University of Twente Student Theses
Automatic building roof plane structure extraction from remote sensing data for lod2 3D city modelling
Campoverde, Carlos (2023) Automatic building roof plane structure extraction from remote sensing data for lod2 3D city modelling.
|
PDF
6MB |
| Abstract: | Roofs are fundamental parts of buildings, and mapping their structure represents an active and emerging research area in urban-related studies. Knowing roof characteristics can lead to more accurate and detailed 3D building models. Creating detailed 3D Building models involves more than just the basic shape of the building; it also includes the structural details of the roof. 3D Building models can derivate into highly detailed 3D city models, opening up a world of applications, such as enhanced urban planning, solar potential estimation, telecommunica-tions planning, transportation, and creating digital twins for urban planning. The rapid advancements in remote sensing technologies have allowed a variety of new datasets for geospatial analysis. In the same way, machine learning has experimented with similar growth, mainly deep learning. The potential for accurately and efficiently deriving object features from images is increasingly promising. While nu-merous deep learning approaches for extracting roof structures have been proposed, challenges persist. These include the regularization of output, false detections and misclassifications, and low computational efficiency. These ongoing issues highlight the need for continued research and innovation. This study explores a cutting-edge deep learning method for planar graph reconstruction applied to building roof structure extraction. A framework has been designed to delineate and extract regularized building roof plane struc-tures, using aerial imagery and the building footprint information across an entire area. The framework is built on top of the work developed by Chen et al. (2022) Holistic Edge Attention Transformer (HEAT), which harnesses the power of an attention-based neural network deployed to detect corners and classify interconnecting edges for planar graph reconstruction in RGB images. In addition, the generated roof plane structures have been tested for their applicability in generating a 3D city model by integrating information from the Digital Surface Model (DSM), Digital Terrain Model (DTM), and Normalized Digital Surface Model (nDSM). The process was performed by using 3D tools from Geographic Information Systems (GIS). The performance of the approach was evaluated using extensive quantitative metrics and qualitative visual analysis. The research is founded on experiments conducted in three distinct geographical locations with different topolo-gies of roof structures: the Stadsveld – 't Zwering neighborhood, and the Oude Markt area, in Enschede, The Netherlands, and the Lozenets neighborhood in Sofia, Bulgaria. The approach started by using the pre-trained HEAT model for outdoor architecture reconstruction to harness the model's learned planar graph reconstruction knowledge. Subsequently, the model underwent training on datasets created from the Stadsveld – 't Zwering neigh-borhood, Lozenets neighborhood, and a combination of both datasets. The results show that the models tailored to specific study areas delineate building inner roof plane structures with the same performance as the model trained on a combined dataset. However, when the models were tested in the Oude Markt, the model trained with a combined dataset demonstrated superior performance with an F score value of 0.43 for building inner roof plane delineation against the F score value of 0.37 from the model trained only on the Stadsveld – 't Zwering neighborhood dataset, and 0.32 from the model trained only on the Lozenets dataset. The obtained building inner roof planes show substantial potential for urban applications and continuous improve-ment. Through this study, we explored new pathways for improving computational efficiency in detecting and extracting roof plane structures, thus contributing to advancing urban-related studies and a step forward in auto-mated frameworks for digital twin cities. |
| Item Type: | Essay (Master) |
| Faculty: | ITC: Faculty of Geo-information Science and Earth Observation |
| Subject: | 74 (human) geography, cartography, town and country planning, demography |
| Programme: | Geoinformation Science and Earth Observation MSc (75014) |
| Link to this item: | https://purl.utwente.nl/essays/96138 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page