Roof Segmentation Towards Digital Twin Generation in LoD2+ Using Deep Learning

Abstract

There is an increasing need for digital twins of cities and their base maps, 3D city models. Creating and updating these twins is not an easy task, so automating and streamlining the process is a field of active research. A significant part of the urban geometry is residential buildings and their roofs. Modeling of roofs for urban buildings can be divided into three main areas - building detection, roof recognition and building reconstruction. The building and roofs are segmented with the help of machine learning and image processing. Afterwards the extracted information is used to generate parametric models for the roofs using methods from computational geometry. The goal is to create correct virtual models of roofs belonging to many different types of buildings. In this study, a supervised deep learning approach is proposed for the segmentation of roof edges from a single orthophoto. The predicted features include the linear elements of roofs. The experiments show that, despite the small amount of training data, even in the presence of noise, the proposed method performs well on semantic segmentation of roofs with different shapes and complexities. The quality of the extracted roof elements for the test area is about 56% and 71% for mean intersection over union (IOU) and Dice metric scores, respectively.