Source-linked AI summary
Automated 3D reconstruction of LoD2 and LoD1 models for all 10 million buildings of the Netherlands
Ravi Peters, Balázs Dukai, Stelios Vitalis, Jordi van Liempt, Jantien Stoter
TL;DR
Different applications need different building-model detail, while existing datasets can vary in geometry, timing, and level of detail. The paper reconstructs multiple LoDs from shared building polygons and LiDAR data, producing consistent models with quality information for buildings across the Netherlands.
Problem
Existing building datasets can differ in geometry, temporal coverage, LoDs, and quality metadata, limiting consistency across areas and applications.
Source: §1 · p. 2, §2.4 · p. 4Method
The workflow partitions building footprints using aligned height points, then extrudes the roof partition into models at multiple LoDs from shared source data.
Source: §2.3 · p. 3, §3.2 · p. 5, §3.1 · p. 5Results
The workflow produces consistent LoD1.2, LoD1.3, and LoD2.2 building models from the same source data and monitors quality information throughout the process.
Source: §5 · p. 11Takeaways & Limitations
The generated data has been used to experiment with urban applications including noise, wind-flow, and energy-consumption simulations.
Source: §5 · p. 11Further work is planned to address point-cloud no-data areas, application-specific LoD needs, and missing building information such as wall types and floor counts.
Source: §5 · p. 12
Abstract
from arXiv · showhide
In this paper we present our workflow to automatically reconstruct 3D building models based on 2D building polygons and a LiDAR point cloud. The workflow generates models at different levels of detail (LoDs) to support data requirements of different applications from one consistent source. Specific attention has been paid to make the workflow robust to quickly run a new iteration in case of improvements in an algorithm or in case new input data become available. The quality of the reconstructed data highly depends on the quality of the input data and is monitored in several steps of the process. A 3D viewer has been developed to view and download the openly available 3D data at different LoDs in different formats. The workflow has been applied to all 10 million buildings of The Netherlands. The 3D service will be updated after new input data becomes available.
1. INTRODUCTION
The paper addresses inconsistent, unevenly updated Dutch 3D building models and the difficulty of producing detailed models automatically at national scale. It investigates a consistent, automated multi-LoD workflow, quality monitoring, and data dissemination.
1. INTRODUCTION
- The paper targets inconsistent models across datasets and LoDs, gaps in building coverage, and temporal differences that can affect downstream applications and processing.It proposes countrywide reconstruction from common source data and principles, with future updates based on updated inputs.Source: §1 · p. 2
- The research investigates one automated process for multiple LoDs, quality assessment at large scale, and efficient visualization and dissemination of the resulting dataset.For block models, users can select among reference heights to suit an application.Source: §1 · p. 2
- The paper presents a workflow for reconstructing LoD1.2, LoD1.3, and LoD2.2 models of all buildings in the Netherlands.The paper describes the research scope, methodology, implementation, and conclusions in Sections 2–5.Source: §1 · p. 2
2. SCOPE OF THE RESEARCH AND PREVIOUS WORK
The paper frames reconstruction choices around application needs and available data, then focuses on three LoDs and reviews large-area 3D city models. Its scope favors consistent, automated outputs while excluding representations that require generalized footprints or manual work.
2.1 LoD in relation to urban applications
- Urban 3D city models have no one-size-fits-all detail level: LoD1 block models suffice for shadow, wind, and noise simulations; LoD2 roof structures with roof-material information are needed for solar-potential or accurate energy-demand estimation.Source: §2.1 · p. 2
2.2 LoD in relation to reconstruction method
- LoD1 models can be generated automatically, but differing height references and calculations can change analytical outcomes; LoD2 reconstruction remains an active research topic.LoD3 elements are difficult to reconstruct automatically and are often generated manually or converted from BIM models.Source: §2.2 · p. 3
2.3 The LoDs in our research
- The study reconstructs LoD1.2, LoD1.3, and LoD2.2; LoD1.3 captures significant height jumps, while generalized-footprint and LoD3 representations are out of scope.LoD1.3 is included for more realistic visualization and more accurate block-shaped simulation inputs; it is harder to generate than LoD1.2 because height discontinuities must be detected.Source: §2.3 · p. 3
2.4 3D city models of large areas
- Existing large-area models are often one-time captures, cover few LoDs consistently, and lack detailed reconstruction and quality metadata.When LoDs use different workflows, their geometry, temporal aspects, and semantics may be inconsistent.Source: §2.4 · p. 4
3. METHODOLOGY FOR AUTOMATED 3D RECONSTRUCTION
The methodology reconstructs multiple LoDs from building polygons and LiDAR height points, using the LoD2.2 process to produce information reused by other reconstructions. The input data and workflow are described as a unified basis for the outputs.
3. METHODOLOGY FOR AUTOMATED 3D RECONSTRUCTION
- The process delivers LoD1.2, LoD1.3, and LoD2.2 models as well as 2D-plus-heights data, and starts with LoD2.2 because its outputs support the other LoD reconstructions.The overview presents six representations reconstructed in the process.Source: §3 · p. 4, Figure 1 · p. 4
3.1 Input data
- Building polygons from BAG and airborne-LiDAR point data from AHN3 are the inputs; BAG buildings are selected to represent the current situation.AHN3 provides classified points, with 10 cm height accuracy; the study uses building and ground classes for building and ground heights, respectively.Source: §3.1 · pp. 4–5
3.2 LoD2.2
- The LoD2.2 method partitions aligned footprints and height points into roof parts, then extrudes the 2D partition into a 3D model.The paper identifies adding LoD2.2 output as the main improvement over its earlier work.Source: §3.2 · p. 5
Footprint partitioning
Footprint partitioning uses detected roof-plane boundaries and intersections to create an initial partition, which is then simplified by assigning roof planes to its faces.
Footprint partitioning
- The initial partition is formed by projecting detected lines into the footprint; roof-plane assignment and optimization produce the final partition shown before the LoD2.2 mesh.The figure depicts building-polygon and surface points, roof-plane and line detection, initial and final partitions, and the resulting mesh.Source: Footprint partitioning · p. 5, Figure 2 · p. 6
- Graph-cut optimization assigns roof planes to partition faces to minimize point-cloud error and the length of edges between different planes.Boundary lines come from roof-plane α-shapes, intersection lines mark adjacent-plane intersections, and duplicate or same-plane edges are removed to reduce complexity.Source: Footprint partitioning · p. 6
Extrusion
The reconstruction produces LoD2.2 meshes and two LoD1 variants from roof-part geometry, with different simplification and height-reference choices. The LoD1 outputs are also supplied as 2D footprints or roof parts with height attributes.
Extrusion
- LoD2.2 roof parts are extruded into a topologically correct mesh of ground, roof, and wall surfaces, without internal walls.The ground plane uses the 5th percentile of ground points within a 4 m building buffer.Source: Extrusion · p. 7
3.3 LoD1.3
- LoD1.3 merges adjacent roof parts with height differences below 3 m, then extrudes remaining parts using the 70th-percentile roof height.It provides 2D roof parts with four height attributes and 3D solids without inner walls; the shared ground height is the 5th percentile of surface points within 4 m.Source: §3.3 · p. 7
3.4 LoD1.2
- LoD1.2 extrudes each original footprint using the 70th-percentile roof height and assigns four reference heights to the 2D footprint.The four reference heights are minimum, maximum, 50th percentile, and 70th percentile; the same surface height is used as for the other LoDs.Source: §3.4 · p. 7
4. IMPLEMENTATION
The implementation tiles buildings for reconstruction and dissemination, then stores the outputs for export or direct use. It provides a public viewer and per-model quality information, including RMSE results for LoD2.2.
4.1 Implementation of reconstruction process
- Quadtree leaves contain at most 3,500 buildings, balancing processing workload while keeping download tiles similar in file size.The tiling also clusters buildings spatially so the corresponding point cloud can be read efficiently.Source: §4.1 · p. 8, Figure 4 · p. 8
- All ten million buildings are reconstructed in about 40 hours using 30 concurrent processes on one machine, and processing scales linearly with building count.Buildings are processed independently, allowing rapid reruns after algorithm improvements or new input data.Source: §4.1 · p. 9
4.2 Visualisation and dissemination
- The 3D viewer supports browsing, querying, and downloading the reconstructed models in several formats.Users can click buildings for height and surface-slope properties; downloads include GeoPackage, PostGIS backup, Wavefront OBJ, and CityJSON.Source: §4.2 · pp. 9–10, Figure 6 · p. 10
4.3 Quality information
- LoD2.2 RMSE is below 31 cm for 95% of models and below 9 cm for 75% of models.Per-model attributes cover source-data quality and reconstruction success; a more extensive evaluation is in progress.Source: §4.3 · p. 11, Figure 7 · p. 11
5. CONCLUSIONS
The paper presents one workflow for reconstructing multiple LoDs from the same source data, with quality information to support users’ fit-for-purpose decisions. The data is used in urban applications, while temporal and geometric version management and pipeline alignment remain topics for future study.
5. CONCLUSIONS
- One process reconstructs LoD1.2, LoD1.3, and LoD2.2 models from the same source data for large areas, supporting consistent data for different applications.The models support different reference heights for block models.Source: §5 · p. 11, Figure 8 · p. 12
- The generated 3D data has been used in noise, wind-flow, and energy-consumption applications.The authors identify temporal and geometric version management and better alignment among process components as subjects for future research.Source: §5 · p. 12
- The AHN4 point cloud yields more reconstructed detail than AHN3 because it has higher resolution.The authors note that the next AHN version is expected to provide more reconstruction possibilities.Source: Figure 8 · p. 12, §5 · p. 12