CR5.8 | Understanding cryospheric processes in the past, present and future using data assimilation and machine learning
EDI
Understanding cryospheric processes in the past, present and future using data assimilation and machine learning
Convener: Celia A. BaumhoerECSECS | Co-conveners: Amber Leeson, Désirée TreichlerECSECS, Jordi BolibarECSECS, Kristoffer AalstadECSECS, Michel Tsamados, Esteban Alonso-GonzálezECSECS
Orals
| Wed, 17 Apr, 16:15–18:00 (CEST)
 
Room L2
Posters on site
| Attendance Thu, 18 Apr, 10:45–12:30 (CEST) | Display Thu, 18 Apr, 08:30–12:30
 
Hall X4
Orals |
Wed, 16:15
Thu, 10:45
Process-level understanding of cryospheric processes is traditionally developed using field-based techniques and earth observation, and future predictions of the state of the cryosphere are usually made using process-based physical or empirical models. In recent years however, data science, machine learning and AI have emerged as a powerful set of tools in Earth system science to complement these traditional techniques. Moreover, the fusion of observations with (typically) mechanistic models through data assimilation (DA) and inverse modeling (IM) has emerged as a promising tool to help infer the state and predict the fate of the terrestrial cryosphere.
Examples of scientific advances achieved using these methods include, but are not limited to, robustly combining data streams into hybrid products, exploring and characterising uncertainty, using physics-informed machine learning methods to improve our understanding of key physical processes, exploiting deep learning to extract new information from earth observation data, implementations of DA/IM techniques for monitoring elements of the cryosphere, describing interactions between observations and process models and exploiting coupled data-process modelling to predict future changes with maximum fidelity.
Here we welcome contributions from anyone using this emerging technology to develop new insight into the physical processes, or produce new forecasts into the future behavior of, cryospheric systems such as ice sheets, ice shelves, sea ice and glaciers.

Orals: Wed, 17 Apr | Room L2

Chairpersons: Celia A. Baumhoer, Jordi Bolibar, Esteban Alonso-González
16:15–16:20
16:20–16:30
|
EGU24-18641
|
ECS
|
solicited
|
On-site presentation
|
Bradley Gay, Neal Pastick, Jennifer Watts, Amanda Armstrong, Kimberley Miner, and Charles Miller

Complex non-linear relationships exist between the permafrost thermal state, active layer thickness, and terrestrial carbon cycle dynamics In Arctic and boreal Alaska. The rate, magnitude, and extent of permafrost degradation remain uncertain, with an increasing recognition of the importance of abrupt thaw mechanisms. Similarly, large uncertainties in the rate, magnitude, timing, location, and composition of the permafrost carbon feedback complicate this issue. The challenge of monitoring sub-surface phenomena, such as the soil temperature and soil moisture profiles, with remote sensing technology further complicates the situation. There is an urgent need to understand how and to what extent permafrost degradation is destabilizing the Alaskan carbon balance and to characterize the feedbacks involved. We employ our artificial intelligence (AI)-driven model GeoCryoAI to quantify permafrost thaw dynamics and greenhouse gas emissions in Alaska. GeoCryoAI uses a hybridized multimodal deep learning architecture of stacked convolutionally layered memory-encoded bidirectional recurrent neural networks and 12.4 million parameters to simultaneously ingest and analyze 13.1 million in situ measurements (i.e., CALM, GTNP, ABoVE ReSALT, FLUXNET, NEON), 8.06 billion remote sensing airborne observations (i.e., UAVSAR, AVIRIS-NG), and 7.48 billion process-based modeling outputs (i.e., SIBBORK-TTE, TCFM-Arctic) with disparate spatiotemporal sampling and data densities. This framework introduces ecological memory components and effectively learns subtle spatiotemporal covariate complexities in high-latitude ecosystems by emulating permafrost degradation and carbon flux dynamics across Alaska with high precision and minimal loss (RMSE: 1.007cm, 0.694nmolCH4m-2s-1, 0.213µmolCO2m-2s-1). GeoCryoAI captures abrupt and persistent changes while providing a novel methodology for assimilating contemporaneous information on scales from individual sites to the pan-Arctic. Our approach overcomes traditional model inefficiencies and seamlessly resolves spatiotemporal disparities.

How to cite: Gay, B., Pastick, N., Watts, J., Armstrong, A., Miner, K., and Miller, C.: Forecasting Permafrost Carbon Dynamics in Alaska with GeoCryoAI, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-18641, https://doi.org/10.5194/egusphere-egu24-18641, 2024.

16:30–16:40
|
EGU24-12922
|
ECS
|
On-site presentation
Steven Wallace, Aiden Durrant, William D. Harcourt, and Georgios Leontidis

Climate change poses a significant global challenge, with its effects manifesting prominently through melting and retreating glaciers in the Arctic and Antarctic. Understanding the dynamics of glacier flow is imperative for predicting the future evolution of the Polar ice sheets. Crevasses play an important role in regulating ice flow by acting as a conduit for surface meltwater to reach the bed and speed up ice flow, as well as providing the line of weakness through which icebergs detach from tidewater glacier termini. Furthermore, this study delves into the potential of computer vision techniques that use deep learning, leveraging foundation models trained using self-supervised learning like the Segment Anything Model (SAM) and DINOv2 from Meta AI, to automate crevasse mapping on glacial surfaces. Manual mapping crevasses on any glacier is currently labour-intensive and time-consuming without automation. Therefore, automating the process will allow scientists to map crevasses automatically over time in the exact location and over larger areas. Notably, this research addresses the scarcity of image segmentation datasets specifically tailored for mapping crevasses in polar regions and explores alternative deep learning methodologies, such as domain adaptation and few-shot learning, to overcome data limitations. The evaluation of foundation models harnesses high-resolution satellite imagery sourced from open-source remote sensing satellites such as Sentinel-1 and Sentinel-2 provided by the European Space Agency (ESA). Using multiple high-resolution image data modalities (e.g. Synthetic Aperture Radar (SAR) and optical satellite images) will provide insights into how different image data types help deep learning models generalise to crevasse mapping segmentation applications. The study seeks to develop advanced technological solutions to automate the mapping of crevasses tens of metres in width in order to address the knowledge gap of the role that crevasses play in modulating ice flow, particularly in response to climate warming.

How to cite: Wallace, S., Durrant, A., Harcourt, W. D., and Leontidis, G.: Automated Crevasse Mapping Using Deep Learning Foundation Models to Analyse Climate Change and Glaciology , EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-12922, https://doi.org/10.5194/egusphere-egu24-12922, 2024.

16:40–16:50
|
EGU24-21497
|
ECS
|
Highlight
|
Virtual presentation
Kim Bente, Roman Marchant, and Fabio Ramos

High-fidelity maps of Antarctica's subglacial bed topography constitute a critical input into a range of cryospheric models. For instance, ice flow models, which inform high-stakes sea level rise projections, rely on truthful and sufficiently detailed Digital Elevation Models (DEMs) of the ice-bedrock interface. Adversely, data collection of bed elevation profiles is extremely challenging, requiring airborne geophysical surveys of the vast landscape. Subsequent processing, as well as interpolation of the limited measurements onto a regular grid, introduce additional layers of uncertainty.

While the prevalent continent-wide gridded data products for Antarctica's bed topography, like BedMachine or upcoming Bedmap3, which arise from laborious ongoing international collaborations, are limited to a spatial resolution of 500 meters, machine learning methods present new opportunities to go beyond existing resolutions. Super-resolution, a class of computer vision techniques, aims to increase the resolution of a given low-resolution image and thereby generate a high-resolution version of that image. With a grid of elevation values interpreted as a special case of a grid of pixel values (i.e. an image), super-resolution approaches can hence be applied to topography grids. 

While existing bed topography super-resolution approaches for Antarctica have been challenged by the lack of available gridded data at dense target resolutions, which are needed to train deep learning architectures, we propose a probabilistic approach based on Gaussian Processes (GPs), that generates more robust and uncertainty-aware high-resolution topographies without the need for gridded target resolution training data. In addition, our proposed method leverages abundant high-resolution ice surface data from satellites by transferring covariance patterns from the ice surface to the bed via a purpose-designed covariance function.

We evaluate our multimodal Bayesian fusion model in a controlled topography reconstruction experiment over mountainous regions of East Antarctica, where we assess various models' skills to reconstruct original 500 m BedMachine topography, given the respective artificially degraded 1000 m, 1500 m, 2000 m, 2500 m, and 3000 m low-resolution input grids. As metrics we use root mean square error (RMSE), peak signal-to-noise ratio (PSNR), and the structural similarity index measure (SSIM) between reconstructions and withheld ground truth topographies. We compare our model against bilinear and bicubic interpolation baselines, DeepBedMap DEM and its multi-branch extension, MB_DeepBedMap DEM, as well as the Hybrid Attention Transformer (HAT), a pretrained state-of-the-art single image super-resolution model, for which we explore various fine-tuning strategies. Our results highlight the utility of our proposed uncertainty-aware and interpretable fusion model for the data-constrained endeavour of mapping Antarctica's subglacial bed topography at high resolutions.

How to cite: Bente, K., Marchant, R., and Ramos, F.: Transfer learning for Antarctic bed topography super-resolution, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-21497, https://doi.org/10.5194/egusphere-egu24-21497, 2024.

Glaciers & Ice Sheets
16:50–17:00
|
EGU24-701
|
ECS
|
On-site presentation
Codrut-Andrei Diaconu, Jonathan L. Bamber, Fabien Maussion, and Harry Zekollari

Machine learning plays an increasingly important role in modelling and better quantifying observed changes in various subcomponents of the Earth system. For instance, in the field of glaciology, machine learning methods have the potential to help unravel the ever-growing datasets on observed glacier changes, allowing for a better understanding of these changes and their driving factors.

In this study, we investigate the benefit of using a non-linear machine learning framework to model the observed recent glacier changes (individual glaciers’ geodetic mass balance over the 2000-2019 period) for nearly all the land-terminating glaciers larger than 2 km^2. To this end, we build a Random Forest model driven by a set of predictors, composed of both topographic (e.g. area, slope, debris coverage) and climatic features (e.g. temperature and precipitation anomalies), which explain up to 70% of the global variance in the observational dataset. Generally, we find that the climatic features are more important, explaining alone approx. 55% of the variance, as compared to the approx. 40% obtained with the topographical ones alone. We further investigate the importance of the topographical predictors within subregions that are assumed to be climatically homogeneous, showing different behaviours across them.

Our study illustrates the benefit of using non-linear models when statistically modelling multi-decadal geodetic mass balances, providing further insights into the drivers of current glacier changes. The proposed framework also has the potential to be used as a gap-filling tool to estimate the geodetic mass balance of unmeasured glaciers or those with uncertain geodetic mass balance observations and to predict future mass balance when forced with CMIP6 climate data or similar Earth System Model output.

How to cite: Diaconu, C.-A., Bamber, J. L., Maussion, F., and Zekollari, H.: Investigating the drivers of global glacier volume changes over the last two decades using machine learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-701, https://doi.org/10.5194/egusphere-egu24-701, 2024.

17:00–17:10
|
EGU24-13710
|
On-site presentation
Niccolò Maffezzoli, Gianluca Lagnese, Sebastiano Vascon, Troels Petersen, Carlo Barbante, and Eric Rignot

Estimating the ice volume of Earth's glaciers is a key challenge in Earth System science, crucial for understanding their evolution, quantifying future global sea level rise and freshwater resources in climate-sensitive regions. Given current global warming that causes glaciers mass loss, precise ice volume estimates become a top priority in face of future climate scenarios. 

Here we present the SKYNET project, which aims to develop a general modeling framework for estimating ice volumes of Earth’s glaciers using generative deep learning. 

The modeling framework comprises two main networks. The first network, a generative adversarial network, reconstructs glacier bedrocks using elevation maps of surrounding ice-denuded regions. Trained on over 1 million Digital Elevation Maps, the model learns key geometrical features and patterns of Earth’s mountain regions. The challenge is addressed as an image inpainting problem, in which the objective is to reconstruct the bedrock altitude in a missing portion of the image, using surrounding information. 

The second network leverages existing ice thickness measurements. Such a network is trained on local features such as ice velocity, slope, distance from glacier boundaries, and other glacier statistics to predict local ice thickness. We use the Glacier Ice Thickness Dataset (GlaThiDa) and other input products as training dataset. We employ a graph neural network (GNN) with an architecture that explicitly accounts for the connectivity of the data. The GNN’s local ice thickness estimates serve as a prior to refine the inpainting network’s generated ice thickness maps. 

We introduce the model, discuss its concept, advantages, limitations, current challenges and present preliminary results and tests on diverse continental glaciers across the globe. 

How to cite: Maffezzoli, N., Lagnese, G., Vascon, S., Petersen, T., Barbante, C., and Rignot, E.: A glacier ice volume modeling framework based on generative adversarial networks and graph neural networks, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-13710, https://doi.org/10.5194/egusphere-egu24-13710, 2024.

17:10–17:20
|
EGU24-16820
|
ECS
|
solicited
|
On-site presentation
Fiona Turner, Jonathan Rougier, Tamsin Edwards, Violaine Coulon, and Ann Kristin Klose

In order to predict future global sea level rise, it is key for us to have better understanding of the changes in the cryosphere, as is being done in the PROTECT project (https://protect-slr.eu). Large uncertainties exist around how these changes will present over the coming centuries, with the Antarctic ice sheet being the most uncertain component with regards to predicted mass changes. It is therefore necessary to turn to statistical techniques to create more robust predictions.

Here, we present results from a random forests emulator simultaneously trained on two ice sheet models, Kori and PISM, forced by four global climate models. We emulate the relationship between inputs, namely climate change and ice sheet model settings, and an output, sea level contribution. The use of random forests allows us to improve on previous Gaussian Process emulators (Edwards et al., 2021) in speed and the treatment of factor inputs. We also transform the multi-centennial output in order to allow us to model the whole time series, rather than each year individually. The emulator allows us to interpolate (and extrapolate slightly) in order to build probabilistic projections of sea level contribution to 2300 that include climate and ice sheet modelling uncertainties under all five Shared Socioeconomic Pathways (SSPs), despite only two being used in the ensemble of simulations.

References

Edwards, T. L., Nowicki, S., Marzeion, B., Hock, R., Goelzer, H., Seroussi, H., Jourdain, N. C., Slater, D. A., Turner, F. E., Smith, C. J., et al. (2021). Projected land ice contributions to twenty-first-century sea level rise. Nature, 593(7857):74–82.

How to cite: Turner, F., Rougier, J., Edwards, T., Coulon, V., and Klose, A. K.: Building probabilistic projections of the Antarctic contribution to global sea level rise using a random forests emulato, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-16820, https://doi.org/10.5194/egusphere-egu24-16820, 2024.

17:20–17:30
|
EGU24-2444
|
ECS
|
Highlight
|
On-site presentation
Kejdi Lleshi, Guillaume Jouvet, Frédéric Herman, and Samuel Cook

Understanding past natural climate variations during the Quaternary period is crucial for understanding the ongoing climate change. Glaciation events during the Quaternary have left visible footprints in today's landscape, such as moraines and trimlines, that could be used to reconstruct paleo glacier extent. 

Reconstructed glacier extent offers great potential to retrieve paleo-climate information during the coldest episodes of the Quaternary by inverting a glacier evolution model. However, current inversion methods are computationally expensive and their forward model relies on simplified physics. Fundamentally, they all assume glaciers are in a stationary state, which is simplistic and fails to capture essential transient features linking climate to glacier response.

Here, we develop a new Machine-Learning (ML)-based inversion technique that overcomes the previously-mentioned limitations to reconstruct the glacier equilibrium line altitude (ELA), a proxy for temperature and precipitation, during a glacial maximum from reconstructed glacier extent. Our forward model consists of  a deep-learning emulator that learns the physical processes of a glacier from climate forcing to the glacier response. This approach has the advantage of being computationally highly-efficient, as well as allowing for automatic (thanks to the automatic differentiation) inversion of reconstructed glacier extent to retrieve a realistic ELA field that informs us about paleoclimates.

When applying our method to the Last Glacial Maximum in the European Alps, our reconstructed ELA fields show  a clear separation between the northern and southern Alps, with northern ELAs being considerably lower as shown in Fig 1. Our results are supported by the glacier footprints reconstructed from geomorphological observations in the northern Alps, which suggest the presence of large glacier lobes. In contrast, the glacial lobes in the southern Alps were noticeably smaller.

Our method is applicable in any formerly glaciated areas, and therefore has a high potential for paleoclimate reconstruction of the Earth’s coldest episodes.

Figure1. The resulting ELA field from inverting the ‘observed’ glacier footprint. There is a visible difference in the climate between the north and the south.

How to cite: Lleshi, K., Jouvet, G., Herman, F., and Cook, S.: Retrieving climatic and temporal information from the last glacial maximum using an invert glacier model, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-2444, https://doi.org/10.5194/egusphere-egu24-2444, 2024.

Snow
17:30–17:40
|
EGU24-20329
|
On-site presentation
Claudia Notarnicola

Remote sensing imagery offers a unique tool to monitor the snow cover evolution both at global and regional scales. However, the availability of the time series of normally 3-4 decades limits the applicability of the data to understand the dynamics of snow processes at longer time scale (Pulliainen et al., 2020). In this perspective the possibility to generate hybrid dataset merging satellite records with model simulations is offering the chance to both cross-validate the model simulations and to extent the time series (Lenton et al 2024).

In this work, an approach proposed by Notarnicola 2022, with the aim to merge earth observation imagery and modelled data was adapted to generate longer time series. The method is a machine learning approach based on Artificial Neural Network (ANN), where the uncertainties on the ANN trained model were obtained through a bootstrap procedure with a resampling technique. As modelled data, the NOAA-CIRES-DOE 20th Century Reanalysis V3 contains land surface and meteorological maps and their uncertainty available in the period 1806-2015 was addressed (Slivinski et al., 2019). For satellite data, the longest time series on snow parameters were used derived from ESA Climate Change Initiative (CCI) Snow project, namely Snow Water Equivalent (SWE) data available from 1979 to 2020 (https://climate.esa.int/en/odp/#/project/snow). As predictors for the ANN training and test, dynamics variables such as air temperature and precipitation were considered as they regulate the snow dynamics and rainfall events. As static parameters, the location in terms of latitude and longitude, mean elevation, the land cover type, and percentage of vegetation cover were inserted. The target variable to be improved is the snow water equivalent value at pixel level. 

Before the application of the ANN approach, a comparison between the SWE values in two data sets for the overlapping period (1979-2015) indicated an averaged correlation coefficient of around 0.52, a bias in the range 35-42 mm, the latter one being strongly depending on the different months. After the application of the model derived by the trained ANN applied to the test dataset, the correlation coefficients raised on average to 0.87, and the RMSE is reduced to around 20 mm. The merged data set will be further compared with ground data where available and then used to derive long-term trends for the SWE variable.

 

References

Lenton, T.M., Abrams, J.F., Bartsch, A. et al. Remotely sensing potential climate change tipping points across scales. Nat Commun 15, 343 (2024). https://doi.org/10.1038/s41467-023-44609-w

Notarnicola, C. Overall negative trends for snow cover extent and duration in global mountain regions over 1982–2020. Sci Rep 12, 13731 (2022). https://doi.org/10.1038/s41598-022-16743-w

Pulliainen, J. et al. Patterns and trends of Northern Hemisphere snow mass from 1980 to 2018. Nature, 582, E18, doi:10.1038/s41586-020-2416-4 (2020).

Slivinski L.C., Compo G.P., Whitaker J.S., et al. Towards a more reliable historical reanalysis: Improvements for version 3 of the Twentieth Century Reanalysis system. Q J R Meteorol Soc. 2019; 145: 2876–2908.https://doi.org/10.1002/qj.3598

How to cite: Notarnicola, C.: Merging Earth Observation imagery and model simulations to monitor long-term global snow cover dynamics, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-20329, https://doi.org/10.5194/egusphere-egu24-20329, 2024.

17:40–17:50
|
EGU24-13035
|
ECS
|
On-site presentation
David R. Casson, Andrew W. Wood, Guoqiang Tang, Karl Rittger, and Martyn Clark

This study investigates probabilistic estimates of fractional Snow Cover Area (fSCA) in mountainous terrain, aiming to bridge the gap between mechanistic hydrological models and operational remote sensing measurements. To capture the spatial variability of snow cover, we generate high-resolution ensemble meteorological forcing datasets from in-situ measurements, employing locally weighted regression and random forest methods. We then discretize a physically-based hydrological model tailored for mountainous terrain, incorporating the dominant factors influencing snow cover. Subsequently, fluxes from an intermediate complexity snowpack model are utilized to simulate fSCA, which we evaluate against an operational data source. This research is a progressive step toward integrating ensemble data assimilation techniques, with the goal of improving hydrological forecast performance.

How to cite: Casson, D. R., Wood, A. W., Tang, G., Rittger, K., and Clark, M.: Probabilistic Estimates of Fractional Snow Cover Area in Mountainous Terrain, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-13035, https://doi.org/10.5194/egusphere-egu24-13035, 2024.

17:50–18:00
|
EGU24-13310
|
ECS
|
On-site presentation
Ross T. Palomaki, Karl Rittger, Sebastien J. P. Lenard, Jeff Dozier, and S. McKenzie Skiles

Snow albedo data are required for various research and applications at a wide range of spatial and temporal scales. Typically, spatially-distributed snow albedo measurements are generated using multispectral satellite data, including MODIS, Sentinel-2, and Landsat imagery. While a number of algorithms can be employed to create snow albedo products from multispectral satellite imagery, a recent MODIS-focused analysis shows that spectrally-based approaches result in the most accurate snow albedo. These approaches use spectral libraries of snow, vegetation, and rock reflectance to solve for snow fraction, grain size, and the impact of light absorbing particles (LAP) on snow albedo; snow albedo is estimated by combining the grain size with darkening due to LAP.

Spectral unmixing algorithms produce more accurate snow albedo measurements when applied to hyperspectral data because the additional spectral information removes ambiguities associated with sparser multispectral imagery. Various airborne sensors and satellite missions EnMAP, EMIT, and PRISMA provide hyperspectral data with spatial resolutions on the order of tens of meters, but depending on the platform have repeat periods between 8-29 days, and may miss important albedo changes related to early season snow accumulation and late season dust events.

In this presentation, we show initial results from a data fusion approach to produce daily snow albedo data at high spatial resolutions using multispectral and hyperspectral imagery. Our model fuses snow albedo measurements directly instead of reflectance data to take advantage of the improved ability of the spectral unmixing algorithm to address mixed pixels and better discern clouds from snow. To demonstrate our approach, we train a random forest model on snow albedo measurements generated from airborne hyperspectral data at 50 m resolution. Predictor variables include daily, 463 m MODIS snow albedo generated using a spectral unmixing algorithm, as well as terrain characteristics and solar illumination. The fused snow albedo data take advantage of the more accurate and finer resolution hyperspectral data will maintaining the daily temporal resolution of multispectral MODIS imagery. Additionally, our fusion approach is flexible and can incorporate snow albedo measurements from additional airborne or satellite sensors, including multispectral VIIRS data and hyperspectral data from the upcoming SBG and CHIME satellite missions.

How to cite: Palomaki, R. T., Rittger, K., Lenard, S. J. P., Dozier, J., and Skiles, S. M.: A data fusion approach to produce daily, high resolution snow albedo using multispectral and hyperspectral imagery, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-13310, https://doi.org/10.5194/egusphere-egu24-13310, 2024.

Posters on site: Thu, 18 Apr, 10:45–12:30 | Hall X4

Display time: Thu, 18 Apr 08:30–Thu, 18 Apr 12:30
Chairpersons: Désirée Treichler, Michel Tsamados, Kristoffer Aalstad
Snow
X4.16
|
EGU24-19594
|
ECS
|
Highlight
Snowpack Stability Prediction with Machine Learning
(withdrawn)
Julia Kaltenborn, Tiziano Di Pietro, David Rolnick, and Martin Schneebeli
X4.17
|
EGU24-7804
|
ECS
Laura Sourp, Simon Gascoin, Vanessa Pedinotti, Lionel Jarlan, and Esteban Alonso-González

Despite its significance, the snow water equivalent (SWE) is poorly characterized in many mountain regions due to i) a lack of in situ measurements and ii) the difficulty to measure the SWE directly from satellite observations. 

We developed a tool to simulate the spatial distribution of SWE at high resolution (typically 100 m) in any region of interest using SnowModel (Liston and Elder, 2006a, 2006b), global meteorological data (ERA5) and satellite observations of the snow cover fraction (SCF). Satellite observations are used to mitigate errors in the model parameterization and the meteorological forcings using the particle batch smoother  (Margulis et al., 2015). This method consists in computing N perturbed simulations on the whole hydrological season and transforming the simulated SWE in a simulated SCF. In this study, we used the formula linking the SWE and the SCF of the Noah Land Surface Model as the measurement operator. After that, the simulations are compared to remotely sensed SCF data and weighted according to their agreements with the observations. 

We implemented this data assimilation method with both MODIS and Sentinel-2 SCF data. We are testing it on the Bassies catchment in the French Pyrenees, and on the Tuolumne River Catchment in the Sierra Nevada, USA. In the Bassies catchment, we compare the results with snow depth maps from Pléiade satellites. We perturbed the precipitations with a log-normal law and found a very good agreement between the posterior simulated SCF and the observed one in Bassies as expected. However, the simulated snow depths in this catchment do not match the Pléiades snow depths observations. We added the perturbation of the air temperature with a normal law and found similar results. We also found a very small sensitivity of the posterior snow depth on the empirical parameters of the measurement operator. These results suggest that the SWE-SCF relationship may not be sufficiently informative in the study site of Bassiès.  We will present the extension of this work to the Tuolumne river basin where the Airborne Observatory provides SWE maps over a larger region.

How to cite: Sourp, L., Gascoin, S., Pedinotti, V., Jarlan, L., and Alonso-González, E.: Evaluation of a tool to simulate high resolution mountain SWE from global datasets, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-7804, https://doi.org/10.5194/egusphere-egu24-7804, 2024.

X4.18
|
EGU24-20649
|
ECS
Pau Wiersma and Grégoire Mariéthoz

Mapping the dynamics of snow water equivalent (SWE) is critical for understanding the hydrology of mountain regions. While methods to reconstruct SWE exist, they usually rely on either remote sensing data or the presence of in-situ observations. Streamflow observations can be considered an indirect and delayed observation of catchment-wide SWE, but despite their broad spatial and temporal availability, their potential for SWE reconstructions has not been explored so far.

In this study, we investigate how much SWE information can be extracted from the streamflow record, both in terms of its total mass as well as its spatial distribution. To this end, we set up an inverse streamflow-based SWE reconstruction framework that can operate without remote sensing data or in-situ snow observations. As a basis, we use a distributed hydrological model with a temperature-index snow model at 1km resolution, which generates SWE reconstructions from a set of prior snow and climate parameters and translates them into streamflow. By using the streamflow observations to optimize these parameters, we obtain SWE reconstructions that better match the streamflow.

However, there is likely a multitude of SWE reconstructions that all lead to the same streamflow, which is defined as an ill-posed inverse problem. In order to find out by how much the streamflow can constrain the prior SWE reconstructions, we perform an experiment with synthetic observations. Using synthetic observations instead of real observations eliminates both model and observation errors, allowing us to focus solely on the nature of this ill-posed problem.

Firstly, we select a set of soil, snow and climate parameters to generate synthetic SWE observations and their corresponding streamflow. All parameters are kept constant over the entire period except the parameter controlling the snowfall bias correction, which is set to fluctuate on a yearly basis and consequently also needs to be optimized for each year separately. Then, we run a single calibration to attempt to rederive these parameters and reconstruct the synthetic observations. Finally, we analyze our posterior ensemble of parameter sets and SWE reconstructions and quantify the resemblance to the synthetic streamflow and SWE observations. We expect to find a near-perfect match with the synthetic streamflow observations, a strong constraint of the total catchment-wide SWE but a weak constraint of the spatial distribution of this SWE.

How to cite: Wiersma, P. and Mariéthoz, G.: Inverse hydrological modeling to infer historical snow mass based on streamflow records, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-20649, https://doi.org/10.5194/egusphere-egu24-20649, 2024.

X4.19
|
EGU24-1069
|
ECS
Cryosphere, Causality, and Couplings of Ocean and Atmosphere: The Secrets of Himalayan Snow
(withdrawn after no-show)
Shairik Sengupta and Rajarshi Das Bhowmik
Glaciers & Ice Sheets
X4.20
|
EGU24-4036
|
ECS
Benjamin Wallis, Anna Hogg, and David Hogg

The subglacial discharge of sediment-rich meltwater plumes can be detected in satellite imagery where plumes reach the ocean surface at the terminus of tidewater glaciers. These meltwater plumes can influence glacier melting and ocean properties in important ways. Studying them provides insights into meltwater pathways in glacial hydrological systems.

Sediment-rich meltwater plumes are observed extensively in Greenland, Svalbard and other regions, however in Antarctica observations have been more limited. Recent detections of seasonal speed variations on tidewater glaciers of the Antarctic Peninsula suggest that surface meltwater reaching the glacier bed may be an important factor influencing ice dynamic behaviour on the Antarctic Peninsula Ice Sheet (APIS), however surface-bed hydrological connections have not been directly observed through fieldwork on the mainland APIS. Therefore, studying the presence and distribution of sediment plumes around the Peninsula can provide insights to understand the factors influencing newly observed seasonal ice speed fluctuations.

Here we develop a remote-sensing approach to map the locations and frequency of sediment plumes on the Antarctic Peninsula coastline using high-resolution multi-spectral imaging from Sentinel-2 satellites and a U-Net based convolutional neural network. This methodology allows us to detect small sediment plumes in images with high cloud and sea-ice densities. We apply our approach to the Antarctic Peninsula north of 65°S, including the South Shetland Islands, to produce a time-series of sediment plumes from 2016 to 2023 covering 150,000 km2.

We use these results combined with outputs from regional climate models and reanalysis to assess the link between surface-visible sediment plumes and surface melt and runoff from the Antarctic Peninsula’s glaciers. We find that the timings and locations of sediment plumes correspond to modelled runoff, providing evidence for widespread surface-bed hydrological connections in the Antarctic Peninsula.

How to cite: Wallis, B., Hogg, A., and Hogg, D.: Mapping glacial sediment plumes in the Antarctic Peninsula using deep learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-4036, https://doi.org/10.5194/egusphere-egu24-4036, 2024.

X4.21
|
EGU24-2495
|
ECS
|
Highlight
Gong Cheng, Mathieu Morlighem, and Sade Francis

Predicting the future contribution of the ice sheets to sea level presents several challenges due to the lack of observations of critical boundary conditions, such as basal sliding. Traditional numerical models often rely on data assimilation methods to determine spatially variable friction coefficients by solving an inverse problem, given an empirical friction law. However, these approaches are not versatile, as they demand sometimes extensive code development efforts when integrating new physics into the model. Furthermore, the requirement for comprehensive data alignment on the computational grid hampers their adaptability, especially in handling sparse data effectively. To tackle these challenges, we propose a transformative approach utilizing Physics-Informed Neural Networks (PINNs) to seamlessly integrate observational data and underlying physical laws into a unified loss function, facilitating the solution of both forward and inverse problems within the same framework. We illustrate the versatility of PINNs by applying the framework to two-dimensional problems on Helheim Glacier in southeast Greenland. By systematically concealing one component within the system, we showcase the ability of PINNs to accurately predict and reconstruct hidden information, emphasizing their adaptability to handle scenarios marked by missing or incomplete datasets. Furthermore, we extend the application to address a challenging mixed inversion problem. We show how PINNs are capable of inferring the basal friction coefficient while simultaneously filling gaps in sparsely observed ice thickness. This mixed inversion problem represents a class of scenarios beyond the reach of conventional numerical methods. Our unified framework offers a promising avenue to enhance the predictive capabilities of ice sheet models, reducing uncertainties and advancing our understanding of intricate ice dynamics.

How to cite: Cheng, G., Morlighem, M., and Francis, S.: A Unified Framework for Forward and Inverse Modeling of Ice Sheet Flow using Physics Informed Neural Networks, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-2495, https://doi.org/10.5194/egusphere-egu24-2495, 2024.

X4.22
|
EGU24-9687
Philipp Immanuel Voigt and Andreas Born

The thickness of englacial isochrones is the combined product of accumulation and dynamical thinning by the flow of ice.Dated ice stratigraphy data provides access to this archive , which in principle holds the potential for improving simulations and reconstructions  of the Greenland ice sheet. However, the combined effects of accumulation and dynamical thinning are convoluted and spatially heterogeneous, making it difficult to separate their respective contributions to the observed thickness of isochrones.

Here we simulate isochrones explicitly using the Englacial Layer Simulation Architecture (ELSA) coupled with a fully transient thermomechanical ice sheet model. This enables us to link accumulation rates to the ice stratigraphy, including a physically consistent representation of dynamical thinning. By ensemble data assimilation techniques, the linear model response to changes in past local accumulation are estimated, enabling the inversion of the model and reconstruction of accumulation rates from stratigraphy. An iterative approach is chosen to account for the nonlinear response of ice flow to anomalous accumulation. The result is an optimized simulation of the Greenland ice sheet with an englacial stratigraphy matching the observations, forced by the reconstructed accumulation. Because the model is fully transient including ice dynamics, our approach also constrains ice sheet stability and sea level contribution.

Here we present preliminary findings from inversions of idealized ice sheet stratigraphy, including some encouraging insights, physical and methodological limitations and challenges yet to be overcome.

How to cite: Voigt, P. I. and Born, A.: Assimilating the Greenland ice sheet stratigraphy for reconstructing accumulation rates, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-9687, https://doi.org/10.5194/egusphere-egu24-9687, 2024.

X4.23
|
EGU24-15332
|
ECS
Wenxue Cao, Louise Steffensen Schmidt, Kristoffer Aalstad, Sebastian Westermann, and Thomas V Schuler

The accurate quantification of glacier mass balance is of vital importance for the evaluation of climate change impact and the management of hydrological resources. However, traditional modeling methodologies on a regional scale are frequently plagued by uncertainties in forcing data, model structure, and parameters. Data assimilation emerges as an effective technique to incorporate observations into modeling, thereby reducing the uncertainty of results. In this study, we evaluate the performance of different ensemble-based schemes, including the Ensemble Smoother (ES) and the Ensemble Smoother-Multiple Data Assimilation (ES-MDA), to incorporate albedo derived from MODIS satellite observations and in-situ mass-balance measurements vis stakes into the full energy balance model CryoGrid applied to Svalbard glaciers. Our primary aim is to enhance the accuracy of both the reconstruction and prediction of glacier mass balance in the Svalbard region through the synergistic use of observational data and model. In a range of experiments, we analyze the performance of different assimilation methods and different observation products. The implementation of ES-MDA has demonstrated marked improvements, while the variations in parameter dynamics have varied effects on the results. We compare the prior and posterior states to help disentangle which process or forcing has the most impact on the uncertainty of the model’s results.

How to cite: Cao, W., Schmidt, L. S., Aalstad, K., Westermann, S., and Schuler, T. V.: Data assimilation in glacier mass balance modeling, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-15332, https://doi.org/10.5194/egusphere-egu24-15332, 2024.

X4.24
|
EGU24-2110
|
ECS
Marijn van der Meer, Harry Zekollari, Matthias Huss, Jordi Bolibar, Kamilla Hauknes Sjursen, and Daniel Farinotti

Estimations of glacier mass balance are commonly made using field techniques, empirical or physically based models, and remote sensing. More recently, data-driven tools like machine learning have become powerful complements to these conventional techniques. This study explores the potential of using machine learning to simulate the individual point mass balance of 30 sites from 13 glaciers in Switzerland spanning over 60 years, sourced from the Glacier Monitoring Switzerland (GLAMOS) network. To this end, we use two machine-learning models: LASSO regression, a linear regression model with L1-regularisation, and eXtreme Gradient Boosting (XGBoost), a gradient-boosted ensemble of decision trees. The models are driven by temperature and precipitation data at 1 km grid resolution from the Federal Office of Meteorology and Climatology (MeteoSwiss). The seasonal point mass balance data are used to train and test the models for each site individually. A comparative analysis is performed in which the performance of the LASSO regression and XGBoost are compared to a standard approach of calculating mass balance from a temperature-index model. In this analysis, we also explore how different temporal frequencies of climate variables, ranging from annual to monthly, affect the performance of the machine learning methods. Beyond their computational efficiency, these machine learning models are particularly suited to provide valuable insights into feature importance. Harnessing this, we study which months’ temperature and precipitation most significantly contribute to explaining individual stake mass balances and compare these findings with commonly assumed drivers of mass balance.

How to cite: van der Meer, M., Zekollari, H., Huss, M., Bolibar, J., Hauknes Sjursen, K., and Farinotti, D.: Glacier point mass balance modeling using machine learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-2110, https://doi.org/10.5194/egusphere-egu24-2110, 2024.

Sea Ice
X4.25
|
EGU24-9175
|
ECS
Weibin Chen, Michel Tsamados, Rosemary Willatt, David Brockley, Marc Deisenroth, Claude De Rijke-Thomas, Alistair Francis, Len Hirata, Thomas Johnson, Isobel Lawrence, Jack Landy, Sanggyun Lee, Wenxuan Liu, Dorsa Nasrollahi Shirazi, Connor Nelson, Julienne Stroeve, and So Takao

In our research, we leverage the capabilities of the Sentinel-3A and Sentinel-3B satellites, launched in February 2016 and April 2018, respectively, to deepen our understanding of the polar regions. These satellites offer a unique blend of high-resolution Ku-band radar altimetry data, synthetic aperture radar (SAR) mode altimetry, and the Ocean and Land Colour Instrument (OLCI) imaging spectrometer. This combination enables the acquisition of both optical imagery and SAR radar altimetry data, extending up to 81 degrees North. Central to our study is the application of deep learning techniques, specifically the Vision Transformers (ViT), which adapt the Transformer algorithm for surface classification in polar environments. This approach is instrumental in distinguishing between sea ice and leads, demonstrating robust performance across various metrics, including accuracy and model roll-out on comprehensive OLCI image datasets. We produce our first lead classification maps at the original OLCI swath level resolution of 300m and a lead fraction prototype mosaic spring pan-Arctic product at gridded level of 1km, 5km and 10km resolution and on daily, weekly and monthly timescales. The use of binned statistics in conjunction with our deep learning classifications provides valuable insights into the spatial distribution and changes of leads within the polar ice. We compare our prototype product with other existing lead products and with auxiliary datasets on thin ice (roughness, thickness). Our work combining different satellite products at pan-Arctic intermediate resolution enhances our capacity to estimate sea ice thickness and aids in forecasting future changes in the Arctic and Antarctic regions, thereby contributing to the field of polar remote sensing with direct applications to the future polar missions CRISTAL and CMIR.

How to cite: Chen, W., Tsamados, M., Willatt, R., Brockley, D., Deisenroth, M., De Rijke-Thomas, C., Francis, A., Hirata, L., Johnson, T., Lawrence, I., Landy, J., Lee, S., Liu, W., Nasrollahi Shirazi, D., Nelson, C., Stroeve, J., and Takao, S.: Co-located OLCI optical imagery and SAR altimetry from Sentinel-3 for enhanced surface classification in sea ice, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-9175, https://doi.org/10.5194/egusphere-egu24-9175, 2024.

X4.26
|
EGU24-22394
ARISGAN: Extreme Super-Resolution of Arctic Surface Imagery using Generative Adversarial Networks
(withdrawn)
Michel Tsamados, christian Au-Boehm, Petru Manescu, and So Takao
X4.27
|
EGU24-10231
|
ECS
Luca Bianchi, Matteo Sangiorgio, Stefano Materia, Doroteaciro Iovino, and Andrea Castelletti

Modelling Arctic sea ice dynamics has proven to be a successful application for machine learning, leveraging its ability to generate accurate and computationally efficient forecasts. Nevertheless, prevailing limitations lie in the need for physical interpretability and the inability to unveil the dynamics and interdependencies between relevant ice-related variables and their drivers. In this study, we provide a two-step framework designed to combine the high accuracy and computational efficiency characteristics of machine learning while ensuring high interpretability.

The first step of our framework entails time series clustering to identify subregions that are homogeneous with respect to the spatiotemporal variability in the considered variable and obtain the barycentric time series of each cluster. We then use an advanced feature selection algorithm, the Wrapper for Quasi Equally Informative Subset Selection, that identifies neural predictors, specifically Extreme Learning Machines, to forecast the future evolution of sea ice. It then provides the most relevant set of inputs necessary for accurately describing the evolution for each cluster.

Our investigation focuses on the monthly evolution of sea ice thickness and uses data from the Pan-Arctic Ice-Ocean Modeling and Assimilation System (PIOMAS). Other PIOMAS variables (i.e., sea ice concentration, snow depth, sea surface temperature, and sea surface salinity) as well as observed discharge from five major Arctic rivers (i.e., Ob, Yenisey, and Lena in Asia, Mackenzie and Yukon in North America, provided by the Arctic Great Rivers Observatory discharge dataset) are considered as potential driving factors. 

Our results indicate the pivotal role of past sea ice thickness values, since the forthcoming state of sea ice seems to be influenced by both the current situation and historical trends and periodicity. Sea surface salinity in the open Arctic Ocean is highly persistent, and therefore is not used by the algorithm to explain the sea ice evolution. On the other hand, the Arctic rivers’ flows are more representative of the processes occurring in the clusters along the coast. Finally, the interaction between sea surface temperature and snow depth controls the interplay between ice formation and melting, and therefore plays a significant role in shaping the sea ice evolution in the short term.

Our framework aims to advance our comprehension of the complex physical processes governing sea ice thickness evolution in the Arctic region. Moreover, its effectiveness in uncovering sea ice related processes is expected to further improve with the inclusion of additional input variables and, possibly, of a longer data record.

How to cite: Bianchi, L., Sangiorgio, M., Materia, S., Iovino, D., and Castelletti, A.: Advancing Physical Interpretability of Arctic Sea Ice Dynamics through Automatic Feature Selection, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10231, https://doi.org/10.5194/egusphere-egu24-10231, 2024.

X4.28
|
EGU24-13556
Yibin Ren and Xiaofeng Li

The Arctic sea ice has been retreating dramatically in recent years in summer and fall. The navigation season for open water vessels along the Northeast Passage has lengthened to sub-seasonal scales. Accurate perditions of Arctic sea ice in sub-seasonal scales are essential for planning shipping activities. The numerical model cannot achieve a high accuracy of daily sea ice predictions on a sub-seasonal scale. The advanced deep learning brings new solutions for the data-driven-based sea ice prediction.

This study proposed a transformer-based deep learning model to predict multiple sea ice parameters, including sea ice concentration (SIC), sea ice thickness (SIT), and sea ice drift (SID), in the Pan-Arctic in a sub-seasonal scale, 90 days’ lead. An encoder and decoder are constructed based on transformer modules to extract spatio-temporal dependencies from daily SIC, SIT, and SID sequences. The spatio-temporal dependencies at different scales are fused to form the final feature maps. Three SIC, SIT, and SID output modules are designed based on the final feature maps to output different parameters for the next 90 days. The satellite-observed sea ice data from the National Sea Ice Data Center (NSIDC) are employed to train the proposed model. We compared our model with anomaly persistence and the European Centre for Medium-Range Weather Forecasts (ECMWF) ensemble predictions to demonstrate the model’s prediction skill.

Further, based on the proposed model, we discuss the effects of typical thermal and dynamic factors on sub-seasonal scale daily sea ice prediction. The selected factors include surface air temperature (SAT), sea surface temperature (SST), surface solar radiation downwards (SSRD), and geopotential height. Finally, we conclude with some scientific guidelines for the sub-seasonal sea ice predictability of the Arctic. 

How to cite: Ren, Y. and Li, X.: Multi-task predictions of the Arctic sea ice by a transformer-based deep learning model, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-13556, https://doi.org/10.5194/egusphere-egu24-13556, 2024.

Supraglacial Lakes
X4.29
|
EGU24-16672
Anja Rösel, Niklas Neckel, and Vytautas Jancauscas

Melt ponds are pools of water that form during summer on the surface of the arctic ice. Due to the lower albedo, melt ponds absorb more solar radiation than surrounding ice and hence have higher temperature. This causes more water to melt, creating a feedback loop. This means that melt pond fraction in ice sheets is an important factor to consider in global climate and sea ice models. In situ measurements are difficult and expensive in terms of time and labor. Furthermore, these measurements can only cover limited areas. This makes using Earth Observation methods for this task particularly attractive.

Until today, there is no sophisticated global melt pond data set available:

Accurate methods may exist for determining melt ponds from Sentinel-2 data. The downside of using Sentinel-2 is that parts of the High Arctic are not covered by this mission.

MODIS data covers the whole globe at least once every three days, but the downside of it is that MODIS resolution is much coarser (250m  vs. 10m). Since melt ponds are in general much smaller than 250m, it means that accurately capturing melt pond fraction from these data is difficult.

We propose to address these issues by employing Deep Learning techniques. Namely, we use Sentinel-2 data to train a model to super-resolve MODIS images to higher resolution and to use all available MODIS bands and their surrounding pixels for information context when predicting melt pond and open water fractions.

In addition, a thorough uncertainty quantification (UQ) will be applied by using the UQ Toolbox.

How to cite: Rösel, A., Neckel, N., and Jancauscas, V.: Arctic Meltponds: Automated Detection Algorithm Using Enhanced Machine Learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-16672, https://doi.org/10.5194/egusphere-egu24-16672, 2024.

X4.30
|
EGU24-14894
|
ECS
Celia A. Baumhoer, Jonas Koehler, Stef Lhermitte, Bert Wouters, and Andreas J. Dietz

Monitoring the dynamics of Antarctic supraglacial lakes is of particular interest in the context of global warming. Supraglacial meltwater accumulation on ice sheets and ice shelves can be a major driver of accelerated ice discharge. This is caused through processes such as surface runoff leading to ice thinning, basal meltwater injection causing basal sliding, and hydrofracture triggering ice shelf collapse and subsequent glacier acceleration. In addition, an increased presence of supraglacial lakes around the Antarctic margin can trigger enhanced melting due to the low albedo of surface lakes, which leads to increased absorption of solar radiation. Hence, continuous monitoring of supraglacial lakes is crucial for improving our understanding of their seasonal variations in extent and their impacts on ice shelf stability and ice sheet surface mass balance. Initially, an automated supraglacial lake mapping approach was developed to create bi-weekly lake extent maps for six Antarctic ice shelves based on fused results from convolutional neural network predictions and a Random Forest (RF) classification trained on Sentinel-1/-2 data. However, regular large-scale monitoring beyond these six selected areas requires a model with higher spatio-temporal transferability and an efficient fully automated data processing workflow. We tested for a potential improvement by replacing the RF-based mapping with an attention-based U-Net, expanding the training and test sites on a total of 23 regions and switching the processing to a more powerful high-performance computing infrastructure. We will discuss how remote sensing-based mapping accuracies can be improved by extending the training/test dataset, selecting the right machine learning model and the choice of processing infrastructure. In the future, the automated processing workflow will provide a regularly updated dataset on supraglacial lake dynamics of 23 Antarctic ice shelves via a web service by exploiting the full archive of available Sentinel-1/-2 satellite imagery.

How to cite: Baumhoer, C. A., Koehler, J., Lhermitte, S., Wouters, B., and Dietz, A. J.: Towards monitoring supraglacial lake dynamics in Antarctica with convolutional neural networks, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-14894, https://doi.org/10.5194/egusphere-egu24-14894, 2024.