Building facade segmentation of oblique aerial images using convolutional neural networks for urban climate modeling

Abstract

In this thesis essential information about the urban surface is derived from oblique aerial images. This information serves as input for a modern and high-efficient urban climate model that is developed within the MOSAIK project. It aims at providing a tool of reliable urban climate models to evaluate and access changes of environmental conditions to support urban planning. First of all, segmentation of building facades allow to determine a window fraction. For this task, the state-of-the-art convolutional neural network architecture Mask R-CNN is trained with labeled terrestrial open-source data. During training image augmentation is applied to simulate the aerial perspectives. The knowledge about the window fraction paves the way to estimate radiation transfer and heat diffusion through windows. Furthermore, predominant colors of the facade are extracted to provide the wall albedo by applying a K-means clustering. The facade color could help to determine the surface material and thereby the heat capacity of the facade.