Klassifizierung von InSAR-generierten Zeitreihen unter Nutzung von maschinellem Lernen

Translated title of the contribution: Classification of InSAR-based Time Series using Machine Learning

Selmin Sakiri

Research output: ThesisMaster's Thesis


In this work, the possibilities of data segmentation (clustering) of time series using algorithms of unobserved learning (machine learning) are examined. The caverns used for storing fossil fuels in the Etzel cavern storage facility cause surface movements. These were measured over a specific period using Radar interferometry methods by satellites, specifically InSAR. The aim of this work was to segment the observed horizontal and vertical displacements. Specific cluster algorithms from the field of machine learning, known as cluster algorithms in Pandas, were used to cluster (segment) the existing datasets based on certain features and properties of the time series. First, the K-Shape method, an adapted version of the well-known KMeans algorithm, was applied. Second, the Gaussian Mixture Models (GMM) method was used, whereby the time series were separated based on statistical metrics. The number of possible clusters was estimated using the Ellbow method and the datasets were segmented with different cluster numbers ranging from n=2-6. The third possible method used was clustering using statistical indicators. Both K-Shape and GMM provided similar results in terms of trends, outliers, and characterization of the datasets. However, differences were observed in terms of quality and numerical results Thus, the two methods deliver a different ¿optimal¿ number of clusters. While both methods classify two clusters as optimal for the horizontal data, the GMM tends towards three to four clusters for the vertical data, while K-Shape again rates two as sufficient. This difference is due to the skewed distribution of the vertical data sets. GMM are used in normal distributions, resulting in a difference to the K-shape results. The distribution of the time series clusters is also similar in the visual results, and seasonal components are evident and similar in the plots of both methods. Through additionally comparing of the results with older data and a forecast model, these two clustering methods can be adapted and verified. Clustering using statistical metrics is helpful for less complex cases as it provides an initial insight, but it offers limited possibilities in terms of evaluating and interpreting the data itself. Clustering of time series in the field of ground movements proves to be a promising approach to attribute characteristics to the datasets. However, the respective results need to be verified with forecasting models and conventional results given the black-box problem.
Translated title of the contributionClassification of InSAR-based Time Series using Machine Learning
Original languageGerman
Awarding Institution
  • Montanuniversität
  • Tost, Michael, Co-Supervisor (internal)
  • Moser-Tscharf, Alexander, Supervisor (internal)
  • Benndorf, Jörg, Supervisor (external), External person
Award date20 Oct 2023
Publication statusPublished - 2023

Bibliographical note

no embargo


  • Mine surveying
  • InSar Measurments
  • Mining subsidience
  • Clustering
  • Time Series Analysis
  • geodesy
  • Machine Learning

Cite this