InSAR Displacement Time Series Mining: A Machine Learning Approach

Kurzfassung

Interferometric Synthetic Aperture Radar (InSAR)-derivedsurface displacement time series enable a wide range of ap-plications from urban structural monitoring to geohazardassessment.With systematic data acquisitions becomingthe new norm for SAR missions, millions of time series arecontinuously generated. Machine Learning provides a frame-work for the efficient mining of such big data. Here, we focuson unsupervised mining of the data via clustering the similartemporal patterns and data-driven displacement signal re-construction from the InSAR time series. We propose a deepLong Short Term Memory (LSTM) autoencoder model whichcan exploit temporal relations in contrast to the commonlyused shallow learning methods, such as Uniform ManifoldApproximation and Projection (UMAP). We also modify theloss function to allow the quantification of uncertainties inthe time series data. The two approaches are applied to theLazufre Volcanic Complex located at the central volcaniczone of the Andes and thereby compared.