AuthentiCity: A Multi-Source Provenance-Aware Knowledge Graph and Benchmark for 3D City Models

AuthentiCity is a multi-source, provenance-aware 3D city knowledge graph spanning five cities on three continents. This record contains the five released graph instances as offline neo4j-admin database dump archives: Hamburg (DE), Helsinki (FI), Zurich (CH), New York (US) and Tokyo (JP).

Each dump is a complete, self-contained Neo4j database: authoritative CityGML/cadastral city-model features ingested into a labeled property graph, fused with OpenStreetMap, and, for Hamburg, additionally enriched with machine-learned roof-material predictions and reconstructed LoD3 facades. Spatial R-tree layers are contained in the dumps; nothing has to be re-indexed after a restore.

Corpus totals

• 179,571,139 nodes and 218,624,128 relationships
• 3,583,114 buildings (from 4,558,196 source building features across 2,435 CityGML files)
• 9,490,346 OpenStreetMap features, of which 2.82 M are matched to a city object and 6.67 M are kept as spatially indexed standalone nodes (net-new knowledge)
• 2,946,150 confidence-weighted ENRICHED_BY correspondence edges and 181,539 HAS_POI attachments
• 176.8 GiB of Neo4j store, distributed here as 26.9 GiB of compressed dumps

Provenance model. The graph is deliberately multi-source and spans a spectrum of trust: authoritative (CityGML/ALKIS, cadastral), crowd-sourced (OSM), machine-learned (roof materials), and reconstructed (LoD3). Derived information never replaces authoritative values. Every node and edge carries a source; OSM-derived scalars live in a separate osm_* namespace beside the authoritative attributes so the two can be compared rather than silently merged; predictions carry confidence (per-material pixel coverage); cross-source correspondences are explicit edges carrying overlap_ratio, jaccard, intersection_area, match_type and is_primary, so many-to-many footprint correspondences (1:1, 1:n, n:1, n:m) resolve into canonical urban entities while every contributing piece of evidence stays traceable.

Per-city contents

Hamburg, Germany — authenticity_hamburg_v1_0.dump (3.07 GiB). LGV Hamburg LoD2 (2025), CityGML 2.0, EPSG:25832; 388,267 buildings. Fused with hamburg-260629.osm.pbf (1,188,139 OSM features, 373,211 ENRICHED_BY, 22,219 HAS_POI). ML roof materials on 194,799 buildings (50.2 %) as coverage-ranked predicted_roof_material_<i> properties; LoD3 facades on 17 anchors. 17,440,924 nodes / 26,299,391 relationships. The flagship instance: the only city carrying all four provenance layers.

Helsinki, Finland — authenticity_helsinki_v1_0.dump (0.15 GiB). Helsinki Region Infoshare LoD2 (2019), CityGML 2.0, EPSG:3879; 2,980 buildings and 31 bridges. Fused with finland-260704.osm.pbf (55,930 OSM features). 414,459 nodes / 619,964 relationships. 61 duplicate-id groups in the source resolved by splitting onto synthesized ids with source_gml_id preserved.

Zurich, Switzerland — authenticity_zurich_v1_0.dump (5.57 GiB). swisstopo swissBUILDINGS3D 3.0, LoD2.3, EPSG:2056; 78 tiles merged into one dataset; 102,668 buildings. Fused with switzerland-260704.osm.pbf (1,860,401 OSM features). 41,508,371 nodes / 45,548,405 relationships. The highest geometric detail per building in the corpus.

New York City, United States — authenticity_nyc_v1_0.dump (6.81 GiB). NYC OTI 3D Building Model (2016), 20 DA tiles (14 CityGML 1.0 upgraded to 2.0), LoD1–2, source EPSG:2263 (US survey feet) reprojected at ingest to metric EPSG:32618; 1,083,437 buildings. Fused with new-york-260703.osm.pbf (2,555,468 OSM features). 45,797,985 nodes / 54,207,134 relationships.

Tokyo, Japan — authenticity_tokyo_v1_0.dump (11.32 GiB). MLIT PLATEAU 2025 (v5), 23 special wards, 2,335 files, B-core modules only (bldg, tran, brid, frn, veg, wtr); source EPSG:6697 (geographic degrees) reprojected at ingest to metric EPSG:6677. 2,980,839 source building elements resolve to 2,005,762 distinct buildings (975,077 byte-identical ward-tile replicas deduplicated, 468 distinct same-id features split), plus 659,006 non-building city objects. Fused with kanto-260703.osm.pbf (3,830,408 OSM features). 74,409,400 nodes / 91,949,234 relationships. The largest instance, and the multilingual axis of the corpus: PLATEAU's Japanese attribute keys are preserved verbatim as queryable keys.

How to restore

Written by Neo4j 2025.10.1. A dump loads only into the same or a newer Neo4j version. Restore each dump into its own database or container instance:

cp neo4j_hamburg_lod2_roofmats_lod3_osm_20260717.dump /backups/neo4j.dump
docker run --rm -v neo4j_data:/data -v /backups:/backups neo4j:2025.10.1 \
  neo4j-admin database load neo4j --from-path=/backups --overwrite-destination

Then start Neo4j (APOC, Neo4j Spatial, GDS and GenAI plugins as in the project's Docker setup) and verify against the counts in the accompanying *_counts.txt file:

MATCH (n) RETURN count(n);
MATCH ()-[r]->() RETURN count(r);
MATCH (b:Building) RETURN count(b);
MATCH (o:OsmFeature) RETURN count(o);
MATCH ()-[r:ENRICHED_BY]->() RETURN count(r);

Sizing note: these are large graphs. Hamburg restores comfortably on a 32 GiB machine; Tokyo (99 GiB store) wants a generous Neo4j page cache — the reference configuration is 16 GiB heap and 32 GiB page cache on a 127 GiB workstation.

Quality assurance

• Lossless CityGML round-trip: ingest to Neo4j and export back reproduces every element, attribute, leaf text and coordinate token of the source (appearance module excluded by design); coordinates are stored verbatim and never rounded.
• OSM ingestion census: for every city, the fused graph was verified against the source .osm.pbf at five levels (feature completeness, lossless tags, verbatim geometry, exact attach-or-orphan partition, spatial-index and provenance registration). All five cities PASSED with zero missing and zero extra features.
• Per-city statistical and anomaly control before and after ingestion (duplicate and placeholder ids, degenerate geometry, coordinate ranges, coverage, fan-out). No building loss in any city: distinct source ids equal graph building nodes throughout.

Known limitations

• Derived layers (ML roof materials, LoD3) exist for Hamburg only. Absence of a layer means "not covered", never a negative observation — coverage flags make this explicit.
• Tokyo excludes land use (luse), terrain (dem), underground (uro) and the PLATEAU ADE extensions by design, and is predominantly LoD1.
• New York's source model carries no measuredHeight.
• Materialized building-proximity edges are deliberately absent; proximity is answered at query time via the spatial R-tree and centroid distances.
• The roof-material model was trained on OSM labels, so OSM-derived features leak for roof-material prediction tasks. Report both protocol variants (with and without the OSM layer as input).
• Source vintages differ (2016–2025) and are not harmonized; each city's graph reflects its own source date.

Source data and attribution

• Hamburg: Landesbetrieb Geoinformation und Vermessung (LGV) Hamburg, 3D-Stadtmodell LoD2 (2025) — dl-de/by-2-0.
• Helsinki: Helsinki Region Infoshare, Helsinki 3D city model, CityGML LoD2 (2019) — CC BY 4.0.
• Zurich: swisstopo, swissBUILDINGS3D 3.0 — swisstopo open government data terms.
• New York: NYC Office of Technology and Innovation, 3D Building Model (2016) — NYC OpenData terms of use.
• Tokyo: MLIT Project PLATEAU, 3D city model (2025) — PLATEAU/Japanese government open data terms.
• All cities: © OpenStreetMap contributors, extracts of 2026-06-29 / 2026-07-03 / 2026-07-04, available under the Open Database License (ODbL 1.0). See openstreetmap.org/copyright.

Because these graphs embed OSM-derived content, the released databases are distributed under ODbL 1.0. Attribution to both OpenStreetMap and the respective authoritative provider is required for any use or redistribution.

Software

The graphs were built with pykci (Python Knowledge Graph for Cities), released separately under the MIT license: github.com/hcu-cml/pykci. The ingestion, fusion and verification pipeline is fully scripted and reproducible from the source files listed above.

Papers

AuthentiCity paper:

@misc{nguyen_authenticity_2026,
      title={{AuthentiCity: A Multi-Source Provenance-Aware Knowledge Graph and Benchmark for 3D City Models}}, 
      author={Huynh Duc An Son Nguyen and Lukas Arzoumanidis and Youness Dehbi},
      year={2026},
      eprint={2607.25243},
      archivePrefix={arXiv},
      primaryClass={cs.DB},
      url={https://arxiv.org/abs/2607.25243}, 
}

AuthentiCity was built using our tool pykci - Python Knowledge Graphs for Cities:

@misc{nguyen_pykci_2026,
      title={{pykci: A Compact Urban Knowledge Graph for Semantic and Spatial Queries using LLMs}}, 
      author={Huynh Duc An Son Nguyen and Lukas Arzoumanidis and Youness Dehbi},
      year={2026},
      eprint={2607.01605},
      archivePrefix={arXiv},
      primaryClass={cs.DB},
      url={https://arxiv.org/abs/2607.01605}, 
}