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

Abstract

In the context of sustainable planning, knowledge about building type is crucial. Yet, this information is scarce and mostly inhomogeneous. In regard to the big-data era, two approaches for building type classification are presented based on different data basis. The first approach shows semantic classification of building footprints using a set of features (simple geometric, morphological and topological features) and the machine learning algorithm Random Forest. Very high accuracies for the federal states of Germany could be achieved with Kappa Coefficients between 0.87 and 0.98. The second framework presents the possibility to conduct semantic labelling of aerial images using Fully Convolutional Neural Networks. The gained accuracy in this case is a Kappa of 0.73 for the federal state of Berlin.