large scale

A supervised machine learning approach to predict the construction year of buildings based on publicly available 2D urban morphology data.

Demo

See prediction demo Jupyter notebook using building data from France, Spain, and the Netherlands.

Dependencies

Using pip:

$ pip install -r requirements.txt

Using conda:

conda env create --file=environment.yml

Reproduce results

• Data: Harmonized European buidling data used from https://eubucco.com/data [1].
• Features engineering: Urban form features crafted with eubucco.
• Experiments: All experiments conducted to answer the research questions are defined in bin/experiments.py.
• Deployment: A Slurm cluster was utilized for model training. Submit a Slurm job to reproduce the experiments using:
  • sbatch bin/slurm-submit/submit-prepare.sh to prepare the data
  • sbatch bin/slurm-submit/submit-preliminary.sh for all preliminary experiments
  • sbatch bin/slurm-submit/submit-exp.sh for all main experiments
• Figures: All figures and tables have been created with notebooks/v0_1_prediction.ipynb (partially outdated).

[1] Milojevic-Dupont, Nikola, and Wagner, Felix, Hu, Jiawei, Zumwald, Marius, Nachtigall, Florian, Biljecki, Filip, Heeren, Niko, Kaack, Lynn, Pichler, Peter-Paul, & Creutzig, Felix. (2022). EUBUCCO (v0.1). Zenodo. https://doi.org/10.5281/zenodo.6524781

Abstract

To stay within 1.5°C of global warming, reducing energy-related emissions in the building sector is essential. Rather than generic climate recommendations, this requires tailored, low-carbon urban planning solutions and spatially explicit methods that can inform policy measures at urban, street and building scale.

Here, we propose a scalable method that is able to predict building age information in different countries using only open urban morphology data.

We find that spatially cross-validated regression models are sufficiently robust to generalize and predict building age in unseen cities with a mean absolute error (MAE) between 15.3 years (Netherlands) and 19.9 years (Spain). Our experiments show that large-scale models improve generalization for predicting across cities, but are not needed to infer missing data within known cities.

We further find that classification outperforms regression for use cases where only the construction period is of interest such as energy modeling. Overall, our results demonstrate the feasibility of generating missing age data in different contexts across Europe, providing important initial results for large-scale data generation projects such as EUBUCCO. We also highlight challenges posed by data inconsistencies and urban form differences between countries that need to be addressed for an actual roll-out of such methods.

Contact

For any questions, please contact [email protected].