MASSIVE – MAchine learning, Surface mass balance of glaciers, Snow cover, In-situ data, Volume change and Earth observation

• Maslov, Konstantin; Persello, Claudio; Schellenberger, Thomas & Stein, Alfred (2025). Globally scalable glacier mapping by deep learning matches expert delineation accuracy. Nature Communications. 16. doi: 10.1038/s41467-024-54956-x. Full text in Research Archive Show summary

  Accurate global glacier mapping is critical for understanding climate change impacts. Despite its importance, automated glacier mapping at a global scale remains largely unexplored. Here we address this gap and propose Glacier-VisionTransformer-U-Net (GlaViTU), a convolutional-transformer deep learning model, and five strategies for multitemporal global-scale glacier mapping using open satellite imagery. Assessing the spatial, temporal and cross-sensor generalisation shows that our best strategy achieves intersection over union >0.85 on previously unobserved images in most cases, which drops to >0.75 for debris-rich areas such as High-Mountain Asia and increases to >0.90 for regions dominated by clean ice. A comparative validation against human expert uncertainties in terms of area and distance deviations underscores GlaViTU performance, approaching or matching expert-level delineation. Adding synthetic aperture radar data, namely, backscatter and interferometric coherence, increases the accuracy in all regions where available. The calibrated confidence for glacier extents is reported making the predictions more reliable and interpretable. We also release a benchmark dataset that covers 9% of glaciers worldwide. Our results support efforts towards automated multitemporal and global glacier mapping.

• Maslov, Konstantin; Persello, Claudio; Schellenberger, Thomas & Stein, Alfred (2024). Towards Global Glacier Mapping with Deep Learning and Open Earth Observation Data. arXiv.org. ISSN 2331-8422. doi: 10.48550/arXiv.2401.15113. Full text in Research Archive Show summary

  Accurate global glacier mapping is critical for understanding climate change impacts. It is challenged by glacier diversity, difficult-to-classify debris and big data processing. Here we propose Glacier-VisionTransformer-U-Net (GlaViTU), a convolutional-transformer deep learning model, and five strategies for multitemporal global-scale glacier mapping using open satellite imagery. Assessing the spatial, temporal and cross-sensor generalisation shows that our best strategy achieves intersection over union >0.85 on previously unobserved images in most cases, which drops to >0.75 for debris-rich areas such as High-Mountain Asia and increases to >0.90 for regions dominated by clean ice. Additionally, adding synthetic aperture radar data, namely, backscatter and interferometric coherence, increases the accuracy in all regions where available. The calibrated confidence for glacier extents is reported making the predictions more reliable and interpretable. We also release a benchmark dataset that covers 9% of glaciers worldwide. Our results support efforts towards automated multitemporal and global glacier mapping.

• Piermattei, Livia; Zemp, Michael; Sommer, Christian; Brun, Fanny; Braun, Matthias H. & Andreassen, Liss M. [Show all 35 contributors for this article] (2024). Observing glacier elevation changes from spaceborne optical and radar sensors - an inter-comparison experiment using ASTER and TanDEM-X data. The Cryosphere. ISSN 1994-0416. 18(7), p. 3195–3230. doi: 10.5194/tc-18-3195-2024. Full text in Research Archive Show summary

  Observations of glacier mass changes are key to understanding the response of glaciers to climate change and related impacts, such as regional runoff, ecosystem changes, and global sea level rise. Spaceborne optical and radar sensors make it possible to quantify glacier elevation changes, and thus multi-annual mass changes, on a regional and global scale. However, estimates from a growing number of studies show a wide range of results with differences often beyond uncertainty bounds. Here, we present the outcome of a community-based inter-comparison experiment using spaceborne optical stereo (ASTER) and synthetic aperture radar interferometry (TanDEM-X) data to estimate elevation changes for defined glaciers and target periods that pose different assessment challenges. Using provided or self-processed digital elevation models (DEMs) for five test sites, 12 research groups provided a total of 97 spaceborne elevation-change datasets using various processing approaches. Validation with airborne data showed that using an ensemble estimate is promising to reduce random errors from different instruments and processing methods but still requires a more comprehensive investigation and correction of systematic errors. We found that scene selection, DEM processing, and co-registration have the biggest impact on the results. Other processing steps, such as treating spatial data voids, differences in survey periods, or radar penetration, can still be important for individual cases. Future research should focus on testing different implementations of individual processing steps (e.g. co-registration) and addressing issues related to temporal corrections, radar penetration, glacier area changes, and density conversion. Finally, there is a clear need for our community to develop best practices, use open, reproducible software, and assess overall uncertainty to enhance inter-comparison and empower physical process insights across glacier elevation-change studies.

• Maslov, Konstantin; Persello, Claudio; Schellenberger, Thomas & Stein, Alfred (2023). GLAVITU: A Hybrid CNN-Transformer for Multi-Regional Glacier Mapping from Multi-Source Data, IGARSS 2023 - 2023 IEEE International Geoscience and Remote Sensing Symposium. IEEE (Institute of Electrical and Electronics Engineers). ISSN 9798350320107. p. 1233–1236. doi: 10.1109/IGARSS52108.2023.10281828. Full text in Research Archive Show summary

  Glacier mapping is essential for studying and monitoring the impacts of climate change. However, several challenges such as debris-covered ice and highly variable landscapes across glacierized regions worldwide complicate large-scale glacier mapping in a fully-automated manner. This work presents a novel hybrid CNN-transformer model (GlaViTU) for multi-regional glacier mapping. Our model outperforms three baseline models—SETR-B/16, ResU-Net and TransU-Net—achieving a higher mean IoU of 0.875 and demonstrates better generalization ability. The proposed model is also parameter-efficient, with approximately 10 and 3 times fewer parameters than SETR-B/16 and ResU-Net, respectively. Our results provide a solid foundation for future studies on the application of deep learning methods for global glacier mapping. To facilitate reproducibility, we have shared our data set, codebase and pretrained models on GitHub at https://github.com/konstantin-a-maslov/GlaViTU-IGARSS2023.

View all works in NVA

• Maslov, Konstantin; Schellenberger, Thomas; Paul, Frank; Goerlich, Franz; Rastner, Philipp & Kohler, Jack [Show all 8 contributors for this article] (2026). Benchmark dataset “ICEmapper Svalbard” - Dataset - Sigma2 Data Catalogue. doi: https:/archive.sigma2.no/dataset/benchmark-dataset-icemapper-svalbard. Full text in Research Archive Show summary

  Sigma2 Data Catalogue

• Maslov, Konstantin; Schellenberger, Thomas; PANDIT, PRASHANT H.; Persello, Claudio & Stein, Alfred (2026). Pan-European dataset for glacier facies classification. doi: 10.5281/ZENODO.18469893. Full text in Research Archive Show summary

  The dataset includes 137592 manually annotated point labels and the corresponding image stacks collected on 92 Landsat and Sentinel-2 scenes covering 31 glaciers across the Alps, Pyrenees, Scandinavia, Iceland and Svalbard. Eight surface classes are provided: ice, snow, firn, debris, refrozen-like ice, shadow, water and cloud. Labels were produced by three expert glaciologists and refined through a cross-check procedure involving harmonisation of ambiguous transitions and addition of points in sparsely sampled areas. Pre- and post-cross-check label sets are included, along with acquisition metadata to facilitate reproducibility. The dataset is intended as a benchmark for automated glacier facies mapping and for developing machine learning or remote sensing algorithms under diverse glaciological and radiometric conditions. Please cite the accompanying paper when using these data. To cite this dataset, please follow the instructions at https://github.com/konstantin-a-maslov/glacier_facies_classification

• Schellenberger, Thomas; Maslov, Konstantin & Persello, Claudio (2025). Space-time dimension swap: a novel deep learning based approach to extract calving fronts of marine-terminating glaciers from time series of Sentinel-1 imagery. Full text in Research Archive Show summary

  Marine-terminating glaciers are important contributors to the total the mass balance of all glaciers in the Northern Hemisphere, despite their small count of only ~1500. Several deep learning-based approaches have successfully been applied to specifically map calving fronts from optical and SAR satellite imagery. Yet, this task poses a specific challenge: accuracy decreases significantly when “fjord ice”, e.g. icebergs, seasonal sea ice or ice melange, with similar spectral and textural signature is present. As published methods are all based on single images, they are prone to artifacts of fjord-ice in the outlines. Reported general accuracies are greater 3-10 pixels under good conditions. The standard approach is to map the termini in each satellite image of a given date separately (2D in space & 0D in time). Stacking all retrieved outlines from a time series of satellite images temporally then results in the 3D calving front time series. Here we propose a novel approach, in which we will swap the time and space dimension in the image segmentation. We first extract backscatter values along a flowline of a given glacier (which crosses the terminus) from all Sentinel-1 images (1D in space & 1D in time). Each column in the resulting “time-series” image represents data along the centreline at a given date. We then manually label pixel coordinates where the flowline crosses the calving front. The front is clearly identifiable either from the strong backscatter as the front acts as a large corner reflector, or in very low backscatter due to the front’s radar shadow. The resulting binary masks of several glaciers are then used to train an adapted version of our vision-transformer convolutional neural network hybrid model GlaViTU. The model is applied to all “time series images” extracted along a dense pixel-by-pixel grid of flowlines parallel to the original flowline, covering the whole calving front of a glacier. The resulting “label-like” images are then stacked spatially to retrieve the full 3D time series of calving front of the glacier, back-swapping the dimensions. The novel approach has four advantages over current methods: Firstly, it allows to retrieve data from a single or multiple complete sensor collection(s) for a glacier or region with high computational efficiency. Secondly, labelling single points drastically increases the quality of the training data and reduces expert workload. Thirdly, “fjord ice” ends up in the “fjord” class, which is highly inhomogeneous in time and hence easier to distinguish from glacier ice. Fourthly, with expert knowledge, the maximum distance between neighbouring points in time can be constrained using a expert-informed loss functions in the deep learning model, reducing the average accuracy to an estimated 1-3 pixels, e.g. 10-30 m for Sentinel-1. We will present a time series of calving fronts of all marine-terminating glaciers in Svalbard and Alaska from 2016-2024 retrieved from Sentinel-1 data.

• Schellenberger, Thomas; Maslov, Konstantin & Persello, Claudio (2025). Space-time dimension swap: a novel deep learning based approach to extract calving fronts of marine-terminating glaciers from time series of Sentinel-1 imagery. doi: https:/2025.ieeeigarss.org/view_paper.php?PaperNum=5567&SessionID=1624. Full text in Research Archive Show summary

  Marine-terminating glaciers are important contributors to the total the mass balance of all glaciers in the Northern Hemisphere, despite their small count of only ~1500. Several deep learning-based approaches have successfully been applied to specifically map calving fronts from optical and SAR satellite imagery. Yet, this task poses a specific challenge: accuracy decreases significantly when “fjord ice”, e.g. icebergs, seasonal sea ice or ice mélange, with similar spectral and textural signature is present. As published methods are all based on single images, they are prone to artifacts of fjord-ice in the outlines. Reported general accuracies are greater 3-10 pixels under good conditions. Here we propose a novel approach, in which we will swap the time and space dimension in the image segmentation. The standard approach is to map the termini in each satellite image of a given date separately (2D in space & 0D in time). Stacking all retrieved outlines from a time series of satellite images temporally then results in the 3D calving front time series. Instead, we first extract backscatter values along a given flowline which crosses the termini from all Sentinel-1 images of a given glacier (1D in space & 1D in time). Each column in the image represents data along the centreline at a given date. We then manually label pixel coordinates where the flowline crosses the calving front. The front is clearly identifiable either from the strong backscatter as the front acts as a large corner reflector, or in very low backscatter due to the front’s radar shadow. The resulting binary masks of several glaciers are then used to train an neural network. The model is applied to all space-time images extracted along a dense pixel-by-pixel grid of flowlines parallel to the red line, covering the whole calving front of a given glacier. The resulting “label-like” images are then stacked spatially to retrieve the full 3D time series of calving front of a given glacier, back-swapping the dimensions.

• Maslov, Konstantin; Schellenberger, Thomas; Persello, Claudio & Stein, Alfred (2025). Deep learning applied to Sentinel-1 data shows doubling of glacier area loss in Svalbard compared to 1980–2010. doi: 10.5194/egusphere-egu25-9898. Full text in Research Archive Show summary

  Glaciers in Svalbard are undergoing rapid changes due to Arctic Amplification that demand frequent and accurate monitoring. We present a deep learning model, Intensity-Coherence-Evolution-mapper (ICEmapper) designed to extract annual glacier outlines from Sentinel-1 time series data regardless of cloud cover. The model combines SAR backscatter intensity and interferometric coherence. In extensive validation tests against manually digitised optical imagery, ICEmapper demonstrates human-expert accuracy, with intersection of union score higher than 0.95, total area discrepancies below 0.5%, median distance deviations under 15 m, and 95th percentile deviations within 250 m. Additionally, we report calibrated uncertainties of our classification results at the pixel level, allowing for detailed analysis of significant changes as well as total area uncertainty estimation. This performance allowed us to construct a continuous inventory of glacier outlines in Svalbard from 2016 to 2023. The results of area change analysis suggest a substantial escalation in the rate of glacier area loss to about 180 km2 a−1 in the last decade, nearly doubling from the previously reported ~80 km2 a−1 (1980–2010; Nuth et al., 2013). This increase is predominantly driven by enhanced calving at tidewater glaciers, although climatic signal also shows a significant correlation with the area loss of land-terminating glaciers. Surging glaciers, particularly in the Nathorstbreen system and Austfonna, exhibited unique behaviours that can temporarily increase the total glacier area. In 2016, the Nathorstbreen system gained 107.76 km2, while Austfonna, Basin-3 expanded by 86.54 km2 as compared to the Randolph Glacier Inventory, jointly offsetting net losses by approximately two years. In the last decade, however, surging glaciers lost area more rapidly (−0.57 km2 a−1 on average) than the non-surging ones (−0.09 km2 a−1). Additionally, our analysis uncovered a surge in Austfonna, Basin-7 starting in 2019 and not reported previously, emphasising the capability of annually updated inventories to complement other methods of surge detection. Our methods have the potential to be transferred to other glacierised regions and enhance monitoring and understanding of glacier area changes on larger scales using Sentinel-1 data.

• Pandit, Prashant; Schellenberger, Thomas; Callegari, Mattia & Bruzzone, Lorenzo (2025). Glacier Snowline Mapping from Sentinel-2 images by Machine Learning. doi: https:/lps25.esa.int/lps25-presentations/presentations/382/_382.pdf. Full text in Research Archive Show summary

  Snow accumulation and ablation play a pivotal role in the mass balance of mountain glaciers. Accurate mapping of snow extends on glaciers enhance our understanding of climate change impacts (Larocca et al.) and serve as valuable input for glacier surface mass balance models (Rabatel et al.). For example, the snowline altitude at the end of the summer season serves as a proxy for the Equilibrium Line Altitude and strongly correlated with the annual mass balance (Rabatel et al.). The largest study mapped 3489 snowlines of 269 glaciers of the ~275.000 glaciers from Landsat data between year 1984 and 2022 and found an increase in snow line elevation of approximately 150 meters. This significant rise indicates a reduction in glacier accumulation zones, suggesting negative mass balance trends driven by rising temperatures and shifting precipitation patterns. Regional variability in snowline changes underscores the complex interplay of climatic and local factors, emphasizing the importance of snowline monitoring for glacier health and climate change studies (Larocca et al.). However, manual delineation is time consuming and band- and index threshold-based approaches often fails in areas of steep and complex mountainous terrain and in homogenous snow conditions. The spectral similarity to firn and ice and varying illumination conditions in steep terrain are posing significant challenges for accurate large-scale and long-term snow monitoring Advanced machine learning techniques offer promising solutions to address these limitations by automating snowline detection and improving accuracy under diverse conditions (Prieur et al.). Designed to overcome these limitations by improving accuracy and scalability, this study presents first steps towards region-scale snow cover extend mapping using machine learning. We manually digitized 312 snowlines on 41 glaciers in Scandinavia (13), Svalbard (9), and the European Alps (19) from Sentinel 2 data in the period 2015 to 2023, encompassing a wide range of seasonal snow conditions., Using this benchmark, we trained several machine learning models, including pixel-based classifiers such as Support Vector Machine, Random Forest, and XGBoost (Chen et al.), and U-Net (Ronneberger et al.), a Fully convolutional neural network , and compared them against threshold-based approaches as baselines. The results from Scandinavia demonstrate the superiority of machine learning methods. While threshold-based approaches, such as the Normalized Difference Snow Index (NDSI > 0.4) and Near-Infrared (NIR > 0.11), achieved an Intersection over Union (IoU) score of 0.7147, U-Net significantly outperformed with an IoU of 0.9456. Random Forest was the next best-performing method with an IoU of 0.8957, followed by XGBoost (0.8899) and SVM (0.8887). Adding elevation models and slope data to the classifiers resulted in only marginal performance improvements. This significant improvement highlights the potential of U-Net to accurately capture fine-scale snowline features, especially in heterogeneous and complex mountainous environments with spectrally similar classes such as firn and ice and paves the way for accurate, low-cost, automated and large-scale snow mapping on glaciers.

• Piermattei, Livia; Ioli, Francesco; Webster, Clare; Kugler, Lucas; Treichler, Désirée & Mattea, Enrico [Show all 7 contributors for this article] (2025). Regional Glacier Elevation Changes Assessment from Optical DEM Time Series. doi: https:/lps25.esa.int/programme/programme-session/?id=1560786C-E6E2-43D2-83A0-3D81BFC41D67&presentationId=CAA543F0-380F-4C74-936A-F48A1374258E. Full text in Research Archive Show summary

  This study is part of the Glacier Mass Balance Intercomparison Exercise (GlaMBIE), which aims to collect, homogenise, and estimate global and regional assessments of glacier mass balance using the main observation methods. Here, we present our assessment of glacier elevation change using the geodetic method (DEM differencing) based on spaceborne optical data. We exploited the potential of the SPOT-5 satellite, operational from 2002 to 2015, which provided global coverage. Since 2021, the SPOT 1-5 image archive has been freely available as part of the SPOT World Heritage program run by CNES. However, observation periods vary across regions, limiting temporal coverage to less than five years in some areas. Iceland is selected as a pilot study due to its extensive SPOT-5 temporal coverage, further complemented by ArcticDEM data. Our methodology is also applied to other regions with sufficient temporal coverage, and historical aerial images are incorporated to extend the analysis back to the last century. The workflow starts with generating DEM time series at a regional scale and homogenising the data, including DEM co-registration, selection, noise filtering and void filling. To address challenges posed by sparse DEM time series, especially when including historical DEMs, we developed a method to extrapolate elevation changes over 10-year intervals using the combined DEM time series. This method relies on the assumption that a relationship exists between elevation change and elevation; therefore, an elevation trend can be derived for elevation bands. We extract median elevations for fixed elevation bands (i.e., 100 m bins) from the DEMs time series and interpolate these values over time using linear regression. Elevation data are then extrapolated for each band and pre-defined periods, and area-weighted mean elevation changes are calculated for each glacier using RGI7.0. For comparability, we also applied our approach to derive elevation changes from time series of ASTER DEMs and compared our results with the pixel-based multi-temporal approach of Hugonnet et al. (2021) over a common observation period. Regional and individual glacier estimates from both methods are evaluated. This work discusses key challenges in using spaceborne optical data for regional glacier elevation change assessments, including limitations in the temporal coverage of SPOT-5, issues with DEM generation, co-registration, noise filtering, void filling, and methods for estimating mean elevation changes. Our findings contribute to improving regional assessments of glacier mass balance and advancing geodetic approaches using optical DEM time series.

• Schellenberger, Thomas; Maslov, Konstantin; Pandit, Prashant H; Persello, Claudio; Mattia, Callegari & Stein, Alfred (2025). Large scale satellite-based datasets for glacier modelling. Full text in Research Archive
• Tjoflot, Gunn Kristin & Treichler, Désirée (2025). Smeltende isbreer øker tapet av ferskvann og hever globale havnivåer. Titan.uio.no. Full text in Research Archive
• Maslov, Konstantin & Schellenberger, Thomas (2024). Globally Scalable Glacier Mapping by Deep Learning Matches Expert Delineation Accuracy. doi: 10.11582/2024.00168. Full text in Research Archive Show summary

  Benchmark dataset for global glacier mapping using deep learning and satellite imagery. Details about the dataset and its use can be found here: Maslov, K.A., Persello, C., Schellenberger, T. And A. Stein: Globally scalable glacier mapping by deep learning matches expert delineation accuracy. Nat Commun 16, 43 (2025). https://doi.org/10.1038/s41467-024-54956-x The related code here: https://github.com/konstantin-a-maslov/scalable_glacier_mapping

• Persello, Claudio (2024). AI technologies for processing multi-source Earth Observation (EO) data. Full text in Research Archive Show summary

  - Introduction to Machine Learning concepts - Deep learning: principle and architectures - Convolutional Neural Networks for various image analysis tasks Deep Learning and Earth Observation to support: o Mapping smallholder farms o Automated extraction of cadastral boundaries o Mapping slums and deprived urban areas o Extraction of geometric building information o Large-scale glacier mapping

• Persello, Claudio (2024). Machine Learning and Earth Observation for the Sustainable Development Goals. Full text in Research Archive
• Casarotto, Christian & Mattia, Callegari (2024). Annual glacier mass balance estimation through ASTER DTMs and snowlines extracted from Landsat and Sentinel-2 images. Full text in Research Archive
• Maslov, Konstantin; Schellenberger, Thomas; Persello, Claudio & Stein, Alfred (2024). Glacier Mapping from Sentinel-1 SAR Time Series with Deep Learning in Svalbard. Full text in Research Archive Show summary

  Glaciers are one of the essential climate variables. Tracking their areal changes over time is of high importance for monitoring the impacts of climate change and designing adaptation strategies. Mapping glaciers from optical remote sensing data might result in a very limited temporal resolution due to the absence of cloud-free imagery at the end of the ablation season. Synthetic aperture radar (SAR) solves this problem as it can operate in almost all weather conditions. Here, we present a deep learning strategy for glacier mapping based solely on Sentinel-1 SAR data in Svalbard. We test two options for integrating SAR image time series into deep learning models, namely, 3D convolutions and long short-term memory (LSTM) cells. Both proposed models achieve an intersection over union (IoU) of 0.964 on the test subset. Our results highlight the applicability of SAR data in glacier mapping with the potential to obtain glacier inventories with higher temporal resolution. We shared our dataset, code-base and pretrained models at https://github.com/konstantin-a-maslov/icemapper. See also: https://ieeexplore.ieee.org/document/10640676

• Hänsch, Ronny; Tuia, Devis & Persello, Claudio (2024). Tutorial on Machine Learning in Remote Sensing - Theory and Applications for Earth Observation. Full text in Research Archive Show summary

  Description Following the success of Deep Learning in Remote Sensing during the last years, new topics arise that address remaining challenges that are of particular importance in Earth Observation (EO) applications including interpretability and uncertainty quantification, domain shifts, label scarcity, and integrating EO imagery with other modalities such as language. The aim of this tutorial is threefold: First, to provide insights and a deep understanding of the algorithmic principles behind state-of-the-art machine learning approaches. Second, to discuss modern approaches to tackle label scarcity, interpretability, and multi-modality. Third, to illustrate the benefits and limitations of machine learning with practical examples, in particular in the context of the sustainable development goals (SDGs). Tutorial Learning Objectives to introduce sophisticated machine learning methods to show applications of ML/DL methods in large-scale scenarios to inform about common mistakes and how to avoid them to provide recommendations for increased performance

• Maslov, Konstantin & Schellenberger, Thomas (2024). Machine Learning in Remote Sensing of Glaciers. Full text in Research Archive
• Schellenberger, Thomas; Maslov, Konstantin & Persello, Claudio (2024). Towards global glacier mapping from satellite data by deep learning. Full text in Research Archive Show summary

  Glaciers are under pressure in the current climate warming trend. The aerial extent and the mass balance, a measure of mass loss or mass gain of a glacier are representative of the “health state” of a glacier. They are recognized as Essential Climate Variables by the Global Climate Observing System (GCOS) with the glacier area being an important ECV product. Due to the sheer number of glaciers of roughly 250.000 worldwide and their inaccessibility, satellite data has proven as a valuable resource to map glaciers around the globe. Nevertheless, mapping large numbers of glaciers remains a highly time-consuming task and is prone to systematic biases and errors and only a limited number of ≈870,000 outlines exist in total today. This presentation will show results from the projects MASSIVE and MATS in which we built a benchmark dataset for global glacier mapping. It is based on open optical and synthetic aperture radar satellite imagery and digital elevation models covering 9% or 19,000 glaciers worldwide. We trained a newly developed hybrid convolutional neural network-vision transformer model, GlaViTU, using glacier outlines from the Global Land Ice Measurements from Space (GLIMS) database. As the dataset contains a large variety of glacier types, we tested several training strategies for global generalisation. The best performing model achieves a high accuracy in most regions, with excellent results for ‘clean ice’ glaciers. As a novelty, we also report calibrated confidence of our estimates, increasing their reliability and interpretability. Challenges remain for calving fronts and debris-covered glaciers. To address this, we are currently investigating the use of additional input datasets such as glacier velocity maps, elevation change maps and Synthetic Aperture Radar satellite image time series as well as other machine learning approaches such as LSTMs and 3D convolutional neural networks with promising results.

• Maslov, Konstantin (2024). Vision Transformers. Full text in Research Archive
• Schellenberger, Thomas; Maslov, Konstantin & Persello, Claudio (2024). Towards global glacier mapping by deep learning and open earth observation data. Full text in Research Archive Show summary

  Glaciers are under pressure in the current climate warming trend. The aerial extent and the mass balance, a measure of mass loss or mass gain of a glacier are representative of the “health state” of a glacier. They are recognized as Essential Climate Variables by the Global Climate Observing System (GCOS) with the glacier area being an important ECV product. Due to the sheer number of glaciers of roughly 250.000 worldwide and their inaccessibility, satellite data has proven as a valuable resource to map glaciers around the globe. Nevertheless, mapping large numbers of glaciers remains a highly time-consuming task and is prone to systematic biases and errors and only a limited number of ≈870,000 outlines exist in total today. This presentation will show results from the projects MASSIVE and MATS in which we built a benchmark dataset for global glacier mapping. It is based on open optical and synthetic aperture radar satellite imagery and digital elevation models covering 9% or 19,000 glaciers worldwide. We trained a newly developed hybrid convolutional neural network-vision transformer model, GlaViTU, using glacier outlines from the Global Land Ice Measurements from Space (GLIMS) database. As the dataset contains a large variety of glacier types, we tested several training strategies for global generalisation. The best performing model achieves a high accuracy in most regions, with excellent results for ‘clean ice’ glaciers. As a novelty, we also report calibrated confidence of our estimates, increasing their reliability and interpretability. Challenges remain for calving fronts and debris-covered glaciers. To address this, we are currently investigating the use of additional input datasets such as glacier velocity maps, elevation change maps and Synthetic Aperture Radar satellite image time series as well as other machine learning approaches such as LSTMs and 3D convolutional neural networks with promising results.

• Maslov, Konstantin & Schellenberger, Thomas (2024). Introduction to Machine Learning. Full text in Research Archive
• Maslov, Konstantin; Persello, Claudio; Schellenberger, Thomas & Stein, Alfred (2023). GLAVITU: A Hybrid CNN-Transformer for Multi-Regional Glacier Mapping from Multi-Source Data. Full text in Research Archive Show summary

  Glacier mapping is essential for studying and monitoring the impacts of climate change. However, several challenges such as debris-covered ice and highly variable landscapes across glacierized regions worldwide complicate large-scale glacier mapping in a fully-automated manner. This work presents a novel hybrid CNN-transformer model (GlaViTU) for multi-regional glacier mapping. Our model outperforms three baseline models—SETR-B/16, ResU-Net and TransU-Net—achieving a higher mean IoU of 0.875 and demonstrates better generalization ability. The proposed model is also parameter-efficient, with approximately 10 and 3 times fewer parameters than SETR-B/16 and ResU-Net, respectively. Our results provide a solid foundation for future studies on the application of deep learning methods for global glacier mapping. To facilitate reproducibility, we have shared our data set, codebase and pretrained models on GitHub at https://github.com/konstantin-a-maslov/GlaViTU-IGARSS2023.

• Maslov, Konstantin; Persello, Claudio; Schellenberger, Thomas & Stein, Alfred (2023). Towards the First Global Multitemporal Data Set of Glacier Outlines. Full text in Research Archive
• Hänsch, Ronny; Rußwurm, Marc; Roscher, Ribana & Persello, Claudio (2023). IGARSS 2023 Tutorial: Machine Learning in Remote Sensing - Theory and Applications for Earth Observation. Full text in Research Archive
• Racoviteanu, Adina E.; Miles, Evan; Davies, Bethan; Bolch, Tobias; Watson, C. Scott & Buri, Pascal [Show all 14 contributors for this article] (2023). Intercomparison of remotely sensed methods for delineation of high mountain debris-covered glaciers using machine learning. Full text in Research Archive Show summary

  Glaciers are important contributors to water resources in many parts of the world, as well as constituting potential hazards. Monitoring of glaciers with remote sensing data is complicated by the presence of supraglacial debris which limits the effectiveness of semi-automated glacier delineation techniques. Furthermore, debris-covered glaciers are notoriously hard to survey in the field due to their complex, chaotic topography, and the presence of ephemeral ice cliffs and supraglacial lakes. As such, debris-covered glaciers are poorly represented in global inventories such as GLIMS and RGI, which consist of an assemblage of outlines from various dates. These inventories are subject to uncertainties due to mapping by multi-analysts using different methods. Despite recent efforts to map supraglacial debris cover at regional level, there remains an urgent need to develop a global, robust automated mapping approach based on open access data. Novel remote sensing data including high-resolution optical and radar data combined with emerging machine learning offer unique opportunities to advance current mapping methods. Here we evaluate current indices used in mapping supraglacial debris-cover and derive a best-practice workflow recommendation from particular combinations of these indices to derive glacier outlines for six representative subregions chosen globally: Khumbu and Manaslu regions (Nepal), Cordillera Blanca (Peru), Northern Patagonian Ice fields, Alaska Wrangell range, Hunza valley (Karakoram) and the Tien Shan. We present preliminary results of machine learning algorithms that combine the various remote sensing indices. The aim is to develop a robust, systematic method to map debris cover in a consistent manner at multi-temporal scales.

• Maslov, Konstantin; Schellenberger, Thomas; Persello, Claudio & Stein, Alfred (2022). Towards Fully-Automated Glacier Mapping in Arctic Regions using Deep Learning. Full text in Research Archive
• Maslov, Konstantin & Schellenberger, Thomas (2022). Deep Learning for Glacier Mapping. Full text in Research Archive
• Schellenberger, Thomas & Maslov, Konstantin (2022). Remote Sensing in Glaciology – the traditional basics. Full text in Research Archive
• Pandit, Prashant H; Persello, Claudio; Aghababaei, Hossein; Schellenberger, Thomas & Maslov, Konstantin (2022). How precisely can we map the glaciers? Full text in Research Archive

View all works in NVA