Building Stock Classification Using Machine Learning: A Case Study for Oslo, Norway

Abstract

This paper describes a new concept to automatically characterize building types in urban areas based on publicly available image databases, making parts of seismic risk assessment more time and cost-effective, and improving the reliability of seismic risk assessment, especially in regions where building stock information is currently not documented. One of the main steps in evaluating potential human and economic losses in a seismic risk assessment, is the development of inventory databases for existing building stocks in terms of load-resisting structural systems and material characteristics (building typologies classification). The common approach for building stock model classification is to perform extensive fieldwork and walk-down surveys in representative areas of a city, and in some cases using random sample surveys of geounits. This procedure is time and cost consuming, and subject to personal interpretation: to mitigate these costs, we have introduced a machine learning methodology to automate this classification based on publicly available image databases. We here use a Convolutional Neural Network (CNN) to automatically identify the different building typologies in the city of Oslo, Norway, based on facade images taken fromin-situfieldwork and from Google Street View. We use transfer learning of state-of-the-art pretrained CNNs to predict the Model Building Typology. The present article attempts to categorize Oslo’s building stock in five main building typologies: timber, unreinforced masonry, reinforced concrete, composite (steel-reinforced concrete) and steel. This method results in 89% accuracy score for timber buildings, though only 35% success score for steel-reinforced concrete buildings. We here classify and define for the first time a relevant set of five typologies for the Norwegian building typologies as observed in Oslo and applicable at national level. In addition, this study shows that CNNs can significantly contribute in terms of developing a cost-effective building stock model.