Doctor of Philosophy
Civil and Environmental Engineering
Bitsuamlak, Girma T.
Sustainable building design requires a clear understanding and realistic modelling of the complex interaction between climate and built environment to create safe and comfortable outdoor and indoor spaces. This necessitates unprecedented urban climate modelling at high temporal and spatial resolution. The interaction between complex urban geometries and the microclimate is characterized by complex transport mechanisms. The challenge to generate geometric and physics boundary conditions in an automated manner is hindering the progress of computational methods in urban design. Thus, the challenge of modelling realistic and pragmatic numerical urban micro-climate for wind engineering, environmental, and building energy simulation applications should address the complexity of the geometry and the variability of surface types involved in urban exposures. The original contribution to knowledge in this research is the proposed an end-to-end workflow that employs a cutting-edge deep learning model for image segmentation to generate building footprint polygons autonomously and combining those polygons with LiDAR data to generate level of detail three (LOD3) 3D building models to tackle the geometry modelling issue in climate modelling and solar power potential assessment. Urban and topography geometric modelling is a challenging task when undertaking climate model assessment. This paper describes a deep learning technique that is based on U-Net architecture to automate 3D building model generation by combining satellite imagery with LiDAR data. The deep learning model used registered a mean squared error of 0.02. The extracted building polygons were extruded using height information from corresponding LiDAR data. The building roof structures were also modelled from the same point cloud data. The method used has the potential to automate the task of generating urban scale 3D building models and can be used for city-wide applications. The advantage of applying a deep learning model in an image processing task is that it can be applied to a new set of input image data to extract building footprint polygons for autonomous application once it has been trained. In addition, the model can be improved over time with minimum adjustments when an improved quality dataset is available, and the trained parameters can be improved further building on previously learned features. Application examples for pedestrian level wind and solar energy availability assessment as well as modeling wind flow over complex terrain are presented.
Summary for Lay Audience
The 3D geometry modelling process of cities and landscape is one of the major challenges in micro-climate studies. This research developed an autonomous workflow that can create 3D models using data collected satellite and laser devices. The workflow applied artificial intelligence principles and has shown a potential to shorten redundant tasks. The resulting models were used for solar power potential assessment in residential buildings, pedestrian level comfort assessment in urban area and for complex terrain analysis.
Alemayehu, Tewodros F., "Autonomous 3D Urban and Complex Terrain Geometry Generation and Micro-Climate Modelling Using CFD and Deep Learning" (2023). Electronic Thesis and Dissertation Repository. 9273.