Tidewater glaciers are critical gateways for global sea level rise, with their stability strongly influenced by complex fjord circulation patterns that control submarine melting. Direct observations of these dynamics with conventional oceanographic instruments remain challenging due to temporal or spatial constraints. However, the fjords themselves contain a distributed sensor network: icebergs. As passive tracers driven by currents, icebergs of different sizes respond to circulation at different depths due to their varying underlying drafts. Deriving quantitative circulation data from these tracers requires tracking hundreds to thousands of similar-looking icebergs simultaneously.
This work presents an automated multi-object-tracking framework that extracts dense, size-stratified velocity fields from time-lapse imagery, providing the observational foundation required to reconstruct depth-dependent circulation patterns within glacier fjords. We introduce a scale-adaptive object detection architecture based on Faster R-CNN that achieves 87.1\% detection recall and successfully captures a large fraction of the iceberg population from only a sparse set of manual labels. To maintain persistent identities in dense scenes, we employ a multi-modal association strategy that combines Kalman-filtered motion priors with appearance similarity learned via Vision Transformers. Evaluated across diverse environmental conditions, the framework demonstrates high stability with 95.7\% identity consistency (IDF1) at 2-minute time intervals and generalizes to unseen glaciers without retraining. By transforming time-lapse imagery into quantitative circulation records, this work provides a robust framework for monitoring the hidden ocean dynamics that drive glacier retreat.
How to cite: Jaeger-Kufel, M., Neumann, A., Vieli, A., Kneib-Walter, A., Welty, E., and Umlauft, J.: Icebergs as a Distributed Sensor Network: Iceberg Tracking in Time-Lapse Imagery for Fjord Circulation Analysis, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13443, https://doi.org/10.5194/egusphere-egu26-13443, 2026.
Deep learning (DL) methods have become a key technique for automatic sea ice type mapping from synthetic aperture radar (SAR) imagery, yet their deployment in operational sea ice charting is still hindered by scarce labelled data, limited adaptive feature extraction, and the lack of interactive mechanisms, which restrict model generalization, accuracy, and usability, especially in hard-to-classify scenes. To address these bottlenecks, we propose an efficient sea ice classification model, ESICM, targeting four ice types: open water (OW), young ice (YI), first-year ice (FYI), and multiyear ice (MYI), and enhance performance and practicality under label-scarce conditions through three key designs. First, we introduce a few-shot learning (FSL) framework to more effectively exploit limited labels and reduce the reliance of traditional supervised learning on large labelled datasets. Second, inspired by classical sea ice parameter retrieval algorithms, we design a lightweight channel multiply–divide convolution module (CMDM) that strengthens adaptive feature extraction with only ~190k parameters, thereby improving discrimination of multi-scale textures and sea ice types with subtle backscattering differences. Third, we incorporate an interactive mechanism based on the Segment Anything Model (SAM) and couple it with the FSL framework, allowing the classifier to be adjusted with minimal human intervention and thus improving operability in difficult SAR scenes. ESICM is trained on 512 scenes from the AI4Arctic sea ice challenge dataset and evaluated on 20 independent test scenes, achieving 91.73% overall accuracy (OA), 91.29% F1 score, 85.61% Cohen’s kappa, and 71.52% mean intersection over union (mIoU), outperforming comparative DL models by at least 1.35, 1.90, 2.54, and 2.53 percentage points on these metrics, respectively. In melting season scenes, particularly those dominated by MYI, ESICM’s F1 and IoU outperform the second-best model by 22.21% and 19.15%, respectively. Further cross-domain experiments demonstrate that, even when trained on only about one quarter of local scenes, ESICM still achieves the highest accuracy, demonstrating strong cross-regional generalization. Meanwhile, its interactive functionality enables users to refine classification results via prompts in hard-to-classify scenes, substantially enhancing classification performance. Overall, ESICM provides a lightweight, high-accuracy, and interactively adjustable DL solution for SAR-based sea ice classification under limited labelled data, offering robust technical support for polar navigation safety and sea ice environmental monitoring.
How to cite: li, Y., Uotila, P., Li, C., Leppäranta, M., and Peng, C.: Efficient Sea Ice Classification Built on Few-Shot Learning Framework, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-518, https://doi.org/10.5194/egusphere-egu26-518, 2026.
Monitoring Arctic sea ice variability is crucial for maritime safety. Synthetic Aperture Radar (SAR) imagery provides an effective means of achieving this through all-weather, day-and-night coverage of the Arctic. Navigation in the Canadian Arctic Archipelago currently relies on operational ice information services, including analyst-derived ice charts, satellite imagery, and ice routing products provided by national ice services. However, the development of machine-learning systems capable of automatically processing large volumes of satellite imagery and accurately identifying ice conditions is constrained by the need for extensive manually labelled datasets. To address this limitation, we developed a self-supervised learning (SSL) approach, which uses unlabelled data to learn general image representations. Specifically, we use Bootstrap Your Own Latent (BYOL), a non-contrastive SSL framework, to pretrain a UNet encoder on unlabelled dual-polarised Sentinel-1 Extra-Wide mode (EW) SAR scenes before fine-tuning with a small set of labelled images. We compare the BYOL-pretrained UNet (called UNet SSL in this study) to four baselines: a control UNet, a fully supervised UNet, a Random Forest classifier, and the Segment Anything Model (SAM). With only three labelled scenes, the BYOL-pretrained UNet achieved higher segmentation accuracy than the fully supervised model trained on seven images, more than twice the number of labelled scenes. The most significant gains occurred in Marginal Ice Zone (MIZ) scenes, where the BYOL-pretrained UNet achieved a Matthews Correlation Coefficient (MCC) of 0.2087, compared with 0.1685 for the fully supervised UNet trained on seven labelled scenes and 0.1449 for the control model trained on three scenes—representing an MCC increase of approximately 24% and 44%, respectively. These improvements were accompanied by a substantial reduction in false negatives and a marked increase in recall, indicating improved discrimination under low-contrast, fragmented floe conditions. Our findings demonstrate that SSL reduces annotation requirements for SAR-based sea ice segmentation, improving model generalisation in both consolidated and fragmented ice conditions. This approach offers a scalable solution to the labelling bottleneck in Arctic monitoring and highlights the potential of BYOL as a general pretraining strategy for SAR-based Earth observation image segmentation.
How to cite: Seston, J., Harcourt, W. D., Leontidis, G., Rea, B., Spagnolo, M., and McWhinnie, L.: Self-supervised learning reduces labelling requirements for sea ice segmentation in Sentinel-1 SAR imagery , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7803, https://doi.org/10.5194/egusphere-egu26-7803, 2026.
Antarctic Sea Ice (ASI) has experienced a sudden and drastic decline since 2016, following decades of gradual growth since the start of satellite observations. This sharp reversal caused suggestions that a regime shift has happened. However, the underlying drivers remain uncertain due to complex atmosphere-ocean interactions and pronounced regional variability. While atmospheric circulation patterns and large-scale teleconnections influence ASI variability, their spatial aggregation limits their ability to explain regional changes. Recent studies point to an increasing role of ocean heat content, yet its contribution relative to atmospheric influences has not been quantified. In this study, we address this gap by developing a framework to identify stable, spatially coherent climate drivers of regional ASI. We first introduce a workflow combining correlation analysis with HDBSCAN clustering to detect global clusters that have persistent correlations with regional ASI and can serve as robust predictors. We then use these clusters as input features in a linear regression model combining atmospheric variables and ocean heat content to assess how well ASI variability can be reconstructed. Finally, we evaluate how the relative importance of atmospheric and oceanic drivers has changed before and after the extreme low-ice events beginning in 2016.
Our results demonstrate that (1) the proposed clustering framework reliably identifies physically meaningful driver regions, (2) a linear model using these drivers can successfully reproduce regional ASI variability, and (3) the contribution of ocean heat relative to atmospheric forcing varies markedly across regions.
How to cite: Öhlckers, N. S., Lorenz, D., and Ionita, M.: Beyond Teleconnections - Uncovering Stable Drivers of Antarctic Sea Ice Anomalies, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-364, https://doi.org/10.5194/egusphere-egu26-364, 2026.
Quantifying iceberg calving is important for understanding ice mass loss of the Greenland Ice Sheet and its subsequent impact on sea level rise, and for refining calving laws that currently represent a major source of uncertainty in global climate models. These calving events, particularly those involving capsizing icebergs, exert time-varying forces on the edge of marine-terminating glaciers that produce distinct seismic signals known as glacial earthquakes.
By processing twelve years of continuous seismic data and employing a Random Forest classifier to distinguish these glacial earthquakes from tectonic events, we generated a comprehensive catalogue of 6263 previously undocumented glacial earthquakes occurring between 2013 and 2024. The detected events are located along the Greenland coast with surface wave magnitudes ranging from MSW 4.1 to 5.4. They cluster at nine major calving glaciers, though the vast majority of activity is concentrated at Sermeq Kujalleq (Jakobshavn Isbræ) and Helheim Gletsjer.
To investigate the driving mechanisms behind these events, we analysed the correlation between calving activity and various environmental variables, including glacier velocity, air temperature, sea ice fraction, sea surface temperature, and wind speed. We train a second Random Forest model to predict monthly calving events and evaluate the relative importance of these environmental features, while applying statistical analyses to investigate correlations on a yearly basis where data points are limited. Our results indicate that the relationship between calving and the environment is highly complex and site-specific, as no single variable serves as a universal driver for all glaciers.
This complexity is further highlighted by scale-dependent correlations between calving events and environmental variables. For instance, while the glacier velocity shows a strong correlation with cumulative yearly calving at Sermeq Kujalleq, its importance diminishes on a monthly scale. Conversely, Helheim Gletsjer exhibits no clear yearly correlation with the glacier velocity, highlighting the site-specific nature of calving dynamics. We will present the spatio-temporal evolution of the detected events and discuss how these diverse environmental correlations quantify the varying sensitivity of individual glaciers to environmental forcing across different temporal scales.
How to cite: Wetter, S., Mangeney, A., Stutzmann, E., Hibert, C., and Lane, S. N.: From seismic signals to calving drivers: Assessing twelve years of glacial earthquakes in Greenland using Random Forest models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14626, https://doi.org/10.5194/egusphere-egu26-14626, 2026.
Dynamical ice-sheet models are among the primary tools used to investigate the evolution of ice sheets. However, their computational cost increases rapidly with spatial resolution, often making long-term or ensemble simulations prohibitively expensive. Here, we investigate whether recent advances in machine-learning-based super-resolution techniques for spatiotemporal data can be leveraged to reduce these computational costs while retaining high-resolution information.
Using one pair of low- and high-resolution simulations of the Greenland ice sheet for the 20th century, generated with the PISM dynamical ice-sheet model, we train a machine-learning-based super-resolution model to learn the mapping from low- to high-resolution states. For subsequent simulations, computationally inexpensive low-resolution model runs are combined with the trained super-resolution model to reconstruct high-resolution fields. We evaluate this hybrid framework by assessing (1) whether the super-resolution model can accurately reproduce the spatial details of high-resolution simulations, and (2) whether it can mitigate deficiencies in the long-term trends produced by low-resolution models. Our results provide insight into the potential of machine-learning-based super-resolution as a cost-effective tool for high-resolution dynamical ice-sheet modeling.
How to cite: Scher, S., Aschwanden, A., Schalamon, F., Trügler, A., and Abermann, J.: A machine-learning-based super-resolution approach for dynamical ice-sheet modeling, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9225, https://doi.org/10.5194/egusphere-egu26-9225, 2026.
Snow cover is a critical component of the earth’s climate and weather system, which exhibits high spatial and temporal variability. Therefore, we predict daily snow cover using spatio-temporal forecasting. Unlike traditional forecasting approaches, that require spatial or temporal aggregation, our approach employs deep learning models specifically designed for spatio-temporal data. Spatio-temporal predictive learning has primarily focused on nowcasting and sub-seasonal forecasts at a daily scale with a lead time up to 15 days. However, we implemented the capability of using such models with long multi-year satellite image time series to predict at a daily scale annually, specifically for daily snow cover. In our research, we use the historical snow cover data from the DLR Global SnowPack remote sensing product, which is a daily cloud free 500m snow cover representation on the ground. We generate high resolution daily snow cover forecasts for up to 365 days (one year ahead) beginning from 1st July. We implemented models such as Convolutional Long Short-Term Memory (ConvLSTM) networks, convolutional encoder decoder architectures with attention mechanisms, and Vision Transformer (ViT) based models and adapted them for our use case. To further enhance our predictions, we also made adaptations to the models to include multivariate spatial and temporal data which are key drivers of snow cover variability into the model. Topographical feature maps derived from elevation, and time series of climatological indices (atmospheric oscillation patterns) are two examples. Validation against reference data demonstrates exceptional accuracy and F1-scores exceeding 84% across forecasts.
How to cite: Shrestha, S. N., Dietz, A., Leibrock, S., and Kuenzer, C.: Deep learning based spatio-temporal forecasting of snow cover in the Alps using remote sensing data , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20480, https://doi.org/10.5194/egusphere-egu26-20480, 2026.
Snow water equivalent (SWE) is an important component of the global hydrological cycle, acting as a primary reservoir for seasonal water storage. Despite its relevance, only few datasets are available that provide long-term daily SWE estimates at global scale. Even amongst the best gridded SWE products, the spatial resolution does not go beyond 10km, a significant limitation considering the large spatial variability of snow. Furthermore, assimilation of snow observations in such products remains another key challenge. Machine learning (ML) models and their combination with process-based simulations, what is known as Hybrid Modelling, offer a promising alternative for producing detailed SWE predictions at large scales, given their high inference speed and adaptability to their training data. Hybrid ML models have already been used for SWE prediction over a small number of sites, improving both pure ML approaches and advanced process-based snow models such as Crocus, but their applicability for long-term spatiotemporal modelling of snow at larger scales remains to be tested.
In this project, we trained an LSTM model using in-situ snow data from roughly 10000 sites throughout the Northern Hemisphere with the aim of creating a 40-year gridded dataset of daily SWE at 1 km resolution. The model incorporates temperature, precipitation, and shortwave radiation as meteorological predictors, alongside a small set of topographic variables and land cover classification. Preliminary results show a good fit to stations excluded from the training set, with an RMSE of 44 mm, where unequal distribution of observation locations was accounted for by a weighting scheme. These findings suggest the suitability of this approach for extending coverage to ungauged regions across the Northern Hemisphere. The use of the ERA5-Land SWE product as a hybrid support promises further improvements in model performance.
Ultimately, this project aims to provide a finer-scale alternative to existing daily SWE products. By improving the spatial resolution to 1km and incorporating available snow measurements, it contributes to a more refined view of seasonal snow storage across the Northern Hemisphere.
How to cite: Pomarol Moya, O., Karssenberg, D., Immerzeel, W. W., Kraaijenbrink, P., Nussbaum, M., and Mehrkanoon, S.: Creating a high-resolution Northern Hemisphere daily SWE dataset (1980-2020) using Machine Learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12034, https://doi.org/10.5194/egusphere-egu26-12034, 2026.
Rock glaciers, the most visible surface expression of permafrost landforms, are found across glacial, periglacial and paraglacial environments. Accurate and consistent mapping of their extent is fundamental for advancing research in geomorphology, hydrology, ecology, geohazard assessment, permafrost dynamics, and climate studies. However, delineating their boundary remains challenging because rock glaciers often occur alongside or merge with other geomorphic equifinal landforms that are difficult to distinguish by their spectral identity in aerial or satellite imagery. Additionally, their boundaries are inherently ambiguous, evolving with changes in topographic and climatic factors. The widely used approach involving manual digitisation through visual interpretation of geomorphic features is time-consuming and subjective. Recent studies have explored deep learning as a means to automate and scale up rock glacier mapping, but existing studies still remain limited in number and geographic scope, with minimal attention to evaluating discrepancies or uncertainties in mapped extents. This study examines the use of a U-Net deep learning model for automated delineation of rock glacier extent, with particular emphasis on associated uncertainties. Using data from the Chile National Glacier Inventory for the Coquimbo region, we trained the model under two strategies: (1) differentiated training based on rock glacier types. A set of models was trained exclusively on landforms with clearly expressed geomorphological features of frontal slopes, lateral margin, and ridge-furrow structures, while another set incorporated all inventoried rock glaciers, including both well-expressed and subdued geomorphological features; (2) different predictor combinations, comparing a configuration that used only RGB + NIR bands from Sentinel 2 or PlanetScope imagery with an expanded set that integrated these spectral bands with DEM derivatives and imagery-derived variables. The highest-performing models from these strategies were then applied to an independent test area, and their outputs were compared against existing inventories to evaluate spatial consistency and assess potential mapping biases. By integrating an automated method with uncertainty assessment, this work contributes to the ongoing advancement of rock glacier detection and delineation methods and highlights the critical need to validate deep learning outputs. Such uncertainty quantification is essential for ensuring the robustness of mapped extents and for supporting applications that depend on accurate and reliable representations of landforms with inherently ambiguous and dynamic boundaries.
How to cite: Tamang, S., MacDonell, S., Brasington, J., Shulmeister, J., and Robson, B. A.: Mapping the Margins: Evaluating Accuracy and Ambiguity in Automated Rock Glacier Delineation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1379, https://doi.org/10.5194/egusphere-egu26-1379, 2026.
Geophysical techniques revealed frozen ground within relict periglacial landforms in which the presence of ice was excluded by traditional geomorphological and topographic approaches. These unexpected frozen bodies, referred to here as cold spots, suggest that permafrost can exist outside traditionally mapped permafrost zones. Under climate change, with retreating glaciers and increasing snow variability, subsurface ice in periglacial landforms becomes a potentially important but overlooked water resource. However, its spatial distribution and climatic controls remain poorly understood.
Here, we develop a methodology to identify cold spots. We focus on the Southern Alps and we assume that cold spots are related to micro-climatic and topographic conditions that allow permafrost to persist. We use a limited set of sites investigated by geophysical surveys, including confirmed cold spots and geomorphologically similar control sites without permafrost. We analyze topographic and climatic remote-sensing data to derive relevant features and examine their relation to cold spots. We then use these features in semi-supervised machine learning classification models to identify areas with conditions similar to known cold spots. The resulting maps highlight potential cold-spot locations targeted for forthcoming geophysical field investigations and provide a practical framework for improving the detection of hidden permafrost.
How to cite: Goldschmidt, Y., Boaga, J., and Marra, F.: Mapping “cold spots” of potential hidden alpine permafrost using semi-supervised machine learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3470, https://doi.org/10.5194/egusphere-egu26-3470, 2026.
There is a distinct difference in the behavior of sea ice extent response to global warming between the Arctic and Antarctic; the former is decreasing while the latter had been increasing slightly until recently. However, satellite data show that Antarctic sea ice has been continuously decreasing since 2016 and reached its minimum in February 2023. The minimization of sea ice extent in Antarctica would have various impacts on the Earth's system. Since ice is more reflective than liquid water, sea ice plays a significant role in maintaining the Earth’s energy balance. Therefore, it is crucial to accurately predict future sea ice response. Here, we aim to predict the sea ice extent for the upcoming season using deep learning models, employing U-Net. Atmospheric and oceanic data related to sea ice, such as sea surface temperature, wind speed, etc., were used as features, while the sea ice extent was set as the target. We trained and tested the models using data from the CESM2 Large Ensemble. We tarined the final model by fine-tuning the model pre-trained on numerical model data with observational data. The performance of the models was compared using ACC and RMSE as evaluation metrics. Additionally, to assess the impact of each variable within the model, we replaced each variable with its climatological mean and observed the changes in the evaluation metrics to determine their importance. These research findings are anticipated to significantly contribute to predicting more accurate changes in Antarctic sea ice and understanding future Antarctic sea ice changes.
How to cite: Baek, G., Ko, J., Jin, E. K., and Kug, J.-S.: Seasonal Predictability of Antarctic Sea Ice based on Deep-learning Approach, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8748, https://doi.org/10.5194/egusphere-egu26-8748, 2026.
Glacier elevation change is a fundamental measure for quantifying glacier mass balance and assessing glacier–climate interactions. Large-scale estimates are commonly derived either from satellite altimetry, which provides robust but spatially sparse measurements, or from digital elevation model (DEM) differencing, which enables spatially continuous mapping but is more sensitive to noise and bias in complex mountain terrain. Machine learning (ML) approaches have increasingly been used to bridge this gap by correcting or reconstructing elevation measurements using climate and topographic predictors. However, because ML-based prediction inherently involves extrapolation beyond directly sampled glaciers, its reliability across heterogeneous glacier systems such as existing in High Mountain Asia (HMA)remains poorly constrained.
In this study, we explore the behaviour of ML-based glacier elevation change predictions trained with ICESat-2 elevation measurements combined with climate and terrain variables across multiple HMA subregions. ICESat-2 footprints provide dense elevation change observations over only a limited subset of glaciers within each subregion. We train subregion-specific XGBoost models and evaluate their performance in relation to glacier sampling characteristics, feature importance, and elevation-dependent behavior.
The results reveal pronounced regional contrasts despite comparable glacier size and sample coverage across regions. In the Karakoram for example, ML-based extrapolation produces spatially coherent and elevation-dependent patterns of glacier elevation change, with predicted dh systematically decreasing from lower to higher elevations, consistent with expected glacier-scale behavior. These structured predictions are associated with robust model performance (R² ≈ 0.7). In contrast, in West Kunlun Shan, extrapolated elevation change fields are spatially uniform and weakly structured, showing little sensitivity to the applied climate and terrain predictors. These results indicate that the effectiveness of ML-based glacier elevation change modeling depends less on sample size or glacier extent alone than on the presence of stable and internally consistent response structures within glacier systems.
How to cite: Huang, Y., Huang, L., and Bolch, T.: Exploring machine-learning extrapolation of glacier elevation change in High Mountain Asia derived from ICESat-2 data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10587, https://doi.org/10.5194/egusphere-egu26-10587, 2026.
Ice-flow inversions aim to infer unobserved controls on glacier and ice-sheet dynamics from limited, noisy surface data but are notoriously ill-posed: multiple parameter fields can reproduce the same observations, solutions are sensitive to priors/regularization, and model nonlinearity amplifies both data and structural errors. Here we target a particularly challenging variant — emulator-based inversion using a machine-learning surrogate for ice flow — where the forward operator is fast and differentiable but only an approximation of the governing physics. We focus on inverting for ice thickness, which remains poorly constrained for most glaciers yet strongly conditions driving stress, basal traction, and therefore hindcast skill and projection uncertainty.
We present emulator-based inversions with the Instructed Glacier Model (IGM), benchmarking against synthetic tests with known truth and contrasting performance with a full-physics ice-flow solver. IGM provides a PINN-based emulator trained by minimizing a energy representing the Blatter–Pattyn equations. This powerful approach has proven very successful in the forward problem but leads to an emulator that may need regular retraining to ensure an accurate solution. We show that this training approach can introduce surrogate error modes that distort gradients and create spurious minima, degrading convergence and reliability of gradient-based optimization used for the inverse problem. To address this, we introduce a hybrid training strategy that augments the physics loss with a data-misfit term against a large training set, with the aim of improving out-of-distribution generalization across glacier geometries. The resulting emulator yields more reliable recovery of unknown fields such as ice thickness and supports the fast, scalable inversions needed for ensemble modelling and robust uncertainty quantification.
How to cite: Rosier, S., Gregov, T., Finley, B., Jouvet, G., and Vieli, A.: Hybrid training for robust emulator-based ice-thickness inversion, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12303, https://doi.org/10.5194/egusphere-egu26-12303, 2026.
Temporal and spatial patterns of snow depth are key predictors for variations in snow water equivalent and snow-hydrological processes. Observations of snow depth distribution are usually scarce either in space or in time. Automatic weather stations can measure snow depth continuously but only for one point with a very small footprint. Campaign based surveys, in contrast, cover larger areas but are limited in spatial coverage due to technical and logistical constraints.
The partial recurrence of snow depth patterns correlated with terrain features is well known. In this work a machine learning approach based on Random Forest and XGBoost is presented to analyze the temporal evolution of snow depth distributions in the Zugspitze region. Input features include elevation and derived terrain features like slope, aspect, curvature, topographic position index and Winstral wind shelter index. Furthermore, simulated energy balance sums and snow occurrence from optical remote sensing data are used. Snow depth data include terrestrial Lidar measurements and photogrammetric data based on airborne and spaceborne platforms including drones, airplanes and the Pléiades satellite constellation.
Interestingly, due to the specific topography of the area featuring a karstic plateau surrounded by steep slopes, no clear elevational gradients were found. Historical information constitutes a useful feature for machine learning but explains only parts of the variability as actual snow depth distributions are altered by wind drift and energy balance. This is reflected by a moderate temporal transferability of the trained machine learning models. Within the study domain, campaign specific machine learning models produce plausible results for areas with data gaps. While Random Forest and XGBoost produce similar results, differences between different sets of input features can be substantial. Meltout patterns based on remote sensing data can partly compensate for a lack of historical snow depth information.
Machine learning proves to be a suitable tool for closing spatial data gaps. The results also highlight the importance of a process-based choice of input features, as inter- and intraannual snow depth distributions differ even in a region with stable snow depth patterns.
How to cite: Schattan, P., Knieß, J., Gascoin, S., Jubanski, J., Facchinetti, R., Rempfer, C., Wetzel, K.-F., Voigt, C., Schulz, K., and Koch, F.: A Random Forest and XGBoost analysis of the temporal and spatial variability of snow depth in the Zugspitze region based on terrain features, simulated energy balance and remote sensing, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13166, https://doi.org/10.5194/egusphere-egu26-13166, 2026.
The absence of data in the existing instrumental record significantly limits our ability to comprehend and forecast tipping points in the Greenland Ice Sheet (GrIS) and Subpolar Gyre (SPG). Multidirectional approaches are therefore required to capture the complexity of systemic changes and support future early warning efforts. In this study we discuss the ongoing work of the research project VERIFY: Out Of Sample Testing For Early Warning Systems Using Past Climate. We combine computational experiments, with physics-informed emulation and insights from expert elicitation to better understand dirvers of paleoclimate regime shifts in the GrIS. Employing uncertainty quantification methods, we make use of machine learning surrogate models to approximate the system's response. Surrogate models can accurately mimic input-output relationships of complex and computationaly expensive models, providing the opportunity to produce large ensembles for fully exploring the range of plausible model inputs. The goal is to understand what drives the exceedance of critical thresholds through the integration of computational experiments, machine learning and current scientific knowledge.
How to cite: Salmanidou, D., Gregoire, L., Snoll, B., Frisby, C., Graham, M., and Guillas, S.: Quantifying the uncertainty of regime shifts in the paleoclimate via physics-informed emulation and expert knowledge, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15167, https://doi.org/10.5194/egusphere-egu26-15167, 2026.
Accurate, high‑resolution Snow Water Equivalent (SWE) information is critical for reliable hazard assessment and effective water resource management. Yet, widely used global reanalysis products provide SWE at coarse spatial scales and exhibit substantial terrain‑ and melt‑related biases. In contrast, dynamically downscaled products offer improved detail but are costly to run and thus, remain limited in availability.
To address this limitation, we introduce the Linear Attention Snow Downscaling Model (LASDM), a lightweight hybrid deep learning architecture designed specifically to enhance the spatial detail and physical realism of SWE fields. LASDM combines convolutional neural networks with linear attention based transformer blocks, enabling efficient representation of synoptic‑to‑local snow processes while remaining highly parameter‑efficient (<1 million parameters).
Applied to the ERA5 → ERA5‑Land downscaling problem over the Great Lakes region (1980–2022), LASDM demonstrates stronger performance than U‑Net, Swin Transformer, and statistical baselines across a range of evaluation metrics. Case studies for two winter storms provide additional context for these differences. More broadly, this work suggests the potential of machine‑learning architectures for downscaling and bias correction. LASDM offers a compact and adaptable framework that may help improve snow representation and support applications that rely on higher‑resolution SWE.
How to cite: Ali, M., Ait-Chaalal, F., Kumar, S., and Dobbin, A.: A Lightweight Hybrid CNN–Transformer Architecture for High‑Resolution Downscaling and Bias Correction of Snow Water Equivalent, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16321, https://doi.org/10.5194/egusphere-egu26-16321, 2026.
Sea ice covers the oceans in polar regions and is closely related to heat circulation between the Sun and the Earth, and the absolute amount of sea ice can represent climate change itself. The research aiming to prepare for future climate conditions by predicting sea ice concentration (SIC), the area ratio of the ocean covered by sea ice, is being actively conducted.
From a long-term perspective, sea ice is influenced by sea surface temperature (SST), 2 m air temperature (t2m), wind fields, and so on, and machine learning or deep learning techniques are used to predict SIC in order to leverage the correlations among variables. However, due to the characteristics of deep learning techniques, there are limitations in identifying how much each variable influences the SIC prediction results.
This study simultaneously predicts SIC, t2m, and SST through a multitask Transformer model, and the predicted t2m and SST are converted into a gate intensity map to correct the bias of SIC. Through this, we interpreted how atmospheric and oceanic environmental factors affected the SIC prediction results. In addition, by comparing the prediction results of SIC and environmental factors under conditions such as specific seasons and regions, where prediction is relatively unstable, we quantified the variable-specific weights under those conditions.
The gate intensity map used for SIC bias correction can itself be used as an uncertainty map, and expresses, as a spatial distribution, regions that are difficult for the deep learning model to predict. In addition, by comparing the impacts of each environmental factor by lead time, the contributions of variables can be identified at long-term and short-term prediction time points.
How to cite: Kim, W., Chung, I., Lee, G., Choo, M., and Im, J.: Enhancing sea ice concentration prediction with multi-task learning and conditional residual refinement, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16558, https://doi.org/10.5194/egusphere-egu26-16558, 2026.
For most glaciated areas, detailed mountain glacier evolution since the last interglacial is largely unknown. Due to limitations of traditional numerical modelling, previous studies have typically operated at a coarse spatial resolution, limited study area size, or focused on major climate events. Here we address these limitations by applying the Instructed Glacier Model (IGM), a deep-learning ice-flow emulator enabling efficient GPU-accelerated transient simulations. We model mountain glacier evolution since 130 ka and up to the present-day at 500 m resolution across eight broad mountain ranges in North America, South America, Eurasia, and Africa. Paleoclimate variables are approximated and regionalised using a combination of global and regional climate proxy datasets. We perform 617 parameter-calibration experiments varying paleoclimate and ice-dynamic parameters, with an average runtime of 21 hours per experiment. Model performance is assessed using combined areal and volumetric validation at two known glacial states; the last glacial maximum and the present-day. We also introduce a reproducible probabilistic model-evaluation framework combining confusion matrix validation score and simulation rank to identify sets of acceptable model parameters rather than a single best-fit solution. Our results show that IGM can model realistic ice-flow patterns, glacier geometries, and transient evolution across full glacial-interglacial cycles, demonstrating that machine-learning models of ice dynamics generalise to new domains and conditions, although performance can decline at coarser spatial resolutions. Together, these results demonstrate the feasibility of global-scale, high-resolution, transient glacier modelling over orbital timescales using a deep learning instructed model, while providing a 100-year interval dataset including glacier extent, ice thickness, and ice flow patterns for the last glacial cycle.
How to cite: Barndon, S., Lima, A. C., Chandler, D. M., Wiersma, A. T., Rentier, E. S., Prats, R. P., and Flantua, S. G. A.: Global-scale modelling of mountain glacier evolution since the last interglacial, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18020, https://doi.org/10.5194/egusphere-egu26-18020, 2026.
Geo foundation models (GFMs) have recently emerged as a new paradigm in Earth observation (EO). They provide a promising approach for enhancing remote sensing analysis. GFMs enable faster and more generalised applications. They are deep learning models trained on large unlabelled datasets to learn general spatial, spectral, and contextual representations of the Earth’s surface. The datasets used are usually diverse in location, seasons, and even sensors. This diversity ensures that the model learns features that are as general as possible. This is vital because labelled data in remote sensing are limited, while high-quality unlabelled data are widely accessible. As a result, GFMs are increasingly viewed as a promising tool for scalable and robust environmental monitoring.
Among various EO tasks, glacier mapping is particularly relevant in the context of GFMs. Glaciers are located in hardly accessible regions, which makes ground-truth (GT) preparation difficult. Delineation of glaciers is often affected by seasonal snow and regional variability. Moreover, debris-covered and rock glaciers are harder to detect due to their complex landforms and their similarity to surrounding terrain. Accurate glacier delineation is crucial for monitoring cryospheric changes, assessing climate change impacts, managing water resources, and mitigating natural hazards.
In this study, we explore the applicability of GFMs for glacier mapping using multispectral Sentinel-2 imagery. We apply fine-tuning of pre-trained GFMs for glacier delineation, with the aim of assessing their potential in comparison with traditional deep learning approaches.
How to cite: Barzegar, F., Kuehtreiber, N., and L. Ullo, S.: Exploring Geo Foundation Models for Glacier Mapping Using Remote Sensing Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18070, https://doi.org/10.5194/egusphere-egu26-18070, 2026.
Arctic snow exerts a critical control on winter soil temperature and carbon exchange, however representation of its properties in Earth System Models (ESMs) remains simplified. In the Community Land Model v5.0 (CLM5.0), recent updates to snow compaction schemes have led to overly dense tundra snow and excessive conductive heat loss, producing a persistent cold-soil bias. Here we developed a Random Forest (RF) regression model to derive tundra snow density from meteorological variables, trained on Arctic SVS2-Crocus (ASC) simulations supported by in-situ observations collected around peak annual SWE from Trail Valley Creek (TVC), Northwest Territories, Canada. The RF model reproduces ASC-simulated density evolution with a mean absolute error of 23 kg m-3 and an R2 of 0.90, matching field measurements more closely than CLM5.0. Future snow density predictions using the RF model driven by bias-corrected NA-CORDEX meteorology (2016 – 2100) indicate bulk snow densities 200 – 450 kg m-3 lower than CLM5.0 and more consistent with tundra conditions. Application of RF-derived snow densities decreases CLM5.0 winter season 10cm soil temperature RMSE by approximately 2 – 3 °C relative to field measurements (2017 – 2023) and increases future winter soil temperature projections (2016 – 2100) by 4 – 7 °C, highlighting the strong sensitivity of CLM5.0’s soil thermal regime to snow physical properties.
How to cite: Rutherford, J., Rutter, N., Wake, L., Woolley, G., Boike, J., and Cannon, A.: Random forest predictions of tundra snow density elevate Arctic soil temperatures in CLM5.0, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19235, https://doi.org/10.5194/egusphere-egu26-19235, 2026.
Faithful modeling of Antarctic climate relies on capturing polar-specific processes at high spatial resolution (~10-30 km). Most CMIP earth system models (ESMs) used for climate projections inadequately represent physical processes that are key drivers in polar climates, and operate at resolutions too coarse to resolve them. Polar-oriented regional climate models (RCMs) are considered the state-of-the-art in modeling the atmosphere at high latitudes, where air-snow interactions are critical, but they are costly to run. This limits their use for exploration of large ensembles and scenarios as well as their potential of integrating into a coupled modeling pipeline.
In order to address these limitations, we develop an affordable surrogate model, or emulator, of the polar-oriented Modèle Atmosphérique Régional (MAR), using deep learning and a variant of the commonly used U-Net convolutional neural network architecture. The emulator is trained to predict 35 km-resolution daily maps of surface mass balance (SMB) components—snowfall, rainfall, run-off and sublimation—over the Antarctic ice sheet from large-scale atmospheric fields of ESMs, effectively learning the downscaling function embedded in MAR. To achieve this, we use a dataset composed of MAR simulations forced by 4 CMIP ESMs over the 1980–2100 period, covering SSPs 1-2.6, 2-4.5 and 5-8.5. We conduct different experiments to assess its best-case performance as well as its transferability to unseen scenarios and ESMs.
The emulator demonstrates strong in-domain skill, displaying high fidelity in reproducing both day-to-day and spatial synoptic variability of the predicted quantities. Long-term SMB trends and interannual variability through 2100 are also well-replicated, with predicted integrated surface mass change over the 1980–2100 period differing by only 1% from MAR. We find that the emulator is robust against unseen emission scenarios, with marginal increase of up to few percent in RMSE. Transferability to other ESMs proves more challenging but results remain promising.
The MAR emulator can be used to generate SMB forcings for ice-sheet models at a negligible computational cost compared to RCMs, allowing century-scale simulations to be produced within minutes and thereby enabling the exploration of a wide range of scenarios and ensemble members. We suggest the general framework of this work could allow for the emulation of MAR in any application where it can be traditionally used. Ongoing work is also investigating the applicability of the emulator within an atmosphere–ice sheet coupled framework.
How to cite: Gellens, A., Agosta, C., N. Legasa, M., Vrac, M., Amory, C., and Kittel, C.: A deep learning-based emulator of the regional atmospheric model MAR for estimation of the Antarctic surface mass balance, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20684, https://doi.org/10.5194/egusphere-egu26-20684, 2026.
Remoteness, harsh environmental conditions, short field seasons, and high operational costs severely constrain the ability to collect observations of the polar cryosphere at scale. These limitations make efficient survey planning an important methodological need: data acquisition strategies must prioritise measurement locations that simultaneously (i) reduce model uncertainty and (ii) maximise scientific utility, for example by tightening constraints on projected ice sheet contributions to sea level rise. We address this need with Bayesian optimisation (BO), a probabilistic machine learning framework for black-box optimisation that uses a Gaussian process surrogate to model the target geospatial field and an acquisition function to formalise the trade-off between uncertainty reduction and scientific utility when proposing subsequent measurement locations. To showcase the approach, we consider a case study on planning airborne geophysical surveys of Antarctic ice thickness and bed topography, for which we introduce a set of novel acquisition functions tailored to Antarctic ice dynamics that translate cryospheric objectives into the BO framework:
- The FluxUCB (Flux Upper Confidence Bound) acquisition function incorporates satellite-derived ice velocity observations to prioritise sampling uncertain, potentially high-flux regions under the current posterior, since such regions can exert a disproportionate influence on ice discharge.
- Alternatively, PBBS (Probability of Bed Below Sea level) prioritises locations with a high posterior probability of marine-based grounding, thereby focusing effort on areas most relevant to assessing marine ice sheet instability (MISI).
In simulation, these objectives reduce posterior uncertainty per flight hour more efficiently than baseline strategies and more consistently target scientifically consequential regions. Together, these acquisition functions illustrate how BO can translate scientific priorities into an uncertainty-aware decision framework for data-efficient polar observation campaigns. More broadly, the framework has strong potential to extract greater value from limited polar field resources beyond airborne surveys, from optimising seismic survey design to informing ice core drilling site selection.
How to cite: Bente, K., Marchant, R., and Ramos, F.: Bayesian Optimisation for Antarctic Survey Planning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22268, https://doi.org/10.5194/egusphere-egu26-22268, 2026.
Posters virtual: Tue, 5 May, 14:00–18:00 | vPoster spot 1a
EGU26-16248 | Posters virtual | VPS20
Explainable Expert-in-the-loop sea-ice classification with statistical modelsTue, 05 May, 14:12–14:15 (CEST) vPoster spot 1a
Sea ice classification is often a crucial step to predict climatic insights and ensure safe marine navigation. In the last few decades, satellite information has been widely used to classify sea ice in broad areas for practical applications. However, common problems are:
1) Low resolution of satellite images to provide precise classification,
2) High computational need, and
3) Scarcity of general models to discover unknown patterns in the data, especially those that enable free selection of satellite sensors to fit the application at hand.
We propose an explainable unsupervised model to integrate ice-experts’ inputs to models so that the problem of having low-resolution data can be overcome. In other words, the results of the models, given as semantic maps, can be further refined using inputs from ice-experts.
Model explainability and visual interpretation of models serve as tools to talk to’ domain experts. The use of Explainable AI in such vital activities ensures trust and easy detection of error. We present an example from a sea ice classification with Sentinel-1 time-series in the scope of the Horizon 2020 project ExtremeEarth.
A further example from the Horizon Europe project dAIEdge demonstrates the use of these explainable models for ‘on-the-edge’ inference.
How to cite: Dumitru, C. O., Karmakar, C., and Wiehle, S.: Explainable Expert-in-the-loop sea-ice classification with statistical models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16248, https://doi.org/10.5194/egusphere-egu26-16248, 2026.
