CrossEarth-SAR: A SAR-Centric and Billion-Scale Geospatial Foundation Model for Domain Generalizable Semantic Segmentation

Ziqi Ye; Ziyang Gong; Ning Liao; Xiaoxing Hu; Di Wang; Hongruixuan Chen; Chen Huang; Yiguo He

cs.CVMar 12, 2026

CrossEarth-SAR: A SAR-Centric and Billion-Scale Geospatial Foundation Model for Domain Generalizable Semantic Segmentation

Authors: Ziqi Ye, Ziyang Gong, Ning Liao, Xiaoxing Hu, Di Wang, Hongruixuan Chen, Chen Huang, Yiguo He,

Organizations: Fudan University · Shanghai Innovation Institute · Shanghai Jiao Tong University · Beijing Institute of Technology ·

Abstract

Synthetic Aperture Radar (SAR) enables global, all-weather earth observation. However, owing to diverse imaging mechanisms, domain shifts across sensors and regions severely hinder its semantic generalization. To address this, we present CrossEarth-SAR, the first billion-scale SAR vision foundation model built upon a novel physics-guided sparse mixture-of-experts (MoE) architecture incorporating physical descriptors, explicitly designed for cross-domain semantic segmentation. To facilitate large-scale pre-training, we develop CrossEarth-SAR-200K, a weakly and fully supervised dataset that unifies public and private SAR imagery. We also introduce a benchmark suite comprising 22 sub-benchmarks across 8 distinct domain gaps, establishing the first unified standard for domain generalization semantic segmentation on SAR imagery. Extensive experiments demonstrate that CrossEarth-SAR achieves state-of-the-art results on 20 benchmarks, surpassing previous methods by over 10% mIoU on some benchmarks under multi-gap transfer. All code, benchmark and datasets will be publicly available.

Figures & tables

Figure 3: Framework of CrossEarth-SAR. (a) Collect large-scale labeled SAR segmentation data to construct CrossEarth-SAR-200K. (b) The continued pre-training structure of CrossEarth-SAR. (c) SAR RS-PEFT by Earth-Adapter. (d) The information of diverse benchmarks.

Domain Gap	Dataset	Source Domain	Target Domain	Abbreviation	Train	Test	Categories
Unseen Region	AIR-PolSAR-Seg 2.0 [ 70 ]	North	South	N2S	600	159	5
	AIR-PolSAR-Seg 2.0 [ 70 ]	South	North	S2N	633	150	5
	DDHR-SK [ 37 ]	KR	CN	K2C	4916	1566	6
	DDHR-SK [ 37 ]	CN	KR	C2K	6265	1229	6
Unseen Polarization	AIR-PolSAR-Seg 2.0 [ 70 ]	VV	Full	VV2F	1233	309	6
		Full	VV	F2VV			6

Table 1: CrossEarthSAR benchmark collection. North/South: northern/southern China. KR/CN: Korea/China. VV/HH: single-polarization. Full: full-polarization. Complex(r)/Complex(i): real or imaginary part of complex-valued PolSAR data.

Method	Backbone	Param.	Unseen Region				Unseen Polarization				Unseen Complex Value				Avg.
Method	Backbone	Param.	N2S	S2N	K2C	C2K	VV2F	F2VV	HH2F	F2HH	C(r)2R	R2C(r)	C(i)2R	R2C(i)	Avg.
S12-MoCo [ 43 ]	ViT-S [ 7 ]	20M	23.2	24.0	8.9	18.6	21.9	16.1	19.8	29.9	29.9	48.1	29.3	37.7	25.6
S12-DINO [ 43 ]	ViT-S [ 7 ]	20M	22.5	27.0	12.6	30.8	23.0	37.5	15.7	35.1	25.0	59.0	23.4	55.0	30.5
S12-MAE [ 43 ]	ViT-S [ 7 ]	20M	22.9	29.9	5.9	34.8	12.8	17.7	12.6	15.6	23.7	33.3	19.5	22.8	21.0
DOFA [ 60 ]	ViT-B [ 7 ]	80M	25.3	27.5	15.8	31.9	20.4	30.7	21.6	30.6	24.6	42.6	24.7	26.3	26.8
SatMAE [ 5 ]	ViT-L [ 7 ]	300M	26.2	32.1	14.5	33.4	22.5	33.4	17.1	30.0	46.4	34.4	43.4	35.6	30.7

Table 2: Performance comparison across one-domain-gap benchmarks. All models are frozen during downstream fine-tuning. Bold and underlined denote the best and second-best results. Performances in

(\cdot)

show changes relative to Baseline. * mark means the use of Earth-Adapter for PEFT. Parameters in

(\cdot)

mean the activated parameters, representing the efficient inference cost of CrossEarth-SAR.

Method	Backbone	Param.	Unseen Region and Polarization		Unseen Region and Platform		Unseen Region and MW Band		Unseen Region and Pol, MW Band		Unseen Region and Platform, MW Band		Avg.
Method	Backbone	Param.	F2A	A2F	O2D	D2O	S2A	A2S	D2F	F2D	W2D	D2W	Avg.
S12-MoCo [ 43 ]	ViT-S [ 7 ]	20M	7.4	14.7	5.5	8.1	45.2	42.2	6.7	17.4	10.3	5.2	16.3
S12-DINO [ 43 ]	ViT-S [ 7 ]	20M	8.9	23.6	4.3	9.5	53.9	51.1	15.5	15.9	13.6	10.1	20.6
S12-MAE [ 43 ]	ViT-S [ 7 ]	20M	12.1	11.4	5.3	8.8	31.3	51.8	19.6	16.1	11.9	14.8	18.3
DOFA [ 60 ]	ViT-B [ 7 ]	80M	10.3	22.3	4.8	10.2	50.9	27.5	8.3	16.0	12.3	12.5	17.5
SatMAE [ 5 ]	ViT-L [ 7 ]	300M	11.8	13.1	12.1	9.9	49.2	48.1	20.3	17.5	10.2	7.7	20.0

Table 3: Performance comparison on benchmarks across two and three domain gaps. Results show that CrossEarth-SAR-S outperforms other larger FMs. CrossEarth-SAR-L ∗ performance also demonstrates further improvement potential in PEFT of CrossEarth-SAR.

Figure 4: Visualizations of predicted segmentation maps on six representative benchmarks. For the top three rows color map, blue is water, green is vegetation, red is ground, cyan is road, yellow is building, and purple is mountain. For the bottom two rows color map, blue is farmland, green is greenery, red is road, cyan is building, yellow is water, and white is background.

Method with CPT dataset	N2S	VV2F	C(r)2R	A2F	O2D	S2A	D2F	W2D	avg.
DINOv2 + 40K real label datasets	32.7	68.6	67.7	16.3	16.1	52.9	20.4	17.1	36.5
DINOv2 + 40K pseudo label datasets (randomly selected)	33.2	66.3	66.9	18.6	17.2	53.6	21.3	17.6	36.8
DINOv2 + CrossEarth-SAR- 200K	36.6	70.6	74.5	23.4	19.7	55.2	21.9	18.9	40.1
CrossEarth-SAR	37.8	73.8	76.4	25.0	23.7	59.1	25.1	22.2	42.9

Table 4: Ablation Study on the Structure and Data Validity of CrossEarth-SAR.

Datasets	Resolution (m)	Band	Polarization	Satellites
QXS-SAROPT [ 15 ]	1m	C	VV	Gaofen-3
WHU-SEN-City [ 46 ]	20-22m	C	VV, VH	Sentinel-1
SEN12 [ 40 ]	10m	C	VV, VH	Sentinel-1
Osdataset [ 58 ]	1m	C	VV	Gaofen-3
SARptical [ 49 ]	1m	X	VV, HH	TerraSAR-X
DDHR-X [ 37 ]	1m	C	VV	Gaofen-3

Table 7: CrossEarth-SAR-200K source datasets information. Res.:Resolution. GF-3: Gaofen-3. S-1: Sentinel-1. T-X: TerraSAR-X. Air: Airborne SAR.

Figure 8: Loss Curve of CrossEarth-SAR in other 20 downstream tasks fine-tuning. CrossEarth-SAR consistently starts and converges at lower losses than Baseline, illustrating the effectiveness of CPT across diverse SAR domains.

Method	Backbone	Domain	Classes					mIoU (%)	Domain	Classes						mIoU (%)
Method	Backbone	Source $\rightarrow$ Unseen	Bldg	Grnd	Road	Veg	Wtr	mIoU (%)	Source $\rightarrow$ Unseen	Bldg	Farm	Green	Otr	Road	Wtr	mIoU (%)
SARATR-X [ 13 ]	HiViT-B [ 59 ]	N2S (Unseen Region)	55.6	5.7	23.1	31.6	20.3	28.4	K2C (Unseen Region)	35.5	0.0	0.0	0.0	0.0	0.0	5.9
S12-MoCo [ 43 ]	ViT-S [ 7 ]		54.1	3.9	7.3	32.3	31.6	23.2		37.4	0.0	16.1	0.0	0.0	0.0	8.9
S12-DINO [ 43 ]	ViT-S [ 7 ]		48.0	4.0	3.6	32.6	35.8	22.5		43.1	0.0	12.9	0.0	0.2	19.6	12.6
S12-MAE [ 43 ]	ViT-S [ 7 ]		48.5	4.1	5.2	28.7	28.6	22.9		35.5	0.0	0.0	0.0	0.0	0.0	5.9
DOFA [ 60 ]	ViT-B [ 7 ]		53.0	12.9	10.3	28.9	21.4	25.3		54.1	0.0	23.3	0.0	0.0	17.3	15.8

Table 8: Performance comparison on benchmark for gaps1-1. Bolds = best; underlines = second best. Bldg:Bulidings. Grnd:Ground. Road:Roads. Veg:Vegetation. Wtr:Water. Farm:Farmland. Green:Greenery. Otr:Others.

Method	Backbone	Domain	Classes						mIoU (%)	Domain	Classes						mIoU (%)
Method	Backbone	Source $\rightarrow$ Unseen	Bldg	Grnd	Mnt	Road	Veg	Wtr	mIoU (%)	Source $\rightarrow$ Unseen	Bldg	Grnd	Mnt	Road	Veg	Wtr	mIoU (%)
SARATR-X [ 13 ]	HiViT-B [ 59 ]	VV2F (Unseen Polarization)	67.3	35.0	81.7	49.3	58.3	51.2	57.1	HH2F (Unseen Polarization)	61.0	23.4	73.3	33.4	55.7	44.6	48.6
S12-MoCo [ 43 ]	ViT-S [ 7 ]		10.5	14.6	27.2	3.5	40.2	35.6	21.9		4.3	11.9	33.6	2.6	38.2	28.5	19.8
S12-DINO [ 43 ]	ViT-S [ 7 ]		21.3	7.4	41.8	2.3	32.1	32.8	23.0		33.0	0.6	0.4	1.6	32.8	25.8	15.7
S12-MAE [ 43 ]	ViT-S [ 7 ]		2.8	3.8	4.0	1.1	36.2	28.6	12.8		3.4	4.8	7.0	0.9	35.4	24.1	12.6
DOFA [ 60 ]	ViT-B [ 7 ]		46.2	0.8	0.3	0.3	45.5	29.4	20.4		52.0	2.4	0.0	0.3	48.6	26.5	21.6

Table 9: Performance comparison on benchmark for gaps1-2. Bolds = best; underlines = second best. Bldg:Bulidings. Grnd:Ground. Mnt:Mountain. Veg:Vegetation. Wtr:Water.

Method	Backbone	Domain	Classes						mIoU (%)	Domain	Classes						mIoU (%)
Method	Backbone	Source $\rightarrow$ Unseen	Bldg	Grnd	Mnt	Road	Veg	Wtr	mIoU (%)	Source $\rightarrow$ Unseen	Bldg	Grnd	Mnt	Road	Veg	Wtr	mIoU (%)
SARATR-X [ 13 ]	HiViT-B [ 59 ]	C(r)2R (Unseen Number Type)	71.0	24.2	0.0	34.7	61.3	60.4	73.8	C(i)2R (Unseen Number Type)	74.0	31.6	22.0	37.8	66.0	63.5	72.3
S12-MoCo [ 43 ]	ViT-S [ 7 ]		59.3	12.4	0.9	7.6	47.0	52.4	29.9		59.1	8.0	1.0	7.7	49.2	51.0	29.3
S12-DINO [ 43 ]	ViT-S [ 7 ]		56.2	4.7	0.0	1.6	46.1	41.5	25.0		54.9	4.3	0.0	2.0	41.3	37.8	23.4
S12-MAE [ 43 ]	ViT-S [ 7 ]		53.8	3.7	0.0	0.8	46.3	37.4	23.7		49.7	0.9	0.0	0.3	31.0	35.4	19.5
DOFA [ 60 ]	ViT-B [ 7 ]		54.3	4.8	0.1	0.6	47.1	40.8	24.6		54.3	8.9	0.0	0.3	45.1	39.5	24.7

Table 10: Performance comparison on benchmark for gaps1-3. Bolds = best; underlines = second best. Bldg:Bulidings. Grnd:Ground. Mnt:Mountain. Veg:Vegetation. Wtr:Water.

Method	Backbone	Domain	Classes					mIoU (%)	Domain	Classes						mIoU (%)
Method	Backbone	Source $\rightarrow$ Unseen	Bldg	Otr	Road	Veg	Wtr	mIoU (%)	Source $\rightarrow$ Unseen	Bldg	Farm	Green	Otr	Road	Wtr	mIoU (%)
SARATR-X [ 13 ]	HiViT-B [ 59 ]	F2A (Unseen Region and Polarization)	0.0	5.6	0.0	0.0	15.2	13.0	O2D (Unseen Region and Platform)	5.4	0.0	22.3	2.1	0.0	2.6	5.4
S12-MoCo [ 43 ]	ViT-S [ 7 ]		0.0	4.3	0.0	17.4	15.1	7.4		0.6	0.0	21.4	0.8	0.0	10.3	5.5
S12-DINO [ 43 ]	ViT-S [ 7 ]		0.0	2.8	0.0	32.6	9.0	8.9		3.6	0.0	15.5	2.5	0.0	4.2	4.3
S12-MAE [ 43 ]	ViT-S [ 7 ]		0.0	3.0	0.0	35.3	22.0	12.1		1.5	0.0	21.2	0.9	0.0	8.7	5.4
DOFA [ 60 ]	ViT-B [ 7 ]		0.0	2.8	0.8	30.3	17.6	10.3		4.5	0.0	20.8	2.6	0.0	0.6	4.8

Table 11: Performance comparison on benchmark for gaps2. Bolds = best; underlines = second best. Bldg:Bulidings. Otr:Others. Veg:Vegetation. Wtr:Water. Green:Greenery.

Method	Backbone	Domain	Classes					mIoU (%)	Domain	Classes						mIoU (%)
Method	Backbone	Source $\rightarrow$ Unseen	Bldg	Otr	Road	Veg	Wtr	mIoU (%)	Source $\rightarrow$ Unseen	Bldg	Farm	Green	Otr	Road	Wtr	mIoU (%)
SARATR-X [ 13 ]	HiViT-B [ 59 ]	D2F (Unseen Region, Polarization and Microwave Band)	12.3	0.0	0.0	0.0	0.0	21.7	W2D (Unseen Region, Polarization and Microwave Band)	0.0	0.0	23.0	0.0	0.0	50.4	12.2
S12-MoCo [ 43 ]	ViT-S [ 7 ]		15.2	0.4	0.0	22.0	1.9	6.7		0.0	0.6	22.2	0.0	0.0	39.0	10.3
S12-DINO [ 43 ]	ViT-S [ 7 ]		18.9	19.4	0.0	25.5	13.6	15.5		7.4	11.8	20.3	0.0	0.0	41.7	13.6
S12-MAE [ 43 ]	ViT-S [ 7 ]		18.3	3.7	0.2	27.1	48.5	19.6		0.0	0.1	22.4	0.0	0.0	48.8	11.9
DOFA [ 60 ]	ViT-B [ 7 ]		15.8	4.9	0.1	13.6	7.0	8.3		0.0	0.0	22.5	0.0	0.0	51.5	12.3

Table 12: Performance comparison on benchmark for gaps3. Bolds = best; underlines = second best. Bldg:Bulidings. Otr:Others. Veg:Vegetation. Wtr:Water. Farm:Farmland. Green:Greenery.

Figure 9: Visualizations of predicted segmentation maps on six representative benchmarks. The blue is the water class, green is the vegetation class, red is the ground class, cyan is the road class, yellow is the building class, and purple is the mountain class.

Figure 10: Visualizations of predicted segmentation maps on six representative benchmarks. The blue is the farmland class, green is the greenery class, red is the road class, cyan is the building class, yellow is the water class, and white is the background class.

Figure 11: Visualizations of predicted segmentation maps on 1-4 benchmarks. Images in (a) and (b) respectively represent N2S and S2N. The blue is the water class, green is the vegetation class, red is the ground class, cyan is the road class, yellow is the building class, and purple is the mountain class. Images in (c) and (d) respectively represent K2C and C2K. The blue is the farmland class, green is the greenery class, red is the road class, cyan is the building class, yellow is the water class, and white is the background class.

Figure 12: Visualizations of predicted segmentation maps on 5-12 benchmarks. Images in (a), (b), (c), and (d) respectively represent VV2F, F2VV, HH2F, F2HH. Images in (e), (f), (g), and (h) respectively represent C(r)2R, R2C(r), C(i)2R and R2C(i). The blue is the water class, green is the vegetation class, red is the ground class, cyan is the road class, yellow is the building class, and purple is the mountain class.

Figure 13: Visualizations of predicted segmentation maps on 13-16 benchmarks. Images in (a) and (b) respectively represent F2A, A2F. The blue is the water class, green is the vegetation class, red is the building class, yellow is the road class, and black is the other class. Images in (c) and (d) respectively represent O2D, D2O. The blue is the farmland class, green is the greenery class, red is the road class, cyan is the building class, yellow is the water class, and white is the background class.

Figure 14: Visualizations of predicted segmentation maps on 17-18 benchmarks. Images in (a) and (b) respectively represent S2A and A2S. The white is the building class, black is the other class.

Figure 15: Visualizations of predicted segmentation maps on 19-22 benchmarks. Images in (a), (b), (c), and (d) respectively represent D2F, F2D, W2D and D2W. For the D2F benchmark, the blue is the water class, green is the vegetation class, red is the building class, yellow is the road class, and black is the other class. For the other three banechmarks, the blue is the farmland class, green is the greenery class, red is the road class, cyan is the building class, yellow is the water class, and white is the background class.

Domain Generalization for Semantic Segmentation

See all 14 papers

GeoSET: Generalist Foundation Model for SAR-to-EO Image Translation
Sep 26, 2026Jeonghyeok Do, Munchurl KimImage-to-Image Translation Geospatial Foundation Models
Korea Advanced Institute of Science and Technology (KAIST)
OptiSAR-Net++: A Large-Scale Benchmark and Transformer-Free Framework for Cross-Domain Remote Sensing Visual Grounding
Mar 25, 2026Xiaoyu Tang, Jun Dong, Jintao Cheng +1Remote Sensing Vision-Language Grounding
School of Electronics and Information Engineering, and Xingzhi College, South China Normal University, Foshan 528225, China · School of Data Science and Engineering, and Xingzhi College, South China Normal University, Shanwei, 516600, China · Department of Electronic and Computer Engineering, Hong Kong University of Science and Technology, Hong Kong SAR, China +1
Unlocking Optical Prior: Spectrum-Guided Knowledge Transfer for SAR Generalized Category Discovery
Apr 24, 2026Jingyuan Xia, Ruikang Hu, Ye Li +3Domain Adaptation Frequency-Domain Feature Learning
College of Electronic Science and Technology, National University of Defense Technology, Changsha, 410073, China · State Key Laboratory of Complex System Simulation and Modeling Technology, Beijing, 100091, China

Sep 26, 2026J/K move · Enter open · S save

GeoSET: Generalist Foundation Model for SAR-to-EO Image Translation

Jeonghyeok Do, Munchurl Kim

Korea Advanced Institute of Science and Technology (KAIST)

Image-to-Image Translation Geospatial Foundation Models Synthetic Aperture Radar

Open paper AskPDF arXiv

Paired synthetic aperture radar (SAR) and electro-optical (EO) imagery is increasingly available across sensors, resolutions, and geographic regions. Yet existing SAR-to-EO image translation (SET) methods are typically trained on a single, limited-scale dataset, producing models specialized to particular sensing conditions. We introduce GeoSET, the first generalist model for SET, built around a single pretrained parent that is adapted to downstream datasets under a common protocol. We curate over 3 million high-quality SAR--EO pairs from a collection of more than 10 million SAR observations, spanning diverse sensors, spatial resolutions, and ground sampling distances. To bridge the modality gap between SAR observations and a pretrained image generator, we develop a speckle-robust SAR encoder and pretrain the conditional generator on this heterogeneous corpus. The resulting parent supports efficient adaptation across downstream datasets through low-rank adaptation (LoRA), updating only 0.60% of the generator parameters and requiring approximately one hour per dataset. Across six downstream benchmarks, GeoSET achieves state-of-the-art results in FID and DISTS with full fine-tuning or LoRA, demonstrating effective transfer across heterogeneous SAR-EO domains.

Abstract

Domain Generalization for Semantic Segmentation

Explore similar work

GeoSET: Generalist Foundation Model for SAR-to-EO Image Translation

OptiSAR-Net++: A Large-Scale Benchmark and Transformer-Free Framework for Cross-Domain Remote Sensing Visual Grounding

Unlocking Optical Prior: Spectrum-Guided Knowledge Transfer for SAR Generalized Category Discovery