@misc{9239903,
  abstract     = {{Three-dimensional building models are becoming increasingly important in areas such as urban planning, simulations, and geographical analyses. At the same time, detailed information about roof structures is often missing from current geospatial datasets, which complicates the automated reconstruction of coherent and functional 3D building models. Point cloud data acquired through laser scanning enables detailed geometric representation of buildings, but classification of roof structures from point clouds remains a complex and resource-demanding task. In addition, the application of machine learning methods is limited by the lack of annotated training data.
This thesis investigates the possibility of classifying real-world roof structures based on point cloud data using a machine learning model trained on synthetically generated data. The study also examines how classified roof types can be represented parametrically and used for reconstruction of simplified 3D building models.
The methodology consists of four main stages:
[list]
[*]Generation of parametric 3D building models
[*]Creation of synthetic point clouds
[*]Training and evaluation of a machine learning model
[*]Reconstruction of 3D buildings based on classified roof types.
[/list]
Parametric building models are generated in CityEngine and FME using CGA rules, where parameters such as roof type, building height, and roof height are varied. The generated models are converted into synthetic point clouds, rasterized into image representations, and used to train a ResNet-based deep learning model. Real-world LiDAR data from Gothenburg were used for testing.
The results show that the model achieves high classification accuracy when both training and testing are performed on synthetic data. However, the accuracy decreases significantly when the model is applied to real-world LiDAR data, indicating that generalization between synthetic and real data remains a major challenge. The study also demonstrates that certain roof types are difficult to distinguish due to similar geometric characteristics, variations in real building structures, and differences between synthetic and real data. The classified roof types were subsequently linked to parametric roof descriptions and used to reconstruct simplified 3D building models.
In conclusion, the study demonstrates that synthetically generated training data combined with parametric modeling and machine learning has potential for automated roof classification and parametric representation of buildings from point cloud data. At the same time, improved generalization towards real-world data and more representative training datasets are identified as important areas for future development.}},
  author       = {{Åberg, Iris and Åsvärd, Thelma}},
  language     = {{eng}},
  note         = {{Student Paper}},
  series       = {{Thesis in geographical information technics}},
  title        = {{Parametric Roof Modelling Using a Machine Learning Approach}},
  year         = {{2026}},
}