Automated building typology clustering and identification using a variational autoencoder on digital land cadastres

Abstract

This study introduces a novel, automated methodology for extracting urban building typologies from digital land cadastres using a Variational Autoencoder (VAE). Unlike traditional shape clustering approaches, that depend on predefined rules or manual labelling, the method employs unsupervised learning to identify building typologies, based solely on geometric features, derived from roof-print shapes. Leveraging a large-scale dataset of over 100,000 buildings from the Seville, Spain cadastre, the VAE has been trained with augmented data to generate a latent space that captures dominant morphological patterns. The analysis has revealed 24 to 26 distinct building typologies, encompassing both prevalent and rare urban forms. The approach effectively filters out non-representative shapes and is scalable for application across entire cities. By automatically identifying representative building shapes, the method facilitates the creation of parametric structural models, which are essential for developing machine learning tools to predict seismic damage. This replicable and automated strategy significantly reduces the time and resources required for typology-based seismic vulnerability assessments, providing valuable support for civil protection agencies and urban planners.