13 February 2020 in The Cryosphere
We present a novel approach to simulate and reconstruct annual glacier-wide surface mass balance (SMB) series based on a deep artificial neural network (ANN; i.e. deep learning). This method has been included as the SMB component of an open-source regional glacier evolution model. While most glacier models tend to incorporate more and more physical processes, here we take an alternative approach by creating a parameterized model based on data science. Annual glacier-wide SMBs can be simulated from topo-climatic predictors using either deep learning or Lasso (least absolute shrinkage and selection operator; regularized multilinear regression), whereas the glacier geometry is updated using a glacier-specific parameterization. We compare and cross-validate our nonlinear deep learning SMB model against other standard linear statistical methods on a dataset of 32 French Alpine glaciers. Deep learning is found to outperform linear methods, with improved explained variance (up to +64 % in space and +108 % in time) and accuracy (up to +47 % in space and +58 % in time), resulting in an estimated r2 of 0.77 and a root-mean-square error (RMSE) of 0.51 m w.e. Substantial nonlinear structures are captured by deep learning, with around 35 % of nonlinear behaviour in the temporal dimension. For the glacier geometry evolution, the main uncertainties come from the ice thickness data used to initialize the model. These results should encourage the use of deep learning in glacier modelling as a powerful nonlinear tool, capable of capturing the nonlinearities of the climate and glacier systems, that can serve to reconstruct or simulate SMB time series for individual glaciers in a whole region for past and future climates.
20 January 2022 in Nature Communications
Glaciers and ice caps are experiencing strong mass losses worldwide, challenging water availability, hydropower generation, and ecosystems. Here, we perform the first-ever glacier evolution projections based on deep learning by modelling the 21st century glacier evolution in the French Alps. By the end of the century, we predict a glacier volume loss between 75 and 88%. Deep learning captures a nonlinear response of glaciers to air temperature and precipitation, improving the representation of extreme mass balance rates compared to linear statistical and temperature-index models. Our results confirm an over-sensitivity of temperature-index models, often used by large-scale studies, to future warming. We argue that such models can be suitable for steep mountain glaciers. However, glacier projections under low-emission scenarios and the behaviour of flatter glaciers and ice caps are likely to be biased by mass balance models with linear sensitivities, introducing long-term biases in sea-level rise and water resources projections.
16 November 2023 in Geoscientific Model Development
Geoscientific models are facing increasing challenges to exploit growing datasets coming from remote sensing. Universal differential equations (UDEs), aided by differentiable programming, provide a new scientific modelling paradigm enabling both complex functional inversions to potentially discover new physical laws and data assimilation from heterogeneous and sparse observations. We demonstrate an application of UDEs as a proof of concept to learn the creep component of ice flow, i.e. a nonlinear diffusivity differential equation, of a glacier evolution model. By combining a mechanistic model based on a two-dimensional shallow-ice approximation partial differential equation with an embedded neural network, i.e. a UDE, we can learn parts of an equation as nonlinear functions that then can be translated into mathematical expressions. We implemented this modelling framework as ODINN.jl, a package in the Julia programming language, providing high performance, source-to-source automatic differentiation (AD) and seamless integration with tools and global datasets from the Open Global Glacier Model in Python. We demonstrate this concept for 17 different glaciers around the world, for which we successfully recover a prescribed artificial law describing ice creep variability by solving ∼ 500 000 ordinary differential equations in parallel. Furthermore, we investigate which are the best tools in the scientific machine learning ecosystem in Julia to differentiate and optimize large nonlinear diffusivity UDEs. This study represents a proof of concept for a new modelling framework aiming at discovering empirical laws for large-scale glacier processes, such as the variability in ice creep and basal sliding for ice flow, and new hybrid surface mass balance models.
21 February 2025 in The Cryosphere
KSKamilla Hauknes Sjursen
Glacier retreat presents significant environmental and social challenges. Understanding the local impacts of climatic drivers on glacier evolution is crucial, with mass balance being a central concept. This study introduces miniML-MB, a new minimal machine-learning model designed to estimate annual point surface mass balance (PMB) for very small datasets. Based on an eXtreme Gradient Boosting (XGBoost) architecture, miniML-MB is applied to model PMB at individual sites in the Swiss Alps, emphasising the need for an appropriate training framework and dimensionality reduction techniques. A substantial added value of miniML-MB is its data-driven identification of key climatic drivers of local mass balance. The best PMB prediction performance was achieved with two predictors: mean air temperature (May–August) and total precipitation (October–February). miniML-MB models PMB accurately from 1961 to 2021, with a mean absolute error (MAE) of 0.417 m w.e. across all sites. Notably, miniML-MB demonstrates similar and, in most cases, superior predictive capabilities compared to a simple positive degree-day (PDD) model (MAE of 0.541 m w.e.). Compared to the PDD model, miniML-MB is less effective at reproducing extreme mass balance values (e.g. 2022) that fall outside its training range. As such, miniML-MB shows promise as a gap-filling tool for sites with incomplete PMB measurements as long as the missing year's climate conditions are within the training range. This study underscores potential means for further refinement and broader applications of data-driven approaches in glaciology.
3 September 2020 in Earth System Science Data
Glacier mass balance (MB) data are crucial to understanding and quantifying the regional effects of climate on glaciers and the high-mountain water cycle, yet observations cover only a small fraction of glaciers in the world. We present a dataset of annual glacier-wide mass balance of all the glaciers in the French Alps for the 1967–2015 period. This dataset has been reconstructed using deep learning (i.e. a deep artificial neural network) based on direct MB observations and remote-sensing annual estimates, meteorological reanalyses and topographical data from glacier inventories. The method's validity was assessed previously through an extensive cross-validation against a dataset of 32 glaciers, with an estimated average error (RMSE) of 0.55 , an explained variance (r2) of 75 % and an average bias of −0.021 . We estimate an average regional area-weighted glacier-wide MB of −0.69 ±0.21 (1σ) for the 1967–2015 period with negative mass balances in the 1970s (−0.44 ), moderately negative in the 1980s (−0.16 ) and an increasing negative trend from the 1990s onwards, up to −1.26 in the 2010s. Following a topographical and regional analysis, we estimate that the massifs with the highest mass losses for the 1967–2015 period are the Chablais (−0.93 ), Champsaur (−0.86 ), and Haute-Maurienne and Ubaye ranges (−0.84 each), and the ones presenting the lowest mass losses are the Mont-Blanc (−0.68 ), Oisans and Haute-Tarentaise ranges (−0.75 each). This dataset – available at https://doi.org/10.5281/zenodo.3925378 (Bolibar et al., 2020 a) – provides relevant and timely data for studies in the fields of glaciology, hydrology and ecology in the French Alps in need of regional or glacier-specific annual net glacier mass changes in glacierized catchments.
25 July 2024 in Journal of Advances in Modeling Earth Systems
Because Antarctic surface melt is mostly driven by local processes, its simulation necessitates high‐resolution regional climate models (RCMs). However, the current horizontal resolution of RCMs (≈25–30 km) is inadequate for capturing small‐scale melt processes. To address this limitation, we present SUPREME (SUPer‐REsolution‐based Melt Estimation over Antarctica), a deep learning method to downscale surface melt to 5.5 km resolution using a physically‐informed super‐resolution model. The physical information integrated into the model originates from observations tied to surface melt, specifically remote sensing‐derived albedo and elevation. These remote sensing data, in addition to a Regional Atmospheric Climate Model (RACMO) run at 27 km resolution, account for the diverse drivers of surface melt across Antarctica, facilitating effective generalization beyond the training region of the Antarctic Peninsula. A comparison of SUPREME with a dynamically downscaled RACMO run at 5.5 km over the Antarctic Peninsula shows high accuracy, with average yearly RMSE and bias of 5.5 mm w.e. yr−1 and 4.5 mm w.e. yr−1, respectively. Validation at five automatic weather stations reveals SUPREME's marked improvement with substantially lower average RMSE (81 mm w.e.) compared to RACMO 27 km (129 mm w.e.). Beyond the training region, SUPREME aligns more closely with remote sensing products associated with surface melt than super‐resolution models lacking physical constraints. While further validation of SUPREME is needed, our study highlights the potential of super‐resolution techniques with physical constraints for high‐resolution surface melt monitoring in Antarctica, providing insights into the impacts of localized melting on processes affecting ice shelf integrity such as hydrofracturing.