CR5.6 | Driving and assisting cryospheric models with observations and artificial intelligence techniques
EDI PICO
Driving and assisting cryospheric models with observations and artificial intelligence techniques
Convener: Irena Vankova | Co-conveners: Lilian SchusterECSECS, Francois Massonnet, Yibin RenECSECS, Elisa Mantelli, Olaf Eisen, Johannes SutterECSECS
PICO
| Thu, 18 Apr, 16:15–18:00 (CEST)
 
PICO spot 4
Thu, 16:15
This interdisciplinary session brings together modelers, observationalists, and data scientists to present results and exchange knowledge and experience in the use of data assimilation and artificial intelligence (AI) techniques in the cryospheric sciences. With advances in observation and computing power, massive data from satellite observations, reanalysis, and simulations have pushed the cryosphere community, historically limited by scarcity of observations, into the era of big data. A large potential for future developments lies at the intersection of observations, models, and AI with the aim to improve prognostic capabilities in space and time.
We invite contributions from a wide range of methodological and topical backgrounds that bring new insights into cryospheric science. The topics span permafrost, sea ice, snow, glaciers, and ice sheets, covering both state characterization, process understanding, and prediction aspects.
The methods include satellite observations and AI-based data products from remote sensing, deep-looking geophysical methods, data assimilation, inverse methods and advancements in numerical techniques, geostatistics, machine learning, AI-based parameter retrieval, AI small scale physics parameterizations coupled to numerical models to enhance cryosphere simulations, and more.

PICO: Thu, 18 Apr | PICO spot 4

Chairpersons: Irena Vankova, Yibin Ren, Olaf Eisen
16:15–16:20
Ice Sheet
16:20–16:22
|
PICO4.1
|
EGU24-17119
|
ECS
|
On-site presentation
Anne Hermann, Clara Henry, Guy Moss, and Reinhard Drews

Uncertainties in surface mass balance (SMB) have the potential to significantly impact modelled ice thickness and ice-flow dynamics. A common approach is to reconstruct the surface mass balance history from the internal ice stratigraphy as imaged by radar. Particularly for intermediate and deeper layers, this requires accounting for deformation by ice flow using ice-dynamic forward models in an inverse framework. Numerous approaches to do so exist, but many of them are tailored to specific stratigraphy datasets and do not include uncertainties.

Previous work used simulation-based inference (SBI) to infer the SMB rates of steady state ice shelves, solving velocities using the shallow shelf approximation [1]. The approach not only estimates the spatially varying SMB field but also their uncertainties. We adapt this framework to infer SMB on grounded ice, where flow is dominated by internal deformation rather than longitudinal stretching. The forward model used in this study calculates velocities by solving the shallow ice approximation. The resulting velocities are used as input to an isochronal tracer scheme which calculates the stratigraphy of the ice [2]. Initial testing of our method is conducted on an idealized ice sheet, namely the Vialov profile. In future work, we aim to infer for the SMB history along a radar transact of Derwael Ice Rise in East Antarctica, where internal stratigraphy data is available. Our method provides a new uncertainty aware approach to estimate the SMB field on grounded ice.

 

[1] Moss et al.: Simulation-Based Inference of Surface Accumulation and Basal Melt Rates of an Antarctic Ice Shelf from Isochronal Layers (2023).

[2] Born: Tracer transport in an isochronal ice-sheet model (2017).

How to cite: Hermann, A., Henry, C., Moss, G., and Drews, R.: Simulation-Based Inference of Surface Mass Balance of Antarctic Ice Sheets, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-17119, https://doi.org/10.5194/egusphere-egu24-17119, 2024.

16:22–16:24
|
PICO4.2
|
EGU24-6433
|
ECS
|
On-site presentation
Beatriz Recinos, Daniel Goldberg, and James R. Maddison

Ice streams in the Amundsen Sea Embayment (ASE) are some of most rapidly thinning in Antarctica, experiencing ocean-driven grounding-line retreat and increased mass loss. Ice streams in this sector are crucial for maintaining the stability of the West Antarctic ice sheet, which contains enough ice to raise global mean sea-level by 5.3 m. Ice sheet models are our main tool to make projections of ice sheet mass loss or volume above floatation (which is equivalent to sea level rise). Standard methodologies consist of constraining model parameter fields using satellite observations, and then simulating model response to climate forcing. However, to date there is no comprehensive assessment of how sensitive these projections are to different satellite products or to spatial variations in ice velocity observations, or changes in the spatial distribution of unknown model parameters.  

Mapping the sensitivity of ice streams to a given set of model inputs can be done using an ensemble of simulations and running the model workflow tens (or hundreds) of thousands of times. Automatic Differentiation (AD) and data assimilation, however, provides an efficient method to perform this analysis and get these sensitivity maps at the scale of the model’s mesh. Here we use the ice sheet model Fenics_ice and its AD capabilities to construct maps of the sensitivity of volume above floatation to changes in the calibrated unknown model parameters, as well as changes in ice velocity observations. Our preliminary results show that the sea level rise forecast is more sensitive to ice velocity changes at the grounding line and to changes to the basal drag coefficient at those same locations.  

How to cite: Recinos, B., Goldberg, D., and Maddison, J. R.: Assessing the Impact of Ice Velocity Observations on Ice Sheet Mass Loss Projections from the Amundsen Sector, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-6433, https://doi.org/10.5194/egusphere-egu24-6433, 2024.

16:24–16:26
|
PICO4.3
|
EGU24-8113
|
ECS
|
On-site presentation
Oskar Herrmann and Johannes Fürst

We are introducing a novel technique for calibrating the unknown parameters of a glacier model by integrating remote sensing data. Our approach involves the fusion of the Instructed Glacier Model (IGM) with an Ensemble Kalman Filter designed explicitly for transient data assimilation. Our primary objective is to assimilate remotely sensed observations at the respective time of acquisition during forward simulations. During the presentation, we explore the Ensemble Kalman Filter's core concept and showcase our approach's effectiveness through twin experiments, fine-tuning model parameters related to ice dynamics and surface mass balance. Utilizing observations from prominent glaciers in the European Alps, our methodology concurrently minimizes uncertainty estimates of crucial parameters such as equilibrium line altitude and ablation/accumulation gradients. The resulting uncertainty estimate is then integrated into future projections.

How to cite: Herrmann, O. and Fürst, J.: Increasing the information flow into glacier system models using an Ensemble Kalman Filter, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-8113, https://doi.org/10.5194/egusphere-egu24-8113, 2024.

16:26–16:28
|
PICO4.4
|
EGU24-19721
|
On-site presentation
|
Wojciech Milczarek, Anna Kopeć, Michał Tymplaski, and Marek Sompolski


Dynamic climate changes are especially noticeable in polar regions, where glaciers are retreating landward at an increased rate. This study emphasizes the application of the feature tracking method for analyzing the ice flow velocity across the entire Svalbard archipelago from 2015 to 2024. We conducted our analysis using the Geogrid and autoRIFT platforms, complemented by Sentinel-1 imagery. To validate our ice velocity products, we employed the Glacier Feature Tracking (GLAFT) aproach. Our analysis culminated in the development of an extensive glacier repository for the archipelago, which includes velocity data spanning the study period. These results facilitated the identification of glaciers demonstrating surging behavior. Additionally, the study introduces a novel approach to analyzing glacier flow velocities, focusing on their acceleration characteristics.

How to cite: Milczarek, W., Kopeć, A., Tymplaski, M., and Sompolski, M.: Quantifying of the ten-year variation in surface velocity of the Svalbard Archipelago glaciers using SAR data, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-19721, https://doi.org/10.5194/egusphere-egu24-19721, 2024.

16:28–16:30
|
PICO4.5
|
EGU24-5964
|
ECS
|
On-site presentation
Therese Rieckh, Andreas Born, Alexander Robinson, Robert Law, and Gerrit Gülle

We are presenting the ice-sheet layer-age tracer ELSA (Englacial Layer Simulation Architecture) — a model that uses a straightforward method to simulate the englacial stratification of large ice sheets — as an alternative to Eulerian or Lagrangian tracer schemes. ELSA’s vertical axis is time; individual layers of accumulation are modeled explicitly and are isochronal. 

ELSA is not a stand-alone ice-sheet model, but requires uni-directional coupling to another model providing ice physics and dynamics (the “host model”). Via ELSA’s layer tracing, the host model’s output can be evaluated throughout the ice sheet interior using ice core or radiostratigraphy data. 

We show results regarding the stability and resolution-dependence of this coupled modeling system using simulations of the last glacial cycle of the Greenland ice sheet with Yelmo as the host model. We present options for making ELSA computationally efficient enough for ensemble runs, as well as requirements for offline forcing of ELSA with output from a range of existing ice-sheet models. 

ELSA is an open source project and available on git (https://git.app.uib.no/melt-team-bergen/elsa) for the ice sheet modeling community.

How to cite: Rieckh, T., Born, A., Robinson, A., Law, R., and Gülle, G.: Introducing ELSA: An isochronal model for ice sheet layer-age tracing, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-5964, https://doi.org/10.5194/egusphere-egu24-5964, 2024.

16:30–16:32
|
PICO4.6
|
EGU24-11310
|
On-site presentation
Marcel Dreier, Nora Gourmelon, Moritz Koch, Thorsten Seehaus, Matthias Braun, Andreas Maier, and Vincent Christlein

Melting glaciers and ice sheets contribute significantly to the global sea level rise, posing an existential threat to coastal regions. For that reason, many expeditions collect radio echo sounding data to track, model, and predict the changes in ice mass and thickness. However, manually analyzing the data collected to determine the thickness of the ice sheet and mountain glaciers is a time-consuming task, as the collected data often extends over tens to hundreds of kilometers. Hence, the development and utilization of automated and semi-automated tools have risen in popularity to alleviate the problem. Especially machine learning-based approaches show promising results as they can capture complex relationships in the data. However, such methods require large labeled datasets to achieve their full potential. Thus, we are currently gathering and cleaning data that will be released as a benchmark dataset for radio echo sounding data to train, test, and compare different approaches. 

In detail, the dataset will consist of 2D radar data with georeferenced labels for the position of the air-ice and ice-bedrock interface. We plan to incorporate radar data from multiple sources since the underlying radar system and expedition area play a significant role in depicting the data and the general difficulty of the task. That way, the performance of a system can be fairly assessed with the benchmark dataset without the evaluation results being skewed by external factors. We will further ensure a fair comparison by dividing the data of every source in an independent training, validation, and test set. This division will also allow us to use the benchmark dataset to train and compare systems specialized on only a single data source. Thus, future work can also utilize our benchmark dataset to develop systems specialized on only a single data source. 
Lastly, we will release a baseline model alongside the benchmark dataset to demonstrate its practicality. Thereby giving an initial point of reference that future work can compare to. 

How to cite: Dreier, M., Gourmelon, N., Koch, M., Seehaus, T., Braun, M., Maier, A., and Christlein, V.: A Benchmark Dataset for Radio Echo Sounding Data, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-11310, https://doi.org/10.5194/egusphere-egu24-11310, 2024.

16:32–16:34
|
PICO4.7
|
EGU24-13789
|
ECS
|
On-site presentation
Shivangini Singh, Duncan Young, Shuai Yan, Gregory Ng, Dillon Buhl, Alejandra Vega Gonzalez, Megan Kerr, Jamin Greenbaum, Scott Kempf, and Donald Blankenship

The Center for Oldest Ice Exploration (COLDEX) aims to unearth a stratigraphically intact ice core record going back to 1.5 million years. The target area is the corridor between South Pole and the southern flank of Dome A. While our current survey has yielded dated stratigraphy extending to 93.9 thousand years ago across the region, a significant portion of the stratigraphy remains undated. Our approach has involved leveraging englacial connections with existing ice cores and dust loggers, yet much of the stratigraphy awaits dating. Rapid ice access tools are capable of sampling the ice sheet using far fewer resources as compared to conventional drilling. By optimizing the site for such sampling, we can preemptively maximize the information that can be extracted eventually.

Our study’s objective in selecting a rapid ice access site is twofold: firstly, to maximize the age-depth scale extraction by dating the hitherto undateable deeper isochrones, and secondly, to strategically sample the pervasive basal layer in the survey region to understand its role in preserving old ice. To achieve an optimal age-depth scale extraction, we aim to target sites containing a larger portion of undated stratigraphy while maintaining the optimum resolution to understand ice age dust cycles. For instance, Ice Diver, a melt probe by COLDEX, houses an optical dust logger capable of counting ice age dust cycles.

Previous research (e.g., Chung et al., 2023) indicates that a basal layer identified through radar sounding may exhibit a distinct flow regime compared to the stratigraphic portion of the ice sheet column. Accessing this layer in advance could provide insights into the interface between the echo-free zone and stratigraphic ice, thereby refining our understanding of current ice sheet models and how to exploit them in the pursuit of old ice. An ideal access site would facilitate quick sampling without requiring deep drilling to reach the basal layer. Our aim is to reconcile these methodologies by identifying the most suitable site(s) for deploying such probes or drills.

How to cite: Singh, S., Young, D., Yan, S., Ng, G., Buhl, D., Vega Gonzalez, A., Kerr, M., Greenbaum, J., Kempf, S., and Blankenship, D.: Optimizing rapid access englacial sampling location to date deep radiostratigraphy for old ice, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-13789, https://doi.org/10.5194/egusphere-egu24-13789, 2024.

16:34–16:36
|
PICO4.8
|
EGU24-21693
|
Highlight
|
On-site presentation
Julien Bodart, Vjeran Višnjević, Antoine Hermant, and Johannes Sutter

Modelling the past evolution of the East Antarctic Ice Sheet (EAIS) in response to climate
and ocean forcing is challenged by the scarcity of observed palaeo boundary conditions. One
of the most spatially extensive records of past ice-sheet conditions comes from radar-detected
isochrones which, if dated precisely at ice cores, can provide a highly accurate temporal and
spatial history of ice-sheet evolution over time. Previous work has highlighted the benefit of
using such isochrones for testing and benchmarking 3-D ice sheet models; however,

uncertainty remains as to which model parameters fare better and how different bed and ice-
flow conditions affect the ability of ice-sheet models to reproduce the observed isochrones.

Here, we make use of previously acquired airborne radar data over the Wilkes Subglacial
Basin (East Antarctica) to connect existing stratigraphies and extract a temporal record of
isochrones at regular time intervals spanning the Holocene to the last interglacial and beyond.
The radar flight lines were carefully selected to provide a record of isochrones crossing
boundaries of different bed and ice-flow conditions situated between the ice divide and the
ice-sheet margins to represent as diverse a set of conditions as possible. The aim of this work
is to ultimately be able to test the ability of dated isochrones to tune ice-sheet model
parameters that will reproduce isochrone elevations in different parts of the catchment and
under different bed and ice-flow conditions.

How to cite: Bodart, J., Višnjević, V., Hermant, A., and Sutter, J.: Extracting Dated Isochrones on Airborne Radar Data Across EastAntarctica for Input into 3-D Regional Ice Sheet Models, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-21693, https://doi.org/10.5194/egusphere-egu24-21693, 2024.

16:36–16:38
|
PICO4.9
|
EGU24-16023
|
ECS
|
On-site presentation
Mylène Jacquemart and Ethan Welty

Ice temperature is an important characteristic of any glacier. It influences glacier dynamics, subglacial hydrology, glacier retreat behavior, and potential glacier hazards. Additionally, ice temperature can serve as an archive of past air temperature changes and be used for climate reconstructions and validation of thermo-mechanical glacier models. Measuring (deep) ice temperatures, however, is very laborious and costly, and most  ice temperature measurements are hidden away in publications that span almost a century.

To overcome the gap between data availability and data need, we have compiled ice temperature data from 132 glaciers in an open-source, version-controlled database available to the scientific community. This global englacial temperature database (glenglat; https://github.com/mjacqu/glenglat) contains temperatures measured in 410 different boreholes in the Americas, Greenland, Eurasia, Africa, and Antarctica between 1938 and 2023. Roughly 20% of all boreholes are known to have reached the glacier bed; the deepest borehole is 743 meters deep; and most measurements are from cold or polythermal glaciers. Data for 369 boreholes were extracted from published literature while data for the remaining 41 boreholes were directly submitted to the authors.

The database is structured following the Frictionless Data Tabular Data Package specification. The data are stored as comma-separated-value (CSV) files and the metadata are provided in a single YAML file – text formats which are both human and machine readable. Following a standard structure provides two key advantages: First, upon any change to the data or metadata, the structure and content of the database can automatically be validated using existing software. Second, documentation and data submission templates can be rendered automatically from the machine-readable metadata, which lowers the bar for data maintainers and future contributors. We hope that glenglat can serve many glaciological applications and become the  repository of choice for future ice temperature measurements.

How to cite: Jacquemart, M. and Welty, E.: A global englacial temperature database (glenglat), EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-16023, https://doi.org/10.5194/egusphere-egu24-16023, 2024.

16:38–16:40
|
PICO4.10
|
EGU24-6532
|
ECS
|
On-site presentation
Niya Shao, Michael Field, Emma MacKie, and Felicity McCormack

While subglacial topography serves a crucial role in ice sheet models, it remains generally sparsely sampled across Antarctica. Subglacial topography is primarily measured by airborne ice-penetrating radar with data gaps exceeding 100s of kilometers. Traditional kriging methods used to interpolate the sparse radar data cause spurious effects on ice flow divergence. Numerically solving for mass conservation equations addresses this issue but may smooth out the roughness of topography observed in the covariance structure of radar data. In this study, we propose a novel approach to generate an ensemble of realistically rough and mass-conserving subglacial topography. We utilize the Monte Carlo Markov Chain algorithm, where in each iteration of the Markov Chain, the topography is perturbed by Sequential Gaussian Simulation to reproduce the covariance structure in the radar data. The perturbed topography is then accepted with a probability constrained by both prior probability based on the radar measurement uncertainty and likelihood indicated by the topography’s deviation from mass conservation law. After the Markov Chain converges to a stable state, an ensemble of topography is sampled from the chain. We tested the method on Denman Glacier and produced posterior distribution of topography constrained by radar measurements and mass conservation. Moreover, the covariance structure of radar data is preserved in every generated topography realization. The method we developed provides a possibility to incorporate realistically rough topography into ice sheet models while avoiding artifacts caused by the violation of mass conservation. Furthermore, multiple subglacial topography realizations allow the propagation of inherent uncertainties in the sparsity of radar measurement to the result of downstream models.

How to cite: Shao, N., Field, M., MacKie, E., and McCormack, F.: Stochastic Simulation of Mass-Conserving Subglacial Topography with Monte Carlo Markov Chain, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-6532, https://doi.org/10.5194/egusphere-egu24-6532, 2024.

Sea Ice
16:40–16:42
|
PICO4.11
|
EGU24-16508
|
Highlight
|
On-site presentation
Zikang He, Yiguo Wang, Julien Brajard, and Xidong Wang

 

Dynamical sea ice predictions have significant biases or systematic errors that are difficult to effectively remove. In this work, we introduce machine learning into the Norwegian Climate Prediction Model (NorCPM, a state-of-the-art dynamical prediction system) to improve Arctic sea ice predictions.

We build a statistics bias-correction methodology employing machine learning techniques. An artificial neural network is trained with NorCPM data. It is then used to predict sea ice concentration biases or systematic errors and correct them either in post-processing of the predictions (offline manner) or during the production of the prediction (online manner). We evaluate the outcomes by assessing sea ice extent (SIE) and comparing them against observational data. Our findings reveal that offline correction markedly reduces the prediction biases in summer (more than 30%), while online correction enhances the variability in sea ice predictions up to four months. These results underscore the potential of machine learning as a potent tool for refining the accuracy of Arctic sea ice seasonal predictions.

How to cite: He, Z., Wang, Y., Brajard, J., and Wang, X.: Improving dynamical seasonal sea ice prediction in the Arctic with machine learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-16508, https://doi.org/10.5194/egusphere-egu24-16508, 2024.

16:42–16:44
|
PICO4.12
|
EGU24-9945
|
ECS
|
On-site presentation
Lorenzo Zampieri, Nils Hutter, and Francesco Cocetta

State-of-the-art sea ice models struggle to accurately simulate historical sea ice thickness changes, which could be partially due to inadequate representation of dynamics and thermodynamic processes. High-resolution observations are fundamental tools for improving our understanding of the sea ice physical processes, validating numerical models, and ultimately formulating better sea ice parameterization. Winter observations collected during the Multidisciplinary Drifting Observatory for the Study of Arctic Climate (MOSAiC) expedition in winter 2019-2020 are unique tools for evaluating our models during the freezing season. However, a shortcoming of these observations is that they cannot be easily combined due to differences in measurement techniques and processing chain, impeding a comprehensive characterization of the sea ice system and limiting their diagnostic employment with models. Here, we present an advanced spatiotemporal colocation algorithm designed to integrate airborne measurements collected during multiple helicopter surveys, which provide a two-dimensional characterization of the sea ice surface temperature through infrared camera images and of the surface elevation through an airborne laser scanner at a resolution of approximately one meter over areas spanning several kilometers. The co-located temperature and elevation fields can be combined with boundary layer observations and ground-based transect surveys via drift correction approaches. These observations put in relation for the first time snow freeboard with the equilibrium skin temperature resulting from the surface energy balance while resolving small-scale thickness features (e.g., snow dunes, ridges, and refrozen leads). We will showcase how this innovative observational dataset enables multi-category sea ice model evaluation and the development of new parameterizations. 

How to cite: Zampieri, L., Hutter, N., and Cocetta, F.: Colocated airborne observations from MOSAiC enable sea ice process understanding and new model parameterization development, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-9945, https://doi.org/10.5194/egusphere-egu24-9945, 2024.

16:44–16:46
|
PICO4.13
|
EGU24-3826
|
Highlight
|
On-site presentation
Julien Brajard, Anton Korosov, Richard Davy, and Yiguo Wang

This work introduces a simulator of high-resolution sea ice thickness in the Arctic based on diffusion models, which is a type of artificial intelligence (AI) generative model. Diffusion models have already shown impressive skill in generating realistic high-resolution images (e.g., DALL-E, Midjourney, Stable Diffusion).

Current satellite-based observations or climate model simulations of sea ice thickness provide valuable data but are limited by their coarse spatial resolution. High-resolution information is crucial for useful predictions and understanding small-scale features such as leads and thin ice, which significantly impact seasonal forecasting and heat flux calculations.

To increase the resolution of sea ice thickness products, we propose the following. First, a physically-based sea ice model, neXtSIM, is employed to generate a synthetic but realistic high-resolution sea ice thickness dataset. This synthetic dataset is then filtered to mimic the resolution of present satellite products or climate model outputs. An AI-based diffusion model is then trained to enhance the low-resolution SIT data. 

By comparing the field produced by the simulator and a high-resolution test image, we will demonstrate that the simulator is able to produce accurate and realistic high-resolution sea ice thickness fields. Accuracy is demonstrated by assessing the error of the reconstruction, while realism is assessed by visual inspection and by computing the power density spectra. For both criteria, accuracy, and realism, our simulator outperforms a linear interpolation of the low-resolution field, which is used as a baseline.

This work is held in the framework of the project SuperIce, funded by ESA.

How to cite: Brajard, J., Korosov, A., Davy, R., and Wang, Y.: Super-resolution of satellite observations of sea ice thickness using diffusion models and physical modeling., EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-3826, https://doi.org/10.5194/egusphere-egu24-3826, 2024.

16:46–16:48
|
PICO4.14
|
EGU24-13621
|
ECS
|
On-site presentation
Yan Huang and Xiaofeng Yang

Sea ice poses a significant threat to high-latitude navigation and offshore operations. Accurate and timely classification of sea ice is crucial for ensuring the safety of maritime activities in polar regions. Synthetic aperture radar (SAR) is widely used for sea ice classification due to its high resolution and all-weather observation capability. However, the Sentinel-1 extra-wide (EW) swath mode images, which are commonly used to monitor sea ice in the polar region, exhibit thermal noise in the cross-polarization images, and it is thought to affect the accuracy of sea ice classification models. In this study, we used Sentinel-1 EW mode images and a deep learning (DL) model, U-Net, to investigate the impact of thermal noise on sea ice classification. Sensitivity experiments were conducted for the U-Net and the comparison models, such as support vector machine (SVM), random forest (RF), and convolutional neural network (CNN), with or without using a denoising method for cross-polarization images. Both co-polarization and cross-polarization images were used to train these models. The experimental results indicate that SVM, RF, CNN, and U-Net achieved classification accuracies of 67.98%, 77.96%, 86.49%, and 90.00% respectively, using undenoised images. The classification accuracies improved to 71.69%, 80.75%, 86.65%, and 90.73% respectively after the denoising method was used. The SVM and RF models show an increase in accuracy of about 3%, while the CNN and U-Net models show an improvement of less than 1%, suggesting that CNN and U-Net are more tolerant to noise when used for sea ice classification.

How to cite: Huang, Y. and Yang, X.: Assessment of thermal noise impact on sea ice classification using Sentinel-1 images and U-Net, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-13621, https://doi.org/10.5194/egusphere-egu24-13621, 2024.

16:48–16:50
|
EGU24-14303
|
ECS
|
Virtual presentation
|
Yunhe Wang, Xiaojun Yuan, Yibin Ren, Mitchell Bushuk, Qi Shu, Cuihua Li, and Xiaofeng Li

Antarctic sea ice concentration (SIC) prediction at seasonal scale has been documented, but a gap remains at subseasonal scale (1-8 weeks) due to limited understanding of ice-related physical mechanisms. To overcome this limitation, we developed a deep learning model named SIPNet that can predict SIC without the need to account for complex physical processes. Compared to mainstream dynamical models like ECMWF, NCEP, and GFDL-SPEAR, as well as a relatively advanced statistical model like the linear Markov model, SIPNet outperforms them all, effectively filling the gap in subseasonal Antarctic SIC prediction capability. SIPNet results indicate that autumn SIC variability contributes the most to sea ice predictability, whereas spring contributes the least. In addition, the Weddell Sea displays the highest sea ice predictability, while predictability is low in the West Pacific. SIPNet can also capture the signal of ENSO and SAM on sea ice.

How to cite: Wang, Y., Yuan, X., Ren, Y., Bushuk, M., Shu, Q., Li, C., and Li, X.: Subseasonal prediction of regional Antarctic sea ice by a deep learning model, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-14303, https://doi.org/10.5194/egusphere-egu24-14303, 2024.

16:50–16:52
|
EGU24-19134
|
Virtual presentation
Yubao Qiu, Yang Li, Shuwen Yu, and Zekai Jin

Sea ice plays a significant role in Arctic research and operations. However, the lack of high spatiotemporal resolution observations on sea ice makes it challenging to accurately depict short-term sea ice changes. This limitation hampers the development of small-scale sea ice change studies and increases uncertainties in Arctic research and operational safety. With advancements in deep learning techniques and the abundance of multi-source remote sensing data such as optical, radar, and passive microwave, reconstructing high spatiotemporal resolution sea ice concentration in the Arctic becomes feasible. Based on multi-source remote sensing data and the integration of sea ice dynamics and thermodynamics, this study proposes a novel deep learning model for high spatiotemporal resolution sea ice concentration reconstruction. Based on this model, we achieved sub-kilometer scale and hourly-level reconstructions of Arctic sea ice concentration from 2021 to 2022, with a mean absolute error of less than 5%, thereby providing data support for Arctic research and operations.

How to cite: Qiu, Y., Li, Y., Yu, S., and Jin, Z.: High spatiotemporal reconstruction of Arctic sea ice concentration based on multi-source remote sensing data, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-19134, https://doi.org/10.5194/egusphere-egu24-19134, 2024.

16:52–16:54
|
EGU24-13988
|
ECS
|
Virtual presentation
Xinran Yang

The Arctic is one of the important drivers of global climate and environmental change. Its atmosphere, oceans, and sea ice movements have direct or indirect impacts on global atmosphere and ocean circulation as well as climate variability. Sea ice plays a crucial role in maintaining climate balance in the Arctic region, preventing more solar radiation from entering the sea through its high reflectivity and limiting air sea heat exchange, reducing sea surface temperature and sea ice melting rate. Sea ice thickness is an important parameter that describes the properties of sea ice. Obtaining the freeboard of sea ice through satellite altimetry measurements is of critical importance for sea ice thickness retrieval and understanding changes in Arctic sea ice. So far, research on satellite observations of pan-Arctic sea ice thickness has been limited to winter months. The key to measuring the freeboard using altimetry lies in distinguishing sea surface types and correctly identifying adjacent inter ice waterways. However, there are a large number of melting pools on the surface of summer floating ice, which make traditional waveform classification schemes unable to accurately distinguish sea surface types and become the main obstacle to retrieval freeboard of summer sea ice. In this study, a one-dimensional convolutional neural network classification model is built using CryoSat-2 summer sea ice classification training and testing dataset to improve summer sea ice freeboard retrieval. The model uses three parameters as input sources, i.e., elevation, pulse peakness, and backscatter coefficient. It achieves an overall accuracy of 84.3%. The sea ice freeboard is calculated from the elevation difference between the ice-covered waterways and its surrounding floating ice, resulting in distributions of 15-day individual sea ice freeboard and sea ice freeboard on an 80-km resolution grid. The results show that although there are many missing data due to noise and other issues, effective sea ice freeboard can be obtained in all months in summer. It demonstrates the feasibility of this method.

How to cite: Yang, X.: Research on Sea Ice Freeboard Retrieval from CryoSat-2 Based on Artificial Intelligence, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-13988, https://doi.org/10.5194/egusphere-egu24-13988, 2024.

16:54–16:56
|
EGU24-20355
|
ECS
|
Virtual presentation
Marlena Reil, Gunnar Spreen, Marcus Huntemann, Lena Buth, and Dennis Wilson

The Arctic is significantly affected by climate change, as evidenced by the constant decline of sea ice since the beginning of satellite observations. One driver of this transformation are melt ponds - pools of water that form as a result of melting sea ice during summer. Due to their darker color, they increase the absorption of incoming sunlight and accelerate ice melt. Accurate determination of melt pond extent and characteristics is considered a main factor in reducing uncertainty in Arctic climate models and sea ice concentration retrievals, but precise large scale observations are not available. Most knowledge to date is based on in-situ measurements, which are restricted to small areas. Satellite retrievals offer Arctic-wide coverage on a regular basis but lack resolution. To validate satellite measurements and allow observation at a moderate scale, helicopter-borne images are used. This ongoing work exploits a new dataset of helicopter-borne thermal infrared (TIR) imagery for melt pond retrieval. The derivation of geophysical parameters requires effective segmentation of different surface classes, which is challenged by temporally and spatially varying surface temperatures. We adapt and fine-tune AutoSAM, a prompt-free Segment Anything (SAM)-based segmentation tool that was introduced by Xinrong Hu et al. for medical imagery (Hu, X., Xu, X., & Shi, Y. (2023). How to Efficiently Adapt Large Segmentation Model (SAM) to Medical Images. arXiv preprint arXiv:2306.13731). Initial results with a limited number of annotated images indicate promising outcomes in the generalization of AutoSAM to cases that are rare in the training set, compared to U-Net and PSP-Net approaches. Beyond the scope of this project, this could serve as an example of how to use SAM as a segmentation tool for the remote sensing domain, which is typically hampered by the lack of labeled training data. Code and first results are provided at https://github.com/marlens123/autoSAM_pond_segmentation.

How to cite: Reil, M., Spreen, G., Huntemann, M., Buth, L., and Wilson, D.: Detection of Melt Ponds on Arctic Sea Ice from Infrared Images using AutoSAM, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-20355, https://doi.org/10.5194/egusphere-egu24-20355, 2024.

16:56–16:58
|
EGU24-18391
|
ECS
|
Virtual presentation
Solving Navier-Stokes Equations with Deep Learning Solvers for Ocean Phenomena Simulation
(withdrawn after no-show)
He Gao
16:58–17:00
|
EGU24-13646
|
Virtual presentation
Deep blue artificial intelligencefor knowledge discovery of theintermediate ocean
(withdrawn after no-show)
Baoxiang Huang
17:00–18:00