Displays

CR2.4

The explosion of data and computing power that is now available to glaciologists presents significant opportunities for advancing our understanding of glacial environments. However, significant barriers exist to achieving this, with the scales and rates at which data are being generated rendering many traditional approaches to analysis impractical.
Researchers across nearly all fields of glaciology are therefore increasingly requiring the development of automated and/or machine learning based approaches to effectively monitor and investigate these environments, in addition to new ways of visualising results. This session will therefore bring together glaciologists who use big data, machine learning and/or artificial intelligence to help share knowledge of different approaches that are currently being taken by the community and where possible demonstrate their potential transferability in this emergent field. Contributions are invited from those involved in developing and/or applying methods that seek to address these data generation, analytical and visualisation challenges with the aim of gaining greater understanding of past, present and future glacier and ice sheet change.

Public information:
During the chat session we will provide the opportunity for our presenters to briefly introduce their work and answer questions. If there is time, we will also facilitate a discussion for authors and attendees to discuss the following topics:

1. What are the challenges experienced by those applying big data/machine learning/AI techniques, and how can those involved in other areas of glaciology help?

2. What are the challenges of experienced by those using big data/machine learning/AI data products, and how can those creating them make them more accessible?

Other topics for discussion are very much encouraged both from those working with big/machine learning/AI data and those who may be interested in the potential of such approaches in glaciology.

Share:
Convener: James Lea | Co-conveners: Celia A. Baumhoer, Stephen BroughECSECS, Soroush REZVANBEHBAHANIECSECS, Leigh Stearns
Displays
| Attendance Wed, 06 May, 10:45–12:30 (CEST)

Files for download

Download all presentations (24MB)

Chat time: Wednesday, 6 May 2020, 10:45–12:30

Chairperson: James Lea
D2486 |
EGU2020-9723
Jordi Bolibar, Antoine Rabatel, Isabelle Gouttevin, Clovis Galiez, Thomas Condom, and Eric Sauquet
Glacier surface mass balance (SMB) and glacier evolution modelling have traditionally been tackled with physical/empirical methods, and despite some statistical studies very few efforts have been made towards machine learning approaches. With the end of this past decade, we have witnessed an impressive increase in the available amount of data, mostly coming from remote sensing products and reanalyses, as well as an extensive list of open-source tools and libraries for data science. Here we introduce a first effort to use deep learning (i.e. a deep artificial neural network) to simulate glacier-wide surface mass balance at a regional scale, based on direct and remote sensing SMB data, climate reanalysis and multitemporal glacier inventories. Coupled with a parameterized glacier-specific ice dynamics function, this allows us to simulate the evolution of glaciers for a whole region. This has been developed as the ALpine Parameterized Glacier Model (ALPGM), an open-source Python glacier evolution model. To illustrate this data science approach, we present the results of a glacier-wide surface mass balance reconstruction of all the glaciers in the French Alps from 1967-2015. These results were analysed and compared with all the available observations in the region as well as another physical/empirical SMB reconstruction study. We observe some interesting differences between the two SMB reconstructions, which further highlight the interest of using alternative methods in glacier modelling. Due to (relatively) recent advances in data availability and open tools (e.g. Tensorflow, Keras, Pangeo) this research field is ripe for progress, with many interesting challenges and opportunities lying ahead. To conclude, some perspectives on data science glacier modelling are discussed, based on the limitations of our current approach and on upcoming tools and methods, such as convolutional and physics-informed neural networks. 

How to cite: Bolibar, J., Rabatel, A., Gouttevin, I., Galiez, C., Condom, T., and Sauquet, E.: Glacier evolution modelling with deep learning: challenges and opportunities, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-9723, https://doi.org/10.5194/egusphere-egu2020-9723, 2020.

D2487 |
EGU2020-1212
Ryan Johnson, Carlos Oroza, Francesco Avanzi, Yamaguchi Satoru, Hiroyuki Hirashimia, and Tessa Maurer

Predicting the occurrence of preferential-flow snowpack runoff as opposed to spatially homogeneous matrix flow has recently become an important topic of cryosphere research, because of its implications for better understanding and forecasting wet-snow avalanche formation, streamflow generation during rain-on-snow events, and the polar-sheet water balance. Using twelve seasons of daily data from nine multi-compartment snow-lysimeters and concurrent weather and snowpack observations, we explored the accuracy of a machine-learning algorithm, Random Forest, in predicting the occurrence of preferential-flow snowpack runoff in a maritime context where sub-freezing conditions are rare (Nagaoka, Niigata prefecture, Japan). The algorithm was trained to predict three metrics of preferential-flow snowpack runoff: the coefficient of variation and standard and maximum deviations from mean spatial snowpack runoff. Two validation scenarios were used: one in which data were randomly subsampled from the entire period of record (66% training data, 33% testing), and a leave-one-year-out scenario, in which the model was trained on 11 years and tested on an unseen year. The latter was intended to represent a more realistic scenario in which limited data are available. Five tiers of features were used as inputs (independent variables) to the algorithm, including concurrent weather and bulk-snow properties, snow-atmosphere energy-balance components, internal snow structure, simulated matrix-flow snowpack runoff, and a selection of the five most important features from all previous groups. Relatively high model performance (Nash-Sutcliffe-Efficiency, NSE, > 0.53) was observed in all all-year scenarios, whereas the leave-one-year-out scenario displayed nearly a 50% reduction in performance, indicative of an inconsistent relationship across weather, snow conditions, and preferential-flow snowpack runoff generation between seasons. Random Forest also underestimated seasonal peaks in preferential flow, indicative of under-sampling in the dataset or unrepresented processes exceeding the spatial scale of multi-compartment lysimeters. This research presents an initial framework for understanding key factors influencing preferential-flow occurrence; improvements in algorithm accuracy could support predictions of preferential-flow snowpack runoff, especially in sparsely monitored regions.

How to cite: Johnson, R., Oroza, C., Avanzi, F., Satoru, Y., Hirashimia, H., and Maurer, T.: A Random-Forest approach to predicting preferential-flow snowpack runoff: early results and outlook for the future, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-1212, https://doi.org/10.5194/egusphere-egu2020-1212, 2020.

D2488 |
EGU2020-11237
Sophie Goliber, Taryn Black, and Ginny Catania

Marine-terminating outlet glacier terminus change mapped from satellite and aerial imagery in Greenland is used extensively in understanding how outlet glaciers adjust to climatic changes over a range of time scales. Numerous studies have digitized termini manually, but this process is labor-intensive and may lead to duplication of efforts. Additionally, these studies use different methods to pick the front (e.g. centerline pick, whole delineation, box method), which makes them difficult to compare. At the same time, machine learning techniques are rapidly making progress in their ability to accurately automate the extraction of glacier termini, with promising developments across a number of satellite sensors. However, limitations still exist: in particular, further high-quality manually-digitized training data are needed to make robust automatic picks. Here we present efforts to produce a database of manually digitized terminus picks and an intercomparison of picking techniques to determine errors and best practices for future efforts in digitization. These data will be cleaned, associated with appropriate metadata, and compiled so they can be easily accessed by scientists. Ultimately, these data will be used to create training data for further automatic picking efforts. We hope to solicit further collaboration with members of EGU and encourage those interested to email the authors.

How to cite: Goliber, S., Black, T., and Catania, G.: IcePicks: a collaborative database of Greenland outlet glacier termini, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-11237, https://doi.org/10.5194/egusphere-egu2020-11237, 2020.

D2489 |
EGU2020-4486
Celia A. Baumhoer, Andreas J. Dietz, Mariel Dirscherl, and Claudia Kuenzer

The Antarctic ice sheet drains ice through its peripheral ice shelves and glaciers making them an important factor for ice sheet mass balance. The extent of ice shelves and their calving front position influences ice sheet discharge and can yield valuable information on ice dynamics. Moreover, glacier fronts can have strong seasonal variations of retreat and advance. Yet, little is known about the seasonal pattern of Antarctic calving front fluctuations and their effect on ice sheet dynamics. The current developments in remote sensing and big data processing allow accurate monitoring of the Antarctic coastline. But to derive monthly calving front positions, the traditional approach of manual delineation is too time-consuming to cope with the temporal and spatial abundance of contemporary satellite missions. To create an up to date monitoring of changes in the Antarctic coastline a fully-automated approach is necessary. Automation of ice front delineation is a very challenging task as conventional edge detection methods fail due to the very low contrast between shelf ice and sea ice. Therefore, to exploit the abundance of available Sentinel-1 imagery over Antarctica, we created an automated workflow to extract monthly ice shelf front positions from Sentinel-1 imagery. The core of our processing chain is the deep learning architecture U-Net trained with about 44.000 image tiles covering parts of the Antarctic coastline during various seasons. Post-processing allows generating shapefiles of front positions and creating time series of seasonal ice shelf front fluctuations. We demonstrate our proposed method on selected ice shelves along the West and East Antarctic coastline and present intra-annual changes of calving front positions. This allows us to investigate seasonal change patterns of Antarctic ice shelves between 2014 and 2019 (depending on Sentinel-1 data availability) and to obtain a better picture on current Antarctic ice shelf front dynamics.

How to cite: Baumhoer, C. A., Dietz, A. J., Dirscherl, M., and Kuenzer, C.: Deep learning reveals seasonal patterns of Antarctic ice shelf front fluctuations, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-4486, https://doi.org/10.5194/egusphere-egu2020-4486, 2020.

D2490 |
EGU2020-19357
AmirAbbas Davari, Thorsten Seehaus, Matthias Braun, and Andreas Maier

Glacier and ice sheets are currently contributing 2/3 of the observed global sea level rise of about 3.2 mm a-1. Many of these glaciated regions (Antarctica, sub-Antarctic islands, Greenland, Russian and Canadian Arctic, Alaska, Patagonia), often with ocean calving ice front. Many glaciers on those regions show already considerable ice mass loss, with an observed acceleration in the last decade [1]. Most of this mass loss is caused by dynamic adjustment of glaciers, with considerable glacier retreat and elevation change being the major observables. The continuous and precise extraction of glacier calving fronts is hence of paramount importance for monitoring the rapid glacier changes. Detection and monitoring the ice shelves and glacier fronts from optical and Synthetic Aperture Radar (SAR) satellite images needs well-identified spectral and physical properties of glacier characteristics.

Earth Observation (EO) is producing massive amounts of data that are currently often processed either by the expensive and slow manual digitization or with simple unreliable methods such as heuristically found rule-based systems. As it was mentioned earlier, due to the variable occurrence of sea ice, icebergs and the similarity of fronts to crevasses, exact mapping of the glacier front position poses considerable difficulties to existing algorithms. Deep learning techniques are successfully applied in many tasks in image analysis [2]. Recently, Zhang et al. [3] adopted the state-of-the-art deep learning-based image segmentation method, i.e., U-net [4], on TerraSAR-X images for glacier front segmentation. The main motivation in using SAR modality instead of the optical aerial imagery is the capability of the SAR waves to penetrate cloud cover and its all year acquisition.

We intend to bridge the gap for a fully automatic and end-to-end deep learning-based glacier front detection using time series SAR imagery. U-net has performed extremely well in image segmentation, specifically in medical image processing community [5]. However, it is a large and complex model and is rather slow to train. Fully Convolutional Network (FCN) [6] can be considered as architecturally less complex variant of U-net, which has faster training and inference time. In this work, we will investigate the suitability of FCN for the glacier front segmentation and compare their performance with U-net. Our preliminary results on segmenting the glaciers demonstrate the dice coefficient of 92.96% by FCN and 93.20% by U-net, which essentially indicate the suitability of FCN for this task and its comparable performance to U-net.

 

References:

[1] Vaughan et al. "Observations: cryosphere." Climate change 2103 (2013): 317-382.

[2] LeCun et al. "Deep learning." nature 521, no. 7553 (2015): 436.

[3] Zhang et al. "Automatically delineating the calving front of Jakobshavn Isbræ from multitemporal TerraSAR-X images: a deep learning approach." The Cryosphere 13, no. 6 (2019): 1729-1741.

[4] Ronneberger et al. "U-net: Convolutional networks for biomedical image segmentation." MICCAI 2015.

[5] Vesal et al. "A multi-task framework for skin lesion detection and segmentation." In OR 2.0 Context-Aware Operating Theaters, Computer Assisted Robotic Endoscopy, Clinical Image-Based Procedures, and Skin Image Analysis, 2018.

[6] Long et al. "Fully convolutional networks for semantic segmentation." CVPR 2015.

How to cite: Davari, A., Seehaus, T., Braun, M., and Maier, A.: Glacier Front Detection at Tidewater Glaciers from Radar Images, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-19357, https://doi.org/10.5194/egusphere-egu2020-19357, 2020.

D2491 |
EGU2020-19996
Melanie Marochov, Patrice Carbonneau, and Chris Stokes

In recent decades, a wealth of research has focused on elucidating the key controls on the mass loss of the Greenland Ice Sheet and its response to climate forcing, specifically in relation to the drivers of spatio-temporally variable outlet glacier change. Despite the increasing availability of high-resolution satellite data, the time-consuming nature of the manual methods traditionally used to analyse satellite imagery has resulted in a significant bottleneck in the monitoring of outlet glacier change. Recent advances in deep learning applied to image processing have opened up a new frontier in the area of automated delineation of glacier termini. However, at this stage, there remains a paucity of research on the use of deep learning for image classification of outlet glacier landscapes. In this contribution, we apply a deep learning approach based on transfer learning to automatically classify satellite images of Helheim glacier, the fastest flowing outlet glacier in eastern Greenland. The method uses the well-established VGG16 convolutional neural network (CNN), and is trained on 224x224 pixel tiles derived from Sentinel-2 RGB bands, which have a spatial resolution of 10 metres. Based on features learned from ImageNet and limited training data, our deep learning model can classify glacial environments with >85% accuracy. In future stages of this research, we will use a new method originally developed for fluvial settings, dubbed ‘CNN-Supervised Classification’ (CSC). CSC uses a pre-trained CNN (in this case our VGG16 model) to replace the human operator’s role in traditional supervised classification by automatically producing new label data to train a pixel-level neural network classifier for any new image. This transferable approach to image classification of outlet glacier landscapes permits not only automated terminus delineation, but also facilitates the efficient analysis of numerous processes controlling outlet glacier behaviour, such as fjord geometry, subglacial plumes, and supra-glacial lakes.

How to cite: Marochov, M., Carbonneau, P., and Stokes, C.: Automated image classification of outlet glaciers in Greenland using deep learning, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-19996, https://doi.org/10.5194/egusphere-egu2020-19996, 2020.

D2492 |
EGU2020-19979
Daniel Cheng, Yara Mohajerani, Michael Wood, Eric Larour, Wayne Hayes, Isabella Velicogna, and Eric Rignot

We present Calving Front Machine (CALFIN), an automated method for extracting calving fronts from satellite imagery. We generate results for 66 glaciers along East/West Greenland from 1972 to 2019. We output these results as a dataset, and provide new constraints on glacial evolution over the time period. This method is uniquely robust to clouds, illumination differences, ice mélange, and Landsat-7 Scan Line Corrector errors. The current implementation offers a new opportunity to explore previous trends, and validate existing models moving forward.

This method utilizes deep learning, in the form of the Google DeeplabV3+ Xception derived CALFIN Neural Network. This approach builds on existing work by Mohajerani et al., Zhang et al., and Baumhoer et al. Additional post-processing techniques allow our method to achieve accurate and useful segmentation of raw images into Shapefile outputs. 

We achieve are often indistinguishable from the manually-curated fronts, deviating from such test data by 1 pixel (about 80 meters) or less XXX% of the time across 162 test images.

CALFIN excels among the current state of the art. We show this by performing a model inter-comparison to evaluate CALFIN's performance against existing methodologies. We also showcase CALFIN's ability to generalize to SAR and MODIS imagery. We achieve a mean error of 2.25 pixels (86.76 meters) from the true front on a diverse set of 162 testing images.

How to cite: Cheng, D., Mohajerani, Y., Wood, M., Larour, E., Hayes, W., Velicogna, I., and Rignot, E.: Calving Front Machine (CALFIN): Automated Calving Front Dataset and Deep Learning Methodology for East/West Greenland, 1972-2019, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-19979, https://doi.org/10.5194/egusphere-egu2020-19979, 2020.

D2493 |
EGU2020-17780
Stephen Brough and James Lea

The drainage of supraglacial lakes provides a fundamental mechanism for the rapid transfer of surface meltwater to the bed of an ice sheet, impacting both subglacial hydrology and ice dynamics. As a consequence, it is crucial to understand where and when these lakes drain, and how or if this has changed through time. Given that lakes are now occurring in greater numbers and at higher elevations, identifying changing modes in behaviour will have significant implications for the future dynamics of the Greenland ice sheet. Nevertheless, previous studies of supraglacial lakes and associated drainage events have been limited in spatial and/or temporal scale relative to the entire ice sheet.

Here we use daily maps of Greenland wide supraglacial lake coverage – derived from MODIS Terra within Google Earth Engine – to investigate the style, pattern and timing of lake drainages between 2000 and 2019. Results from this study: i) add to the understanding of how supraglacial hydrology and its coupling to the bed has changed in response to more extensive supraglacial lake cover over the last 20 years; and ii) provide insight into how these lakes and associated drainage events can be expected to respond to increased surface meltwater production under a warming climate.

How to cite: Brough, S. and Lea, J.: Greenland ice sheet supraglacial lake drainages between 2000 and 2019, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-17780, https://doi.org/10.5194/egusphere-egu2020-17780, 2020.

D2494 |
EGU2020-1155
Maite Lezama Valdes, Marwan Katurji, and Hanna Meyer

Anthropogenic Climate Change is expected to take a toll on the Antarctic environment and its biodiversity, which is concentrated on the continent’s few ice-free areas, such as the McMurdo Dry Valleys (MDV). To model the current terrestrial habitat distribution and predict possible climate induced changes, high spatial and temporal resolution abiotic variables, especially land surface temperature (LST) and soil moisture are needed, but are currently unavailable.

The aim of this project is to fill this gap and create a high resolution LST dataset of the Antarctic Dry Valleys. This variable is acquired in a high temporal resolution (sub-daily) by the MODIS sensor aboard Terra and Aqua satellites. However, as LST varies greatly in space, the spatial resolution provided by this data source (1000 m) is too low to give a meaningful impression of LST and to study biodiversity patterns. Therefore, we use data from Landsat and ASTER sensors as a reference to downscale MODIS LST to a spatial resolution of 30 m. 7 year’s worth of satellite data as well as terrain-derived auxiliary variables went into the development of the model, which predicts 30 m LST for the Antarctic Dry Valleys. 

To model complex relations between terrain, radiation, land cover and LST, machine learning models are used. Multiple algorithms (Random Forest, NN, SVM, Gradient Boosting) are compared to find the best approach for predicting high resolution LST based on MODIS data. Using the best performing model, a daily dataset is created that provides LST for the Antarctic Dry Valleys from 2002 on.

How to cite: Lezama Valdes, M., Katurji, M., and Meyer, H.: High spatial and temporal resolution land surface temperature for the Antarctic Dry Valleys , EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-1155, https://doi.org/10.5194/egusphere-egu2020-1155, 2020.

D2495 |
EGU2020-12949
Eleanor Bash, Christine Dow, and Greg McDermid

Recent advances in camera sensors, data storage, and structure-from-motion (SfM) processing are opening new possibilities for monitoring glacier processes through time series imagery. With SfM processing, internal and external camera parameters can be estimated in a bundle adjustment, alleviating problems associated with camera stability in the field. Orienting points in real world coordinates, however, still requires manual intervention in the form of ground control identification in imagery when dealing with two camera systems. We are introducing a new automated method of orienting point clouds from two-camera time lapse set ups to allow for fast processing of large numbers of frames. We accomplish this by leveraging several algorithms developed for computer vision and apply them to an analysis of glacier surface elevation change. Two time lapse systems were installed overlooking Nùłàdäy (Lowell Glacier), Yukon, Canada, on July 13, 2019. Each system consisted of a Nikon D5600 and a DigiSnap Pro, recording images at 2-hour intervals. On July 1, 2019 a manned aircraft flight collected imagery of the glacier using a Nikon D850 with a differential GPS collecting high precision locations for each image. The July 1 imagery was processed using Agisoft Photoscan Professional through the Python API to produce a target point cloud for orientation of unregistered time lapse imagery. Using Photoscan Professional’s 4D capability, a time series of images from each time lapse camera were aligned in a one-step bundle adjustment to produce a series of dense point clouds at each time step. Point clouds from time lapse imagery were coregistered to the target point cloud using a Fast Point Feature Histograms and a color-enhanced point cloud alignment based on Rusu et al. (2009) and Park et al. (2017). The M3C2 algorithm (Lague et al., 2013) was used to difference point clouds in the timeseries and derive a time series of elevation change at Nùłàdäy with an uncertainty of 1.5 m2.  All steps in the workflow are executed through Python, allowing for easy automated execution of data processing. With streamlined processing it is possible to integrate more time steps into SfM analysis of glacier surface elevation change and integrate the data into modelling efforts of glacier evolution.

How to cite: Bash, E., Dow, C., and McDermid, G.: Dealing with “too much data”: Automated Structure-from-Motion Processing of Time Lapse Imagery at Nùłàdäy Glacier, Yukon, Canada, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-12949, https://doi.org/10.5194/egusphere-egu2020-12949, 2020.

D2496 |
EGU2020-13513
Mauro Werder, Matthias Huss, Frank Paul, Amaury Dehecq, and Daniel Farinotti

Accurate estimations of ice thickness and volume are indispensable for ice flow modelling, hydrological forecasts and sea-level rise projections. We present BITE, a new ice thickness estimation model based on a mass-conserving forward model and a Bayesian inversion scheme. The forward model calculates flux in an elevation-band flow-line model, and translates this into ice thickness and surface ice speed using a shallow ice formulation. Both ice thickness and speed are then extrapolated to the map plane. The model assimilates observations of ice thickness and speed using a Bayesian scheme implemented with a Markov chain Monte Carlo method, which calculates estimates of ice thickness and their error. We illustrate the model's capabilities by applying it to a mountain glacier, validate the model using 733 glaciers from four regions with ice thickness measurements, and demonstrate that the model can be used for large-scale studies by fitting it to over 30 000 glaciers from five regions. The results show that the model performs best when a few thickness observations are available; that the proposed scheme by which parameter-knowledge from a set of glaciers is transferred to others works but has room for improvements; and that the inferred regional ice volumes are consistent with recent estimates.

How to cite: Werder, M., Huss, M., Paul, F., Dehecq, A., and Farinotti, D.: BITE, the Bayesian Ice Thickness Estimation model, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-13513, https://doi.org/10.5194/egusphere-egu2020-13513, 2020.