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

Abstract

This study aims to extend the building anomaly detection method using self-supervised learning (SSL) developed in the RepreSent project [1] from Rome (Italy) to Bucharest (Romania). By incorporating European-wide Building Footprint Datasets, we seek to understand the patterns of building anomalies in Bucharest's urban environment. Analysis-ready Dataset Creation The Municipality of Bucharest has an extensive list of buildings at risk based on in-situ analysis, primarily located in the city center [2]. As seen in Figure 1, the buildings are categorized based on their urgency for demolition, seismic fragility, and other factors. The categories include: Urgent demolitions Less urgent ones (based on seismic fragility) No-risk buildings Reinforced buildings However, the main limitation of this dataset is that some addresses are missing, incorrect, or changed. All of this information is stored in a tabular format. Data Preparation Steps: Plain text addresses are converted into coordinates using the Nominatim service of OpenStreetMap. Each coordinate point is paired with the closest building footprint to identify buildings at risk, utilizing EUBUCCO building footprint records [3]. InSAR time series data, fetched from the European Ground Motion Service, are filtered for each building footprint to finalize the dataset. Some missing buildings requires manual annotation, which is targeted for the next stage of the project. Model Development Our current framework leverages Moment-FM [4], a foundational model for time series data, as illustrated in Figure 4. In the task of building anomaly classification, we employ Moment-FM as a feature embedder, utilizing its encoder weights without additional fine-tuning. The embedding vectors derived from the building InSAR time series are then used to train a classifier using traditional machine learning algorithms. To assess the performance of the embeddings obtained via SSL from Moment-FM, we compare them against a widely used conventional time series encoding technique, Rocket [5]. [1] Kuzu, R. S., Bagaglini, L., Wang, Y., Dumitru, C. O., Braham, N. A. A., Pasquali, G., ... & Zhu, X. X. (2023). Automatic Detection of Building Displacements Through Unsupervised Learning From InSAR Data. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. [2] https://www.nordimo.ro/lista_imobilelor_expertizate_cu_grad_seismic.pdf [3] Milojevic-Dupont, N., Wagner, F., Nachtigall, F., Hu, J., Brüser, G. B., Zumwald, M., ... & Creutzig, F. (2023). EUBUCCO v0. 1: European building stock characteristics in a common and open database for 200+ million individual buildings. Scientific Data, 10(1), 147. [4] Goswami, M., Szafer, K., Choudhry, A., Cai, Y., Li, S., & Dubrawski, A. (2024). Moment: A family of open time-series foundation models. arXiv preprint arXiv:2402.03885. [5] Dempster, A., Petitjean, F., & Webb, G. I. (2020). ROCKET: exceptionally fast and accurate time series classification using random convolutional kernels. Data Mining and Knowledge Discovery, 34(5), 1454-1495.