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

Conference: CVPR2026 arXiv: 2603.12008 Code: VisionXLab/CrossEarth-SAR Area: Semantic Segmentation Keywords: SAR semantic segmentation, vision foundation model, Mixture of Experts (MoE), domain generalization, remote sensing

TL;DR

This paper presents CrossEarth-SAR, the first billion-scale SAR vision foundation model, which integrates a physics-guided sparse MoE architecture with SAR physical descriptors. It achieves state-of-the-art performance on 20 out of 22 cross-domain semantic segmentation benchmarks, surpassing prior methods by over 10% mIoU in certain multi-gap scenarios.

Background & Motivation

1. Advantages of SAR all-weather observation: SAR is unaffected by weather and illumination conditions, making it a critical tool for time-sensitive applications such as disaster monitoring, environmental surveillance, and urban management. However, semantic understanding in the SAR domain is substantially more challenging than in the optical domain.
2. Inherent SAR imaging challenges: Coherent imaging introduces multiplicative speckle noise; side-looking geometry causes spatial distortions such as layover, foreshortening, and shadowing; and radar backscattering rather than color creates semantic ambiguity. These three factors collectively violate the fundamental assumptions of modern vision models.
3. Extreme domain fragmentation: SAR data suffers from severe domain fragmentation due to variations in sensor platform (Sentinel-1/ALOS-2/Capella), frequency band (C/L/X), polarization mode (HH/HV/VH/VV), and incidence angle, leading to catastrophic performance degradation during cross-domain transfer.
4. Inadequacy of existing foundation models for SAR: Geospatial foundation models such as SatMAE and SkySense are primarily designed for optical multispectral data; their architectures and pre-training strategies do not account for SAR backscattering physics or noise characteristics.
5. Lack of unified DG evaluation standards: The SAR cross-domain semantic segmentation community lacks a systematic domain generalization benchmark, hindering fair comparison and methodological progress.
6. Scarcity of large-scale annotated SAR data: Acquiring high-quality semantic segmentation annotations for SAR imagery is difficult, constraining the training of large-scale models.

Method

Overall Architecture

CrossEarth-SAR integrates a physics-guided sparse MoE into a DINOv2 ViT backbone, replacing the original FFN with MoE modules consisting of a router and multiple experts. Input SAR images are replicated into 3-channel tensors and encoded by the ViT; simultaneously, three SAR physical descriptors are computed to assist routing. A Mask2Former decoder then generates the final segmentation predictions. The model is offered in three scales—S/B/L—with 90M/300M/1.3B total parameters and 20M/80M/300M activated parameters, respectively.

SAR Physical Descriptors

A SAR Physical Operator \(g_{\text{sar}}(\cdot)\) is designed to compute three physical descriptors that provide the router with stable physical priors:

• Directional Entropy \(H_{DE}\) (imaging geometry): Quantifies structural regularity via the entropy of a Sobel gradient-direction histogram. Low values indicate clear linear edge structures; high values indicate complex irregular structures.
• Equivalent Number of Looks ENL (radar system): \(\text{ENL} = (\mu/\sigma)^2\), measuring speckle noise intensity. High values indicate weak speckle and statistical stability; low values indicate large noise fluctuations.
• Local Roughness \(R_{LR}\) (target scattering): Block-wise variance of local means, capturing textural variability. High values indicate complex textures; low values indicate smooth textures.

The three descriptors are concatenated as \(s = [H_{DE}, \text{ENL}, R_{LR}] \in \mathbb{R}^3\) and, at each ViT block, appended to the token embeddings before being fed into the router.

Physics-Guided Sparse MoE

• Router: Concatenates token embeddings \(Z\) with descriptors \(S\) and computes softmax scores \(\pi = \text{softmax}(W_r[Z \| S] + b_r)\), selecting the Top-\(k\) experts.
• Token-level MoE aggregation: \(\tilde{z} = \sum_{k \in \mathcal{I}} g_k \cdot E_k(z)\), where gating weights are normalized routing scores.
• Optimal configuration: \(n=6\) experts with Top-\(k=1\) activation, balancing computational efficiency and model capacity.

Loss & Training

\[\mathcal{L} = \mathcal{L}_{\text{seg}} + \mathcal{L}_{\text{BC}}\]

• \(\mathcal{L}_{\text{seg}}\): Mask2Former segmentation loss.
• \(\mathcal{L}_{\text{BC}} = \lambda_{\text{BC}} \cdot n \sum_{k=1}^{n} f_k p_k\) (load-balancing loss, \(\lambda_{\text{BC}}=0.005\)), preventing expert collapse.

Data & Benchmarks

• CrossEarth-SAR-200K: Integrates publicly available SAR data (40K fully supervised) with collected data (160K weakly supervised via pseudo-labels), covering hundreds of cities across six continents, with all images uniformly cropped/resized to \(512\times512\).
• 22 DG benchmarks: Spanning 8 types of domain gaps (region / polarization / complex-valued / region+polarization / region+platform / region+microwave-band / region+polarization+band / region+platform+band), constructed from 6 public datasets.

Key Experimental Results

Single Domain Gap (Tab. 2)

Method	Params	N2S	VV2F	HH2F	C(r)2R	12-bench Avg
DINOv2 (Baseline)	300M	32.3	65.7	56.8	71.3	55.5
DINOv3	300M	33.7	48.3	50.6	69.9	53.0
MTP	300M	30.6	30.4	36.0	70.8	44.7
CrossEarth-SAR-L	1.3B(300M)	37.8	73.8	72.3	76.4	61.9
CrossEarth-SAR-L*	1.3B(300M)	38.0	73.9	71.8	76.9	62.7

• Average over 12 single-gap benchmarks: CrossEarth-SAR-L* reaches 62.7%, surpassing the baseline by +7.2%.
• Largest gain on HH2F: +15.5% mIoU (56.8→72.3).

Multiple Domain Gaps (Tab. 3)

Method	Params	F2A	A2F	O2D	S2A	D2F	W2D	10-bench Avg
DINOv2 (Baseline)	300M	13.4	15.5	17.8	55.9	26.0	16.7	24.3
CrossEarth-SAR-L*	1.3B(300M)	16.1	27.0	23.1	57.9	26.5	25.6	28.5

• Average over 10 multi-gap benchmarks: +4.2%; largest gain on A2F: +11.5%.

Ablation Study (Tab. 5–6)

• Pseudo-label effectiveness: Fully supervised 40K only → 45.1%; adding 160K weakly supervised → 59.4% (+14.3%).
• MoE design: No load balancing, no descriptors → 61.1%; full proposal → 62.4% (+1.3%).
• Physical descriptors: Each of the three descriptors contributes independently to different domain gaps; combined use yields the best performance.
• Number of experts \(n\): Monotonically increasing from \(n=3\) to \(6\) (60.9→62.4).
• Top-\(k\) selection: \(k=1\) is optimal (62.4); \(k=2/3\) degrades performance.

Highlights & Insights

• First billion-scale SAR foundation model: Sparse MoE scales the parameter count to the billion level while keeping inference cost manageable (only 300M parameters activated).
• Physics-guided routing mechanism: Three SAR physical descriptors address routing instability in MoE on heterogeneous SAR data, with a design that is both elegant and physically interpretable.
• Systematic contributions: The work simultaneously introduces a 200K large-scale pre-training dataset, 22 DG benchmarks, and three model scales (S/B/L), forming a complete research infrastructure for the community.
• State-of-the-art on 20 of 22 benchmarks: Comprehensive evaluation covering single-, dual-, and triple-gap scenarios, with improvements exceeding 10% mIoU in certain settings.

Limitations & Future Work

• Pseudo-labels are generated by the CrossEarth model trained on paired optical images; their quality is bounded by optical–SAR matching accuracy, and annotation reliability may be questionable in some scenarios.
• Performance gains in multi-gap settings (e.g., D2O, D2F) are limited and occasionally fall below the baseline, indicating that three-gap domain generalization remains an open problem.
• Pre-training requires 16× A100 80GB GPUs, posing a prohibitively high computational barrier for community reproduction and deployment.
• The physical descriptors are hand-crafted 3-dimensional vectors with limited information capacity; learnable physical feature extractors are worth exploring.
• Top-\(k=1\) means only one expert is activated per token, leaving the potential of multi-expert collaboration largely untapped.
• Evaluation is restricted to semantic segmentation; generalization to other SAR tasks such as object detection and change detection has not been validated.

Related Work & Insights

Dimension	CrossEarth-SAR	CrossEarth (Optical DG)	SARATR-X (SAR Object Recognition)	SatMAE/SkySense (Optical FM)
Modality	SAR-specific	Optical	SAR	Optical/multispectral
Task	Cross-domain semantic segmentation	Cross-domain semantic segmentation	Object recognition	Multi-task
Architecture	MoE + ViT	ViT	HiViT	ViT
Parameters	1.3B (sparse)	300M	60M	300M
Physical prior	SAR descriptor-guided routing	None	None	None
DG benchmarks	22 / 8 gap types	Optical DG	No systematic DG evaluation	No systematic DG evaluation

Rating

• Novelty: ⭐⭐⭐⭐ — Physics-guided MoE routing is a first in the SAR domain; the three-descriptor design demonstrates clear physical insight.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ — 22 benchmarks covering 8 gap types, comprehensive ablations, and thorough visualizations.
• Writing Quality: ⭐⭐⭐⭐ — Clear structure and well-articulated motivation, though the large number of tables makes the presentation slightly verbose.
• Value: ⭐⭐⭐⭐ — The complete ecosystem (model + data + benchmarks) represents a significant contribution to the SAR community.

Related Papers

• [CVPR 2026] SARMAE: Masked Autoencoder for SAR Representation Learning
• [CVPR 2026] Generalizable Knowledge Distillation from Vision Foundation Models for Semantic Segmentation
• [CVPR 2026] A Mixed Diet Makes DINO An Omnivorous Vision Encoder
• [CVPR 2026] Prompt-Driven Lightweight Foundation Model for Instance Segmentation-Based Fault Detection in Freight Trains
• [CVPR 2026] CA-LoRA: Concept-Aware LoRA for Domain-Aligned Segmentation Dataset Generation