Semantic labelling of building types A comparison of two approaches using Random Forest and Deep Learning

Kurzfassung

Knowledge about semantic building types is crucial for various applications, especially in the context of sustainable regional and urban planning. Official information containing semantic building type is scarce, most often only available for urban areas and inhomogeneous. Hence, the present thesis presents two methodological approaches to conduct semantic classification of building types based on different available data sources to generate homogeneous information on semantic building types for large-scale applications. Based on the assumption that footprint data of buildings containing height information is available, the first framework evaluates the performance of semantic building type classification using the machine learning classifier Random Forest. Semantic classification is conducted for residential buildings, differentiating four semantic classes using morphometric and topological features: Single-Family Homes - Detached Buildings, Semi-Detached Buildings, Terraced Buildings, and Multi-Family Homes - Apartment Buildings. Several set-ups were carried out using different sets of features and applying the algorithm on data with different geographical, and hence structural, context. Classification results showing high accuracies for almost every employed set-up are promising. However, if no building footprint data is available another approach for semantic building type classification is presented. Using Fully Convolutional Neural Networks, a Deep Learning architecture, semantic segmentation of aerial images, based only on the spectral characteristics, is conducted. Three semantic classes are differentiated: Auxiliary Buildings, Non-Residential Buildings and Residential Buildings. A first evaluation of this approach for semantic buildings type classification shows very promising results.