Topological and geometric measures derived from space floor polygons for classification of spaces in buildings by function and form

Across the building life cycle, applications generate substantial, potentially unstructured data about architectural spaces. Our study aims to automatically enrich unstructured space data to classify architectural spaces by function and form. Our first contribution is a space partitioning method to derive the complete topology for unstructured space layouts represented as sets of space floor polygons in a robust manner. We use the same method to derive unique convex decompositions of floor polygons that model their internal topology. The second and third contributions are space function and form classifiers that leverage advanced measures computed from floor polygon topology and geometry. Topology derivation, graph and geometric measures computation, and space classification are steps in an automated data processing pipeline. We evaluated the pipeline with a space layout dataset comprising 1760 spaces of apartment buildings. We successfully performed topology derivation and convex decomposition for each floor polygon set. Graph and geometric measures were generally weakly or moderately correlated. We used these measures as features to compare six space function classification models based on graph deep learning. Classification performance was higher for space adjacency graph measures and lower for subspace adjacency graph and floor polygon measures, indicating that these measures were less informative than anticipated. We used subspace adjacency and floor polygon measures as shape descriptors in a space form classification method based on shape matching. We defined twelve non-convex form descriptors and successfully detected ten floor polygon forms in the dataset, indicating that the measures are highly informative for space form classification. • Adapted a space partitioning method to derive the complete topology for polygon sets. • Application to room polygon sets and convex decomposed non-convex room polygons. • Computation of graph and geometric measures from derived topology. • Measures used as features for space function classification with graph deep learning. • Measures used as shape descriptors for space form classification with shape matching.