Exploration of latent space of LOD2 GML dataset to identify similar buildings
share
Explainable numerical representations of otherwise complex datasets are vital as they extract relevant information, which is more convenient to analyze and study. These latent representations help identify clusters and outliers and assess the similarity between data points. The 3-D model of buildings is one dataset that possesses inherent complexity given the variety in footprint shape, distinct roof types, walls, height, and volume. Traditionally, comparing building shapes requires matching their known properties and shape metrics with each other. However, this requires obtaining a plethora of such properties to calculate similarity. In contrast, this study utilizes an autoencoder-based method to compute the shape information in a fixed-size vector form that can be compared and grouped with the help of distance metrics. This study uses "FoldingNet," a 3D autoencoder, to generate the latent representation of each building from the obtained LOD2 GML dataset of German cities and villages. The Cosine distance is calculated for each latent vector to determine the locations of similar buildings in the city. Further, a set of geospatial tools is utilized to iteratively find the geographical clusters of buildings with similar forms. The state of Brandenburg in Germany is taken as an example to test the methodology. The study introduces a novel approach to finding similar buildings and their geographical location, which can define the neighborhood's character, history, and social setting. Further, the process can be scaled to include multiple settlements where more regional insights can be made.
READ FULL TEXT
