Cryo-Bench: Benchmarking Foundation Models for Cryosphere Mapping

Saurabh Kaushik; Lalit Maurya; Beth Tellman; Swalpa Kumar Roy; Valerio Marsocci; Gustau Camps-Valls; Jocelyn Chanussot

cs.CVMar 2, 2026

Cryo-Bench: Benchmarking Foundation Models for Cryosphere Mapping

Authors: Saurabh Kaushik, Lalit Maurya, Beth Tellman, Swalpa Kumar Roy, Valerio Marsocci, Gustau Camps-Valls, Jocelyn Chanussot

Organizations: Center for Sustainability and the Global Environment (SAGE), University of Wisconsin–Madison USA · Portsmouth AI and Data Science Centre (PAIDS), School of Computing, University of Portsmouth Portsmouth, UK · ESA, ESRIN, 𝜑-lab, Frascati Italy

Abstract

Geo-Foundation Models (GFMs) have been evaluated across diverse Earth observation tasks and domains, showing strong potential to produce reliable maps even with sparse labels. However, systematic benchmarking of GFMs for Cryosphere applications remains limited, primarily because suitable evaluation datasets are scarce. We address this gap by introducing Cryo-Bench, a benchmark comprising six semantic segmentation datasets covering five cryospheric components: supraglacial debris, glacial lakes under two sensing configurations, sea ice, calving fronts and Antarctic ice-shelf extent. The benchmark includes multispectral, RGB, and synthetic aperture radar observations from regions underrepresented in existing pretraining archives. We evaluate thirteen GFMs alongside U-Net and Vision Transformer baselines trained from scratch under a unified evaluation protocol. With frozen encoders, the U-Net achieves the highest six-dataset average mean intersection over union (mIoU) of 68.22%, slightly exceeding TerraMind (67.86%). However, the paired difference of 0.36 percentage points has a 95% confidence interval of [-0.15, +0.89], indicating that the observed ordering is not statistically significant. Fine-tuning with a fixed learning rate produces mixed outcomes, with leading models such as TerraMind and DOFA experiencing declines in average mIoU (-1.2 and -1.7). In contrast, learning-rate optimization substantially improves fine-tuning performance: DOFA reaches 93.97% mIoU on the RGB glacial lake task, ranking the U-Net fourth, while Scale-MAE and GFM-Swin surpass U-Net on calving fronts. In the few-shot setting, five GFMs outperform U-Net, retaining 91.1% of their full-label accuracy compared with 86.1% for U-Net.

Figures & tables

Fig. 1: Cryo-Bench covers six datasets and five cryospheric components across multispectral, RGB, and radar imagery. Every encoder is paired with the same UPerNet decoder and evaluated under the same three regimes. Dataset colors indicate the input modality.

Benchmark	Domains / tasks	Task types	Modality	Temporality
EarthNets	Land use / land cover	Classification, regression	MSI	Single
SustainBench	Sustainable-development goals	Classification, regression	MSI, street-view	Single
GEO-Bench	Urban, agriculture, forest	Classification, semantic segmentation	MSI	Single
PhilEO Bench	Urban	Classification, semantic segmentation	MSI	Single
FoMOBench	Urban, forest	Semantic segmentation	MSI	Single
PANGAEA	Urban, agriculture, forest, marine, disaster	Semantic segmentation, change detection, regression	MSI, SAR, DSM	Multi

TABLE I: Cryo-Bench compared with existing benchmarks used to evaluate geospatial foundation models.

Dataset	Component	Sensor	Classes	Train	Val	Test	10% tiles
GSDD	Supraglacial debris	Sentinel-2 + terrain	2	1,496	186	198	~150
GLID	Glacial lakes	Multi-source RGB	2	14,400	1,600	2,367	~1,440
GLB	Glacial lakes	S2 + S1 + terrain (11 band)	2	15,300	1,912	1,913	~1,530
SICD	Sea ice	Sentinel-1 SAR	6	461	51	20	~46
CaFFe	Calving fronts	Multi-mission SAR	4	504	55	122	~50
Shelf-Bench	Ice-shelf extent	SAR	2	42,424	6,974	6,545	~4,242

TABLE II: Datasets associated with component, sensor, class count and split sizes in Cryo-Bench.

Fig. 2: Sample pairs of images and labels for each of the six Cryo-Bench datasets. When the dataset permitted, we chose scenes to show more than one class. When optical inputs make the target class more visible than true color, the inputs are shown as false-color composites: GSDD uses B5, B4, B3 and GLB uses B8, B4, B3 (it lacks B5); GLID is true color (B4, B3, B2), matching Figures 13 to 15. Binary tasks are drawn with a black background and a white target class, as in Figures 13 to 18 ; multi-class tasks maintain the categorical palette and use grey marks to indicate pixels outside the labeled area/region.

Dataset	Class composition of the test split
GSDD	Debris 18.8%, background 81.2%
GLID	Lake 1.9%, background 98.1%
GLB	Lake 4.5%, background 95.5%
SICD	Old ice 45.1%, open water 33.1%, thick FY 16.8%, young ice 2.3%, thin FY 1.5%, new ice 1.1%
CaFFe	Glacier 49.4%, rock 35.7%, N/A 8.0%, ocean/ice melange 6.9%
Shelf-Bench	Ice shelf 48.7%, non-shelf 51.3%

TABLE III: Class composition of the test splits. Foreground share is 1.9% on GLID and 4.5% on GLB, against 48.7% on Shelf-Bench.

Situation	Handling
Dataset bands match pretraining	Matching channels are used directly
Optical-only encoder on a SAR dataset	SAR replicated into a three-band proxy
Single-channel SAR (CaFFe), radar encoder	Fed directly
Single-channel SAR (CaFFe), optical encoder	Repeated three times
Multi-band dataset, three-channel encoder	Terrain and derived layers dropped
U-Net and ViT baselines	All available bands, full tile for U-Net

TABLE IV: Band adaptation per dataset and encoder family.

Model	GSDD	GLID	GLB	SICD	CaFFe	Shelf-Bench	Avg mIoU	Rank
U-Net	73.89	91.58	80.46	20.61	59.82	82.97	68.22	1
TerraMind	74.63	88.26	79.91	33.27	46.64	84.46	67.86	2
DOFA	72.96	92.61	79.39	20.41	50.71	87.08	67.19	3
RemoteCLIP	73.42	90.89	78.09	14.51	56.64	83.59	66.19	4
GFM-Swin	73.00	89.69	78.42	11.55	58.12	85.63	66.07	5
Scale-MAE	73.47	90.91	79.16	6.02	58.19	83.28	65.17	6

TABLE V: Frozen-encoder results with 100% of the available labels, reported as test mIoU (%). Baselines are trained from scratch.

Fig. 3: Frozen-encoder results. (a) Six-dataset average mIoU, ranked, with 95% bootstrap intervals from resampling test scenes; the shaded band and dashed line mark the leading interval and estimate, and TerraMind’s interval overlaps it, so the lead is not resolved. (b) Distribution of those scores across the fifteen models within each dataset, showing how much each task separates representations overall.

Model	GSDD	GLID	GLB	SICD	CaFFe	Shelf-Bench	Avg mIoU	Rank
DOFA	71.36	88.22	73.12	13.67	47.22	86.50	63.35	1
RemoteCLIP	65.56	86.74	70.92	18.37	55.00	80.80	62.90	2
TerraMind	69.90	79.82	74.88	21.44	38.15	82.67	61.14	3
GFM-Swin	64.86	84.79	71.19	9.97	53.04	82.01	60.98	4
Scale-MAE	66.40	84.86	71.25	8.81	53.06	80.29	60.78	5
U-Net	72.00	81.42	75.42	8.51	36.48	78.40	58.71	6

TABLE VI: Few-shot results with 10% of labels and frozen encoders.

Fig. 4: Model robustness under changing evaluation regimes, measured relative to each model’s frozen performance: (a) the proportion of full-label accuracy retained when only 10% of the labels are available; and (b) the change in average mIoU after fine-tuning with the fixed default learning rate of $10^{-4}$ . Since the baseline models do not have a frozen–fine-tuned distinction, they are assigned a value of zero in (b).

Model	GSDD	GLID	GLB	SICD	CaFFe	Shelf-Bench	Avg mIoU	Rank
U-Net	73.89	91.58	80.46	20.61	59.82	82.97	68.22	1
GFM-Swin	73.77	83.76	75.98	25.21	57.28	86.53	67.09	2
TerraMind	75.89	87.14	79.04	18.57	56.14	82.93	66.62	3
DOFA	77.22	84.33	79.50	10.55	56.25	84.87	65.45	4
S12-MoCo	75.51	86.08	77.92	22.36	45.11	82.30	64.88	5
SatlasNet	76.74	85.83	77.58	12.61	52.04	83.72	64.75	6

TABLE VII: Fine-tuning at a fixed learning rate of

10^{-4}

, on all six datasets. This is the sensitivity reference against which per-model learning-rate selection (Table VIII-A , Table VIII-B ) is compared.

	GLID			CaFFe			GSDD
Model	LR	mIoU	vs $10^{-4}$	LR	mIoU	vs $10^{-4}$	LR	mIoU	vs $10^{-4}$
DOFA	10 -5	93.97	+9.63	10 -5	60.15	+3.90	10 -5	77.08	-0.14
TerraMind	10 -5	93.32	+6.18	10 -5	60.31	+4.18	10 -5	77.08	+1.19
GFM-Swin	10 -5	91.75	+7.98	10 -5	63.49	+6.21	10 -5	74.87	+1.10
U-Net	10 -3	90.44	-1.15	10 -4	59.82	+0.00	10 -4	73.89	+0.00
RemoteCLIP	10 -5	90.28	+12.44	10 -5	58.86	+33.55	10 -5	72.52	+20.35

TABLE VIII-A: Fine-tuning with the learning rate selected per model on the validation split; test mIoU is reported for the selected configuration. RGB glacial lake, calving-front and supraglacial debris datasets, the first group swept; continued in Table VIII-B for the multi-source lake, sea-ice and ice-shelf datasets.

	GLB			SICD			Shelf-Bench
Model	LR	mIoU	vs $10^{-4}$	LR	mIoU	vs $10^{-4}$	LR	mIoU	vs $10^{-4}$
TerraMind	10 -5	83.97	+4.93	10 -5	22.19	-2.61	10 -5	88.47	+5.54
CROMA	10 -5	82.38	+5.28	10 -5	21.88	+5.13	10 -5	84.09	+1.55
SatlasNet	10 -5	82.01	+4.42	10 -5	22.22	-9.37	10 -5	84.72	+0.99
GFM-Swin	10 -5	81.38	+5.40	10 -5	20.79	-9.75	10 -5	88.03	+1.50
Scale-MAE	10 -5	80.69	+7.53	10 -5	19.25	+4.07	10 -5	86.85	+2.89

TABLE VIII-B: Fine-tuning with the learning rate selected per model on the validation split; test mIoU is reported for the selected configuration. Multi-source lake, sea-ice and ice-shelf datasets, the second group swept, continuing Table VIII-A .

Fig. 5: Learning rate sensitivity. (a) Change in test mIoU from selecting the learning rate per model, relative to the fixed default; models with no visible bar selected the default. (b) Rank correlation between regimes across the fifteen models.

Fig. 6: Average mIoU by pretraining objective across the three regimes. Error bars are the standard deviation across models within a group; groups with one member have none.

Fig. 7: Left: parameter count against frozen-encoder accuracy. Right: parameter count against the learning rate selected on validation, for the three swept datasets (GLID filled, GSDD half-filled, CaFFe hollow).

Fig. 8: Frozen-encoder performance on the three radar tasks, grouped by whether the encoder saw radar during pretraining. Error bars are standard deviations across models.

Model	Open water	New ice	Young ice	Thin FY ice	Thick FY ice	Old ice	mIoU
TerraMind	77.82	5.04	3.13	5.39	36.35	71.87	33.27
CROMA	68.24	1.73	2.92	3.90	20.83	33.27	21.81
S12-MoCo	45.65	0.19	2.73	0.78	12.05	62.57	20.66
U-Net	49.10	0.37	0.00	0.00	9.28	64.91	20.61
S12-DINO	53.42	0.12	2.25	6.74	8.17	52.84	20.59
DOFA	73.44	3.92	2.99	5.44	19.28	17.36	20.41

TABLE IX: Per-class IoU on the sea-ice task with frozen encoders and full-scene evaluation.

Model	mIoU (%)	Weighted IoU (%)	Change (pts)
TerraMind	33.27	64.64	+31.38
U-Net	20.61	47.31	+26.70
S12-MoCo	20.66	45.65	+24.99
SatlasNet	19.93	44.03	+24.10
S12-DINO	20.59	43.16	+22.57
CROMA	21.81	41.02	+19.20

TABLE X: Frozen-encoder sea-ice results under unweighted and frequency-weighted aggregation, ordered by weighted IoU.

Model	GSDD	GLID	GLB	Shelf-Bench	Avg foreground IoU
DOFA	57.20	85.58	60.65	86.24	72.42
U-Net	58.77	83.56	62.83	82.05	71.80
Scale-MAE	57.83	82.26	60.25	82.25	70.65
TerraMind	60.11	77.11	61.61	83.31	70.53
GFM-Swin	57.66	79.87	58.77	85.00	70.32
RemoteCLIP	57.95	82.22	58.12	82.59	70.22

TABLE XI: Foreground-class IoU on the four binary tasks with frozen encoders.

Encoder	8-bit statistics	Matched statistics	Change (pts)
ssl4eo_moco	5.76	20.66	+14.90
satlasnet_si	6.34	19.93	+13.59
unet_encoder	10.07	20.61	+10.54
prithvi	8.07	15.37	+7.30
spectralgpt	5.43	12.57	+7.13
ssl4eo_data2vec	12.74	17.25	+4.50

TABLE XII: Frozen-encoder sea-ice mIoU under 8-bit and matched input standardization, with full-scene evaluation in both cases.

Fig. 9: Per-class IoU on sea ice, with the reported mIoU, the mean over all six classes, in the final column. The three rare classes are close to zero for every model.

Fig. 10: Per-class IoU on the four calving-front labels with frozen encoders, with the reported mIoU, the mean over all four, in the final column. The N/A region is not a surface class yet scores highest.

Model	Params (M)	Window (px)	GFLOPs/window	Windows/tile	GFLOPs/tile	ms/window
U-Net	14.8	full tile	32.4	1	32.4	1.7
RemoteCLIP	126.8	224	14.0	4	56.1	4.2
SatlasNet	121.2	128	18.8	4	75.1	15.3
GFM-Swin	120.4	192	19.2	4	76.8	11.6
S12-MAE	53.5	224	45.9	4	183.4	4.4
S12-MoCo	53.5	224	45.9	4	183.4	4.4

TABLE XIII: Parameters, per-window and per-tile computational cost, and inference latency, measured on 256 by 256 tiles.

Fig. 11: Frozen-encoder six-dataset average mIoU versus the computational cost of segmenting one 256 by 256 tile, shown on a logarithmic axis. The frozen regime is used to ensure that the vertical axis is directly comparable with Table V and Figure 3 .

Fig. 12: Per-dataset accuracy after learning-rate selection plotted against per-tile computational cost, for the four datasets with a fixed native tile size (RGB and multi-source glacial lakes, supraglacial debris, ice-shelf extent). Calving fronts and sea ice are omitted because their tiles are assembled from variable-sized raw scenes, making per-tile cost less directly comparable across models there. Unlike Figure 11 , the vertical axis represents performance on an individual dataset under tuned fine-tuning, resulting in a different efficiency frontier.

Fig. 13: Predictions on the RGB glacial lake task (GLID). In the third scene, all models except the U-Net fail to detect the western lake.

Fig. 14: Predictions on supraglacial debris (GSDD), the least discriminative task in the benchmark. The input is displayed as a false-color composite (B5, B4, B3), which clearly separates debris from ice better than the natural-color rendering in Figure 2 .

Fig. 15: Predictions on glacial lakes in the multi-source (GLB) configuration. The input is a false-color composite (B8, B4, B3); Since GLB does not include B5, B8 substitutes the standard NIR band, causing water to appear near-black due to strong NIR absorption.

Fig. 16: Predictions on Antarctic ice-shelf extent (Shelf-Bench).

Fig. 17: Predictions on calving-front zones (CaFFe), coloured by zone. Rectangular discontinuities in the foundation model predictions correspond to sliding-window boundaries.

Fig. 18: Predictions on sea ice (SICD). Prithvi assigns a single class to the entire first scene.

Explore similar work

Jul 20, 2026cs.LG

Now We Know? A Systematic Comparison of TerraMind and THOR

Benchmarks for Geospatial Foundation Models (GFMs) increasingly rank models by aggregate score, but such rankings obscure why models differ: how much of the gap is architecture, how much is decoder capacity, and how much is a use-case-specific artefact? This study addresses that gap through a controlled comparison of two GFMs developed under European Space Agency's

Φ

-lab with contrasting design philosophies: THOR, which introduces a compute-adaptive architecture supporting variable patch sizes and unifies Sentinel-1, -2, and -3 data at their native resolutions; and TerraMind, a multimodal generative GFM pretrained with a dual-scale token/pixel objective that enables any-to-any cross-modal generation (Thinking-in-Modalities) to infer missing sensors at inference time. Rather than reporting a single leaderboard, we investigate the axes along which the two architectures actually differ - patch size, decoder complexity, finetuning regime, input modality, and model scale - across ten use cases spanning segmentation and regression in diverse domains, including climate disaster response, methane leak detection, snow monitoring, or sea ice mapping. We find that architectural design choices - patch size and decoder type in particular - explain more performance variance than model identity itself, that the two models embody complementary investment strategies (pretraining-time scale for TerraMind versus inference-time tokenisation for THOR), and that correctly interpreting results requires dataset-level characterisation. The resulting picture is not a single winner but a set of hypotheses and a diagnostic ablation methodology that we expect to generalise to future GFMs beyond THOR and TerraMind.

Frederick Schindlegger, Kenzo Bounegta, Eva Gmelich Meijling +5

University of Münster, Münster, Germany · 2ESA Φ-lab, ESRIN, Frascati, Italy · 3Rubicon, Paris, France +2

Geospatial Foundation Models Wireless Foundation Models Comparison

Ask PDF

May 19, 2026eess.IV

CryoNet: A Deep Learning Framework for Multi-Modal Debris-Covered Glacier Mapping. A Case Study of the Poiqu Basin, Central Himalaya

Glaciers play a critical role as freshwater reserves and indicators of climate change, yet their automatic delineation, especially for debris-covered glaciers, remains challenging due to spectral similarity with surrounding terrain. This study introduces CryoNet, a deep learning framework that leverages a rich multi-modal dataset combining Sentinel-2 optical imagery, DEM-derived topographic variables, spectral indices, Principal Component Analysis (PCA), InSAR coherence and phase, tasseled-cap features, and GLCM texture to discriminate clean-ice glaciers, debris-covered glaciers, and glacial lakes. CryoNet is an encoder-decoder CNN with nested skip connections and spatial-channel Squeeze-and-Excitation (scSE) attention, built upon a ResNet101 encoder to capture hierarchical contextual and spatial features. The study is conducted in the Poiqu Basin in the central Himalaya, and transferability is evaluated by applying the trained model to the Mont Blanc Massif in the Alps. We additionally analyse the importance of each data layer in improving glacier mapping performance. The proposed model achieves an overall IoU of 90.52%, mean Recall of 98.08%, and mean Precision of 92.26%. For debris-covered glaciers specifically, CryoNet obtains an IoU of 90.46%, a recall of 95.79%, and a precision of 94.21%. Across both per-class and overall metrics, CryoNet surpasses DeepLabV3+, SegFormer, and U-Net, taken as state-of-the-art (SOTA) references, demonstrating its effectiveness for robust glacier mapping in complex high-mountain environments.

Farzaneh Barzegar, Tobias Bolch, Norbert Kuehtreiber +1

Institute of Geodesy, Graz University of Technology, Graz, Austria · Engineering Department, University of Sannio, Benevento, Italy

Sea-Surface Temperature Data Sentinel-2 Deep Learning Freeze Ridge

Ask PDF

Aug 3, 2026cs.CV

GEOID-Flood: A Large-Scale Multi-Modal Benchmark Dataset for Flood Segmentation

Geospatial foundation models aim to learn representations that transfer across regions and sensors, yet evaluating them on specific tasks requires large, high-quality, multi-modal benchmarks that measure how well such models extract value from data. Concerning flood mapping, existing datasets rarely combine bi-temporal SAR and co-registered optical imagery at scale, leaving the value of foundation models for this downstream task largely untested. We introduce GEOID-Flood, a large-scale multi-modal flood segmentation benchmark, derived from Copernicus Emergency Management Service activations, spanning 219 events across 65 countries over ten years. The dataset provides more than 14,000 tiles with co-registered pre- and post-event Sentinel-1, in GRD and RTC format, pre-event Sentinel-2 composite, and DEM, including manually validated labels that separate background from permanent water and flooded water. Using this benchmark, we evaluate foundation models against conventional encoders across single-image, multi-temporal, and multi-modal protocols. We report three main findings: foundation models offer a consistent but modest advantage; optical-SAR fusion with finetuning best resolves transient flooding; and models trained on GEOID-Flood transfer to unseen events better than those trained on existing datasets. Dataset and code available at https://github.com/links-ads/geoid-flood.

Gaetano Chiriaco, Luca Barco, Andrea Bragagnolo +2

Fondazione LINKS, Via Pier Carlo Boggio, 61, 10138 Torino, Italy · Politecnico di Torino, Corso Duca degli Abruzzi 24, 10129 Torino, Italy

Flood Detection Geospatial Foundation Models Sentinel-2 Multimodal Dataset Geogs-Slam

Ask PDF