The monitoring of ice floe is essential for mapping marine ecosystems, ensuring safe ship navigation, and ice hazard forecasting. Satellite imagery, such as Synthetic Aperture Radar (SAR), is a prime candidate for capturing information related to ice floes, due to the ability to discern sea ice conditions in this imagery in cloudy or poor lighting conditions. This SAR imagery can then be passed along to image processing algorithms to extract quantities of interest such as floe size distribution (FSD). Whilst considerable research has used fully supervised machine learning models in this domain, such models require an abundant amount of annotated data for training. The time-consuming, subjective, and costly process of annotating limits the amount of available data that can be used during training and, thus, reduces the performance of the trained model. To alleviate this problem, we turn towards the burgeoning field of generative modeling to create synthetic labeled data.
An important class of generative models, known as diffusion models, has been shown to be particularly efficient. Over the years, a rich plethora of techniques and architectures have been developed to enable these diffusion models to provide realistic samples from an approximate distribution of the training data. Moreover, such models can also be conditioned by additional information, such as texts or images, offering an interesting degree of flexibility to explore and enhance the sampling process. More pertinently, diffusion models have been employed to generate synthetic images of semi-natural areas captured by drones, as well as satellite imagery of rural and urban scenes. However, to date, their application to SAR imagery of the cryosphere remains unexplored.
In this work, we describe a process whereby we use a diffusion model, namely a Denoising Diffusion Probabilistic Model, to model the joint distribution over the space of SAR images and their corresponding labels. In addition to standard error metrics, we use FSD to demonstrate that the synthetic SAR data is consistent with the real data. Furthermore, we show that using a dataset composed of both the real data and the synthetic data results in better performance for segmentation modeling. Additional experiments are performed to show performance as a function of the amount of real and synthetic data.
How to cite: Bunker, J., Rogers, M. S. J., van Zeeland, L., Wilkinson, J., and Girolami, M.: Ice Floe Data Augmentation Using Diffusion Models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9315, https://doi.org/10.5194/egusphere-egu25-9315, 2025.
Sea ice is critical to map for safe and efficient maritime navigation, to mitigate ship trapping and capsizing. Sea ice is also vital to monitor to assess the state of the changing climate and a critical component in climate and weather models, reflecting sunlight towards space and acting as an insulating material between the ocean surface and the atmosphere.
Professional sea ice analysts at national ice services map sea ice based on Synthetic Aperture Radar (SAR) images acquired by satellites, such as the Sentinel-1 (S1) satellite constellation. The ice analysts manually interpret the SAR images using their in-depth knowledge and experience to create sea ice charts with information on the sea ice conditions.
A challenge for the S1 5.4 GHz SAR measurements is that the radar wave does not penetrate deep into the sea ice and is scattered/reflected by the surface. Therefore, the SAR images provide information primarily about the sea ice surface, useful for identifying and classify sea ice conditions. The charts describe, among others, the sea ice’s stage of development - the type of sea ice - an indicator of its thickness. The manual charting process apply sea ice classes, defined by the International Ice Charting Working Group (IICWG) on behalf of the World Meteorological Organization (WMO). Considerable uncertainties are associated with the ice classes that can vary from, e.g. 30-200cm or 70-120cm in thickness. Deep-learning models that produce stage-of-development information from S1 radar images exist but has the same inherent limitations of the sea ice charts in the model outputs.
Current state-of-the-art sea ice thickness retrieval methods relies on altimeter satellites, such as the CryoSat-2 (CS2) satellite. The distance between the ocean and the sea ice is measured, known as the sea ice freeboard. For a Ku-band radar altimeter like CS2, it is assumed that the radar response penetrates the snow and returns from the sea ice surface. As the true penetration is unknown, and the radar wave propagation is delayed when the signal passes through snow, the measured quantity is known as the radar freeboard.
The sea ice thickness can be estimated with an accuracy of 20-40% using the radar freeboard by calculating the sea ice's buoyancy based on snow and ice density estimates, and auxiliary snow depth information. However, CS2 only measures 1600m across the orbit and can thus only monitor sea ice thickness in the Arctic monthly - insufficient for many applications, such as maritime navigation, and leaves data record gaps. S1 SAR on the other hand, cover 400km in Extra Wide mode across the orbit with repeating coverage every week.
Here, we present our preliminary results of circumventing the limitations of CS2 and S1 by training supervised deep-learning convolutional neural network (CNN) models to recognise sea ice textures in S1 SAR images and assign sea ice radar freeboard estimates acquired by CS2. This approach transfers information acquired by CS2 to S1, which we call Cryo2S1. A Cryo2S1 dataset is curated, containing several thousand collocated S1 SAR images and along-track CS2 measurements during 2020-2021.
How to cite: Stokholm, A., Landy, J. C., Saldo, R., Wulf, T., Korosov, A., and Hvidegaard, S. M.: Cryo2S1: Mapping sea ice radar freeboard in Sentinel-1 SAR imagery from CryoSat-2 using deep learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14145, https://doi.org/10.5194/egusphere-egu25-14145, 2025.
The response of ice sheets and shelves to climate change profoundly influences global human activities, ecosystems, and sea-level rise. As such, ice sheets are a vital component of the Earth system, making them a cornerstone for developing a future Digital Twin Earth. Here, we present the initial steps toward an Earth Observation (EO)-driven Digital Twin Component (DTC) for Ice Sheets, marking an effort to understand and predict the behaviour of the Greenland Ice Sheet and Antarctic ice shelves under user-defined “what-if” scenarios.
To meet the diverse needs of stakeholders, DTC Ice Sheets will adopt a modular design comprising 10 Artificial Intelligence/Machine Learning (AI/ML) and Data Science modules. All targeted four initial use cases that will drive the development of DTC Ice sheets. These initial use cases are: (1) Greenland Hydropower Potential: By modelling and monitoring ice sheet hydrology and meltwater runoff, the DTC ice sheets will evaluate Greenland’s renewable energy opportunities and provide actionable insights for sustainable hydropower development. (2) EU Sea Level Response Fingerprint: The DTC Ice Sheets will deliver region-specific insights into how ice sheet mass loss will contribute to global sea level rise, focusing on the implications for coastal infrastructure across Europe. (3) State and Fate of Antarctic Ice Shelves: Through detailed stability analysis, the DTC Ice Sheets will investigate the vulnerability of Antarctic ice shelves to climatic and oceanic changes, shedding light on their role in regulating ice sheet mass loss and global sea level. (4) Enhanced Surface Climate: Leveraging EO data and climatology, the DTC Ice Sheets will improve understanding of surface climate interactions, advancing predictions of feedback loops between ice sheets, the atmosphere, and the ocean.
The DTC Ice sheet implementation on the DestinE Core Service Platform (DESP) will consist of interconnected modules to serve the use cases. Still, it will also, when fully implemented, provide a holistic view of an ice sheet digital twin. Hence, DTC Ice Sheets aims to provide high-resolution insights into ice sheets' past, present, and future states, align with stakeholders, and foster interdisciplinary collaboration by interfacing with other thematic Digital Twin Earth systems, such as ocean and coastal processes. The DTC ice sheets will empower stakeholders to explore What-if scenarios to address climate change's impacts and feedback mechanisms. All are found in current state-of-the-art EO data of ice sheets.
How to cite: Simonsen, S., Karlsson, N., and DTC Team, T.: A first view of the EO-driven digital twin for ice sheets, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-8017, https://doi.org/10.5194/egusphere-egu25-8017, 2025.
The surface mass balance (SMB) is projected to become the main driver of mass changes for the Greenland Ice Sheet (GrIS) by the end of this century. Therefore, it is crucial to have realistic projections of the SMB for future estimates of mass loss and sea-level rise.
To date, estimates of the surface mass balance are most often provided by either (i) stand-alone parameterization schemes, such as positive degree days (PDD) or energy balance approaches, (ii) direct outputs from Earth system models (ESMs), or (iii) regional climate models (RCMs) forced by boundary conditions from ESMs. Each of these approaches has its disadvantages. Stand-alone parameterization schemes are often overly simplified and unable to capture smaller-scale processes at the surface. ESMs often provide forcing fields that are too coarse compared to the resolution required for ice sheets. Meanwhile, regional climate models are expensive to run and computationally slow.
In this study, we address these issues by employing a generative model-based approach to realistically downscale the SMB directly from ESM fields to a 5 km resolution. We train a diffusion-based model on historical and future SMB fields from the regional climate model MAR. This allows us to generate high-resolution SMB fields in a fraction of the time required by a regional climate model. We condition our diffusion model on an initial estimate of the SMB derived from ESMs. Specifically, we add noise to the initial ESM estimate and subsequently de-noise the SMB field at different noise levels. By selecting the noise level during inference, we can effectively choose the spatial scale at which ESM features should be preserved.
Our approach enables fast, simple, and probabilistic downscaling of the SMB and potentially other climate fields.
How to cite: Bochow, N., Hess, P., and Robinson, A.: Generative Model-Based Downscaling of the Surface Mass Balance of the Greenland Ice Sheet, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11686, https://doi.org/10.5194/egusphere-egu25-11686, 2025.
Rock glaciers are tongue-shaped complex landforms that indicate current or past permafrost conditions. They are commonly found in high-latitude and/or high-elevation environments and consist of poorly sorted angular debris and ice-rich sediments formed by gravity-driven creep. In the Austrian Alps, it is estimated that over 5700 rock glaciers exist (Kellerer-Pirklbauer et al., 2022). Knowing the location, extent and characteristics of rock glaciers is important for several reasons. These include estimating their hydrological importance as a water resource (e.g., for alpine huts) and assessing the geohazard potential because of the destabilisation of rock glaciers due to climate change. Unlike other cryosphere features, such as snow and glaciers, rock glaciers are spectrally inseparable from the surrounding terrain. This makes them difficult to automatically detect and delineate from Earth observation (EO) data. As a result, rock glaciers are usually mapped by labour-intensive, subjective manual interpretation of EO data. This often leads to inhomogeneous, incomplete, and inconsistent mapping. Therefore, there is a need for automated and efficient methods to map rock glaciers. This can be achieved by using globally applicable satellite data sets such as Sentinel-2.
Modern machine learning methods, such as deep learning (DL), provide new opportunities to automate mapping tasks and address the challenges of detecting rock glaciers from EO data. However, research on DL-based rock glacier mapping remains limited, and there is no consensus on the best-suited parameters for this application. In addition, features with surface textures similar to rock glaciers, such as landslides, avalanche deposits, or fluvial deposits, may be misclassified by DL models. Hence, a thorough investigation of the DL model architectures and input data types is necessary to determine the most effective approach for mapping rock glaciers. In the project “ROGER - EO-based rock glacier mapping and characterisation”, we test different DL models (e.g. Unet, DeepLABV3) with different settings (backbones, input layers (including optical imagery and DEM-derived information)) to identify the most suitable model for rock glaciers delineation in Austria. We evaluate the performance, robustness, and reliability of the different DL models for automated EO-based mapping of rock glaciers in different study areas in Austria, and quantify the accuracy of the results in comparison with reference data.
Through our study, we aim to make a substantial contribution to cryospheric research by evaluating methods for the automated identification of rock glaciers, thereby enhancing our understanding of the potential of DL to efficiently map complex natural phenomena using EO data. The results will also contribute to increase the trustworthiness of DL methods, which is critical for various applications and particularly in communicating and explaining results to stakeholders and decision makers.
Kellerer-Pirklbauer, A., Lieb, G.K., Kaufmann, V. (2022). Rock Glaciers in the Austrian Alps: A General Overview with a Special Focus on Dösen Rock Glacier, Hohe Tauern Range. In: Embleton-Hamann, C. (eds) Landscapes and Landforms of Austria. World Geomorphological Landscapes. Springer, Cham.
How to cite: Streifeneder, V., Robson, B. A., Hölbling, D., Nafieva, E., Dabiri, Z., Hauglin, E., and Abad, L.: Deep learning-based rock glacier mapping using Earth observation data, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4229, https://doi.org/10.5194/egusphere-egu25-4229, 2025.
Glacier retreat poses significant environmental and societal challenges. Understanding the local impacts of climate drivers on glacier evolution is essential, with glacier mass balance being a central concept. This study uses the Mass Balance Machine (MBM; Sjursen et al., 2025), an open-source, data-driven model based on eXtreme Gradient Boosting (XGBoost) that reconstructs glacier mass balance with high spatiotemporal resolution at regional scales. Trained on point mass balance data from multiple glaciers, MBM captures both intra- and inter-glacier variability, enabling the identification of transferable patterns and applications to glaciers without direct observations. Here, we applied MBM to reconstruct the mass balance of Swiss glaciers. The model was trained using a comprehensive dataset of approximately 34,000 winter and annual point mass balance measurements from 35 Swiss glaciers in diverse climate settings from 1951 to 2023. Using MBM, we generated high spatial resolution reconstructions of seasonal and annual mass balance for these 35 glaciers. When validated on independent unseen glaciers, MBM demonstrated robust performance across spatial scales (point to glacier-wide) and temporal scales (monthly to annual). This study underscores how MBM can be effectively used in Switzerland to generalize across diverse glaciers and climatic conditions, highlighting the model's versatility and broad applicability.
How to cite: van der Meer, M., Zekollari, H., Sjursen, K. H., Huss, M., Bolibar, J., and Farinotti, D.: High-resolution mass balance reconstructions for Swiss glaciers using machine learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16155, https://doi.org/10.5194/egusphere-egu25-16155, 2025.
Snowpacks are important elements of the Earth’s cryosphere and are composed of layers with unique physical properties. Snow stratigraphy, the study of distinct snow layers and their properties, provides essential data for climate modeling, water resource management, and avalanche prediction. However, existing methods for characterizing snowpacks with near-infrared (NIR) photography are based on manually segmenting layers from images, which is a laborious and time-consuming task. In this work, we develop an approach to automate snowpack layer segmentation based on fine-tuning Segment Anything (SAM), a state-of-the-art deep learning segmentation model. We use a small set of expert-labeled NIR snowpack images and explore different task representations. We approach the problem through the lens of 1) edge detection, which focuses on detecting snowpack layer boundaries and 2) region detection, which focuses on predicting the area occupied by the layers. Our results indicate that deep learning segmentation is promising for automating the segmentation of snowpacks. This ultimately leads to facilitating snow stratigraphy analysis to improve applications such as avalanche forecasting and snowpack modeling.
How to cite: Reil, M., Mastikhina, O., Marks, J., Navarro, K. F., Davari, M. R., Mewes, L., Kaltenborn, J., and Rolnick, D.: Automated Analysis of Snowpack Stratigraphy NIR Images Using Deep Learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14736, https://doi.org/10.5194/egusphere-egu25-14736, 2025.
Posters on site: Thu, 1 May, 14:00–15:45 | Hall X4
With the ongoing effects of global warming, supraglacial meltwater in polar regions plays a critical role in ice sheet dynamics, influencing global sea levels. In Antarctica, the accumulation of meltwater on ice surfaces not only reduces albedo—accelerating melting in a self-reinforcing cycle—but also drives processes such as meltwater injection and basal lubrication, with possible destabilizing effects for ice sheets. Monitoring the seasonal evolution and dynamics of supraglacial lakes is essential for understanding these processes, yet the vast and remote nature of the Antarctic ice sheet presents significant challenges. Spaceborne remote sensing offers the best solution, providing continuous, large-scale, and long-term observations. However, extracting reliable information from optical and synthetic aperture radar (SAR) data remains complex due to limitations in spatial transferability, cloud cover, polar night, and the spectral similarities of frozen lakes with surrounding ice. The Sentinel mission bridges these gaps, enabling the combination of optical and SAR data to achieve the best possible accuracy for mapping and monitoring supraglacial lakes.
This study evaluates whether a deep learning-based mapping approach outperforms a pixel-based Random Forest (RF) classification algorithm for supraglacial lake (SGL) detection in Antarctica. As a benchmark, we utilized an RF model trained on 14 regions and 24 input channels, including Sentinel-2 spectral bands, spectral indices, and topographic variables. To work toward a circum-Antarctic, operational SGL mapping product, we reduced the input channels by selecting the four most important features identified by the RF approach and trained a convolutional neural network (CNN) on partially labeled data from 16 Sentinel-2 scenes, including more images with cloud cover. Both models were validated using the same 16 test areas across eight Antarctic ice shelves.
The RF approach achieved a producer’s accuracy, user’s accuracy, and F1 score of 0.750, 0.945, and 0.837, respectively, whereas the CNN-based workflow achieved scores of 0.915, 0.912, and 0.913, respectively. In scene-specific comparisons, the CNN outperformed the RF approach in 13 of the 16 validation scenes. Key advantages of the CNN approach include its ability to detect lakes under thin clouds and over floating ice, resulting in less fragmented lake area estimates and requiring fewer input features. However, challenges persist in transition zones between lakes and slush, where spectral details outweigh the benefits of shape-based detection.
How to cite: Baumhoer, C. A., Koehler, J., and Dietz, A.: From Decision Trees to Deep Learning: Enhanced Supraglacial Lake Detection in Antarctica, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4329, https://doi.org/10.5194/egusphere-egu25-4329, 2025.
Monitoring of glaciers is crucial as they are an important source of freshwater, an indicator of global warming, and a contributor to sea level rise. Accurate delineation of glaciers plays a crucial role in glacier monitoring and remote sensing is the most appropriate tool to map glaciers.
Existing glacier inventories have shortcomings such as unavailability in recent years and data quality. Traditional glacier mapping methods using remote sensing often rely on spectral band ratio techniques or manual digitizing. However, glacier boundaries achieved from manual digitizing are highly affected by human errors. Moreover, in the band ratio technique challenges arise in mapping debris-covered glaciers as traditional optical methods fail to distinguish debris-covered ice from surrounding rock due to their spectral similarities. Therefore, automatic mapping of glaciers is still challenging.
Advanced deep learning methods have demonstrated significant advancements in automatic glacier mapping. However, the potential of state-of-the-art deep learning methods in glacier mapping has not yet been fully explored. When it comes to deep learning, one of the challenges is the amount of training data. With the low amount of training data, the results won't be of the desired accuracy. However, it is still possible to obtain good results using a lower amount of training data and the transfer learning technique.
This study focuses on glacier mapping in Poiqu Basin (Central Himalaya), using U-Net and transfer learning. To this purpose, Sentinel-2 images and the Shuttle Radar Topography Mission (SRTM) Digital Elevation Model are deployed.
The results indicate that transfer learning leads to considerably better results than training the deep learning network from scratch. Moreover, trying different backbones does not considerably affect the results. This study highlights the efficiency of the transfer learning technique, emphasizing its potential and effectiveness in regions with limited training data.
How to cite: Barzegar, F. and Bolch, T.: Mapping of Glaciers in the Poiqu Basin (Central Himalaya) Using U-Net and Transfer Learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-5204, https://doi.org/10.5194/egusphere-egu25-5204, 2025.
The floating ice shelves surrounding Antarctica play a crucial role in regulating ice sheet mass loss by providing mechanical buttressing, which regulates ice discharge into the ocean. The recent collapses of Larsen-A and B along with the rapid retreat of Thwaites were followed by accelerated ice discharge, underlining the importance of understanding the full dynamic of this process. Enhanced damage through fracturing has recently been shown to play a critical role in ice shelf weakening, reducing its buttressing capacity and potentially accelerating their collapse. Despite its significance, the processes governing ice shelf damage remain poorly understood. Damage manifests itself as large crevasses, rifts, and shearing regions clearly visible on satellite imagery. Historically, the mapping of fractures has been challenging due to the labor-intensive nature of manual delineation. Rapid advancements in machine learning, however, have revolutionized damage mapping, enabling the automatic detection of damage features. Although SAR backscatter imagery from ESA's Sentinel-1 has been the primary source of data in recent studies, it suffers from limited temporal coverage (2013-present), which does not capture the entire damage dynamic of ice shelves that destabilized in the early 2000s. Other available products, also exhibited significant discrepancies with modeled changes in ice viscosity, suggesting that critical features of ice damage are not fully captured (e.g. basal fracturing). To address these gaps, this study presents a novel methodology leveraging multisensor optical imagery and supervised/semi-supervised machine learning algorithms to identify damage features. A U-Net algorithm was trained on manually annotated images from 10 acquisitions from the USGS/NASA's Landsat satellite, across diverse Antarctic ice shelves. These annotations represented various types of damage to ensure broad applicability. The model was then refined using a human-in-the-loop approach with additional Landsat and Sentinel imagery datasets, enhancing prediction accuracy. We demonstrate the capability of our model to map comprehensively the evolution of damage in the Amundsen Sea Embayment, one of Antarctica's most vulnerable regions, from the 1990s to the present. The results are compared with existing damage products derived from machine learning and radon transform methods using Sentinel-1 SAR images, on the period 2013 to present. We map the dynamic evolution of surface and basal fractures, along with their morphological characteristics such as maximum length and area, and compare this evolution with dynamical changes over the same time period. We complement our analysis by comparing our result to damage modeling using an ice flow model on the Pine Island ice shelf. We use the Shallow Shelf Approximation within the Elmer/ice model to invert for damage and ice viscosity evolution since 1992, by assimilating a long record of satellite-derived surface flow velocity and thickness. We finally analyze the spatial correlations between modeled and observed damaged and draw conclusions on the features of importance regarding ice sheet stability through time. We demonstrate the potential of multisensor optical imagery, which offers broader temporal coverage dating back to the 1970s, to address critical gaps in understanding ice shelf damage and its evolution.
How to cite: Shahateet, K., Millan, R., Bacchin, L., Mosbeux, C., and Surawy-Stepney, T.: Mapping and analyzing ice shelf damage using multisensor imagery and machine learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-8433, https://doi.org/10.5194/egusphere-egu25-8433, 2025.
Arctic sea ice, the vast body of frozen water near the North Pole, has been in steady decline since satellite observations began. While state-of-the-art models attempt to project future scenarios, they often show significant discrepancies, even though the sea ice system is generally considered to decline linearly with rising temperatures. Machine learning models, although they may lack the ability to fully explain the underlying physical processes, offer a complementary approach. By training these models on existing data, we can generate plausible future predictions that are less influenced by the biases inherent in traditional modeling methods. In this study, we evaluate several machine learning architectures to identify the most effective ones. Using the best-performing model, we explore the stability and potential hysteresis behaviors of the Arctic sea ice system.
How to cite: Cotronei, A., Gallicchio, C., and Graversen, R.: Machine learning for prediction of sea ice stability, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-8973, https://doi.org/10.5194/egusphere-egu25-8973, 2025.
In a rapidly changing permafrost environment driven by climate change and anthropogenic disturbances, tracking geomorphological dynamics is a crucial task, not only to provide hazard monitoring, but also to evaluate climatological feedback processes. Yet, the impact on rapid permafrost disturbances on the Earth system is still uncertain, making the availability of reliable, long term data a very important building block to understand the interconnections and feedbacks between several environmental subsystems.
Specifically, Retrogressive Thaw Slumps (RTS) are a major mass-wasting phenomenon and a rapid disturbance in ground-ice rich permafrost landscapes. They can mobilize large quantities of formerly frozen ground and consequently sediment, carbon, and nutrients. Once initiated they can grow and develop broader erosion disturbances. Over years and decades they can undergo polycyclic behaviour of initialization, growth, stabilization, and re-activation. The spatial distribution and temporal dynamics of RTS are generally poorly quantified so far on a pan-arctic scale, except for some regions covered by more intensive research.
Multiple methods and data are used to map permafrost disturbances like RTS, including in-situ mapping. However, due to the remoteness and reduced accessibility, earth observation data is the primary source of RTS inventories. While RTS mapping is also done manually utilizing expert knowledge from high-resolution remote sensing imagery, machine learning techniques are increasingly used to segment permafrost features from satellite images. However, due to the requirement to process large amounts of data and also the reduced availability of suitable image data, especially in the high-latitudes, these datasets still often lack the temporal and spatial coverage to derive insights related to the recent global environmental changes. Current advancements in artificial intelligence based inference methods make feature segmentation now much more feasible and efficient, so activities for mapping RTS based on high resolution PlanetScope images and deep-learning methods, such as the DARTS dataset, already cover large RTS affected regions. Nevertheless, a full pan-arctic coverage over multiple time-steps is still lacking, thus far.
To expand the existing body of RTS inventories, we use a convolutional neural network to detect these permafrost features from Sentinel 2 imagery to create a multi-year dataset of detected thaw slumps in the circumpolar arctic. The comparison with existing manually labelled and automatically derived high resolution thaw slump inventories provides a quantifiable verification to estimate uncertainties. This is crucial for evaluating Sentinel-2 as a high resolution dataset with favourable properties in terms of data availability and processing requirements compared to commercial and access restricted VHR imagery. Our work can underpin downstream tasks to extend RTS classification, understanding trigger mechanisms and improve vulnerability mapping. Also, time series of RTS disturbance data may be used for the temporal and spatial correlation with climate reanalysis and atmospheric datasets for large scale climate change impact modelling and feedback evaluation over the permafrost domain. Additionally, the open architecture of the processing pipeline can be used to implement near real-time monitoring services based on the Sentinel-2 data release stream for public access. We present ongoing work on the RTS segmentation dataset and current key downstream results.
How to cite: Küpper, J., Hölzer, T., Nicholson, T., Marini, L., von Chamier, L., Hänzelmann, S., Nitze, I., Liljedahl, A., and Grosse, G.: Creating a Pan-Arctic Retrogressive Thaw Slump Dataset with Harmonized Sentinel-2 Data and Deep Learning Methods, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10699, https://doi.org/10.5194/egusphere-egu25-10699, 2025.
Glacier retreat is a key indicator of climate change and requires regular updates of the glacier area. The most recent inventory for the European Alps, released in 2020, showed that glaciers retreated approximately 1.3% per year from 2003 to 2015. This ongoing retreat underscores the urgent need for accurate and efficient monitoring techniques.
Recent advancements in Deep Learning have led to significant progress in the development of fully automated glacier mapping techniques. In this work, we use DL4GAM, a multi-modal Deep Learning-based framework for Glacier Area Monitoring, to assess the change in glacier area in the European Alps over 2015-2023. The main data modality used for training is based on Sentinel-2 imagery, combined with additional features derived from a Digital Elevation Model, along with a surface elevation change map, which is particularly useful for debris-covered glaciers. The framework provides an area (change) estimate independently for each glacier, with uncertainties quantified using an ensemble of models. Region-wide, we estimate a retreat of -1.90 ± 0.71%, which is greater than the rate observed during the previous decade. Our estimates also present a significant inter-glacier variability which we analyze with respect to various topographical parameters such as slope, aspect, or elevation.
Several challenges persist, including model limitations, data availability issues, and the impact of debris, cloud cover, and seasonal snow. We discuss these challenges, the design choices made to address them, and the remaining open issues.
How to cite: Diaconu, C.-A., Bamber, J. L., and Zekollari, H.: Glacier Area Change Assessment over 2015-2023 in the European Alps with Deep Learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15293, https://doi.org/10.5194/egusphere-egu25-15293, 2025.
In its special report on Ocean and Cryosphere in a Changing Climate (SROCC) from 2019, the Intergovernmental Panel on Climate Change (IPCC) highlights clear knowledge gaps concerning the extent and ice content of permafrost in mountain regions. We present results from a study on the distribution of mountain permafrost that includes an improved understanding of its characteristics and an estimation of the sub-surface ice reserves in mountainous regions under climate scenarios. We explore the feature space of mountain permafrost using a range of statistical and machine learning techniques in an uncertainty-aware setting. This space consists of topographic and climatic features such as topographic masks, elevation models, potential incoming solar radiation, seasonal ground temperatures and snow accumulation. We combine these features with existing inventories of rock glaciers, as these are good visible indicators of mountain permafrost, and in addition typically ice-rich. Based on such datasets we create a data-driven model to predict the probability of potential rock glaciers occurrence in order to obtain a first estimate of ground ice content. In addition, output from a numerical permafrost model, the CryoGrid community model, provides synthetic observations. We further investigate the vulnerability of these potentially ground-ice rich areas under climate change by including forcing data from climate models based on various RCP scenarios.
How to cite: Røste, J. and Kääb, A.: Towards quantifying ice contents in mountain permafrost environments, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11399, https://doi.org/10.5194/egusphere-egu25-11399, 2025.
The rapid decline of Arctic sea ice driven by climate change poses significant challenges and opportunities for global shipping, ecosystems, and coastal communities. Understanding and mapping sea ice variability is crucial for assessing its implications on navigability and ensuring maritime safety in this dynamic region. One of the most significant challenges in applying machine learning (ML) to cryospheric sciences is the reliance on large quantities of human-labelled data, which is both costly and time-intensive to produce, particularly in remote and harsh environments like the Arctic. This contribution addresses this challenge by leveraging self-supervised learning (SSL) techniques and Convolutional Neural Network (CNN) to reduce the dependency on labelled data while maintaining high levels of model performance. We used the well-known UNet model, a CNN designed for pixel-wise segmentation tasks, and integrate BYOL (Bootstrap Your Own Latent), an SSL technique that leverages unlabelled data to learn features without requiring explicit labels. BYOL trains the model to match representations of the same image under different transformations, allowing it to learn useful features from unlabelled data without needing explicit labels.
We apply these models to Sentinel-1 SAR imagery in the Canadian Arctic Archipelago, a region of critical importance due to its role in global shipping routes, where sea ice variability directly impacts navigability and maritime safety.
We created binary ice and open water labels to serve as a benchmark for evaluating model performance. Early preliminary results suggest that using BYOL reduces the labelling requirement by approximately 50% compared to models trained without self-supervised pretraining. By pretraining the UNet model on unlabelled Sentinel-1 SAR imagery and fine-tuning it for sea ice segmentation, this approach demonstrates how leveraging unlabelled data can significantly minimise the need for human annotation while maintaining robust segmentation accuracy. These methods optimise the use of limited labelled datasets, enabling efficient and scalable models that potentially generalise to sea ice segmentation tasks where high-quality labels are often scarce or imprecise. These techniques enhance the adaptability of ML models, allowing them to be applied to new datasets and tasks with minimal retraining, further reducing the computational and data requirements. By reducing reliance on labelled data, this approach improves efficiency and opens up possibilities for tackling broader challenges, such as real-time ice monitoring, assessing shipping route viability, and conducting long-term trend analysis.
How to cite: Seston, J., Harcourt, W. D., Leontidis, G., Rea, B., Spagnolo, M., and McWhinnie, L.: Leveraging Self-Supervised Learning for Sea Ice Segmentation in the Arctic to Reduce on Labelling, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18799, https://doi.org/10.5194/egusphere-egu25-18799, 2025.
Marine-terminating glacier dynamics play a crucial role in understanding the climate system. They connect large ice sheets, oceans, and the atmosphere; thus their changes might deliver important information about the relationship between those systems. One of the factors describing ice dynamics is velocity. Its changes can reflect the processes occurring on and underneath the ice sheet surface. Nowadays, that information is delivered mainly by remote-sensing sensors, including satellite radar images (SAR), which provide timely and continuous data even in isolated areas. Plenty of offset-based algorithms already exist to deliver reliable velocity maps based on satellite products. However, these methods require setting a bunch of processing parameters, and they are usually suitable for only one sensor type. This study investigates possible machine learning solutions for finding corresponding areas on satellite images in order to provide velocity maps in an alternative way. In this work, SAR datasets from Sentinel-1 satellite were used to test two machine learning approaches for glacier velocity retrieval. The first approach is based on utilising convolutional neural networks (CNN) to select similar areas on the image pairs. The input data consist of only two coregistered SAR intensity images, which are augmented in the next processing step. As the model output, the most similar image patch is returned. After selecting corresponding image patches, the offsets in both image axes are determined and calculated into velocity values based on a pixel size and temporal baseline. The second approach investigates the possibility of applying the LightGlue image matching technique to the analysis of SAR data in order to detect similar features and determine their movement. The same input products are used, and methods performance and reliability are assessed. Both techniques are tested on two glaciers with different ice dynamics and locations: one in Greenland and one in Svalbard. The methods are compared in terms of efficiency, information density, and velocity values reliability. The final maps are validated by offset-tracking results processed for the same input images.
How to cite: Łucka, M. and Sumara, M.: Comparison of selected machine learning algorithms to derive glacier velocity maps, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16392, https://doi.org/10.5194/egusphere-egu25-16392, 2025.
The layered nature of snow is a key characteristic of the seasonal alpine snowpack. In fact, snow stratigraphy influences all physical processes e.g., mechanical or thermal behavior. In order to describe these physical processes precisely, a profound and objective representation of the snow stratigraphy is paramount. The Snow-Micro Penetrometer (SMP) is a rod-driven snow penetrometer that provides resistance-force profiles across snow depth, offering an objective method to measure vertical snow stratigraphy. These submillimeter-scale profiles facilitate the derivation of a micro-mechanical snow model. These derivatives have the potential to initialize complex, physics-based snow cover models (e.g., SNOWPACK). While many parameters for snowpack simulations can be derived directly, determining grain type remains challenging due to the absence of a clear physical correlation. To address this, machine learning (ML) approaches have been investigated. However, prior ML models are limited in their number of snow grain type classes and datasets, which prevents the operational use of these models. Recently, Kaltenborn et al. introduced Snowdragon, a ML benchmark for automated classification and segmentation of SMP profiles. The current version of Snowdragon is trained on SMP profiles collected during the MOSAiC expedition and contains only specific non-standardized grain types typically observed for snow on Arctic sea ice. In this work, we re-trained the supervised models of the Snowdragon benchmark on Alpine snow. To enable the usage of Snowdragon for a broader community, we adapted the classification of grain types according to the international standard for seasonal snow. Our dataset comprises 52 manually labeled SMP profiles recorded in Alpine snow in Switzerland. Previously identified high-performing ML models were re-trained without additional hyperparameter tuning and subsequently evaluated. We found that the ML model Random Forest performed best but nevertheless had difficulties in recognizing faceted crystals, similar to the other models. Additionally, all models react sensitive to minor force changes in the SMP profiles, often leading to predictions of alternating micro-classes between two grain types. These preliminary results demonstrate the feasibility of this approach for grain type classification, but underscore the limitations posed by the small dataset size. Future work will focus on expanding the training dataset and developing a robust interface for operational use of the prediction output. This work marks a step toward more reliable and generalizable snow grain classification of SMP signals for operational use, like snowpack modeling and avalanche assessment.
How to cite: Lehnert, J., Hofmann, M., Kaltenborn, J., Schneebeli, M., and Mitterer, C.: Advancing snow grain classification for snow micro-penetrometer signals using machine learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-469, https://doi.org/10.5194/egusphere-egu25-469, 2025.
To make accurate projections of future sea level rise, small-scale ice-sheet and ice-shelf processes must be included in global climate models. Since high-resolution fully-coupled ice-sheet--ocean models are computationally expensive, multi-centennial simulations use lower resolution grids combined with simple parameterizations of the ice-ocean interface. However, these simple parameterizations do not fully reproduce observed melt patterns and have low sensitivity to warmer conditions. Instead, neural networks can be used to improve models by emulating the ice-ocean interactions simulated by high resolution models. We present a framework for training neural networks to emulate small-scale Antarctic basal melt processes within a global low-resolution model (here the NEMO ocean model). We employ a multi-layer perceptron which is trained with a variety of model simulations on a grid with quarter degree resolution, and aim to assess the performance of the neural network, particularly in warmer conditions representative of potential future climate states. This simple framework provides a springboard for future work using more complex architectures, and offers the potential to run computationally affordable long-period global simulations while still capturing crucial ice-shelf--ocean interactions.
How to cite: Ockenden, H., Burgard, C., Jourdain, N., and Mathiot, P.: Neural network emulators of high resolution melt processes under Antarctic ice shelves, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-5944, https://doi.org/10.5194/egusphere-egu25-5944, 2025.
Snow water equivalent (SWE) is an important component of the hydrological cycle but still faces large uncertainties in its quantification due to its high temporal and spatial variability. While machine learning (ML) has been applied to multiple domains in hydrology, its use for SWE prediction has been hindered by limited observational training data beyond the local scale. Hybrid models that integrate simulated data from physics-based models with a ML setup may overcome this lack of observations, outperforming both physics-based models and conventional ML approaches in data-scarce regions.
In this project, we tested two different hybrid ML setups that predict the daily change in SWE using Crocus snow model simulations together with data from ten meteorological and snow observation stations throughout the northern hemisphere containing 7-20 years of data. The first setup follows a common post-processor approach where the outputs and state variables from Crocus are fed as additional predictors to the ML model at each time step. The second setup follows the concept of data augmentation, where Crocus is used to simulate SWE for stations for which no observations are available. These simulations are then fed as additional data points to the ML model, but are weighted in the loss function to control their influence during training.
The obtained results show that the post-processor approach is best suited for predicting SWE in years excluded during training. However, when predicting SWE in untrained stations the data augmentation setup achieves the largest increase in performance, reducing the root mean squared error by 22% compared to Crocus and by 42% compared to the measurement-based ML model. A feature importance analysis reveals that the hybrid model predictions are influenced the most by the current SWE status, incoming radiation, snowfall and air temperature. These results showcase the potential of hybrid models for predicting variables that suffer from data scarcity such as SWE.
How to cite: Pomarol Moya, O., Karssenberg, D., Immerzeel, W. W., Kraaijenbrink, P., Nussbaum, M., Mehrkanoon, S., and Gouttevin, I.: Bridging machine learning and physics-based models for improving snow water equivalent predictions in the northern hemisphere, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11168, https://doi.org/10.5194/egusphere-egu25-11168, 2025.
Sea ice is a crucial component of the polar marine environment. A contiguous piece of sea ice is called an ice floe, and the size variation in these floes across a region is described as the floe size distribution (FSD). Analysis of FSD provides information on the physical processes associated with sea ice dynamics, which is needed for calibrating and validating numerical sea ice models. For example, the size and shape of sea ice floes is predominantly controlled by wind and ocean wave conditions, thus the FSD metric provides crucial insight into these environmental conditions. Consequently, the automatic detection of floes, and hence FSD, is required to improve our understanding of these conditions over large spatial-temporal scales. Here, we present a method to automatically segment sea ice from Synthetic Aperture Radar (SAR) images for downstream applications. Our method uses an autoencoder architecture, minimising dual losses concurrently to guide the training on a large number of SAR images.
For machine learning (ML) to assist in automatic labelling of sea ice, traditional supervised learning models require the provision of a sufficiently large, labelled dataset to train the model. Manual interpretation and identification of sea ice in satellite imagery is a time consuming and tedious process, frustrating the development of annotations over large spatial areas. Additionally, manually labelled data are subject to unintentional human variability thus potentially introducing bias. It is not a scalable solution.
Feature learning or representation learning is a ML technique that automatically guides its own training to extract useful information without the need for labelled data. Instead of using optical images as many other works done on FSD with supervised learning techniques, we use SAR images here with representation learning. Using SAR images allows us to monitor sea ice conditions year-round, including during periods of polar darkness and cloudy conditions, where the detection of sea ice conditions in optical images is problematic. As this autoencoder model does not require labelled data, it can be scaled both spatially and temporally. It also has the potential to be extended to detect other features and to learn beyond ice-water segmentation.
How to cite: van Zeeland, L., Rogers, M. S. J., Hughes, N., Evans, B. R., Strickson, O., Veyssière, G., Fleming, A., Hosking, S., and Wilkinson, J.: Label-free ice floes segmentation in SAR images for floe size distribution in the Antarctic, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12902, https://doi.org/10.5194/egusphere-egu25-12902, 2025.
Global warming threatens to cause irreversible planetary changes and is accelerated in the polar regions, warming at nearly four times the global average. Warmer temperature exacerbates ice sheet ice loss, increasing the freshwater discharge into oceans and contributing to rising sea levels and regional changes in ocean salinity, threatening a collapse of ocean currents. The number of humans living below sea level is projected to rise by 73% by the turn of the century. Therefore, accurately determining the ice loss and the freshwater discharge is paramount to enable decision-makers to take necessary actions.
Ice sheet ice loss can be estimated using a satellite altimeter, measuring the spatial ice sheet surface height at many time instances. The apparent elevation change can be converted into mass change by accounting for bedrock movement and snow/firn processes. An obstacle in utilising satellite altimeter data is the unstructured nature of the data points resulting from elevation observations at different time instances. We propose to treat these altimeter data as cloud points in the space-time domain and utilise implicit neural representation (INR) to encode the target field as a continuous function varying both in time and space. Compared to traditional interpolation methods such as trilinear interpolation or kriging, the INR method can capture non-linearities and long-term trends while providing a compact encoding of the target field, allowing for scalable dissemination of the product.
We present a feasibility study of utilising INR to reconstruct the surface elevation of the Petermann glacier, northwest Greenland, from CryoSat-2 radar altimeter elevation observations. We carried out many model training experiments, consisting of ablation studies on additional loss terms as well as model architectures (SIREN, RFF, KAN and MFN) and hyperparameters (number and width of layers and loss term weights), to find the best combination. The main difficulty is correctly capturing the glacier temporal dynamics. In addition, we trained models with varying quantities of data (5 months, 1 year, 2 years and 12.5 years) to investigate whether more data improved the model performance. Results are evaluated using Operation IceBridge (OIB) LIDAR, and GeoSAR elevation measurements. OIB allows for evaluation of model elevation over a large temporal and geographical area, whereas GeoSAR allows for comparing high resolution elevation data on a single day over a small area.
Results indicate that we achieve the best performance using the SIREN INR architecture coupled with high temporal and spatial loss weights. In addition, models perform best when using CryoSat-2 data from the entire 12.5 year time frame. The models perform particularly well in regions with high data point density but struggle at the outer rims of the ice sheet where the point density is low. The feasibility study presents a promising direction in modelling the spatiotemporal evolution of the ice sheet at a sub kilometre resolution with a daily temporal time step using INR. We foresee these methods being applicable to many geoscience applications with irregular data sampling in space and time.
How to cite: Naylor, P., Stokholm, A., Dionelis, N., Andersen, N. H., and Simonsen, S. B.: Temporal Evolution of the Petermann Glacier Surface Elevation with Implicit Neural Representation in High Spatiotemporal Resolution using CryoSat-2 Data, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15407, https://doi.org/10.5194/egusphere-egu25-15407, 2025.
In recent years, Artificial Intelligence (AI) has been a game-changer in climate modeling, providing innovative and adaptable approaches while improving accuracy and computational efficiency. For instance, hybrid models can preserve the robustness of physical modeling while emulating components that are computationally expensive or poorly represented. The relatively two-dimensional and localized nature of sea ice makes it an ideal candidate for AI-based emulation, offering a solution to the significant computational burden it imposes on ocean models. Here, we present a sea ice emulator for the Finite Element Sea Ice-Ocean Model (FESOM), capable of predicting the evolution of sea ice thickness (SIT), concentration (SIC), and drift (SID) on timescales ranging from weeks to months.
First, an adaptive U-Net-based model is trained to predict sea ice state (SIT, SIC, SID) increments at one (or multiple) lead times ahead, using corresponding atmospheric forcing and past and current sea ice states. The model is driven by multi-decadal series of daily to sub-daily atmospheric forcing and 2D sea ice and ocean outputs from FESOM, which have been preprocessed and re-interpolated onto a regular grid. To ensure scalability, training sequences are divided into chunks, managed by a custom mapper that balances their usage during training and supports compatibility with multi-GPU configurations. The model is trained by minimizing a penalized mean square error loss function, with an adaptive learning rate controlled via a dedicated scheduler, until convergence. The quality and accuracy of the training process are systematically assessed prior to inference.
Emulation of sea ice can then be performed using recursive inference of the trained models for rollouts spanning from some weeks to a year. Subsequent sea ice states are occasionally clipped into their physical range in order to prevent non-physical behaviors. Rolling predictions can be eventually generated daily or weekly along the test sequences, similar to operational forecasting.
Apart from SIT, SIC, and SID maps, metrics including Integrated Ice Edge Error, root mean square error, mean ice thickness, and Sea Ice Extent are implemented to evaluate the quality of the prediction in comparison to the actual FESOM outputs and some predefined baselines. The emulator demonstrates robust predictions up to 100 days, while still maintaining a realistic representation of various sea ice states beyond this time. Both training and inference are scalable and have been deployed on GPUs, although rolling predictions can be run on a single CPU without incurring prohibitive costs. Computation times for both steps will be estimated, along with the time required for a standard FESOM simulation including sea ice, to assess the potential gain in computational efficiency.
How to cite: Birrien, F., Hutter, N., and Koldunov, N.: On emulating Sea Ice in the Finite Element Sea Ice-Ocean Model (FESOM), EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18155, https://doi.org/10.5194/egusphere-egu25-18155, 2025.
