Automatic reconstruction and efficient visualization of 3D city models

Today, migration from rural areas, continuous change in cities and the increasing complexity of their structure yielded the need of new methods to increase efficiency in their management. The demand for 3 dimensional (3D) city models is increasing and they are actively used by countries and municipalities at different scales. 3D city models are not only visual models, but also allow analysis and different visualization applications with the help of their semantic data. These models can be produced in different levels of detail (LoD), and as the levels increase, the amount of modeled objects/details of the building and roof also increases. 3D models with a high level of detail are produced manually by photogrammetry operators, usually with the help of very high resolution stereo aerial photographs. However, this process is costly in terms of labor and time. There exist different approaches in the literature to automatically generate highly detailed 3D city models, but the topic is still an active research area being investigated by several researchers. On the other hand, efficient visualization of the produced models also involves optimization issues and depending on the platform, and different approaches exist. Within the scope of this thesis, deep learning-based solutions have been developed for automatic classification of building roof types from very high resolution optical imagery, automatic extraction of building footprints, automatic extraction of roof details at LoD2.2 level and their use in the production of 3D building models. The study sites were selected from different regions of Türkiye and the training data were prepared in accordance with the requirements of the deep learning methods. The results are presented and suggestions for potential improvements are discussed. In addition, different solutions for the visualization of LoD2 and LoD3 city models are developed and discussed. For this purpose, web-based visualization with Cesium library and virtual reality supported Unity game engine were employed to reveal various advantages and disadvantages of both approaches.