CR2.8

The temporal variability of glacier calving front positions provides essential information about the state of marine-terminating glaciers. These positions can be extracted from Synthetic Aperture Radar (SAR) images throughout the year. To automate this extraction, we apply deep learning techniques that segment the SAR images into different classes: glacier; ocean including ice-melange and sea-ice covered ocean; rock outcrop; and regions with no information like areas outside the SAR swath, layover regions and SAR shadow. The calving front position can be derived from these regions during post-processing.   
A downside of deep learning is that hyper-parameters need to be tuned manually. For this tuning, expert knowledge and experience in deep learning are required. Furthermore, the fine-tuning process takes up much time, and the researcher needs to have programming skills.
    
In the biomedical imaging domain, a deep learning framework [1] has become increasingly popular for image segmentation. The nnU-Net can be used out-of-the-box. It automatically adapts the U-Net, the state-of-the-art architecture for image segmentation, to different datasets and segmentation tasks. Hence, no more manual tuning is required. The framework outperforms specialized deep learning pipelines in a multitude of public biomedical segmentation competitions.   
We apply the nnU-Net to the task of glacier segmentation, investigating whether the framework is also beneficial in the domain of remote sensing. Therefore, we train and test the nnU-Net on CaFFe (https://github.com/Nora-Go/CaFFe), a benchmark dataset for automatic calving front detection on SAR images. CaFFe comprises geocoded, orthorectified imagery acquired by the satellite missions RADARSAT-1, ERS-1/2, ALOS PALSAR, TerraSAR-X, TanDEM-X, Envisat, and Sentinel-1, covering the period 1995 - 2020. The ground range resolution varies between 7 and 20 m². The nnU-Net learns from the multi-class "zones" labels provided with the dataset. We adopt the post-processing scheme from Gourmelon et al. [2] to extract the front from the segmented landscape regions. The test set includes images from the Mapple Glacier located on the Antarctic Peninsula and the Columbia Glacier in Alaska. The nnU-Net's calving front predictions for the Mapple Glacier lie close to the ground truth with just 125 m mean distance error. As the Columbia Glacier shows several calving front sections, its segmentation is more difficult than that of the laterally constrained Mapple Glacier. This complexity of the calving fronts is also reflected in the results: Predictions for the Columbia Glacier show a mean distance error of 635 m. Concludingly, the results demonstrate that the nnU-Net holds considerable potential for the remote sensing domain, especially for glacier segmentation.
    
[1] Isensee, F., Jaeger, P.F., Kohl, S.A.A. et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods 18, 203–211 (2021). https://doi.org/10.1038/s41592-020-01008-z 

[2] Gourmelon, N., Seehaus, T., Braun, M., Maier, A., Christlein, V.: Calving Fronts and Where to Find Them: A Benchmark Dataset and Methodology for Automatic Glacier Calving Front Extraction from SAR Imagery, In Prep.

How to cite: Gourmelon, N., Seehaus, T., Braun, M., Maier, A., and Christlein, V.: Dissecting Glaciers - Can an Automated Bio-Medical Image Segmentation Tool also Segment Glaciers?, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-2726, https://doi.org/10.5194/egusphere-egu22-2726, 2022.

The cryosphere is a highly active and dynamic environment that rapidly responds to changing climatic conditions. In particular, the physical processes behind glacial dynamics are poorly understood because they remain challenging to observe. Glacial dynamics are strongly intermittent in time and heterogeneous in space. Thus, monitoring with high spatio-temporal resolution is essential.

In course of the RESOLVE (‘High-resolution imaging in subsurface geophysics : development of a multi-instrument platform for interdisciplinary research’) project, continuous seismic observations were obtained using a dense seismic network (100 nodes, Ø 700 m) installed on Glacier d’Argentière (French Alpes) during May in 2018. This unique data set offers the chance to study targeted processes and dynamics within the cryosphere on a local scale in detail.

To identify seismic signatures of ice beds in the presence of melt-induced microseismic noise, we applied the supervised ML technique gradient tree boosting. The approach has been proven suitable to directly observe the physical state of a tectonic fault. Transferred to glacial settings, seismic surface records could therefore reveal frictional properties of the ice bed, offering completely new means to study the subglacial environment and basal sliding, which is difficult to access with conventional approaches.

We built our ML model as follows: Statistical properties of the continuous seismic records (variance, kurtosis and quantile ranges), meteorological data and a seismic source catalogue obtained using beamforming (matched field processing) serve as features which we fit to measures of the GPS displacement rate of Glacier d’Argentière (labels). Our preliminary results suggest that seismic source activity at the bottom of the glacier strongly correlates with surface displacement rates and hence, is directly linked to basal motion. By ranking the importance of our input features, we have learned that other than for reasonably long monitoring time series along tectonic faults, statistical properties of seismic observations only do not suffice in glacial environments to estimate surface displacement. Additional beamforming features however, are a rich archive that enhance the ML model performance considerably and allow to directly observe ice dynamics.

How to cite: Umlauft, J., Roux, P., Lecointre, A., Gimbert, F., Nanni, U., Walpersdorf, A., Rouet-LeDuc, B., Hulbert, C., Trugman, D., and Johnson, P.: Mapping Glacier Basal Sliding with Beamforming and Artificial Intelligence, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-6948, https://doi.org/10.5194/egusphere-egu22-6948, 2022.

The vast majority of the Antarctic ice sheet is covered with snow that compacts under its own weight and transforms into ice below the surface. However, in some areas, this typically blue-colored ice is directly exposed at the surface. These so-called "blue ice areas" represent islands of negative surface mass balance through sublimation and/or melt. Moreover, blue ice areas expose old ice that is easily accessible in large quantities at the surface, and some areas contain ice that extends beyond the time scales of classic deep-drilling ice cores.

Observation and modeling efforts suggest that the location of blue ice areas is related to a specific combination of topographic and meteorological factors. In the literature, these factors are described as (i) enhanced katabatic winds that erode snow, due to an increase of the surface slope or a tunneling effect of topography, (ii) the increased albedo of blue ice (with respect to snow), which enhances ablative processes, and (iii) the presence of nunataks (mountains protruding the ice) that act as barriers to the ice flow upstream, and prevent deposition of blowing snow on the lee side of the mountain. However, it remains largely unknown which role the physical processes play in creating and/or maintaining  blue ice at the surface of the ice sheet.

Here, we study how a combination of environmental and topographic factors lead to the observation of blue ice. We also quantify the relevance of the single processes and build an interpretable model aiming at not only predicting blue ice presence, but also explaining why it is there. To do so, data is fed into a convolutional neural network, a machine learning algorithm which uses the spatial context of the data to generate a prediction on the presence of blue ice areas. More specifically, we use a U-Net architecture that through convolutions and linked up-convolutions allows to obtain a semantic segmentation (i.e., a pixel-level map) of the input data. Ground reference data is obtained from existing products of blue ice area outlines that are based on multispectral observations. These products contain considerable uncertainties, as (i) the horizontal change from snow to ice is gradual and a single threshold in this transition is not applicable uniformly over the continent, and (ii) the blue ice area extent is known to vary seasonally. Therefore, we train our deep learning model with a loss function with increasing weight towards the center of blue ice areas.

Our first results indicate that the neural network predicts the location of blue ice relatively well, and that surface elevation data plays an important role in determining the location of blue ice. In our ongoing work, we analyze both the predictions and the neural network itself to quantify which factors posses predictive capacity to explain the location of blue ice. Eventually this information may allow us to answer the simple yet important question of why blue ice areas are located where they are, with potentially important implications for their role as paleoclimate archives and for their evolution under changing climatic conditions.

How to cite: Tollenaar, V., Zekollari, H., Tuia, D., Kellenberger, B., Rußwurm, M., Lhermitte, S., and Pattyn, F.: What determines the location of Antarctic blue ice areas? A deep learning approach, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-1294, https://doi.org/10.5194/egusphere-egu22-1294, 2022.

The scarcity and limited accuracy of snow and precipitation observation and estimation in high-mountain regions reduce our understanding of climatic-cryospheric processes. Thus, we compared the snow water equivalent (SWE) from winter mass balance observations of 95 glaciers distributed over the Alps, Canada, Central Asia and Scandinavia, with the cumulative gridded precipitation data from the ERA-5 and the MERRA-2 reanalysis products. We propose a machine learning model to downscale the gridded precipitation from the reanalyses to the altitude of the glaciers. The machine learning model is a gradient boosting regressor (GBR), which combines several meteorological variables from the reanalyses (air temperature and relative humidity are also downscaled to the altitude of the glaciers) and topographical parameters. Among the most important variables selected by the GBR model, are the downscaled relative humidity and the downscaled air temperature. These GBR-derived estimates are evaluated against the winter mass balance observations by means of a leave-one-glacier-out cross-validation (site-independent GBR) and a leave-one-season-out cross-validation (season-independent GBR). The estimates downscaled by the GBR show lower biases and higher correlations with the winter mass balance observations than downscaled estimates derived with a lapse-rate-based approach. Finally, the GBR estimates are used to derive SWE trends between 1981 and 2021 at high-altitudes. The trends obtained from the GBRs are more enhanced than those obtained from the gridded precipitation of the reanalyses. When the data is regrouped regionwide, significant trends are only observed for the Alps (positive) and for Scandinavia (negative), while significant positive or negative trends are observed in all the regions when looking locally at single glaciers and specific elevations. Positive (negative) SWE trends are typically observed at higher (lower) elevations, where the impact of rising temperatures is less (more) dominating.

How to cite: Guidicelli, M., Gabella, M., Huss, M., and Salzmann, N.: Snow accumulation over the world's glaciers (1981-2021) inferred from climate reanalyses and machine learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-3701, https://doi.org/10.5194/egusphere-egu22-3701, 2022.

The last few years have seen an increasing number of studies modeling glacier evolution using deep learning. Most of these techniques have focussed on artificial neural networks (ANN) that are capable of providing a regressed value of mass balance using topographic and meteorological input features. The large number of parameters in an ANN demands a large dataset for training the parameter values. This is relatively difficult to achieve for regions with a sparse in-situ data measurement set up such as the Himalayas. For example, of the 14326 point mass balance measurements obtained from the Fluctuations of Glaciers database for the period of 1950-2020 for glaciers between 60S and 60N, a mere 362 points over four glaciers exist for the Himalayan region. These are insufficient to train complex neural network architectures over the region. We attempt to overcome this data hurdle by using transfer learning. Here, the parameters are first trained over the 9584 points in the Alps following which the weights were used for retraining for the Himalayan data points. Fourteen meteorological from the ERA5Land monthly averaged reanalysis data were used as input features for the study. A 70-30 split of the training and testing set was maintained to ensure the authenticity of the accuracy estimates via independent testing. Estimates are assessed on a glacier scale in the temporal domain to assess the feasibility of using deep learning to fill temporal gaps in data. Our method is also compared with other machine learning algorithms such as random forest-based regression and support vector-based regression and we observe that the complexity of the dataset is better represented by the neural network architecture. With an overall normalized root mean squared loss consistently less than 0.09, our results suggest the capability of deep learning to fill the temporal data gaps over the glaciers and potentially reduce the spatial gap on a regional scale.

How to cite: Anilkumar, R., Bharti, R., and Chutia, D.: Point Mass Balance Regression using Deep Neural Networks: A Transfer Learning Approach, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5317, https://doi.org/10.5194/egusphere-egu22-5317, 2022.

In this talk, we propose to use neural networks in a hybrid modelling setup to learn sub-grid-scale dynamics of sea-ice that cannot be resolved by geophysical models. The multifractal and stochastic nature of the sea-ice dynamics create significant obstacles to represent such dynamics with neural networks. Here, we will introduce and screen specific neural network architectures that might be suited for this kind of task. To prove our concept, we perform idealised twin experiments in a simplified Maxwell-Elasto-Brittle sea-ice model which includes only sea-ice dynamics within a channel-like setup. In our experiments, we use high-resolution runs as proxy for the reality, and we train neural networks to correct errors of low-resolution forecast runs.

Since we perform the two kind of runs on different grids, we need to define a projection operator from high- to low-resolution. In practice, we compare the low-resolution forecasted state at a given time to the projected state of the high resolution run at the same time. Using a catalogue of these forecasted and projected states, we will learn and screen different neural network architectures with supervised training in an offline learning setting. Together with this simplified training, the screening helps us to select appropriate architectures for the representation of multifractality and stochasticity within the sea-ice dynamics. As a next step, these screened architectures have to be scaled to larger and more complex sea-ice models like neXtSIM.

How to cite: Finn, T., Durand, C., Farchi, A., Bocquet, M., Chen, Y., Carrassi, A., and Dansereau, V.: Learning and screening of neural networks architectures for sub-grid-scale parametrizations of sea-ice dynamics from idealised twin experiments, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5910, https://doi.org/10.5194/egusphere-egu22-5910, 2022.

Machine Learning (ML) has become an increasingly popular tool to model the evolution of sea ice in the Arctic region. ML tools produce highly accurate and computationally efficient forecasts on specific tasks. Yet, they generally lack physical interpretability and do not support the understanding of system dynamics and interdependencies among target variables and driving factors.

Here, we present a 2-step framework to model Arctic sea ice dynamics with the aim of balancing high performance and accuracy typical of ML and result interpretability. We first use time series clustering to obtain homogeneous subregions of sea ice spatiotemporal variability. Then, we run an advanced feature selection algorithm, called Wrapper for Quasi Equally Informative Subset Selection (W-QEISS), to process the sea ice time series barycentric of each cluster. W-QEISS identifies neural predictors (i.e., extreme learning machines) of the future evolution of the sea ice based on past values and returns the most relevant set of input variables to describe such evolution.

Monthly output from the Pan-Arctic Ice-Ocean Modeling and Assimilation System (PIOMAS)  from 1978 to 2020 is used for the entire Arctic region. Sea ice thickness represents the target of our analysis, while sea ice concentration, snow depth, sea surface temperature and salinity are considered as candidate drivers.

Results show that autoregressive terms have a key role in the short term (with lag time 1 and 2 months) as well as the long term (i.e., in the previous year); salinity along the Siberian coast is frequently selected as a key driver, especially with a one-year lag; the effect of sea surface temperature is stronger in the clusters with thinner ice; snow depth is relevant only in the short term.

The proposed framework is an efficient support tool to better understand the physical process driving the evolution of sea ice in the Arctic region.

How to cite: Sangiorgio, M., Bianco, E., Iovino, D., Materia, S., and Castelletti, A.: Arctic sea ice dynamics forecasting through interpretable machine learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-10386, https://doi.org/10.5194/egusphere-egu22-10386, 2022.

Ice lead analysis is an essential task for evaluating climate change processes in the Arctic. Ice leads are narrow cracks in the sea-ice, which build a complex network. While detecting and modeling ice leads has been performed in numerous ways based on airborne images, the dynamics of ice leads over time remain hidden and largely unexplored. These dynamics could be analyzed by interpreting the ice leads as more than just airborne images, but as what they really are: a dynamic network. The lead’s start, end, and intersection points can be considered nodes, and the leads themselves as edges of a network. As the nodes and edges change over time, the ice lead network is constantly evolving. This new network perspective on ice leads could be of great interest for the cryospheric science community since it opens the door to new methods. For example, adapting common link prediction methods might make data-driven ice lead forecasting and tracking feasible.
To reveal the hidden dynamics of ice leads, we performed a spatio-temporal and network analysis of ice lead networks. The networks used and presented here are based on daily ice lead observations from Moderate Resolution Imaging Spectroradiometer (MODIS) between 2002 and 2020 by Hoffman et al. [1].
The spatio-temporal analysis of the ice leads exhibits seasonal, annual, and overall trends in the ice lead dynamics. We found that the number of ice leads is decreasing, and the number of width and length outliers is increasing overall. The network analysis of the ice lead graphs reveals unique network characteristics that diverge from those present in common real-world networks. Most notably, current network science methods (1) exploit the information that is embedded into the connections of the network, e.g., in connection clusters, while (2) nodes remain relatively fixed over time. Ice lead networks, however, (1) embed their relevant information spatially, e.g., in spatial clusters, and (2) shift and change drastically. These differences require improvements and modifications on common graph classification and link prediction methods such as Preferential Attachment and EvolveGCN on the domain of ice lead dynamic networks.
This work is a call for extending existing network analysis toolkits to include a new class of real-world dynamic networks. Utilizing network science techniques will hopefully further our understanding of ice leads and thus of Arctic processes that are key to climate change mitigation and adaptation.

Acknowledgments

We would like to thank Prof. Gunnar Spreen, who provided us insights into ice lead detection and possible challenges connected to the project idea. Furthermore, we would like to thank Shenyang Huang and Asst. Prof. David Rolnick for their valuable feedback and support. J.K. was supported in part by the DeepMind scholarship, the Mitacs Globalink Graduate Fellowship, and the German Academic Scholarship Foundation.

References

[1] Jay P Hoffman, Steven A Ackerman, Yinghui Liu, and Jeffrey R Key. 2019. The detection and characterization of Arctic sea ice leads with satellite imagers. Remote Sensing 11, 5 (2019), 521.

How to cite: Kaltenborn, J., Ramesh, V., and Wright, T.: Ice Lead Network Analysis, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-8945, https://doi.org/10.5194/egusphere-egu22-8945, 2022.