Identification of Potential Sinkhole Signatures: Employing Time Series Clustering for Anomaly Detection in InSAR Time Series Over Limburgs Mining District

This paper presents a machine learning-based approach for detecting ground deformation patterns from time series indicative of potential sinkholes in the mining district of Lim-burg, the Netherlands. Our research uses large datasets of un-labeled, high-dimensional InSAR time series from Sentinel-1 and RADARSAT-2 satellites to identify anomalies that signal ground subsidence risks. Our methodology involves an anomaly detection process using dimensionality reduction, clustering time series data and cluster labeling. One of the challenges we face in the Limburg mining district is that there are no annotated datasets and no target labels that are required for a supervised learning approach, therefore we label the clusters by expert judgement. By adopting unsupervised learning, we aim to uncover unknown patterns in the time series, grouping similar temporal features into clusters for further analysis.