Data-Aware 3D Model Enrichment: A Multimodal Pipeline Combining LiDAR Point Clouds and Geotagged Imagery

Abstract

Enhancing 3D city models with facade-level semantics is essential for urban analyses such as flood damage estimation, energy performance modelling, and accessibility assessment. Many existing enrichment methods rely on a single data source and do not systematically produce standardised CityJSON geometry. This study presents a data-aware pipeline that enriches LoD2 building models with windows, doors, and floor levels through two interchangeable workflows, one operating on Mobile Laser Scanning (MLS) point clouds and one on geotagged street-level imagery, that converge to a shared floor estimation and CityJSON enrichment backend. Both workflows employ the Grounding DINO open vocabulary detector in zero shot mode, eliminating the need for task specific training data. Evaluated on data from the Vesdre valley (Belgium), the LiDAR workflow processed 206 buildings and achieved 95.5% and 97.0% within one count accuracy for window and door counts, with 90.8% exact match floor count accuracy. The image workflow processed 338 buildings from a separate area, reaching 93.5% and 95.0% within one count accuracy for windows and doors, with comparable floor estimation, the lower opening accuracy reflecting street-level occlusions. The resulting enriched CityJSON models contain Window, Door, and FloorSurface semantics at LoD3, demonstrating a scalable approach to urban model enrichment that adapts to locally available data.