Preprints
https://doi.org/10.5194/egusphere-2022-1095
https://doi.org/10.5194/egusphere-2022-1095
 
04 Nov 2022
04 Nov 2022
Status: this preprint is open for discussion.

AutoTerm: A "big data" repository of Greenland glacier termini delineated using deep learning

Enze Zhang1, Ginny Catania1,2, and Daniel Trugman3 Enze Zhang et al.
  • 1The University of Texas at Austin, Institute of Geophysics, TX 78758, USA
  • 2The University of Texas at Austin, Department of Geological Sciences, TX 78712, USA
  • 3University of Reno, Nevada, Nevada Seismological Laboratory, NV 89557, USA

Abstract. Ice sheet marine margins via outlet glaciers are susceptible to climate change and are expected to respond through retreat, steepening, and acceleration, although with significant spatial heterogeneity. However, research on ice-ocean interactions has continued to rely on decentralized, manual mapping of features at the ice-ocean interface, impeding progress in understanding the response of glaciers and ice sheets to climate change. The proliferation of remote sensing images lays the foundation for a better understanding of ice-ocean interactions and also necessitates the automation of terminus delineation. While deep learning (DL) techniques have already been applied to automate the terminus delineation, none involve sufficient quality control and automation to enable DL applications to “Big Data” problems in glaciology. Here, we build on established methods to create a fully automated pipeline for terminus delineation that makes several advances over prior studies. First, we leverage existing manually-picked terminus traces (16,440) as training data to significantly improve the generalization of the DL algorithm. Second, we employ a rigorous automated screening module to enhance the data product quality. Third, we perform a thoroughly automated uncertainty quantification on the resulting data. Finally, we automate several steps in the pipeline allowing data to be regularly delivered to public databases with increased frequency. The automation level of our method ensures the sustainability of terminus data production. Altogether, these improvements produce the most complete and high-quality record of terminus data that exists for the Greenland Ice Sheet (GrIS). Our pipeline has successfully picked 278,239 termini for 295 glaciers in Greenland from Landsat-5, -7, -8, Sentinel-1, and -2 images, spanning from 1984 to 2021 with an average uncertainty of ~37 meters. The high sampling frequency and the controlled quality of our terminus data will enable better quantification of ice sheet change and model-based parameterizations of ice-ocean interactions.

Enze Zhang et al.

Status: open (until 30 Dec 2022)

Comment types: AC – author | RC – referee | CC – community | EC – editor | CEC – chief editor | : Report abuse
  • RC1: 'Comment on egusphere-2022-1095', Anonymous Referee #1, 30 Nov 2022 reply
  • RC2: 'Comment on egusphere-2022-1095', Anonymous Referee #2, 04 Dec 2022 reply

Enze Zhang et al.

Viewed

Total article views: 401 (including HTML, PDF, and XML)
HTML PDF XML Total Supplement BibTeX EndNote
286 104 11 401 30 4 3
  • HTML: 286
  • PDF: 104
  • XML: 11
  • Total: 401
  • Supplement: 30
  • BibTeX: 4
  • EndNote: 3
Views and downloads (calculated since 04 Nov 2022)
Cumulative views and downloads (calculated since 04 Nov 2022)

Viewed (geographical distribution)

Total article views: 394 (including HTML, PDF, and XML) Thereof 394 with geography defined and 0 with unknown origin.
Country # Views %
  • 1
1
 
 
 
 
Latest update: 06 Dec 2022
Download
Short summary
Glacier termini are essential for studying why glaciers retreat, but they need to be mapped automatically due to the volume of satellite images. Existing automated mapping methods have been limited due to limited automation, lack of quality control, and inadequacy in highly diverse terminus environments. We design a fully automated, deep-learning-based method to produce termini with quality control. We produced 278,239 termini in Greenland and provided a way to deliver new terminus regularly.