Integrating Heterogeneous Geospatial Data Sources: Semantic Enrichment of Building Functions via Geometric Matching

Kurzfassung

This thesis develops and implements a transparent, rule-based pipeline to enrich remotely sensed building footprints with semantic use types. Using CNN-derived building complex footprints as the target geometry and OpenStreetMap (OSM) buildings, cadastral buildings, and OSM land-use polygons as sources, we follow a union-and-resolve strategy: all intersecting candidate labels are retained and deterministically resolved into a proportion vector over four classes (residential, work, shopping, other). This design explicitly represents mixed-use rather than forcing a single label. Applied to Berlin, the combined sources substantially increase labeling coverage compared to any single source. A considerable share of buildings is multi-class (mixture rate at E =0.05: 8.22%), and the spatial pattern of normalized entropy shows expected centers of mixed use. External validity checks support plausibility: area weighted residential share correlates with census population density (Spearman r=0.52), and zones designated as mixed use exhibit distinctly higher mixture rates than single-use zones. A case study uses the pipeline to enrich building footprints with tenure proxies and links them to block-level migrant shares in a multilevel model. We find that houseownership patterns are dominated by centrality and built form, with a modest positive association with migrant share. We provide a well-documented, reproducible Python implementation. Limitations include sensitivity to small geometric overlaps and city-specific thresholds. The resulting dataset supports applications in infrastructure planning, exposure modeling, and urban policy.