CityJSON LOD3.3 Enrichment Using Zero-Shot Learning on Mobile Mapping Data

Abstract. This paper presents an automated workflow for enriching existing LoD2 building models to LoD3 by integrating aerial LiDAR, mobile mapping imagery, and zero-shot vision–language models. The approach combines TU Delft’s Geoflow for geometric reconstruction with Grounding DINO for façade element detection, followed by homography-based perspective correction and spatial reasoning filters to merge redundant detections. Parameter studies demonstrate that optimized Geoflow configurations achieve sub-decimeter accuracy, while the zero-shot detector reaches an average detection score of 83% with a false alarm rate below 10%. The final CityJSON models, validated through CJVal, show 95% geometric and semantic compliance with international standards. The proposed proof of concept demonstrates scalable, data-driven LoD3 reconstruction without retraining, bridging computer vision and geospatial modeling for large-scale urban digital twins.