Building Anomaly Detection with Self-supervised Learning. Case study: the city of Bucharest, Romania

Abstract

Building anomaly and displacement detection are critical for ensuring the safety and longevity of structures. Based on the progress of the RepreSent project, the unsupervised building anomaly detection methods based on GNN autoencoders and LSTM autoencoders using PS-InSAR have been successfully developed and demonstrated their effectiveness in detecting three types of building anomalies caused by step, noise, and trend displacements for Rome (Italy). The purpose of the current study is to enhance the ability to detect building anomalies. Given the varied and changing nature of urban environments, we aim to expand the area of study from Rome (Italy) to Bucharest (Romania). This expansion allows us to better understand the patterns of anomalies across different urban landscapes. By using the recently released European-wide Building Footprint Datasets in our models, we expect to deepen our knowledge of the relationship between various building attributes (e.g., construction year, height, seismic risk level) and the anomalies detected. We also plan to refine our anomaly detection by applying signal decomposition techniques to minimize prediction errors, particularly those associated with noise. Furthermore, our goal is to advance our detection methodology by not only identifying the occurrence of anomalies but also predicting their timing and duration. The dataset focuses on Bucharest, the capital of Romania, which faces a significant challenge due to numerous buildings from the late 19th century that have structurally deteriorated over time and do not comply with current seismic standards. According to the latest statistics released on March 29th, 2024, by the Bucharest Municipal Administration, over 2700 buildings are at risk of collapse in the event of an earthquake. This work is supported by the European Space Agency with contract as part of the RepreSent project under the Grant 4000137253/22/I-DT.