Investigating Automatic Detection Methods for Retrogressive Thaw Slumps with TanDEM-X-Derived Digital Elevation Models

• Main Authors: Maier, Kathrin, Bernhard, Philipp, Hajnsek, Irena, id_orcid:0 000-0002-0926-3283
• Format: Conference Object
• Language: English
• Published: ETH Zurich, Institute for Environmental Engineering 2023
• Subjects: Arctic; Canada; Ice; permafrost; Thermokarst
• Online Access: https://hdl.handle.net/20.500.11850/655440; https://doi.org/10.3929/ethz-b-000655440

Permafrost underlies approximately 15 % of the landmass in the Northern Hemisphere and is becoming more susceptible to rapid thawing as the climate continues to warm (Obu et al. 2019). When ice-rich permafrost thaws it can alter the surface characteristics of a landscape which is commonly referred to as thermokarst. Retrogressive Thaw Slumps (RTS) are emerging as one of the most dynamic types of thermokarst, varying strongly in shape and thawing behavior. The prevalence and distribution of RTSs on a pan-Arctic scale are not well understood and so is its potential contribution in the Arctic carbon-climate feedback (Kokelj et al. 2013). High-resolution Digital Elevation Models (DEMs) are a valuable tool for monitoring surface characteristics of thermokarst features and track changes over time, especially being able to directly measure volumetric changes in RTS development essential for deriving carbon mobilisation rates which in turn has a great potential to improve our understanding of the impact of rapid thaw to carbon release and climate feedback mechanisms (Turetsky et al. 2020). The TanDEM-X missions merits investigation for a comprehensive monitoring of rapid permafrost thaw and directly retrieve information about volumetric change rates and thus carbon mobilization by differencing multi-temporal DEMs due to the full Arctic coverage and suitable spatial resolution. This approach has already been successfully applied to single-pass InSAR-based time-series DEM analysis to detect and quantify volumetric change rates and potential carbon mobilization of RTSs in several test sites in the Arctic permafrost region between 2011 and 2017 (Bernhard et al. 2020, Bernhard et al. 2022b). In this paper we present the learnings from a case study in the Northwestern Territories, Canada investigating the potential of TanDEM-X derived DEMs for automatic detection of RTSs over large geographic scales based on deep learning. We developed a detection method using the instance segmentation model Mask R-CNN (He et al. 2017) and ...