Mountain glaciers and ice sheets are undergoing rapid loss of ice mass. To assess their current volume and ice-thickness distribution, and to accurately forecast their future, ground-penetrating radar systems are employed to measure ice thickness. These instruments produce radio-echo sounding data that reveal the internal structure of the ice and its boundaries at the surface and beneath the ice body. However, delineating these interfaces is time-consuming as the surveyed area can span several hundred kilometers. To accelerate the labelling process, automated methods have emerged, but they often struggle to delineate the ice bottom, particularly in temperate glaciers. The high water content inherent in temperate ice increases both scattering and attenuation of the radar signal, resulting in weak and ambiguous reflections at the ice bottom and generally noisier data. As a result, the visual features corresponding to the ice bottom become ambiguous. To address this issue, we propose several extensions to deep learning models that incorporate additional contextual information into the prediction. Specifically, we integrate modifications that incorporate additional spatial and geo-spatial contextual information into the network. With these enhancements, we achieve a new state-of-the-art performance on the IceAnatomy benchmark dataset, with mean errors of 10.9 m for the FAU subset and 43.4 m for the CReSIS subset in a single combined system. We also demonstrate that our method generalises well to previously unseen data, outperforming the baseline by 49.6 % on new campaign data from Patagonia. The model code and the new dataset are available at https://github.com/ki7077/ICIBS/tree/main and https://zenodo.org/records/23392580.
Figures & tables
Fig. 1: Comparing how deep learning models commonly view the problem in contrast to humans: Deep learning models typically approach radargram interpretation using a restricted spatial context to reduce computational complexity, relying primarily on local visual features for boundary prediction. In contrast, human annotators interpret radar data within a substantially broader spatial framework and additionally incorporate information from geospatially proximate radargrams to guide their decisions. The corresponding ice bottom is marked in red . The borders of the radargrams are color-coded according to their corresponding flight in gray and dark yellow .
Fig. 2: Overview of our proposed architecture, adapted from Dreier et al. [ 11 ] . The model receives three geospatially proximate radargram sections, from which three context windows are cropped, as indicated by the outline ( small , medium , and large ). To enable joint processing, the nine radargrams are normalized to a consistent size. For each radargram, the network predicts both the ice surface and the ice bottom interfaces. While both outputs are generated during training, the ice surface prediction task serves exclusively as an auxiliary task and is not considered for the evaluation.
Fig. 3: Illustration of the proposed sampling strategy for geospatially adjacent radargrams using flight lines acquired over the Perito Moreno Glacier. The background schematically represents the glacier environment, distinguishing glacier ice , rock , and water . Individual flight lines are shown in green , gray , blue , pink , and dark yellow . The sampling procedure first selects an anchor point and then samples two additional reference points within a predefined radius r . Candidates from nearby, distinct flight lines are prioritized to promote diversity among the sampled radargrams. If no suitable candidates are available within this neighborhood, the procedure falls back to sampling more distant points along the same flight line.
Fig. 4: geoconv module: The inputs are three feature maps. The module first iteratively samples eight traces that maximize the geographical distance from the previously sampled traces (marked with X ). For each sampled trace, the eight nearest traces (outlined in black) are then sampled to approximate the local region. A ResBlock aggregates the local information into the eight initial traces. Finally, two additional residual blocks propagate the aggregated information to every input trace, weighting the eight sampled traces according to their geographical distance from the corresponding input trace.
Fig. 5: Preprocessing of the spatial context. An initial context patch of size 2048×1024 pixels is sampled from the dataset. Two smaller subregions of size 1024×1024 and 512×1024 pixels are then extracted from its center. Finally, all three patches are resized to 512×1024 pixels to ensure a consistent input format.
Fig. 6: Illustration of the proposed hooking module. The module takes as input three vertically centered radargrams with identical dimensions but different granularity levels. First, the corresponding feature maps are rescaled to match the highest granularity among the three inputs. The central regions of the two larger rescaled feature maps are then cropped to match the spatial extent of the finest-resolution feature map. Next, the aligned feature maps are normalized and processed with a one-dimensional convolution to exchange information across the different context windows. The resulting feature representation is subsequently merged with the uncropped feature maps, and the outputs are reshaped to the original input dimensions for further processing by the network.
Ice Boundary Delineation
FAU
CReSIS
Model
MAE ↓
MME ↓
AP-1% ↑
AP-5% ↑
MAE ↓
MME ↓
AP-1% ↑
AP-5% ↑
Baseline
25.8±5.0
19.1±4.0
67.9±2.4
91.8±1.6
16.4±4.7
84.5±22.7
86.3±1.3
93.0±1.1
Ours
18.8±4.2
13.7±3.4
70.5±5.3
94.0±1.1
11.0±1.9
56.9±9.2
88.8±1.3
95.3±1.0
+Ensemble
15.3
10.9
73.4
95.2
8.3
43.4
91.0
96.3
TABLE I: This table summarizes the quantitative comparison of our proposed model with the baseline on the standardized splits of the IceAnatomy dataset. Both approaches were trained on the FAU and CReSIS subsets simultaneously across five independent runs, and the reported values represent the mean and standard deviation. The ensembling model consists of the combination of the five trained models. A more detailed version of the results can be found in the Appendix B in Tab. IV and V .
Fig. 7: Qualitative results of the baseline and the ensemble model. Left: results from the third flight over the Perito Moreno Glacier portion of the FAU subset. Right: section from the TSK2-07 mission in the West Antarctica portion of the CReSIS subset. The yellow line indicates the ice bottom prediction, while the red dotted line indicates the ground truth.
Ice Boundary Delineation
FAU
CReSIS
Model
MAE ↓
MME ↓
AP-1% ↑
AP-5% ↑
MAE ↓
MME ↓
AP-1% ↑
AP-5% ↑
Baseline
50.9±8.1
42.7±6.8
48.5±0.7
81.4±1.1
15.4±1.4
87.2±8.2
82.3±1.2
91.8±0.7
+GeoConv
42.7±2.3
35.8±1.9
49.4±1.2
83.7±0.5
14.7±1.2
83.4±6.2
82.7±2.0
92.5±0.7
+Hooking
36.7±5.0
30.8±4.2
50.8±1.9
85.9±1.4
13.5±1.0
75.9±6.1
84.2±2.1
93.4±1.0
+Ensemble
26.1
21.9
55.2
89.0
11.1
61.4
87.5
94.8
TABLE II: Results on the IceAnatomy dataset using the newly defined region-exclusive splits. To reduce the influence of statistical outliers, each experiment is repeated five times, and the mean and standard deviation are reported. For the ensemble, results are reported for a single ensemble instance, as repeating the experiment five times would require training 5×5 individual models and incur substantial computational cost. A detailed breakdown of the results is provided in Appendix B , Tables VI and VII .
Fig. 8: Qualitative comparison of model predictions as the proposed modifications are progressively incorporated. The left panels show radargrams from the first flight over the Perito Moreno Glacier, included in the FAU subset, while the right panels show a radargram from the PEN2-01 mission in the CReSIS subset. The yellow line denotes the predicted ice-bottom, whereas the red dotted line denotes the ground-truth annotation.
Region
MAE ↓
MME ↓
AP-1% ↑
AP-5% ↑
Baseline
80.8±17.4
67.8±14.6
50.9±1.0
80.9±0.6
+GeoConv
57.8±8.0
48.5±6.7
55.8±2.2
86.2±1.0
+Hooking
51.5±13.9
43.2±11.7
52.7±4.3
86.8±2.1
+Ensemble
40.8
34.2
57.2
89.4
TABLE III: Quantitative evaluation of our subset model on radio-echo sounding data from the Upsala Glacier. Results are reported as the mean and standard deviation across five independent runs. The ensemble model combines the predictions of the five individually trained models.
Fig. 9: Qualitative comparison of the baseline and ensemble model predictions on the Upsala Glacier. The yellow line denotes the predicted ice-bottom, whereas the red dotted line denotes the ground-truth annotation.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Mission
Baseline
Ensemble
mme ↓
map1 ↑
mme ↓
map1 ↑
West Antarctica
TSK2-04
192.5
68.7
103.8
74.0
TSK2-07
373.6
74.7
146.3
85.2
PIG4-01
60.1
83.7
21.1
93.6
PIG4-02
68.9
86.6
57.0
88.9
Appendix
TABLE IV: Comparison of the proposed ensemble model and the baseline method across different geographical settings in the CReSIS subset. Both models are trained on the FAU and CReSIS subsets using the official train-test splits of the Ice-Anatomy benchmark dataset.
Mission
Baseline
Ensemble
mme ↓
map1 ↑
mme ↓
map1 ↑
Perito Moreno 3
46.2
42.6
22.2
51.9
Viedma 2
26.3
58.0
15.8
64.4
Gourdon 2
5.0
83.3
3.8
86.8
Appendix
TABLE V: Comparison of the proposed modifications and the baseline method across different geographical settings in the FAU subset. The models are trained on both the FAU and CReSIS subsets using the official splits of the Ice-Anatomy benchmark dataset.
Mission
Baseline
+GeoConv
+Hooking
+Ensemble
mme ↓
map1 ↑
mme ↓
map1 ↑
mme ↓
map1 ↑
mme ↓
map1 ↑
PEN4-01
36.7
92.6
33.7
93.5
27.1
93.8
18.0
94.6
PEN5-01
69.8
82.3
71.2
81.3
62.6
82.8
50.4
86.8
PEN5-02
90.6
76.8
93.1
78.0
81.1
80.2
68.0
84.7
PEN3-01
56.7
87.8
53.8
88.5
45.8
89.6
31.2
94.0
PEN3-02
197.7
65.0
230.6
60.0
208.8
64.0
186.9
69.5
Appendix
TABLE VI: Effect of the proposed modifications on the CReSIS subset model evaluated using region-exclusive splits across different flights. The best-performing configurations are highlighted in bold .
Mission
Baseline
+GeoConv
+Hooking
+Ensemble
mme ↓
map1 ↑
mme ↓
map1 ↑
mme ↓
map1 ↑
mme ↓
map1 ↑
Perito Moreno 1a
42.9
49.6
41.7
50.0
34.9
50.8
27.6
55.7
Perito Moreno 1b
37.4
51.4
26.1
51.2
20.5
53.3
11.6
60.8
Perito Moreno 2a
54.2
47.4
42.9
48.5
36.1
50.5
23.5
55.6
Perito Moreno 2b
30.9
45.1
26.5
45.6
23.3
48.4
20.5
49.9
Perito Moreno 3
36.5
50.2
24.5
53.3
26.2
52.9
13.0
55.6
Appendix
TABLE VII: Effect of the proposed modifications on the FAU subset model evaluated using region-exclusive splits across different flights. The best-performing configurations are highlighted in bold .