When GeoFMs Meet VHR: Benchmarking Foundation Models on Very High Resolution Earth Observation Imagery

Kshitij Raj Sharma; Pierre Adorni; Minh-Tan Pham; Dirk Tiede; Sébastien Lefèvre

cs.CVNov 5, 2026

When GeoFMs Meet VHR: Benchmarking Foundation Models on Very High Resolution Earth Observation Imagery

Authors: Kshitij Raj Sharma, Pierre Adorni, Minh-Tan Pham, Dirk Tiede, Sébastien Lefèvre

Organizations: Université Bretagne Sud, IRISA, France. · University of Salzburg, Department of Geoinformatics, Austria. · UiT The Arctic University of Norway, Tromsø, Norway.

Abstract

Geospatial foundation models are a widely adopted starting point for remote sensing tasks. Most are pretrained on Sentinel-1/2 and Landsat imagery at a ground sampling distance (GSD) of 10m or coarser. Many applications instead rely on very high resolution (VHR) imagery, where individual buildings, roads, and amenities are resolved. We study which pretraining regimes transfer well to this scale. We introduce VHR-Bench, a benchmark suite of ten datasets at 1.5m GSD or finer spanning five task families, eight application domains, and both satellite and aerial platforms. We evaluate twelve encoders in three pretraining tiers with frozen backbones at two training-data fractions. We find that both pretraining on large natural-image collections and VHR-specific pretraining transfer effectively to VHR tasks, with complementary strengths. SkySense leads when labels are scarce; at the full label fraction the web-pretrained DINOv3 ViT-L edges ahead, with the two close together there. Encoders pretrained only on Sentinel-scale imagery transfer least well under RGB input. Guided by these complementary results, we continue the self-supervised pretraining of DINOv3 ViT-L on 1.02 million held-out VHR tiles and evaluate the resulting encoder without changing the downstream protocol. The continued model gives a clear improvement over the released backbone when labels are scarce and a modest one at the full fraction, at no increase in inference parameters or compute. These findings support a generalize-then-specialize strategy: already strong, broad natural-image pretraining can be further improved by a short VHR continuation stage while preserving its general visual features.

Figures & tables

#	Dataset	Platform	Sensor	GSD	Region	Task	Train / Val / Test
1	SpaceNet 2 [ 1 ]	Satellite	WorldView-3	0.3m	4 cities	Segmentation (buildings)	5186 / 1461 / 2961
2	SpaceNet 3 [ 1 ]	Satellite	WorldView-3	0.3m	4 cities	Segmentation (roads)	1782 / 510 / 257
3	xView2 (xBD) [ 2 ]	Satellite	Mixed (Maxar)	$<$ 0.8m	Global	Change detection (damage)	13384 / 917 / 933
4	FLAIR 2 [ 3 ]	Aerial	Aerial RGBN	0.2m	France	Segmentation (land cover)	4049 / 1022 / 3022
5	Open-Canopy-Small [ 4 ]	Satellite	SPOT 6/7	1.5m	France	Regression (canopy height)	6634 / 737 / 13675
6	EverWatch [ 5 ]	Aerial	Aerial RGB	0.1m	United States	Object detection (birds)	4429 / 500 / 196

TABLE I: Datasets in the VHR suite with sensor, platform, resolution, and region. Some datasets carry extra bands (near-infrared and elevation in FLAIR 2, near-infrared in PureForest and Open-Canopy-Small); only their RGB is evaluated.

Dataset	License	Original source	Obtained from
SpaceNet 2	CC BY-SA 4.0	SpaceNet AWS Open Data	GEO-Bench-2 [ 22 ]
SpaceNet 3	CC BY-SA 4.0	SpaceNet AWS Open Data	Publisher release
xView2 (xBD)	CC BY-NC-SA 4.0	xview2.org	PANGAEA [ 21 ]
FLAIR 2	Open Licence 2.0 (Etalab)	IGN France	GEO-Bench-2 [ 22 ]
Open-Canopy-Small	Open Licence 2.0 (Etalab)	AI4Forest	Size-comparable subset
EverWatch	CC0 1.0	Zenodo	GEO-Bench-2 [ 22 ]

TABLE II: Licenses, Original and Obtained Source for the ten datasets.

Task family	Count	Domain breakdown
Semantic segmentation	4	2 buildings (SpaceNet 2, HOTOSM Buildings), 1 roads (SpaceNet 3), 1 land cover (FLAIR 2, 13 classes)
Change detection	1	1 building damage (xView2)
Regression	1	1 forestry / canopy height (Open-Canopy-Small)
Object detection	2	1 ecology / wading birds (EverWatch), 1 agriculture / cattle (NZ Cattle)
Classification	2	1 energy / PV panels (m-pv4ger, binary), 1 forestry / tree species (PureForest, 13 classes)

TABLE III: Task-family and application-domain composition of VHR-Bench; the eight domains are buildings, roads, land cover, building damage, forestry, ecology, agriculture, and energy.

Fig. 1: One training chip per dataset, drawn through the same loader the encoders use.

Fig. 2: Geographic coverage of VHR-Bench. Chip centroids are coloured by task family and bubble size reflects the number of training chips in each area, so small sites stay visible alongside the largest.

#	Tier	Encoder	Params (M)	GFLOPs	Input (px)	Patch	Tokens	Pretraining data
1	HR-FM	Clay v1.5 ViT-L [ 28 ]	311	723.0	256	8	1024	Sentinel-1/2, NAIP, LINZ
2		DINOv3 ViT-L (SAT) [ 18 ]	303	164.7	256	16	256	SAT-493M (Maxar, 0.6m)
3		DOFA ViT-B [ 14 ]	111	35.4	224	16	196	Multi-sensor (sat + aerial)
4		EoS-FM [ 17 ]	78	206.9	256	–	–	Sentinel-1/2 and HR optical
5		SkySense SwinH [ 16 ]	655	301.6	256	4	–	WorldView 3/4, Sentinel-1/2
6	CV-FM	ConvNeXt-XL [ 29 ]	348	159.1	256	–	–	ImageNet-22k to 1k

TABLE IV: Evaluated encoders. A dash in the Patch or Tokens column means the quantity does not apply, as for the convolutional encoders and for SkySense, whose windowed attention makes the token count vary by stage.

Task	Decoder	Loss	Metric
Segmentation	UPerNet	Cross-entropy	mIoU
Change detection	Siamese UPerNet	Cross-entropy	mIoU
Regression	UPerNet head	MSE	RMSE $\downarrow$
Detection	Faster R-CNN	Cross-entropy + box reg.	mAP@50
Classification	Linear probe	Cross-entropy	Macro F1

TABLE V: Fixed decoder, loss, and primary metric per task family.

Source	Tiles	Platform	Sensor	GSD	License	Obtained from
fMoW [ 32 ]	500,008	Satellite	QuickBird, GeoEye, WorldView	0.3–0.7m	fMoW Challenge (NC)	AWS Open Data
Maxar Open Data [ 33 ]	120,000	Satellite	WorldView	0.3m	CC BY-NC 4.0	Maxar STAC
OpenAerialMap [ 34 ]	100,000	Drone/aircraft	Mixed	0.05–1m	CC BY 4.0	OAM STAC
NAIP [ 35 ]	100,000	Aerial	Aerial RGB	1.25m	Public domain	SatlasPretrain [ 36 ]
RAMP [ 37 ]	95,333	Satellite	WorldView, GeoEye	0.3–0.5m	CC BY-NC 4.0	Source Cooperative
WorldStrat [ 38 ]	62,822	Satellite	SPOT 6/7	1.5m	CC BY-NC 4.0	Zenodo [ 38 ]

TABLE VI: VHR continued-pretraining corpus. Planet tiles are open disaster releases from Venezuela, the Philippines, and Jamaica.

Fig. 3: Geographic coverage of the continued-pretraining corpus. Tile footprints are aggregated on a coarse grid; bubble area is proportional to the number of tiles and colour identifies the dominant source in each cell.

Setting	Value
Initialization	DINOv3 ViT-L (LVD-1689M)
Input	$256\times 256$ RGB
Effective batch	512
Objective	DINO + iBOT + 0.1 KoLeo
Schedule	Freeze 1,000; warm up 500 steps
Backbone learning rate	$5\times 10^{-5}$

TABLE VII: Continued-pretraining configuration.

#	Tier	Encoder	10%	100%	Mean
1	CON	DINOv3 ViT-L (CON)	65.2 ±0.3	70.0 ±0.2	67.6
2	HR	SkySense SwinH	65.8 ±0.4	69.2 ±0.1	67.5
3	CV	DINOv3 ViT-L (LVD)	64.0 ±0.3	69.5 ±0.1	66.8
4	HR	DINOv3 ViT-L (SAT)	63.4 ±0.2	69.1 ±0.3	66.3
5	CV	DINOv3 ConvNeXt-L	62.5 ±0.1	68.2 ±0.2	65.3
6	CV	ConvNeXt-XL	62.4 ±0.0	67.5 ±0.2	65.0

TABLE VIII: Overall ranking by geometric-mean closeness to a perfect score. The mean over three seeds and standard deviation are reported, with the best result per fraction in bold. CON designates our continual pretraining experiment (LVD

\rightarrow

HR). Italics and ∗ denote fully fine-tuned models; otherwise, backbones are frozen. Closeness is scaled by 100.

Fig. 4: Aggregate closeness versus forward GFLOPs; marker size scales with parameter count, colour with the pretraining tier. Patch size sets the token count: Clay v1.5 (patch 8) carries 1024 tokens and 723 GFLOPs against DINOv3 ViT-L (patch 16) at 256 tokens and 164.7 GFLOPs.

Fig. 5: Overall geometric-mean closeness at both training fractions, coloured by pretraining tier. Error bars are one standard deviation over three seeds. Encoders are ordered by the mean of the two fractions. (ft)* means fully fine-tuned. CON designates our continual pretraining experiment (LVD

\rightarrow

HR).

Tier	Encoder	HOT	SN2	SN3	FLR	xV2	OpC	PF	PV	EvW	NZC
CV $\rightarrow$ VHR	DINOv3 ViT-L (LVD $\rightarrow$ HR)	66.5 ±0.4	87.6 ±0.0	69.7 ±0.4	53.9 ±0.6	57.4 ±0.6	39.8 ±0.1	66.2 ±0.4	93.4 ±0.2	33.9 ±0.8	67.0 ±2.8
HR-FM	Clay v1.5 ViT-L	61.2 ±0.8	79.9 ±0.3	58.8 ±0.5	43.9 ±0.1	51.8 ±0.6	42.8 ±0.1	45.7 ±0.3	82.2 ±0.8	28.3 ±1.5	66.8 ±1.1
	DINOv3 ViT-L (SAT)	65.4 ±0.7	85.3 ±0.3	68.6 ±0.4	51.2 ±0.3	55.9 ±0.3	38.2 ±0.1	63.0 ±1.5	90.2 ±0.2	31.9 ±0.7	66.0 ±1.5
	DOFA ViT-B	63.1 ±1.0	83.8 ±0.2	65.5 ±0.5	48.9 ±0.3	53.7 ±0.3	42.7 ±0.1	61.4 ±1.6	92.3 ±0.9	32.2 ±0.5	67.6 ±0.8
	EoS-FM	61.9 ±0.5	82.9 ±0.1	66.8 ±0.7	41.6 ±0.8	53.2 ±0.3	43.3 ±0.1	49.1 ±1.2	89.4 ±0.7	27.3 ±0.6	68.8 ±1.2
	SkySense SwinH	65.9 ±0.3	86.5 ±0.1	67.5 ±0.5	52.0 ±0.2	57.1 ±0.6	40.5 ±0.1	67.1 ±2.2	93.6 ±0.7	35.9 ±0.9	75.6 ±0.7

TABLE IX: Primary metric per dataset at 10 percent training data, frozen backbone except indicated. Segmentation and change detection report mIoU, classification macro F1, detection mAP@50, Open-Canopy-Small RMSE in decimetres (lower is better). Mean over three decoder seeds, subscript is the standard deviation; best frozen-encoder mean in bold. The fully fine-tuned baseline is not counted when bolding. Datasets: HOT: HOTOSM Buildings; SN2, SN3: SpaceNet 2, 3; FLR: FLAIR2; xV2: xView2; OpC: Open-Canopy-Small; PF: PureForest; PV: m-pv4ger; EvW: EverWatch; NZC: NZ Cattle.

Fraction	Seed 42	Seed 1234	Seed 2025	Mean $\pm$ std
10%	65.1	64.7	64.7	64.8 ±0.2
100%	69.9	69.9	70.1	70.0 ±0.1

TABLE X: Aggregate closeness of three independent continued-pretraining runs. The decoder seed is fixed at 42 to isolate pretraining variability.

Layer	5	11	17	23
CKA	0.998 ±0.000	0.996 ±0.000	0.922 ±0.012	0.789 ±0.016

TABLE XI: Patch-token CKA between the continued and released DINOv3 ViT-L encoders; mean and standard deviation over three continued-pretraining runs.

Tier	Encoder	HOT	SN2	SN3	FLR	xV2	OpC	PF	PV	EvW	NZC
CV $\rightarrow$ VHR	DINOv3 ViT-L (LVD $\rightarrow$ HR)	66.7 ±0.2	88.9 ±0.0	73.2 ±0.1	55.4 ±0.1	61.4 ±0.3	36.1 ±0.1	72.6 ±0.5	95.3 ±0.1	43.2 ±0.5	78.1 ±0.4
HR-FM	Clay v1.5 ViT-L	61.6 ±0.5	84.9 ±0.1	67.8 ±0.1	45.9 ±0.1	58.5 ±0.3	37.6 ±0.0	57.5 ±0.0	89.1 ±0.1	42.8 ±0.6	77.7 ±0.7
	DINOv3 ViT-L (SAT)	64.8 ±0.4	88.0 ±0.1	72.5 ±0.1	52.4 ±0.2	60.9 ±0.2	35.1 ±0.1	70.7 ±1.4	92.1 ±0.2	43.5 ±0.9	79.8 ±0.4
	DOFA ViT-B	62.8 ±1.2	86.8 ±0.1	70.7 ±0.1	49.9 ±0.2	59.3 ±0.1	38.7 ±0.0	65.7 ±0.7	93.8 ±0.1	40.1 ±0.2	77.3 ±0.3
	EoS-FM	63.0 ±0.8	85.3 ±0.0	71.1 ±0.1	44.3 ±0.8	56.9 ±1.6	39.2 ±0.1	54.9 ±0.9	90.9 ±0.7	36.7 ±1.1	78.6 ±1.3
	SkySense SwinH	66.1 ±0.4	88.4 ±0.0	72.4 ±0.1	53.1 ±0.2	61.6 ±0.1	36.6 ±0.2	66.2 ±0.4	94.3 ±0.7	43.9 ±0.4	81.4 ±0.5

TABLE XII: Primary metric per dataset at 100 percent training data, frozen backbone except indicated. Metrics and dataset abbreviations as in Table IX . Mean over three decoder seeds, subscript is the standard deviation; best frozen-encoder mean in bold. The fully fine-tuned baseline is not counted when bolding.

Removed	$\rho$	Leader	$\rho$	Leader
	10%		100%
HOT	0.995	–	0.989	–
SN2	0.995	–	0.995	–
SN3	0.995	–	0.995	–
FLR	0.995	–	0.989	–
xV2	0.995	–	0.995	–

TABLE XIII: Leave-one-out sensitivity after adding the continued model.

\rho

is the Spearman correlation with the full 13-encoder ranking. The leader column names an encoder only where it differs from the full-suite leader at that fraction; a dash means unchanged. No encoder moves by more than one position.

Geospatial Foundation Models

Geospatial foundation models turn satellite, map, and Earth-system data into reusable representations for spatial analysis and prediction.

See all 129 papers

A Composition-Aware Pretraining Framework for Geospatial Foundation Models
Aug 31, 2026Aryan Kashyap Naveen, Abhishek Srinivas, Pranav Moothedath +1Remote Sensing Image Understanding Representation Learning
National Institute of Technology, Karnataka · Surathkal, Karnataka, India · Visa Inc +2
Reuse, Don't Retrain: Efficient, Modality-Versatile Geospatial Foundation Models through Feature Distillation
Oct 23, 2026Mario Hamon, Pierre Adorni, Minh-Tan PhamGeospatial Foundation Models Multimodal Foundation Models
IRISA Universit´e Bretagne Sud, UMR 6074
Clustering Guided Domain-Specific Pretrained Foundation Model for Very High-Resolution Arctic Remote Sensing
May 28, 2026Amal S. Perera, Chandi Witharana, Elias Manos +2Domain-Adaptive Pretraining Vision Foundation Model Adaptation
Department of Natural Resources and the Environment, University of Connecticut, Storrs, CT · Woodwell Climate Research Center, Falmouth, MA.

Aug 31, 2026J/K move · Enter open · S save

A Composition-Aware Pretraining Framework for Geospatial Foundation Models

Aryan Kashyap Naveen, Abhishek Srinivas, Pranav Moothedath, Shrutilipi Bhattacharjee

National Institute of Technology, Karnataka · Surathkal, Karnataka, India · Visa Inc+2

Remote Sensing Image Understanding Representation Learning Geospatial Foundation Models Geospatial Representation Learning Compositional Representation Learning

Open paper AskPDF arXiv

Geospatial foundation models have emerged as state-of-the-art methods for downstream Earth observation tasks. However, existing pretraining methodologies process imagery through a single-concept lens, failing to capture the highly compositional nature of complex satellite scenes. We propose a composition-aware pretraining framework that explicitly encodes fractional land-cover mixtures. Each satellite image cell is mapped to a histogram representing its fractional land-cover distribution, which we term the "composition target". These targets serve as the primary prediction objective and are distilled into the backbone using Earth Mover's Distance. Experimental evaluation shows that composition-aware pretraining yields substantial gains on region-level understanding tasks requiring semantic similarity judgment, including zero-shot image retrieval and scene classification, while remaining competitive on tasks requiring fine-grained spatial precision, such as segmentation and object detection. With a 36.8M-parameter backbone, our framework outperforms SatMAE and Prithvi-EO-2.0, which contain 303M and 600M parameters, respectively, in most retrieval and scene classification settings. On the fine-grained ForestNet-12 dataset, a rigorous testbed for compositional discrimination, our method boosts baseline mAP@10 from 0.279 to 0.434, a 55.6% relative improvement, providing direct evidence for the effectiveness of explicit composition modeling. The code implementation can be found at https://github.com/05kashyap/GFM_Composition_Pretraining

Abstract

Geospatial Foundation Models

Explore similar work

A Composition-Aware Pretraining Framework for Geospatial Foundation Models

Reuse, Don't Retrain: Efficient, Modality-Versatile Geospatial Foundation Models through Feature Distillation

Clustering Guided Domain-Specific Pretrained Foundation Model for Very High-Resolution Arctic Remote Sensing