Classification of building roof types through point cloud with deep learning

Since the automatic detection and extraction of buildings has a significant impact on many areas such as urban area planning, solar potential analysis, land and disaster management, urban mapping and three-dimensional (3D) city model creation, detection and extraction of buildings has become an active research topic in the fields of Remote Sensing, Photogrammetry, Geographic Information Systems (GIS) and computer vision. Classification of building roof types from point cloud or aerial and satellite image data is an important step for all these areas as it is meaningful for defining roof surfaces separately, reconstructing and determining building shapes. In recent years, with the importance of deep learning methods for the creation of 3D building models, the need for high quality and reasonably sized data that can train deep learning algorithms has increased. Especially in the classification of building roof types, this data gap in deep learning methods using 3D point clouds has been tried to be eliminated with the RoofN3D data set. In this thesis, an artificial neural network architecture has been developed for the classification of roof types using the RoofN3D dataset, which directly processes the point clouds and produces the class label of each roof as output. In order to organize the data to be used in the developed architecture, the histograms of the point clouds belonging to the roofs were used, and it was decided how many points should be in each point cloud. The data set is made suitable for the neural network by reducing the number of points for frames containing more than the specified number of points, and by increasing the number of points by interpolation for frames containing a small number of points. The hyperparameters of the artificial neural network architecture are optimized using Bayesian optimization. In the accuracy analyzes performed using Precision, Recall and F1-Score metrics, the classification accuracy was obtained as 95% for the saddleback, two-sided hip and pyramid roof types.