I present my ongoing work using the emulator from the Instructed Glacier Model (IGM) (https://github.com/jouvetg/igm) to invert for ice thickness at the 200,000 glaciers in the world outside the polar ice sheets. The basis of the emulator is a convolutional neural network trained on the outputs of full-Stokes simulations of real glaciers. Provided with surface velocities – taken from the new global dataset compiled by Millan et al. (2022) – and surface DEMs, this emulator can invert for thickness at any glacier in the world with a comparable accuracy to traditional full-Stokes inversion, but at a fraction of the computational cost. This allows us to greatly improve our estimates of global glacier volume, vital both for prediction of sea-level rise, but also for local communities in mountainous areas, who often rely on glacier melt for a large proportion of their water resources. I will discuss the rationale and methods behind my work, as well as preliminary results and the problems I’m currently working on.
Research on surge-type glaciers, although they constitute a small percentage of all glaciers, significantly contributes to understanding glacier flow mechanisms. Recent studies have utilized remote sensing techniques to unravel these processes, highlighting the potential of combining high-performance computing and earth observation. While modeling surge events has gained popularity, there is a lack of comprehensive spatial and temporal data on surge timing. To address this, an algorithm has been developed that not only detects surge-type glaciers but also determines the onset of surges. The algorithm relies on time series analysis of glacier surface velocity using Sentinel-1 data, involving seasonal and trend decomposition and outlier detection via the General Studentized Extreme Deviate Test. Cluster analysis is then applied to identify outlier clusters associated with glacier surges. The method’s effectiveness was demonstrated in the Svalbard archipelago between 2015 and 2021, where 18 glacier surges and their timing were successfully identified.
This article discusses the importance of tracking changes in glacier calving front positions as a means of assessing glacier status. Remote sensing imagery is a valuable tool for this purpose, but manual monitoring for all global calving glaciers is impractical due to time constraints. The article introduces a novel framework called AMD-HookNet, designed for the segmentation of glacier calving fronts in synthetic aperture radar (SAR) images. AMD-HookNet enhances feature representation by leveraging an attention mechanism and interactions between low-resolution and high-resolution inputs. The experiments conducted on a benchmark dataset demonstrate that AMD-HookNet outperforms the current state of the art by achieving a mean distance error of 438 meters to the ground truth, confirming its effectiveness.
The concept of the sequential sampling is well known and still widely used in industrial automation radar applications due to its hardware simplicity, low cost and low power consumption. However, there is only a limited amount of publications that describe this concept and its variants in detail. This tutorial introduces the typical sequential sampling impulse radar concept step by step and presents several key characteristics, such as correlation properties and SNR considerations. In addition to the system theory, selected applications are presented to illustrate the attractiveness and elegance, but also the limits of this radar concept. The shown applications range from those in industrial automation to radar concepts in the areas of automotive radar, security scanners, biomedical radar systems and ground penetrating radar. The latter application is among others including the analysis of glaciers, e.g. to determine the ice thickness and stratification in order to provide a better understanding of the glacier and its behavior in a changing climate.
Right to the beginning of the IDP M3OCCA, we organized a “kick-off” workshop at FAU Erlangen in June 2022.
The workshop was the first gathering of all IDP M3OCCA Ph.D. students, including affiliated Ph.D. students. Many subproject supervisors were also around, so the workshop provided the fundament for a fruitful exchange and collaboration between all different institutions involved in the IDP M3OCCA.
Targetting snow drift and refreezing in glacier mass budgets with machine learning
Glacier surface mass balance (SMB) modelling has traditionally focused on improving the determination of surface energy balance conditions. However, water routing in the firn including refreezing, as well as the complex patterns of surface accumulation caused by snow drift, are poorly, or not at all, represented in most SMB models. Studies in the literature have often noted the uncertainties due to neglecting such mechanisms. This sub-project aims to improve the model representation of (i) the internal water budget and (ii) the snow drift contribution to the SMB using deep learning.
This aim links ideally with ongoing snowdrift measurements on Hintereisferner in the frame of a running joint project between FAU and the University of Innsbruck (http://schism-project.info/). It furthermore links very well with our strong background in SMB modelling and developing process-based model parameterizations. We hypothesize, however, that the accuracy of such parameterizations can be increased drastically by including machine learning algorithms and new measurements on snow drift dynamics.
In this subproject, we will explore the associated added value by sensitivity simulations and analyses of the resultant mass and energy balance components. An improved knowledge of the mechanisms in question is also important for the precise estimation of sea level contributions from glaciers measured from altimetry and DEM differencing (SP2.2 and 2.3), since the measured elevation changes have to be converted to mass using a certain density assumption. The SP also links to SP1.2 and SP1.3 where internal structures of the snow and firn cover are mapped.
For specific information on the sub-project please contact: Prof. Dr. Thomas Mölg, Institut für Geographie, Wetterkreuz 15, 91058 Erlangen, T: 09131-85-22633, thomas.moelg@fau.de, https://www.geographie.nat.fau.de/
Co-PIs: L. Nicholson, R. Prinz (Univ. Innsbruck), R. Hock (Univ. Oslo)
Want to learn more about Manuel’s research? Click here to access a short video! (the webcam video is from www.foto-webcam.eu)
My name is Manuel Saigger and within the IDP MOCCA I will be working on the influence of snow drift and internal refreezing on the glacier mass balance. Before I started working in Erlangen, I studied atmospheric sciences at the University Innsbruck. There I mainly focused on weather forecasting and weather simulations in alpine terrain and got really fascinated by the complex nature of flow fields in the surrounding of mountains. These complex wind patterns can in return lead to very complex patterns in snow accumulation, for example by redistribution of the snow.
Here my project comes into play. Because despite possibly being very important, this redistribution is not reflected in most glacier mass balance models, which on the one hand creates uncertainty in these models, but on the other hand gives us the opportunity to work on this problem and learn more about these fascinating processes.
Glacier outlines from optical and SAR imagery by deep learning
Glacier extent is an ECV and various international attempts exist to harmonize a global glacier inventory data set with outlines if possible from multiple data sources (GLIMS, IASC Randolph Glacier Inventory, RGI).
However, to date no real operational large-scale repeat mapping capabilities exist. Glacier outlines are required at different time intervals e.g. with matching observational intervals for specific mass balance computations out of remote sensing measurements (e.g. for comparison with in-situ observations or on regional scale) or to validate ice dynamic modelling. In particular for small fast changing glaciers area updates are important since errors from area mismatch are highest.
While for calving glaciers various deep learning approaches exist outlines of land-terminating glaciers are often more difficult to delineate, in particular when the glacier tongue is debris-covered. We aim at an approach that integrates the advantages of synthetic aperture radar and optical observation capabilities and analysis techniques. The candidate is supposed to advance our existing processing setup and to develop and systematically test a new method. Within this doctoral project we aim at training a machine learning algorithm using multi-temporal SAR coherence images jointly with other data. This SP has strong thematic links to SP3.2 and 3.3 and methodological to SP 2.2.
For specific information on the sub-project please contact: Prof. Dr. Matthias Braun, Institut für Geographie, FAU, Wetterkreuz 15, 91058 Erlangen, T: +49 9131-85-22015, matthias.h.braun@fau.de, https://www.geographie.nat.fau.de/
Co-PIs: E. Bänsch (FAU Mathematics), P. Rizzoli (DLR HR), M. Zemp (Univ. ZH)
Originally coming from Michigan in the United States, I received my bachelor’s degree in Mechanical Engineering from the University of Michigan in April 2017. After subsequently working in the aviation field for a year, I moved to Munich to pursue my master’s degree in Earth Oriented Space Science and Technology at the Technical University of Munich. This program explored the uses of remote sensing, the technology behind it, and the geophysical processes to be observed. Through this, my desire to pursue research in the cryosphere was inspired and I started research in the Institute of Geography at the Friedrich-Alexander-Universität Erlangen-Nürnberg in May 2021.
This research focused on using deep learning techniques to detect supraglacial lakes in Northeast Greenland using Sentinel-2 data. Having been accepted to the IDP M3OCCA program, I am excited to now pursue my doctoral degree with the focus on the improvement of glacier outline segmentation using various remote sensing data while using the strengths of artificial intelligence.
Reconciling machine learning and glacier system modelling
Within this sub-project, we envisage to combine model-driven simulations of a coupled glacier system with data-driven machine learning techniques at specific benchmark sites in the Alps (e.g., Hintereisferner, Vernagtferner). A well-established ice-flow model will first be employed to produce observation-informed training and test data, which will then inform machine learning components that can replace such classical methods. The ultimate goal is to reduces computational efforts and prepare such hybrid techniques for long timescales and/or regional scales.
Detailed Information and position requirements of this SP are formulated in this PDF.
For specific information on the sub-project please contact: Dr. Johannes Fürst (NGL)
Department of Geography and Geosciences, Institute of Geography
Wetterkreuz 15, 91058 Erlangen, T: 09131-85-26680, Johannes.fuerst@fau.de Prof. Dr. Harald Köstler
Department of Computer Sciences, System Simulation
Cauerstraße 11, 91058 Erlangen, T: 09131-85-29359, harald.koestler@fau.de
Co-PIs: I. Tabone (FAU Geography)
O. Galiardini (Univ. Grenoble)
F. Maussion (Univ. Innsbruck)
Get to know our project affiliated PhD students
Mamta KC
mamta.kc@fau.de
Want to learn more about Mamta’s research? Click here to access a short video!
After graduating with a master’s degree in Environmental and Natural Resources from Kathmandu University, Nepal, I opted for the second master’s program in Climate and Environmental Science at the FAU Institute of Geography to broaden my horizon and enhance my environmental data analysis knowledge. During my master’s program, my interest in glacier environment, climatology, and machine learning grew.
The potential of deep learning applications to understand real-world physical phenomena captured my attention. Using efficient deep learning-based models to learn physics-based glacier models is a promising field of research. Thus, I look forward to incorporating machine learning into glacier modeling during my Ph.D. in the IDP M3OCCA program. My research will focus on developing physics-based deep learning emulators that will learn to predict the glacier dynamics. The emulators will be trained using the simulated data generated from the different glacier-based physical models.
