Pith. sign in

REVIEW 4 major objections 5 minor 63 references

DeepTopoNet: A Framework for Subglacial Topography Estimation on the Greenland Ice Sheets

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A residual CNN with a radar-density-aware loss reconstructs Greenland's subglacial bed from surface observations with a reported MAE of 12.49 m, outperforming interpolation and deep-learning baselines.

desk verdict DeepTopoNet's headline accuracy numbers are circular: the model trains on BedMachine and validates against BedMachine, so the R^2 of 0.99 measures self-agreement, not skill at recovering the true bed. read the letter →

arxiv 2505.23980 v2 pith:4HOT5QVY submitted 2025-05-29 cs.CV cs.LGeess.IV

Bayu Adhi Tama , Mansa Krishna , Homayra Alam , Mostafa Cham , Omar Faruque , Gong Cheng , Jianwu Wang , Mathieu Morlighem
show 1 more author
Vandana Janeja
This is my paper
classification cs.CVcs.LGeess.IV
keywords subglacialtopographyGreenlandIceSheetbedelevationestimationsparseradardataMachinedeeplearningconvolutionalneuralnetworkdynamiclossbalancing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

DeepTopoNet sets out to infer the bed topography beneath the Greenland Ice Sheet where airborne radar tracks are separated by tens of kilometers, using a convolutional network trained on surface observables such as elevation, ice velocity, thickening rate, and surface mass balance, along with gradient and trend-surface features. Its central mechanism is a dynamic loss that re-weights two targets: direct radar-derived bed measurements where they exist, and BedMachine's modeled bed everywhere else, with the balance shifting according to the radar density in each training batch. The paper reports that on the Upernavik Isstrøm region this approach reconstructs the full bed grid with a mean absolute error of 12.49 m, RMSE of 19.38 m, and $R^2=0.99$, and that it outperforms interpolation, pixel-based machine learning, and U-Net-family baselines on error and structural similarity. If true, this would give ice-sheet models a scalable way to replace tens-of-kilometer interpolation gaps with data-driven bed estimates, with direct consequences for projections of ice flow and sea-level rise.

What carries the argument

The load-bearing object is the dynamic loss-balancing mechanism. Define the radar loss $L_r$ as the mean squared error between predictions and radar-derived bed heights on radar-covered pixels, and the BedMachine loss $L_m$ as the mean squared error between predictions and BedMachine bed heights on non-radar pixels; the total is $L=\gamma_r L_r + \gamma_m L_m$ with $\gamma_r = L_m/(L_r+L_m+\epsilon)$ and $\gamma_m = L_r/(L_r+L_m+\epsilon)$. This ratio makes the radar term dominate when a batch is radar-sparse and the BedMachine term dominate when a batch is radar-dense, so the network automatically follows direct observations where they exist and falls back on BedMachine's interpolated bed in gaps, without an explicit density computation. Supporting machinery includes a residual CNN with 3x3 convolutions and skip connections, an input feature set of surface observables plus gradient and trend-surface covariates, and overlapping 16x16 patches extracted with stride 8 to capture local and contextual spatial information.

What would settle it

Take a set of radar flight lines that are never shown to the model, train DeepTopoNet on everything else, and compare its predictions on those withheld lines to the radar-measured bed; if the error there is much larger than the reported MAE of 12.49 m and RMSE of 19.38 m, the claim that the model recovers true bed topography in radar-free terrain fails.

Watch

Extended reading notes

Core claim

The paper's claim is that subglacial bed elevation can be predicted at the scale of a 150 m grid from surface data alone, provided the network is trained with a loss that adapts to the local density of radar coverage. DeepTopoNet predicts the bed $b(x,y)$ from surface elevation, ice velocity components, thickening rate, and surface mass balance, augmented by gradient features and degree-2 polynomial trend surfaces, through a residual CNN called BedTopoCNN. The total loss is $L=\gamma_r L_r + \gamma_m L_m$, where $L_r$ penalizes error against radar-derived bed heights on radar-covered pixels and $L_m$ penalizes error against BedMachine heights on non-radar pixels, with dynamic weights $\gamma_r = L_m/(L_r+L_m+\epsilon)$ and $\gamma_m = L_r/(L_r+L_m+\epsilon)$. Because $L_r$ accumulates over radar-covered pixels and $L_m$ over non-radar pixels, the weights automatically shift as batches become radar-dense or radar-sparse, so the model follows direct observations where they exist and leans on BedMachine's interpolated bed in the gaps. The authors report that this design yields the best MAE, RMSE, $R^2$, SSIM, PSNR, and terrain ruggedness agreement across four sub-regions of Upernavik Isstrøm, and that ablations removing the BedMachine loss or the stride/gradient/trend components degrade accuracy sharply.

Load-bearing premise

The evaluation treats BedMachine's map as the true bed over the whole grid, including areas where no radar exists, and the training loss pushes predictions toward that map in exactly those areas; if BedMachine is inaccurate in the gaps, the reported accuracy is inflated.

Editorial extensions

If this is right

  • If the reported accuracy holds, airborne radar flight-line gaps of tens of kilometers in Greenland's interior could be filled at 150 m resolution from surface observations, without new fieldwork.
  • The dynamic loss provides a template for fusing a sparse high-fidelity measurement with a dense model prior in other geophysical mapping tasks.
  • Ablations imply the BedMachine term is not optional: removing $L_m$ raises RMSE sharply, for example to 106.00 m in Sub-region IV, so the method's success depends on the quality of the BedMachine prior outside radar coverage.
  • Grid-based spatial holdout results suggest the network extrapolates to unmeasured transects, with an MAE of 15.90 m on Sub-region I, supporting deployment in regions without any radar tracks.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because $L_m$ trains directly against BedMachine in radar-free pixels, the full-grid metrics likely overstate skill at recovering the true bed where BedMachine itself is a simple interpolation; the honest test is held-out radar tracks, not the full-grid BedMachine comparison.
  • A natural next experiment, not reported in the paper, is to train on several catchments and test on a radar-rich catchment held out in full, or to use BedMachine only as a regularizer rather than as a training target, and then compare the two regimes.
  • The dynamic-loss ratio is specific to mean squared error losses; transferring the idea to other error metrics or to probabilistic predictions would require a new balancing rule, so the mechanism is a proof of concept rather than a universal recipe.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes DeepTopoNet, a deep convolutional framework for estimating subglacial bed elevation in the Upernavik Isstrøm region of Greenland. The method combines surface elevation, ice velocity, thickness-change rate, and surface mass balance with gradient and trend-surface features, and it trains a residual CNN using a dynamic loss that balances radar-derived ice thickness observations against BedMachine topography. The paper reports full-grid metrics (e.g., MAE 12.49 m, RMSE 19.38 m, R² 0.99 on Sub-region I), an ablation study, a spatial-slicing generalization test, and qualitative comparisons against interpolation, pixel-based, and patch-based baselines. Source code is stated to be available.

Significance. The problem addressed is important: better bed topography in radar-sparse regions would directly improve ice-sheet modeling and sea-level projections. The framework is clearly presented, the dynamic loss-balancing idea is simple and potentially transferable, and the comparison against several baseline families is a strength. The code availability is another positive feature. However, the paper's central accuracy claim is not established by the reported evaluation because the reference map used for scoring is also a training target. The significance of the contribution therefore remains conditional on independent held-out radar validation, which is currently absent.

major comments (4)
  1. [Sec. 4.4 and Sec. 5.3 (Table 2)] The headline accuracy claim is circular. In Sec. 4.4, L_m = (1/|Ω_m|) Σ (pred − m)^2 explicitly minimizes error to BedMachine in non-radar regions Ω_m, while Table 2 computes MAE, RMSE, R², SSIM, and PSNR over the full grid against the same BedMachine map. A model that simply reproduced BedMachine's non-radar interpolation would score near-perfectly under this protocol. The reported R² = 0.99 therefore measures self-agreement with the training target, not skill against true bed elevation. This is not a minor subtlety: Sec. 4.1 itself states that BedMachine is highly uncertain in slower-moving regions where it relies on simple interpolation, and Table 1 shows that Sub-region III has only 19,509 radar points. The authors must add an evaluation at held-out radar observations and report metrics separately for radar-covered and non-radar pixels.
  2. [Sec. 5.4 (Table 3)] The ablation result for 'w/o L_m' is consistent with the circularity concern rather than evidence of physical skill. Removing L_m removes supervision over the vast majority of the prediction grid, so the drastic degradation in Sub-region IV (RMSE rising to 106.00 m and R² falling to −2.16) mainly shows that the model relies on the BedMachine target. This does not demonstrate that L_m improves accuracy against true bed elevation, because true bed elevation is not independently known in Ω_m. The conclusion that L_m is 'essential for balancing high-resolution radar-derived observations with global-scale topographic trends' is therefore not supported by the reported experiment.
  3. [Sec. 5.6 (Table 6)] The spatial generalization test still scores against BedMachine, so it inherits the same circularity. Horizontally slicing the domain into odd and even bands does not create independent labels: the BedMachine values in even bands are themselves partly interpolation of radar data and are generated by the same modeling process that produced the training target. The correct test is to hold out complete radar transects or individual radar points and compare predictions at those locations against the actual radar-measured ice thickness. As written, the R² = 0.98 in Table 6 is a measure of spatial extrapolation relative to BedMachine, not a measure of accuracy against true subglacial topography.
  4. [Sec. 5.1.4] The baseline comparison appears to be unequal and possibly leaking test information. For Random Forest, the text states that 'the objective function optimized the RMSE on the test set,' which means the test set was used for hyperparameter selection. In addition, the tuning effort is not uniform across baselines, with some methods receiving extensive Optuna searches and others fixed settings. The conclusion that DeepTopoNet 'outperforms' all baselines is therefore only a claim about reproducing BedMachine under unequal tuning conditions; it does not establish superiority in estimating true bed elevation.
minor comments (5)
  1. [Sec. 4.4] There is a typo in the sentence 'The objective is to train a a deep convolutional neural network'; it should read 'train a deep convolutional neural network.'
  2. [Abstract and Conclusion] The Abstract contains an unmatched parenthesis in 'Upernavik Isstrøm)' and the Conclusion misspells the method name as 'DeepTopNet' instead of 'DeepTopoNet.'
  3. [Sec. 5.1.4] The paper does not describe how overlapping patch predictions are aggregated into the full 600×600 grid; the stitching or averaging procedure should be specified for reproducibility.
  4. [Eq. (2)] The TRI relative-difference formula uses the predicted TRI in the denominator, which can inflate or deflate the metric in smooth regions; a symmetric or reference-denominated formula would be easier to interpret.
  5. [Sec. 5.1.4] Reporting 20,000 epochs with early stopping is not enough; the actual number of epochs used, the selected hyperparameters after tuning, and representative training/validation loss curves should be reported to demonstrate convergence and avoid overfitting.

Circularity Check

3 steps flagged · score 8.0 of 10

Headline accuracy is circular: BedMachine is both the training target (L_m) and the evaluation reference, so the reported errors measure self-agreement with the training map rather than independent bed-elevation accuracy.

  1. fitted input called prediction [Sec 4.4 (Loss Function), Sec 5.1.2 (Evaluation Metrics), Sec 5.3 (Table 2)]
    "L_m = 1/|Ω_m| Σ_{(x,y)∈Ω_m}(b̂(x,y) − m(x,y))² is the BedMachine loss, evaluating prediction accuracy in non-radar regions Ω_m"

    The same map m is the supervised target in L_m over non-radar cells and the reference for the full-grid metrics in Table 2 (MAE, RMSE, R2, SSIM, PSNR), Table 4 (TRI), and Table 5 (difference maps). The reported accuracy therefore quantifies how closely DeepTopoNet reproduces its own training target, not how well it estimates independent bed elevation. A model that copied BedMachine in non-radar cells would score near-perfectly under this protocol, and baselines that are not trained with L_m are compared at a structural disadvantage, so the claim of outperforming baselines conflates fidelity to BedMachine with accuracy.

  2. self definitional [Sec 4.1 (Prior Physical Model)]
    "In this study, BedMachine data serves as a reference for the target variable, subglacial bed topography (b(x,y))."

    The target variable b in Eq. (1) is operationalized as the BedMachine map m, which is also the quantity minimized by L_m. The prediction task is therefore defined as reproducing BedMachine, while the abstract and conclusion present this as reconstructing real subglacial terrain. Since b is defined through m, the claim that DeepTopoNet infers bed topography beyond BedMachine is not supported by the reported target definition; the paper itself admits BedMachine is highly uncertain in slower moving regions where it is based on a simple interpolation scheme.

1 more flagged steps
  1. other [Sec 5.6 (Spatial Generalization Analysis), Table 6]
    "Odd-numbered slices (1, 3, 5, ...) are used for training and even-numbered slices (2, 4, 6, ...) for testing. This design mimics realistic extrapolation to unobserved transects, reflecting operational deployment where radar coverage is spatially sparse."

    The spatial slicing changes which cells are in the training set, but the even-slice metrics are still computed against the same BedMachine map that L_m was designed to reproduce. No held-out radar transect is used as independent ground truth, so the experiment still measures agreement with a model-generated target rather than with true bed elevation measurements. The slicing protocol therefore does not break the loss/evaluation circularity; it only relocates it to unseen spatial bands.

full rationale

The central accuracy claim is not independently validated. DeepTopoNet's loss function explicitly trains the network to match BedMachine m in non-radar regions (L_m, Sec 4.4), and every reported full-grid metric (MAE, RMSE, R2, SSIM, PSNR in Table 2; TRI in Table 4; qualitative difference maps in Table 5) is computed against the same m. This is a textbook fitted-input-called-prediction reduction: the evaluation target is identical to the training target, so the headline numbers (MAE 12.49 m, RMSE 19.38 m, R2 0.99) measure self-agreement with the training map, not skill at recovering true bed elevation. The ablation in Table 3 strengthens this reading: removing L_m collapses Sub-region IV from RMSE 7.65 m and R2 0.98 to RMSE 106.00 m and R2 −2.16, showing that the non-radar predictions are inherited from the BedMachine target rather than learned from radar or physics. The spatial-slicing experiment (Sec 5.6) uses BedMachine as its reference as well, so it does not provide independent radar-based validation. The paper contains no held-out radar evaluation, no comparison to in-situ ice-thickness measurements outside the training set, and no external bed-elevation ground truth. The BedMachine references [37,40] are not themselves the circularity problem; the problem is that this paper uses the same map as both the training label and the evaluation reference. Score 8 reflects that the reported accuracy claim reduces by construction to a fit against the training target; only the radar-covered L_r term provides independent signal, and full-grid metrics are dominated by the non-radar cells where L_m operates.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on treating BedMachine as ground truth while simultaneously training against it, which is the key circularity. The remaining free parameters are standard architecture and feature choices; none are fitted to independent physical data.

free parameters (4)
  • Trend surface polynomial degree = 2
    Chosen by hand for trend surface features; affects global spatial context.
  • Patch size and stride = 16x16, stride 8
    Overlapping patch extraction hyperparameters; ablation shows stride matters for performance.
  • Network width = 32 to 256 filters, 5 residual blocks
    Architecture hyperparameters chosen without systematic search; central to model capacity.
  • Loss smoothing epsilon = small (not specified)
    Added to avoid division by zero in dynamic weights; value not reported.
assumptions (4)
  • domain assumption BedMachine is sufficiently accurate to serve as ground truth for full-grid bed elevation evaluation.
    Invoked in Sec 4.1 and Sec 5.3; if false, the reported accuracy does not reflect true bed elevation.
  • domain assumption Radar-derived ice thickness values at flight lines are accurate bed elevation measurements.
    Used in L_r in Sec 4.4; radar data are sparse and have their own errors, but the paper treats them as ground truth.
  • domain assumption Input features (surface elevation, velocity, dh/dt, SMB) are co-registered and contemporaneous.
    Dataset described in Sec 5.1.1 without alignment details; misalignment would degrade predictions.
  • ad hoc to paper The inverse-loss weighting gamma_r = L_m/(L_r+L_m) improves sparse-region generalization.
    Proposed in Sec 4.4 without theoretical justification or comparison to fixed weights; only ablation on one region.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DeepTopoNet: A Framework for Subglacial Topography Estimation on the Greenland Ice Sheets." pith.science (2026). https://pith.science/paper/4HOT5QVY

@misc{pith2026250523980,
  author       = {Pith},
  title        = {Pith review of: DeepTopoNet: A Framework for Subglacial Topography Estimation on the Greenland Ice Sheets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4HOT5QVY}},
  note         = {Machine review of arXiv:2505.23980}
}
read the original abstract

Understanding Greenland's subglacial topography is critical for projecting the future mass loss of the ice sheet and its contribution to global sea-level rise. However, the complex and sparse nature of observational data, particularly information about the bed topography under the ice sheet, significantly increases the uncertainty in model projections. Bed topography is traditionally measured by airborne ice-penetrating radar that measures the ice thickness directly underneath the aircraft, leaving data gap of tens of kilometers in between flight lines. This study introduces a deep learning framework, which we call as DeepTopoNet, that integrates radar-derived ice thickness observations and BedMachine Greenland data through a novel dynamic loss-balancing mechanism. Among all efforts to reconstruct bed topography, BedMachine has emerged as one of the most widely used datasets, combining mass conservation principles and ice thickness measurements to generate high-resolution bed elevation estimates. The proposed loss function adaptively adjusts the weighting between radar and BedMachine data, ensuring robustness in areas with limited radar coverage while leveraging the high spatial resolution of BedMachine predictions i.e. bed estimates. Our approach incorporates gradient-based and trend surface features to enhance model performance and utilizes a CNN architecture designed for subgrid-scale predictions. By systematically testing on the Upernavik Isstr{\o}m) region, the model achieves high accuracy, outperforming baseline methods in reconstructing subglacial terrain. This work demonstrates the potential of deep learning in bridging observational gaps, providing a scalable and efficient solution to inferring subglacial topography.

Figures

Figures reproduced from arXiv: 2505.23980 by the authors.

Figure 1
Figure 1. Overview of the proposed DeepTopoNet framework, including feature extraction, BedTopoCNN architecture, and [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the proposed BedTopoCNN model. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visualization of BedMachine topography with radar [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 54 canonical work pages

  1. [1]

    Frits Agterberg. 2023. Trend surface analysis. In Encyclopedia of Mathematical Geosciences. Springer, 1567–1575

    work page 2023

  2. [2]

    Henning Åkesson, Mathieu Morlighem, Matt O’Regan, and Martin Jakobsson

    work page

  3. [3]

    Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. 2019. Optuna: A next-generation hyperparameter optimization frame- work. In Proceedings of the 25th ACM SIGKDD international conference on knowl- edge discovery & data mining . 2623–2631

    work page 2019

  4. [4]

    Gabriel Appleby, Linfeng Liu, and Li-Ping Liu. 2020. Kriging convolutional networks. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 34. 3187–3194

    work page 2020

  5. [5]

    Jonathan L Bamber, JA Griggs, RTW Hurkmans, JA Dowdeswell, SP Gogineni, Ian Howat, Jeremie Mouginot, John Paden, Steven Palmer, Eric Rignot, et al. 2013. A new bed elevation dataset for Greenland. The Cryosphere 7, 2 (2013), 499–510

    work page 2013

  6. [6]

    Jonathan L Bamber, Russell L Layberry, and SP Gogineni. 2001. A new ice thickness and bed data set for the Greenland ice sheet: 1. Measurement, data reduction, and errors. Journal of Geophysical Research: Atmospheres 106, D24 (2001), 33773–33780

    work page 2001

  7. [7]

    Jowan M Barnes, Thiago Dias dos Santos, Daniel Goldberg, G Hilmar Gudmunds- son, Mathieu Morlighem, and Jan De Rydt. 2021. The transferability of adjoint inversion products between different ice flow models.The Cryosphere 15, 4 (2021), 1975–2000

    work page 2021

  8. [8]

    Sara Beery, Grant Van Horn, and Pietro Perona. 2018. Recognition in terra incognita. In Proceedings of the European conference on computer vision (ECCV) . 456–473

    work page 2018

Show all 63 references
  1. [9]

    Ian C Briggs. 1974. Machine contouring using minimum curvature. Geophysics 39, 1 (1974), 39–48

    1974

  2. [10]

    Yiheng Cai, Yanliang He, Shinan Lang, Xiangbin Cui, Xiaoqing Zhang, and Zijun Yao. 2025. Siamese topographic generation model: A deep learning model for generating Antarctic subglacial topography with fine details. Computers & Geosciences (2025), 105857

    2025

  3. [11]

    John Canny. 1986. A computational approach to edge detection.IEEE Transactions on pattern analysis and machine intelligence 6 (1986), 679–698

    1986

  4. [12]

    Gong Cheng, Mathieu Morlighem, and Sade Francis. 2024. Forward and inverse modeling of ice sheet flow using physics-informed neural networks: Application to Helheim Glacier, Greenland.Journal of Geophysical Research: Machine Learning and Computation 1, 3 (2024), e2024JH000169....

    2024

  5. [13]

    Haoan Feng, Yunting Song, and Leila De Floriani. 2024. Critical Features Tracking on Triangulated Irregular Networks by a Scale-Space Method. In Proceedings of the 32nd ACM International Conference on Advances in Geographic Information Systems. 54–66

    2024

  6. [14]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition . 770–778

    2016

  7. [15]

    Alfa RH Heryudono and Tobin A Driscoll. 2010. Radial basis function interpola- tion on irregular domain through conformal transplantation. Journal of Scientific Computing 44 (2010), 286–300

    2010

  8. [16]

    Alain Hore and Djemel Ziou. 2010. Image quality metrics: PSNR vs. SSIM. In 2010 20th international conference on pattern recognition . IEEE, 2366–2369

    2010

  9. [17]

    Ian M Howat, A Negrete, and Benjamin E Smith. 2014. The Greenland Ice Mapping Project (GIMP) land classification and surface elevation data sets. The Cryosphere 8, 4 (2014), 1509–1518

    2014

  10. [18]

    Huimin Huang, Lanfen Lin, Ruofeng Tong, Hongjie Hu, Qiaowei Zhang, Yutaro Iwamoto, Xianhua Han, Yen-Wei Chen, and Jian Wu. 2020. Unet 3+: A full- scale connected unet for medical image segmentation. In ICASSP 2020-2020 IEEE international conference on acoustics, speech and sig...

    2020

  11. [19]

    Michael F Hutchinson. 1995. Interpolating mean rainfall using thin plate smooth- ing splines. International journal of geographical information systems 9, 4 (1995), 385–403

    1995

  12. [20]

    David Jacobs. 2005. Image gradients. Class Notes for CMSC 426 (2005), 1–3

    2005

  13. [21]

    Kevin Johnston, Jay M Ver Hoef, Konstantin Krivoruchko, and Neil Lucas. 2001. Using ArcGIS geostatistical analyst. Vol. 380. Esri Redlands

    2001

  14. [22]

    Ian Joughin, Ben E Smith, Ian M Howat, Ted Scambos, and Twila Moon. 2010. Greenland flow variability from ice-sheet-wide velocity mapping. Journal of Glaciology 56, 197 (2010), 415–430

    2010

  15. [23]

    Alex Kendall, Yarin Gal, and Roberto Cipolla. 2018. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition . 7482–7491

    2018

  16. [24]

    TN Kipf and M Welling. 2017. Semi-Supervised Classification with Graph Con- volutional Networks.. In ICLR

    2017

  17. [25]

    Nina Siu-Ngan Lam. 1983. Spatial interpolation methods: a review. The American Cartographer 10, 2 (1983), 129–150

    1983

  18. [26]

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. 2015. Deep learning. nature 521, 7553 (2015), 436–444

    2015

  19. [27]

    Wei Ji Leong and Huw Joseph Horgan. 2020. DeepBedMap: a deep neural network for resolving the bed topography of Antarctica. The Cryosphere 14, 11 (2020), 3687–3705

    2020

  20. [28]

    Jin Li and Andrew D Heap. 2008. A review of spatial interpolation methods for environmental scientists. (2008)

    2008

  21. [29]

    Jia Li, Yanyan Shen, Lei Chen, and Charles Wang Wai Ng. 2023. Rainfall spatial interpolation with graph neural networks. In International conference on database systems for advanced applications . Springer, 175–191

    2023

  22. [30]

    Xin Li, Guodong Cheng, and Ling Lu. 2000. Comparison of spatial interpolation methods. Advances in Earth science 15, 3 (2000), 260

    2000

  23. [31]

    Yang Li, Si Si, Gang Li, Cho-Jui Hsieh, and Samy Bengio. 2021. Learnable fourier features for multi-dimensional spatial positional encoding. Advances in Neural Information Processing Systems 34 (2021), 15816–15829

    2021

  24. [32]

    Douglas R MacAyeal. 1989. Large-scale ice flow over a viscous basal sediment: Theory and application to ice stream B, Antarctica. Journal of Geophysical Research: Solid Earth 94, B4 (1989), 4071–4087

    1989

  25. [33]

    Daniel F Merriam and RH Lippert. 1966. Geologic model studies using trend- surface analysis. The Journal of Geology 74, 3 (1966), 344–357

    1966

  26. [34]

    Hanna Meyer, Christoph Reudenbach, Stephan Wöllauer, and Thomas Nauss

  27. [35]

    Romain Millan, Jérémie Mouginot, Antoine Rabatel, and Mathieu Morlighem

  28. [36]

    LW Morland. 1987. Unconfined ice-shelf flow. Dynamics of the West Antarctic ice sheet 4 (1987), 99–116

    1987

  29. [37]

    Mathieu Morlighem, Eric Rignot, Tobias Binder, Donald Blankenship, Reinhard Drews, Graeme Eagles, Olaf Eisen, Fausto Ferraccioli, René Forsberg, Peter Fretwell, et al. 2020. Deep glacial troughs and stabilizing ridges unveiled beneath the margins of the Antarctic ice sheet. Na...

    2020

  30. [38]

    Mathieu Morlighem, Eric Rignot, Jeremie Mouginot, Helene Seroussi, and Eric Larour. 2014. Deeply incised submarine glacial valleys beneath the Greenland ice sheet. Nature Geoscience 7, 6 (2014), 418–422

    2014

  31. [39]

    Mathieu Morlighem, E Rignot, Hélene Seroussi, Eric Larour, H Ben Dhia, and Denis Aubry. 2011. A mass conservation approach for mapping glacier ice thickness. Geophysical Research Letters 38, 19 (2011)

    2011

  32. [40]

    Mathieu Morlighem, Chris N Williams, Eric Rignot, Lu An, Jan Erik Arndt, Jonathan L Bamber, Ginny Catania, Nolwenn Chauché, Julian A Dowdeswell, Boris Dorschel, et al. 2017. BedMachine v3: Complete bed topography and ocean bathymetry mapping of Greenland from multibeam echo so...

    2017

  33. [41]

    Brice Noël, Willem Jan Van De Berg, J Melchior Van Wessem, Erik Van Meijgaard, DIrk Van As, Jan Lenaerts, Stef Lhermitte, Peter Kuipers Munneke, CJP Smeets, Lambertus H Van Ulft, et al. 2018. Modelling the climate and surface mass balance of polar ice sheets using RACMO2–Part ...

    2018

  34. [42]

    Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazu- nari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. 2018. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999 (2018)

    2018 arXiv

  35. [43]

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. 2019. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computa- tional physics 378 (2019), 686–707

    2019

  36. [44]

    J Reily Shawn, D DeGloria Stephen, and A Elliot Robert. 1999. Terrain Ruggedness Index That Quantifies Topographic Heterogeneity. Intermountain Journal of Science 5, 1-4 (1999), 23–27

    1999

  37. [45]

    Edgar Riba, Dmytro Mishkin, Daniel Ponsa, Ethan Rublee, and Gary Bradski

  38. [46]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolu- tional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceeding...

    2015

  39. [47]

    Aleksandar Sekulić, Milan Kilibarda, Gerard BM Heuvelink, Mladen Nikolić, and Branislav Bajat. 2020. Random forest spatial interpolation. Remote Sensing 12, 10 (2020), 1687

    2020

  40. [48]

    Irwin Edward Sobel. 1970. Camera models and machine perception . stanford university

    1970

  41. [49]

    Xavier Soria, Yachuan Li, Mohammad Rouhani, and Angel D Sappa. 2023. Tiny and efficient model for the edge detection generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 1364–1373

    2023

  42. [50]

    Lee, and J

    Core Writing Team, H. Lee, and J. Romero (Eds.). 2023. Climate Change 2023: Synthesis Report. Intergovernmental Panel on Climate Change (IPCC), Geneva, Switzerland. https://www.ipcc.ch/report/ar6/syr/

    2023

  43. [51]

    The IMBIE Team. 2020. Mass balance of the Greenland Ice Sheet from 1992 to

    2020

  44. [52]

    Hans Wackernagel and Hans Wackernagel. 2003. Ordinary kriging. Multivariate geostatistics: an introduction with applications (2003), 79–88

    2003

  45. [53]

    Hans Wackernagel and Hans Wackernagel. 2003. Universal kriging. Multivariate Geostatistics: An Introduction with Applications (2003), 300–307

    2003

  46. [54]

    Lijing Wang, Luk Peeters, Emma J MacKie, Zhen Yin, and Jef Caers. 2023. Unrav- eling the uncertainty of geological interfaces through data-knowledge-driven trend surface analysis. Computers & Geosciences 178 (2023), 105419

    2023

  47. [55]

    Ximei Wang, Jinghan Gao, Mingsheng Long, and Jianmin Wang. 2021. Self-tuning for data-efficient deep learning. In International Conference on Machine Learning . PMLR, 10738–10748

    2021

  48. [56]

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing 13, 4 (2004), 600–612

    2004

  49. [57]

    Katherine Yi, Angelina Dewar, Tartela Tabassum, Jason Lu, Ray Chen, Homayra Alam, Omar Faruque, Sikan Li, Mathieu Morlighem, and Jianwu Wang. 2023. Evaluating Machine Learning and Statistical Models for Greenland Subglacial Bed Topography. In 2023 International Conference on M...

    2023

  50. [58]

    Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. 2018. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Worksho...

    2018

  51. [2018]

    Nature 579, 7798 (2020), 233–239

    2020

  52. [2019]

    Ecological Modelling 411 (2019), 108815

    Importance of spatial predictor variable selection in machine learning applications–Moving from data reproduction to spatial prediction. Ecological Modelling 411 (2019), 108815

    2019

  53. [2020]

    In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Kornia: an open source differentiable computer vision library for pytorch. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 3674–3683

  54. [2021]

    Journal of Geophysical Research: Earth Surface 126, 6 (2021), e2020JF005921

    Future projections of Petermann Glacier under ocean warming depend strongly on friction law. Journal of Geophysical Research: Earth Surface 126, 6 (2021), e2020JF005921

    2021

  55. [2022]

    Nature Geoscience 15, 2 (2022), 124–129

    Ice velocity and thickness of the world’s glaciers. Nature Geoscience 15, 2 (2022), 124–129

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.