Whereas most of the continent of Antarctica is covered by snow, in some areas, blue-colored ice emerges to the surface. In these blue ice areas (BIAs), mass is removed at the surface through ablative processes. This mass removal exposes deeper layers of ice that are normally located closer to the underlying bedrock. As a result, we can find old ice at the surface of BIAs, as well as the material contained within the ice, such as meteorites and terrestrial rocks. BIAs are unique locations for sampling old ice for palaeoclimatic purposes and collecting meteorites (about ⅔ of all meteorites ever retrieved on Earth come from Antarctica BIAs). Hence, a high-quality BIA map is essential for meteorite searches, the quest for the oldest ice, and surface mass balance modeling.

Prior efforts to map BIAs across the Antarctic continent using remote sensing are single-sensor based, introducing biases related to temporary snow coverage of the exposed ice, and sensor-dependent conditions such as solar illumination angles, anisotropic reflectance, or cloud coverage. To overcome these challenges, we opt for using multi-sensor observations in a deep learning framework to create a new BIA map. The observations we use are (i) radar backscatter, (ii) surface morphology, (iii) elevation, and (iv) multi-spectral reflectance. The deep learning algorithm consists of the well-established convolutional neural network U-Net, which allows for an efficient training process and inclusion of spatial context. The algorithm outputs a pixel-level prediction of blue ice presence. Moreover, by training multiple, randomly initialized models and rotating and flipping data, we obtain multiple predictions for each pixel. Thanks to this data augmentation at test time, we estimate the variation in the predictions, which we then use as an indication of uncertainty. 

We use an existing dataset of BIA outlines as reference for training the model. It is known that these existing labels are noisy due to i) large uncertainties related to the use of a single sensor, and ii) biases as a result of applying a threshold that is based on local observations over the entire continent. However, convolutional neural networks, combined with regularization methods like weight decay and batch normalization, can learn from underlying ‘clean’ patterns of noisy labels during initial epochs of training (i.e., at the start of the training process). Here, we demonstrate this noise-eliminating property by assessing the algorithm's performance on noisy pixels that are used for training, where we see that over 80% of these noisy instances are attributed correctly. Furthermore, we optimize the performance of the neural network based on a reduced set of "noise-free", hand-labeled validation data. Last,  we test the performance of our model on hand-labeled test data, therefore having a realistic estimate of the model performance on precise, so far unused data. These tests indicate that it is possible for the neural net to learn how to map blue ice from the noisy data, leading to an improved map of BIAs in Antarctica.

How to cite: Tollenaar, V., Zekollari, H., Tuia, D., Rußwurm, M., Kellenberger, B., Lhermitte, S., and Pattyn, F.: A new blue ice area map of Antarctica, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-88, https://doi.org/10.5194/egusphere-egu23-88, 2023.

Understanding how the presence of fractured ice alters the dynamics, hydrology and energy balance of glaciers and ice shelves is important in determining the future evolution of the Antarctic Ice Sheet (AIS). However, these processes are not all well understood, and large-scale quantitative observations of fractures are sparse. Fortunately, the large amount of satellite data covering Antarctica gives us the opportunity to change this.

The Sentinel-1 satellite cluster, from the European Space Agency's Copernicus programme, has acquired synthetic aperture radar (SAR) data over the AIS with a repeat period of 6-12 days for the last 8 years. A broad range of crevasse types are visible in this imagery: rifts, surface crevasses and some basal crevasses on ice shelves, and fine surface crevasses on grounded ice streams - even those bridged by snow or pixel-scale in width.

In this study, we use machine learning to automatically map crevasses in this imagery; producing monthly composite maps over the AIS at 50m resolution. We developed algorithms to partition crevasses into those on grounded and floating ice, and extract these features in parallel using a mixture of convolutional neural networks, trained in a weakly supervised way, and more traditional computer vision techniques.

Having developed parallelisable routines for the large-scale batch processing of SAR data, we have processed every Sentinel-1 acquisition over the Antarctic Ice Sheet. The resulting dense timeseries of fracture maps allows us to assess the evolution of crevasses during the Sentinel-1 acquisition period. By measuring the density of fractures we develop a method to quantify structural change on ice shelves, and investigate those of the Amundsen Sea Embayment. We show an increase in crevassing in buttressing regions of the Pine Island and Thwaites ice shelves over the last 8 years, with observed changes elsewhere in the Amundsen Sea dominated by the advection of existing crevasses.

Finally, we develop methods demonstrating how our fracture data can be assimilated into numerical modelling experiments aiming to quantify the impact of ice shelf fracture on glacier dynamics.

How to cite: Surawy-Stepney, T., Hogg, A. E., Cornford, S. L., and Hogg, D.: Mapping Antarctic Crevasses at High Spatiotemporal Resolution with Deep Learning applied to Synthetic Apertur Radar Data, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-14207, https://doi.org/10.5194/egusphere-egu23-14207, 2023.

Icebergs account for half of all ice loss from Antarctica. Their melting affects the surrounding ocean properties through the intrusion of cold, fresh meltwater and the release of terrigenious nutrients. This in turn influences the local ocean circulation, sea ice formation and biological production. To locate and quantify the fresh water flux from Antarctic icebergs, we need to track them and monitor changes in their area and thickness. While the locations of large icebergs are tracked operationally by manual inspection, delineation of iceberg extent requires detailed analysis – either also manually or through automated segmentation of high resolution satellite imagery.

In this study, we apply three machine learning techniques to 191 Sentinel-1 images between 2014 and 2020 and assess their skill to segment seven giant Antarctic icebergs between 54 and 1052 km² in size. Most previous studies to detect icebergs have focused on smaller bergs. In contrast, we aim to segment selected giant icebergs with the goal to automate the calculation of their changing area, volume, and freshwater input. Two of our techniques are standard segmentation techniques (k-means and Otsu thresholding) and the third one is a deep neural network (U-net). It is the first study to apply a deep learning algorithm to iceberg detection.

We analyse the strengths and weaknesses of the different machine learning approaches across a range of challenging environmental conditions: These include scenes where the iceberg is surrounded by deformed sea ice, when other big bergs are present and when berg fragments are close to the main iceberg. We also cover cases when the iceberg drifts close to the coast and summer images with surface thawing conditions, which invert the backscatter contrast between iceberg and ocean.

How to cite: Braakmann-Folgmann, A., Shepherd, A., Hogg, D., and Redmond, E.: Delineating giant Antarctic icebergs with Deep Learning, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-14307, https://doi.org/10.5194/egusphere-egu23-14307, 2023.

Accurate estimates of iceberg populations, disintegration rates and iceberg movements are essential to understand ice sheet contributions to global sea level change and freshwater and heat balances. Knowledge and prediction of iceberg distributions is also important for the safety and efficiency of shipping operations in polar seas. The dynamics, persistence, fragmentation rates, melt rates and dispersal of icebergs are, however, poorly understood due to a lack of automated approaches for monitoring them. Better monitoring of icebergs would help parameterise the locations and quantities of freshwater and nutrient inputs within hydrographic and ecological models respectively and help mitigate collision hazards for navigation.

Here we present a combination of Bayesian approaches to the identification of icebergs in synthetic aperture radar imagery and their subsequent tracking across multiple years. For detection we use a Dirichlet Process Mixture Model, while for tracking we adapt Bayesian Tracker, a probabilistic multi-object tracking algorithm originally developed for cell microscopy applications. We are able to reconstruct iceberg paths and lineages, which we validate against synthetic data and manual annotations. We demonstrate that icebergs across the size distribution can be tracked successfully from their point of calving in dense fields of objects, through dispersal and fragmentation, to distal locations.

How to cite: Evans, B., Hosking, S., Fleming, A., and Lowe, A.: Probabilistic detection and tracking of icebergs in the Amundsen Sea embayment, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-6495, https://doi.org/10.5194/egusphere-egu23-6495, 2023.

Surface melt occurs on most ice shelves in Antarctica each summer, with potential impacts on their strength and stability and thus on the ice sheet's contribution to global sea level rise. However, many questions remain regarding the spatiotemporal variability of surface melt and the processes driving it, particularly in East Antarctica where few in situ observations exist. Previous work in this field has largely relied on remote sensing observations to monitor the occurrence and extent of surface melt, often using metrics such as the onset and freeze-up dates of melt each summer, the number of melt days, or the cumulative melting area. Whilst such metrics are often necessary to handle the sheer volume of data produced by satellite observations, much of the information contained within the datasets is lost, hindering attempts to build a more complete picture of melt variability at different spatial and temporal scales, and thus of disentangling the different processes driving melt.

To help address this problem, we use the machine learning approach of a Self-Organising Map (SOM) and nearly two decades (2002/03–2020/21) of daily observations from the AMSR-E and AMSR-2 passive microwave sensors, gridded at a spatial resolution of 12.5 km. Here, we present results focused on the Shackleton Ice Shelf in East Antarctica, but our code, implemented in the R programming language, is openly available and can be applied to any Antarctic ice shelf, or adapted for use with other melt datasets.

Our results show that the daily distribution of surface melt on the Shackleton Ice Shelf can be described by nine representative spatial patterns of melt. These patterns demonstrate the potential for heterogeneous melt behaviour across the shelf, and thus provide insight into the influence of surface topography, katabatic winds, and surface albedo in driving surface melt. A sensitivity analysis of the SOM algorithm shows that the same general spatial patterns are returned repeatedly regardless of the parameter values used, strengthening confidence in our results and interpretation, and demonstrating the suitability of our approach. We further examine the temporal variability of the nine melt patterns, both within and across melt seasons, finding that there are no significant trends in any of the patterns. Instead, our analysis identifies a number of summers with unusual melt behaviour and also reveals correlations with shelf-wide, summer-averaged surface air temperatures, highlighting that both local and large-scale controls are important for driving surface melt in Antarctica.

How to cite: Saunderson, D., Mackintosh, A., McCormack, F., Jones, R. S., and Picard, G.: Self-organising maps and surface melt on East Antarctic ice shelves, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-10430, https://doi.org/10.5194/egusphere-egu23-10430, 2023.

The AutoICE Competition, launched on ESA’s AI4EO platform, brings together AI and Earth Observation practitioners to address the challenge of “automated sea ice mapping” from Sentinel-1 SAR data. Traversing the polar waters safely and efficiently requires up-to-date maps of the constantly moving and changing sea ice conditions showing the current sea ice extent, local concentration, and auxiliary descriptions of the ice conditions. For several decades, sea ice charts have been manually produced by visually inspecting and analysing satellite imagery.

The objective of the AutoICE challenge is to advance the state-of-the-art for automatic sea ice parameter retrieval from SAR data to derive more robust and accurate sea ice maps. The challenge design and evaluation criteria have been created with input from machine learning experts and members of the International Ice Charting Working Group (IICWG). In this competition, participants are tasked to build machine learning models using the available state-of-the-art challenge dataset and to submit their model results for each of the three sea ice parameters: sea ice concentration, stage of development and floe size. The dataset made available in this challenge contains Sentinel-1 active microwave (SAR) data and corresponding Microwave Radiometer (MWR) data from the AMSR2 satellite sensor to enable challenge participants to exploit the advantages of both instruments and to create data fusion models. Label data in the challenge datasets are ice charts produced by both the Greenland ice service at the Danish Meteorological Institute (DMI) and the Canadian Ice Service (CIS). The challenge datasets also contain other auxiliary data such as the distance to land and numerical weather prediction model data. Two versions of the challenge dataset are available, a raw dataset and a ready-to-train dataset. The datasets each consist of the same 513 training and 20 test (without label data) scenes, however, the ready-to-train version has been further prepared for model training. In addition, a number of tools are made available to help the participants get started quickly, including access to machine learning computing resources on the ESA Polar Thematic Exploitation Platform (Polar TEP). The competition was initiated on the 23rd of  November 2022 and is expected to conclude on the 17th of April 2023.

Here, we present the overall challenge, the underlying objective, the available state-of-the-art dataset and resources, the progress of the challenge and its results, as well as a sneak peek of our upcoming ASID-v3 dataset. 

How to cite: Stokholm, A., Buus-Hinkler, J., Wulf, T., Korosov, A., Saldo, R., Arthurs, D., Solberg, R., Longépé, N., and Kreiner, M.: The AutoICE Competition: Automatically Mapping Sea Ice in the Arctic, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-13038, https://doi.org/10.5194/egusphere-egu23-13038, 2023.

Sea ice plays a vital role in Earth’s human-climate system. It regulates the Earth’s overall energy balance by seasonally increasing surface albedo and reflecting solar radiation; it governs thermodynamic exchanges between the ocean and atmosphere and thereby impacts mid-latitude weather patterns; it buttresses key continental ice shelves in Greenland and Antarctica; it provides an ecosystem in which land, marine, and airborne species thrive; it enables the livelihoods of indigenous populations across the Arctic; it poses a major obstacle to global shipping logistics; and it serves as a key indicator of climate change given the sensitivity of the polar regions to anthropogenically-induced warming. Regular and automated monitoring of sea ice concentration and type may therefore prove valuable to a broad and diverse set of parties. Conventional approaches in sea ice monitoring involve the use of remotely sensed microwave radiometer data with low resolution of 6-25 km and high instrumental sensitivities to environmental factors such as atmospheric water vapour, near-surface brightness temperature, and wind-induced surface roughening. Dual-polarity synthetic aperture radar (SAR) imagery offers a higher resolution alternative, which can also distinguish between sea ice and open water year-round independent of weather conditions. However, manual interpretation of such imagery is time-consuming. In this work, we develop a deep learning system to automatically generate high-resolution maps of sea ice concentration and type using 40m-resolution SAR imagery obtained from the Sentinel-1 mission between 2017 and 2021. Focusing on the East Weddell Sea, a region where compacted sea ice is renowned for inhibiting ship navigation and an active area of iceberg calving, we train the system against reference sea ice charts produced through manual interpretation by experts. We identify strengths and weaknesses of the system and discuss implications for future research at the intersection of machine learning and polar science.

How to cite: McDonald, A., Dimasaka, J., Plumridge, M., Torry, J., Zúñiga González, A. C., van Zeeland, L., Rogers, M., and Hosking, S.: Classifying sea ice in high-resolution SAR imagery using deep learning, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-9816, https://doi.org/10.5194/egusphere-egu23-9816, 2023.

How to cite: Chen, W., Tsamados, M., Willatt, R., Takao, S., Nelson, C., Lawrence, I., Lee, S., Brockley, D., Landy, J., De Rijke-Thomas, C., Shirazi, D., Stroeve, J., and Francis, A.: Discrimination of sea ice leads and floes using Deep Learning applied to Sentinel-3 Ocean and Land Colour Instrument (OLCI) imaging spectrometer, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-17318, https://doi.org/10.5194/egusphere-egu23-17318, 2023.

ESA has developed a series of seven TEPs on different subjects to provide insight into how our oceans, atmosphere, land and ice operate and interact as part of an interconnected earth system by exploiting the unprecedented flow of high-quality global data on the state of our planet, combined with long-term EO archives, in-situ networks and models. The Polar Thematic Exploitation Platform (Polar TEP) was developed to address the particular needs of the polar community.

Polar TEP provides a complete working environment where users can access algorithms and data remotely to obtain computing resources and tools that they might not otherwise have and avoid the need to download and manage large volumes of data. This new approach removes the need to transfer large Earth Observation data sets around the world, while increasing the analytical power available to researchers and operational service providers. Polar TEP provides new ways to exploit EO and other large datasets for research scientists, industry, operational service providers, regional authorities, and policy analysts. Polar TEP provides:

• Data Discovery - Polar TEP makes satellite and other polar data easily accessible for browsing or analysis within the cloud or within the user’s own environment. The infrastructure takes care of the complexity of handling satellite imagery archives and makes the data available via web services. Users can instantly access petabytes of Sentinel, Landsat, and other Earth observation imagery, both historic and the latest acquisitions.

• Interactive Development Environment - Polar TEP offers a managed JupyterLab instance with curated base images. The platform provides different flavors of computational resources and a network file system for persistent data storage. Headless notebook execution is supported.

• Machine Learning - Polar TEP has implemented the MLflow platform to support machine learning activities. MLflow manages all stages of the machine learning lifecycle, including experimentation, reproducibility, deployment, and a central model registry.

• Execution Environment - Docker containers are used to provide processors with a separate custom environment having minimal execution overhead. The computing resources used by the execution environment are scaled to the current demand.

• Application Hosting Environment - Users can host their own applications on a VM within the Polar TEP environment.

• Story Telling - Polar TEP provides tools to communicate analysis results to other researchers or the public.

Polar TEP is an integral part of the wider polar data ecosystem, contributing to data interoperability and fostering the use of information about the polar regions to support environmental protection, safety, and sustainable economic development.

This presentation will illustrate how the power of Polar TEP to process massive amounts of data is being applied to topics such as machine learning for operational sea ice charting, daily calculations of Greenland ice sheet albedo, and providing information to support traditional ways of life in the Arctic.

How to cite: Arthurs, D.: Polar TEP – A Platform for Polar Big Data Analytics and Machine Learning, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-8706, https://doi.org/10.5194/egusphere-egu23-8706, 2023.

The emulator used for projections of the land ice contribution to sea level by 2100 in the Intergovernmental Panel on Climate Change Sixth Assessment Report (Edwards et al., 2021) used a novel approach for incorporating structural uncertainty from the underlying computer models. This statistical (Gaussian Process) emulator represented entire multi-model ensembles at once -- for the Greenland and Antarctic ice sheets, and the world’s glaciers, under the model intercomparison projects ISMIP6 and GlacierMIP, respectively -- by using a noise (nugget) term to allow for multiple estimates of sea level contribution for a given set of model input values, analogous to kriging of spatial data.

However, the emulator was rather simple in other respects: in particular, the sea level projection for each year from 2015-2100, and from each region of land ice (splitting Antarctica into 3 regions, and the glaciers into 19), were modelled independently, so temporal correlations emerged only upon smoothing the projections. The emulator was also not formally evaluated with observations, because the underlying simulations were only driven with meaningful climate forcings from 2015. These limitations have presented difficulties for users, who often need continuous time series projections and prefer, of course, these to be assessed with observations.

Here the IPCC land ice emulator is improved for interpretation and use by decision-makers by estimating spatio-temporal correlations directly from the underlying simulations (Rougier, 2008; Rougier et al., 2009), to produce meaningful trajectories of sea level contribution from each land ice source. The extent to which the land ice emulator can be evaluated with data, now and in future, is also discussed.

References:

Edwards et al. (2021) Projected land ice contributions to twenty-first-century sea level rise, Nature, 593, 74–82.

Rougier, J. (2008), Efficient Emulators for Multivariate Deterministic Functions, Journal of Computational and Graphical Statistics, 17(4):827–843.

Rougier, J.C. et al. (2009), Expert Knowledge and Multivariate Emulation: The Thermosphere-Ionosphere Electrodynamics General Circulation Model (TIE-GCM), Technometrics, 51(4), 414-424.

How to cite: Edwards, T., Rougier, J., and Turner, F.: Improving and evaluating the IPCC land ice emulator, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-6961, https://doi.org/10.5194/egusphere-egu23-6961, 2023.

Small-scale, subgrid processes on the ice sheets, such as localized surface melt, remain unnoticed by current coarse-resolution Regional Climate Models (RCMs), leading to uncertainties in climate reanalyses and projections. Deep learning allows us to enhance the spatial resolution of RCMs but requires sophisticated model development. Earlier studies have shown that rudimental techniques, such as single-image super-resolution, have failed to capture Antarctic surface melt patterns accurately, because the spatial transferability of these models is low. In this study, we add remote sensing data to a super-resolution model: daily observations of surface albedo from MODIS are used to guide the downscaling of low-resolution surface melt (RACMO2, 27 km) to a high-resolution version (RACMO2, 5.5 km) for a 20-year period, between 2001-2019. We extend a conventional SRResNet and add the MODIS data in different configurations (i.e., spatial-channel communication, content communication, and empirical-physical activation). The models are trained over the Antarctic Peninsula, for which RACMO2 simulations are available at 5.5 km resolution (Van Wessem et al., 2016). We verify the performance of the models with three independent datasets to inspect (1) the overall performance (using QuickSCAT); (2) spatial patterns (using Sentinel-1); and (3) temporal patterns (using automatic weather stations). Our work shows the potential of adding remote sensing data to deep learning-based downscaling models, leading to improved spatial transferability compared to single-image downscaling models.

How to cite: de Roda Husman, S., Hu, Z., Kuipers Munneke, P., van Tiggelen, M., Lhermitte, S., and Wouters, B.: The Added Value of Remote Sensing Data in Downscaling Regional Climate Models, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-12927, https://doi.org/10.5194/egusphere-egu23-12927, 2023.

How to cite: Turner, F., Edwards, T., and Rougier, J.: The use of multivariate Gaussian process emulation in making projections of land ice contributions to sea level rise, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-12361, https://doi.org/10.5194/egusphere-egu23-12361, 2023.

There is growing evidence that ice mélange, the granular mixture of sea ice and icebergs at the termini of tidewater glaciers, impacts ice sheet discharge through physical buttressing forces and alterations to fjord circulation via iceberg melting. However, ice mélange is a highly dynamic, fragmented and mobile phenomenon which varies over a range of timescales (e.g. hours, days, weeks) and hence is difficult to monitor using traditional ground-based and spaceborne sensors. In this contribution, we utilise high spatio-temporal satellite imagery acquired from the ICEYE small satellite constellation to assess correlations between ice mélange characteristics and tidewater glacier dynamics. ICEYE is a growing constellation of 20+ small satellites each equipped with an X-band Synthetic Aperture Radar (SAR) and capable of mapping the entire globe at least once a day with fine spatial resolution (1-3 m). We utilised the ICEYE SAR imagery to study the perennial mélange matrix at the terminus of Helheim Glacier in southeast Greenland. ICEYE SAR imagery was acquired during summer and winter to assess how seasonal ice mélange conditions impact tidewater glacier dynamics. Sentinel-1 SAR imagery and ground-based TLS 3D data from two autonomous terrestrial laser scanners (ATLAS) were used to validate remote sensing analysis and provide additional data sources for interpretation of the glaciological processes. We will report on the following: (1) a spatial texture analysis (e.g. Grey Level Co-occurrence Matrix (GLCM), Gabor Transforms) of ice mélange at the terminus of Helheim Glacier using high resolution ICEYE SAR imagery; (2) results of hierarchical and random forest classifiers to map icebergs, sea ice and open water within the ice mélange matrix; (3) quantification of glacier and mélange flow variability at daily to weekly timescales; and (4) the development of observational models correlating ice mélange texture, iceberg distributions, mélange/glacier flow rates, and tidewater glacier stability. Our case study at Helheim Glacier aims to demonstrate a new approach to rapidly monitor ice mélange conditions and tidewater glacier stability using high resolution SAR imagery. In particular, this study pushes forward our Earth Observation capabilities and will help us better understand the complex processes operating at the ice-ocean interface which is critical for improved predictions of how the Greenland Ice Sheet will evolve under a warming climate.

How to cite: Harcourt, W. D., Stearns, L. A., Shahin, M. G., and Shankar, S.: Assessment of ice mélange impacts on tidewater glacier dynamics using high resolution ICEYE imagery, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-9812, https://doi.org/10.5194/egusphere-egu23-9812, 2023.

Monitoring glacier change processes leads to a better understanding of how glaciers respond to various external forcings. For sub-annual remote sensing, Synthetic Aperture Radar (SAR) imagery is indispensable, as polar night and cloud cover limit the temporal continuity of optical imagery. Especially with the launch of the Sentinel-1 mission, the availability of suitable SAR imagery has increased substantially. This high amount of data leads to another challenge: Manual inspection is no longer feasible. Therefore, in recent years, many studies applied deep learning to automate the segmentation of calving fronts in satellite imagery. To ensure comparability between different deep learning models, they must be trained and tested on the same data with a predefined test set and evaluated with the same metrics. A dataset intended to be used for this purpose is called a benchmark dataset and needs to provide the ready-to-use satellite imagery and the corresponding calving front labels. Gourmelon et al. [1] provide such a dataset for SAR imagery of calving fronts called CaFFe (https://doi.pangaea.de/10.1594/PANGAEA.940950). CaFFe includes multi-mission data (ERS-1/2, RADARSAT 1, Envisat, ALOS, Sentinel-1A/B, TerraSAR-X, and TanDEM-X), providing a spatial resolution between 6 and 20 meters and covering the period from 1996 to 2020. It contains images of seven glaciers from Antarctica to Greenland and Alaska. For each of the 681 images contained in the benchmark, two labels are provided: One displaying the calving front versus background and the other showing different zones (ocean, rock outcrops, glacier area, and no information available). CaFFe is split into a train set and a predefined out-of-sample test set, which comprises all images from two of the seven glaciers. A split of the train set into training and validation is not specified, as different approaches like cross-validation shall be possible. The benchmark covers a wide variety of different conditions to capture the variability of SAR calving front images. For example, images with open oceans and images with ice-melange-covered oceans are included in the dataset. Especially including images featuring ice-melange is of great importance, as deep learning models have shown difficulties in accurately segmenting calving fronts under this condition. Including images with ice-melange in the train set helps models to learn accurate predictions even under these circumstances.  Adding such images to the test set ensures that evaluated models are able to cope with ice melange. The test set of CaFFe is specifically designed to be challenging, such that the generalizability of models to different conditions and spatial transferability even to other continents can be verified. Gourmelon et al. provide baselines (one for each of the available labels) complementing the benchmark dataset. Current collaborative work aims to evaluate recently published deep learning techniques for calving front extraction on CaFFe and compare it with the baselines.

[1] N. Gourmelon, T. Seehaus, M. Braun, A. Maier, and V. Christlein: "Calving Fronts and Where to Find Them: A Benchmark Dataset and Methodology for Automatic Glacier Calving Front Extraction from SAR Imagery," Earth System Science Data, vol. 14, no. 9, pp. 4287-4313, 2022, doi: 10.5194/essd-14-4287-2022.

How to cite: Gourmelon, N., Seehaus, T., Braun, M., Maier, A., and Christlein, V.: Comparability of Deep Learning Techniques for Calving Front Segmentation in SAR Imagery, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-5069, https://doi.org/10.5194/egusphere-egu23-5069, 2023.

The Arctic has been warming four times faster than the global mean over the last forty years. In response, glaciers across the Arctic have been retreating and losing mass at accelerated rates in recent years, including Greenland, Alaska, Canadian Arctic, Iceland, Svalbard and Russian Arctic. To predict their evolution with confidence, it is important to understand the mechanisms driving mass loss across the Arctic, especially the interconnected relationships between glacier retreat, ice dynamics, and mass imbalance. Over the past several decades, satellite remote sensing has been used to image glaciers over large spatial scales and at high temporal resolution. The volume of data produced, however, has challenged the traditional manual-based approaches to quantify glacier calving dynamics at a sub-annual scale across the whole Arctic. To address this limitation, we use a fully automated deep learning approach to generate a new calving front dataset for pan-Arctic glaciers at a high temporal resolution, by harmonizing multiple satellite missions that are available from the 1970s onwards, including optical missions such as Landsat, ASTER and Sentinel-2, and various SAR missions such as ERS-1/2, Envisat, RADARSAT-1, TerraSAR-X and Sentinel-1. We first present a new training dataset for the Arctic glaciers. We then present a new deep-learning framework for mapping the pan-Arctic glacier calving fronts. We show the interannual and seasonal variability of glacier termini positions by applying this method at scale and investigate the responses of Arctic glaciers to climate change.

How to cite: Li, T., Heidler, K., Mou, L., Igneczi, A., Zhu, X., and Bamber, J.: Towards pan-Arctic glacier calving front variability with deep learning, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-5656, https://doi.org/10.5194/egusphere-egu23-5656, 2023.

Global mean sea level rise has been accelerating significantly over the past decades, a substantial part of which is attributed to increased surface melting from the Greenland ice sheet (GrIS). Climate models project the GrIS to contribute 9-18 cm to global mean sea level rise by 2100 for the Shared Socioeconomic Pathway SSP5-8.5. The significant uncertainty in this projection prevents accurate mitigation of the effects of sea level rise. The uncertainty stems from a not-comprehensive understanding of the physical processes controlling surface melting. In particular, we lack understanding of ice albedo evolution/variability, a crucial factor in surface melt processes. Ice albedo is a complex and highly variable property of the ice surface that is not well represented in climate model projections, leading to imprecise predictions of sea level rise. The high complexity and number of drivers and feedbacks responsible for ice albedo variability prevent us from building a comprehensive predictive ice albedo model that accurately incorporates all these processes.

From this point of view, we adopt a machine learning-based approach to predict ice albedo variability on the GrIS. We use daily regional climate model output of atmospheric, radiative, and glaciological variables from the Modèle Atmosphérique Régional (MAR) as input data and daily broadband albedo data from the Moderate Resolution Imaging Spectroradiometer (MODIS) as output data. From these data, we construct a Convolutional Long Short-Term Memory (CNN-LSTM) network that models daily ice albedo variability on 6.5 km spatial resolution. A CNN is a neural network that works particularly well for extracting patterns from spatial data. An LSTM is a special kind of recurrent neural network (RNN) that is well-suited for finding patterns and trends in temporal data on a much longer time scale than classic RNNs. Preliminary results show a significant improvement of the correlation between observed and simulated bare ice albedo, with the CNN-LSTM outperforming MAR. Besides the predictive ability of this physically-based machine learning ice albedo model and its suitability for implementation in climate models, it also allows us to gain understanding of what variables drive ice albedo variability on the GrIS, now and in the future.

How to cite: Antwerpen, R., Tedesco, M., Gentine, P., Alexander, P., and van de Berg, W. J.: Predicting Greenland Ice Albedo Using A Physically-Based Convolutional Long Short-Term Memory Network, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-10377, https://doi.org/10.5194/egusphere-egu23-10377, 2023.

While a majority of mass loss from the Greenland Ice Shelf is attributed to glacial terminus retreat via calving, the superimposed force factors of the ice-ocean interface create a challenge for physically modeling terminus change. Here we use time series of environmental and glacial data, input as features into a machine learning regression model, to forecast terminus retreat for marine-terminating glaciers in Greenland. We then identify the critical features that most impact a glacier’s likelihood of retreat using feature importance analysis. We further analyze the heterogeneous outcomes for individual glaciers to classify them by their terminus change profile.  By better understanding the parameters impacting glacial retreat, we inform physical models to reduce uncertainty in mass change projections.

How to cite: Shionalyn, K., Catania, G., Trugman, D., Felikson, D., and Stearns, L.: Predicting Glacier Terminus Retreat Using Machine Learning, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-9122, https://doi.org/10.5194/egusphere-egu23-9122, 2023.

Greenland and other polar regions are highly sensitive to global warming. The impact of climate change on Greenland's glaciers can be seen through an increase in calving events. To better understand this impact, it is important to quantify and document calving activity. However, direct observations are difficult to perform repeatedly over long periods of time due to the hostile climatic conditions, the lack of human witnesses and of the possibility to install in-situ sensors in these remote areas.

With the installation of a regional seismological network in Greenland in the 2000’ ,and the densification of the one in north-eastern Canada, seismic signals caused by large volume calving events, known as glacial earthquakes, were recorded at distances of hundreds of km from the source. These signals have a wide range of frequencies, making it hard to distinguish them from tectonic events, anthropogenic noise, and other natural noise. Using two catalogs of known events - one of 444 glacial earthquakes that occurred between 1993 and 2013, and one for 400 earthquakes that occurred during the same time period selected from USGC - we trained and tested a detection algorithm based on the STA/LTA method to extract event signals from continuous data. We then trained a supervised machine learning algorithm (Random Forest) to automatically classify these signals into two different classes : glacial earthquakes and earthquakes, with a probability of belonging to each class.

With a workflow designed to limit the false alarm rate based on the probability scores of each events, we finally analyzed over 800 days of data from the Greenland regional seismic network and identified almost 1500 new glacial earthquake events using the trained machine learning model. Our detection methods makes it possible to detect four times more ice-quake than the original catalogue.

How to cite: Pirot, E., Hibert, C., and Mangeney, A.: Seismological Monitoring of Calving Events in Greenland: A Machine Learning Approach, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11891, https://doi.org/10.5194/egusphere-egu23-11891, 2023.

Polar ice sheets Greenland and Antarctica are integral parts of the climate system. Understanding their history, dynamics and past accumulation rates determines projections of sea level change. Ice englacial stratigraphy is used to assign ages, taken from ice cores, to radar reflections and subsequently connect these known layers over large areas. One of the main methods to investigate these characteristics is radar reflections. Ground-penetrating radar (GPR) has been used as the primary technique to detect internal ice architecture.

Mapping the internal reflection horizons in order to study and investigate the features, accumulation rates, and ice streams is an important step, which is conventionally done through a semi-automatic process. Such methods are prone to shortcomings in terms of continuity and layer geometry. Moreover, it is highly time-consuming to map an entire profile, the abundance of unmapped radar profiles especially from antarctic ice sheet is an evidence for this. Thus, there is the need for more comprehensive and efficient methods.

The use of machine learning to perform this task automatically will make a significant difference for internal layer detection in terms of efficiency and accuracy. Such machine and deep learning methods would be a suitable fit for radar surveys with different properties, such as center frequencies, making them appropriate for both ice, firn and snow data. In this project, apart from classical computer vision methods and image processing, deep learning methods are used to map the internal reflection horizons (IRH). Convolutional Neural Networks (CNN) are a powerful tool to learn features and track the IRHs continuously.

In this talk, the implemented classical computer vision methods are enumerated, and the machine learning methods that have been used (the specific pre-processing methods unique to this project, labeling method, architecture and hyperparameters) are explained. The results from some more promising architectures such as U-net are presented and compared to the results from image processing methods. The main challenges in this project are lack of complete training data, unknown number of IRHs in a profile, and abundance of features in a single radargram. These challenges are shown and possible solutions are presented.

How to cite: Moqadam, H., Steinhage, D., Eisen, O., and Wilhelm, A.: Tracing Extended Internal Stratigraphy in Ice Sheets using Computer Vision Approaches, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-9532, https://doi.org/10.5194/egusphere-egu23-9532, 2023.

Climate change and the resulting accelerating melt on the Greenland and Antarctic ice sheets are causing dramatic and irreversible changes at a global scale, significantly contributing to sea-level rise. In this scenario, monitoring the evolution of diagenetic snow facies can provide valuable insights to better comprehend climate-related variables and trends. Previous studies of the Greenland ice sheet led to the definition of four main snow facies, depending on the amount of snow melt and on the properties of the snow pack itself: the inner dry snow zone, where melt does not occur; the percolation zone, where a limited amount of melt per year occurs, leading to the generation of larger snow grains and the formation of small ice structures; the wet snow zone, where a substantial part of the snow melt drains off during summer and is characterized by the presence of multiple ice layers; and the outer ablation zone, where the previous year accumulation completely melts during summer, resulting in a surface of bare ice and surface moraine. By exploiting X-band TanDEM-X interferometric synthetic aperture radar (InSAR) acquisitions, previous works explored the idea of classifying different snow facies of the Greenland ice sheet utilizing an unsupervised machine learning clustering approach. The analysis was performed using data acquired in winter 2010/2011 only, under the assumption of stable climatic conditions and similar acquisition geometries. In this paper, we further investigate the evolution of the snow facies of Greenland throughout the last decade of TanDEM-X observations, proposing unsupervised machine learning strategies for snow facies characterization by using InSAR features such as backscatter, volume decorrelation, the incidence angle and height of ambiguity. We use TanDEM-X data acquired during the winter of 2010/2011, 2015/2016, 2016/2017, 2020/2021, and 2021/2022, where full or partial coverage of the Greenland ice sheet is available. The challenges and caveats of such approaches for different image acquisition geometries will be presented. Finally, the potential of TanDEM-X for investigating large-scale interannual changes in the dry snow zone over Greenland will be investigated as well.  

How to cite: Becker Campos, A., Rizzoli, P., Gonzalez, C., Bueso-Bello, J.-L., and Braun, M.: The Evolution of the Snow Facies on the Greenland Ice Sheet Observed by the Last Decade of TanDEM-X Interferometric SAR Data, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-2096, https://doi.org/10.5194/egusphere-egu23-2096, 2023.

In order to improve the forecasting of hydraulic production, EDF uses optical satellite images to evaluate the snow cover [1]. These images are acquired daily by the MODIS instrument of the Terra satellite and provide a snow product through the Normalized Difference Snow Index (NDSI). However, part of the information on the snow cover is lost due to clouds. To complete those gaps, radar satellite images can be interesting because it does not depend on weather conditions.

Dry snow and wet snow have different SAR signature. Wet snow can be detected since its backscatter decreases [2]. Dry snow detection is more challenging. It may be performed with a polarimetric approach [3], and the snow depth (SD) can be estimated using optical images as auxiliary inputs [4]. In this work, wet snow was detected and SD was estimated over the Guil basin in the Alps (420 km²) for the years 2018 and 2019 on three relative orbits of Sentinel-1: the D66 (descending, 87 images), A88 (ascending, 119 images), and D139 (descending, 90 images). The results show an accumulation of snow in autumn on the SD and a peak of snowmelt in spring on the detection of wet snow.

Then we propose to detect the snow from SAR images using a convolutional neural network trained with optical images from MODIS as labels. For the dataset, a smaller area is chosen around Abriès (of approximately 59km²) and we select 36 images for each of the three orbits to study the winter 2018-2019. A binary semantic segmentation is computed from two SAR inputs: R_wet from [2], and R_dry a polarimetric ratio inspired from [3]. The trained model, called SESAR U-net, gives a snow detection with an overall accuracy of 80% for our test set. This low accuracy result can be explained by the fact that MODIS images have a resolution 25 to 100 times coarser than the SAR images, which hinder both the training and the evaluation of the model. Further works will consider the uncertainty of the MODIS label in the loss computation to improve the training.

[1] M. Le Lay et al., “Use of snow data in a hydrological distributed model: different approaches for improving model realism,” in EGU General Assembly Conference Abstracts, EGU General Assembly Conference Abstracts, p. 14545, Apr. 2018.
[2] T. Nagler et al., “Advancements for snowmelt monitoring by means of sentinel-1 sar,” Remote Sensing, vol. 8, 04 2016.
[3] A. Reppucci et al., “Estimation of snow-pack characteristics by means of polarimetric SAR data,” in Remote Sensing for Agriculture, Ecosystems, and Hydrology XIV (C. M. U. Neale and A. Maltese, eds.), vol. 8531, p. 85310Z, International Society for Optics and Photonics, SPIE, 2012.
[4] H. Lievens et al., “Snow depth variability in the Northern Hemisphere mountains observed from space,” Nature Communications, vol. 10, p. 4629, Dec. 2019.

How to cite: Montginoux, M., Weissgerber, F., Monteil, C., and Girard, A.: Snow cover estimation by deep-learning segmentic segmentation of radar images based on optical image references, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-7240, https://doi.org/10.5194/egusphere-egu23-7240, 2023.

Recent work has demonstrated how Gaussian Process Regression (GPR) can be used to interpolate Pan-Arctic radar freeboard of sea ice as measured by satellites. Sea ice freeboard is crucial to measuring sea ice thickness, and thus sea ice volume, which can play an important role in climate models. Similarly sea surface heights from altimetry are essential to determine the geostrophic currents from space. Using GPR can be computationally burdensome for modest dataset sizes and prohibitive for large datasets. To avoid having to deal with a large dataset the raw satellite observations were binned (averaged) onto a regularly spaced grid. We look at how these calculations can be reduced in terms of run time by utilising a Graphical Processing Unit (GPU), a dedicated GPR python package and by making practical adjustments to the methodology. We find by adopting these changes the overall run time of a single day’s interpolation can be greatly reduced by a factor of over 60, making it practical to run such calculations on an environment with a GPU. We then extend the method to use raw satellite observation data (no binning), which greatly increases the number of training points, requiring the use a sparse method for GPR. We conclude with recommendations for further work on this subject as it has the potential for widespread use in remote sensing applications.

How to cite: MacEachern, R., Tsamados, M., Gregory, W., Lawrence, I., and Takao, S.: Fast interpolation of satellite altimetry data with probabilistic machine learning and GPU, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-17323, https://doi.org/10.5194/egusphere-egu23-17323, 2023.

Atmospheric errors in interferometric synthetic aperture radar (InSAR)-derived estimates of surface deformation often obscure real signals, especially in mountainous terrain. By taking advantage of the differing spatial characteristics of periglacial landforms and atmospheric noise, we trained a deep convolutional neural network (CNN) to remove atmospheric noise from individual interferograms. Unlike existing corrections, which rely on coarse climate reanalysis or radiometer data, this computer vision correction is applied at the spatial and temporal resolution of the interferogram. We processed Sentinel 1 interferograms of the Colorado Rocky Mountains using the Alaska Satellite Facility's Hybrid Pluggable Processing Pipeline (ASF HyP3) and used the Miami INsar Time-series software in PYthon (MintPy) package to generate low-noise line-of-sight (LOS) velocity maps containing primarily rock glacier and hillslope motion. These maps were combined with noisy short temporal-baseline interferograms to contrive a training dataset. Model performance was assessed using the structural similarity index measure (SSIM) and compared to that of other widely used corrections. We find that our CNN significantly outperforms standard corrections and that previously hidden intraseasonal kinematic behavior is revealed in Colorado rock glaciers. We suggest that insights from external validation against GNSS data and sensitivity analysis could be used to further improve model performance and assess model scalability and transferability. 

How to cite: Brencher, G., Henderson, S., and Shean, D.: Removing Atmospheric Noise from Interferograms in Mountainous Regions with a Deep Convolutional Neural Network, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-10088, https://doi.org/10.5194/egusphere-egu23-10088, 2023.

Advanced environmental technologies have made available an increasing amount of data from remote sensing satellites, and more sophisticated ground data. Their assimilation into dynamic models is progressively becoming the most frequent, and conceivably the most successful, solution to estimate snow water resources. Models reliability is therefore bounded to data quality, which is often low in mountain, high-elevation, and unattended settings. To add new value to snow-depth sensor measurements, we developed a machine-learning algorithm to automatize the QA/QC procedure of near-surface snow depth observations collected through ground stations data. Starting from a consolidated manual classification, based on the expert knowledge of hydrologists in Valle D'Aosta, a Random Forest classifier was developed to discriminate snow cover from grass or bare ground data and detect random errors (e.g., spikes). The model was trained and tested on Valle d’Aosta data and then validated on 3 years of data from 30 stations on the Italian territory. The F1 score was used as scoring metric, being it most suited to describe the performances of a model in case of a multiclass imbalanced classification problem. The model proved to be robust and reliable in the classification of snow cover and grass/bare ground discrimination (F1 values above 90%), yet less reliable in random error detection, mostly due to the dataset imbalance. No clear correlation with single year meteorology was found in the training domain, and the promising results from the generalization to a larger domain corroborates the model robustness and reliability.This machine learning application of data quality assessment provides more reliable snow ground data, enhancing the quality of snow models.

How to cite: Blandini, G., Avanzi, F., Gabellani, S., Ponziani, D., Stevenin, H., Ratto, S., and Ferraris, L.: A Random Forest approach to quality-chacking automatic snow-depth sensor measurements, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-180, https://doi.org/10.5194/egusphere-egu23-180, 2023.