InSAR time series analysis for sinkhole detection using deep learning

A. Kulshrestha*

*Corresponding author for this work

Research output: ThesisPhD Thesis - Research UT, graduation UT

244 Downloads (Pure)

Abstract

Sinkholes are hazards that are usually small in size, yet can cause serious damage to life and property. Sinkholes may exhibit precursory spatio-temporal deformation patterns, which can be mapped by Interferometric Synthetic Aperture Radar (InSAR) time series. To efficiently detect sinkholes using InSAR time series observations, this work develops and demonstrates methodologies to learn spatio-temporal patterns from previous sinkhole events using conventional and deep-learning based models, and to use them to detect sinkhole-related anomalies in SAR datasets.

Firstly, I developed a least squares estimation based method called sinkhole scanner that tests the best fit for sinkhole shapes in InSAR deformation time series. The sinkhole shaped models fitted with ~40% lower posterior variance over the sinkhole area than other areas. Secondly, I developed a Long-Short Term Memory (LSTM) based deformation time series classification method that classified sinkhole-related anomalies and separated them non-anomalous classes. The model separated anomalous and non-anomalous deformation classes with ~99% accuracy. Thirdly, I devised a novel method to radarcode geographical data so as to use it with SAR data to generate training data for deep learning applications. Fourthly, I classified sinkholes in InSAR wrapped phase images using U-Net and CNN-LSTM methods. The sinkhole fringes were classified with an F-score of >90%. We conclude that these methods could be used and further worked upon to detect sinkhole and other deformation related hazards.
Original languageEnglish
QualificationDoctor of Philosophy
Awarding Institution
  • University of Twente
  • Faculty of Geo-Information Science and Earth Observation
Supervisors/Advisors
  • Stein, Alfred, Supervisor
  • Chang, Ling, Co-Supervisor
Award date14 Jul 2023
Place of PublicationEnschede
Publisher
Print ISBNs978-90-365-5727-6
Electronic ISBNs978-90-365-5728-3
DOIs
Publication statusPublished - 14 Jul 2023

Fingerprint

Dive into the research topics of 'InSAR time series analysis for sinkhole detection using deep learning'. Together they form a unique fingerprint.

Cite this