MaRS: A Multi-Modality Very-High-Resolution Remote Sensing Foundation Model with Cross-Granularity Meta-Modality Learning

Ruoyu Yang, Yinhe Liu, Heng Yan, Yiheng Zhou, Yihan Fu, Han Luo, Yanfei Zhong

Figure 1. Overall framework of MaRS and examples of downstream tasks.

Abstract

We propose MaRS, a multi-modality very-high-resolution (VHR) remote sensing foundation model designed for cross-modality, cross-granularity interpretation of complex scenes. We construct MaRS-16M, a large-scale VHR SAR optical paired dataset (16M+ pairs) via collection and semi-automated processing. MaRS tackles two core VHR SAR optical SSL challenges: imaging discrepancy and modal representation gap. To this end, we introduce Cross-Granularity Contrastive Learning (CGCL) to alleviate alignment inconsistencies by linking patch- and image-level semantics, and design Meta-Modality Attention (MMA) to unify heterogeneous physical characteristics across modalities. Compared with existing RSFMs and general VFMs, MaRS serves as a strong pretrained backbone across nine multi-modality VHR downstream tasks. Dataset and code are available at the project page.

Highlights & Contributions

MaRS-16M: a globally distributed, precisely paired VHR SAR optical dataset with 16,785,168 patch pairs (0.35 m GSD).
MaRS model: dual-encoder (SwinV2) with CGCL (patch/global & patch-to-global alignment) and MMA (alternating intra-/cross-modality attention) for robust fusion.
State-of-the-art results on nine VHR multi-modal tasks (registration, translation, mapping under modality-missing, change/damage, detection, height estimation, roads, etc.).

Capabilities

Handles VHR SAR & optical with robust cross-modal alignment and fusion.
Generalizes to fine-grained urban/disaster scenarios and low-visibility conditions.
Serves as a pretrained backbone across diverse downstream tasks with strong transfer.

Method Overview

Architecture. MaRS adopts dual encoders (E_RGB, E_SAR, both SwinV2) to extract modality-specific tokens, followed by a Meta-Modality Attention (MMA) Transformer that alternates intra-modality and cross-modality attention to form a unified representation. Light decoders are used for dense prediction.

Pretraining. We combine three self-supervised strategies: (1) CGCL at patch-, image-, and patch-to-global levels to mitigate local distortions while preserving global semantics; (2) masked image modeling per modality branch; (3) continued pretraining on VHR optical to further strengthen representation quality. Training uses 512×512 inputs, 60% masking, and runs on 8×A800 GPUs.

Figure 3. MaRS pretraining with CGCL and MMA.

Dataset Distribution

Figure 2. MaRS-16M is composed from Umbra & Capella (X-band, HH/VV), globally distributed.

Key Parameters

Pairs: 16,785,168 SAR optical patch pairs (≈0.35 m GSD).
SAR Sources: Umbra & Capella; X-band; HH or VV polarization.
Collection: 4,225 VHR SAR images collected (to 2024-12-30), quality-filtered; corresponding VHR optical retrieved.
Preprocess: registration checker → resampling → registration; standardized 512×512 patches for training/eval.
Coverage: diverse land cover, urban patterns, disaster scenarios; globally distributed.
Use: designed for large-scale SSL and robust cross-modal alignment under geometric distortion/noise.

Experiments & Results (VHR Multi-Modal)

MaRS is evaluated across nine representative VHR tasks and achieves strong results compared with RSFMs/VFMs and task-specific methods:

Cross-Modality Registration (GUSO): RMSE ≈ 2.83.
Modality-Missing Mapping (EarthMiss): mIoU ≈ 49.90.
Cross-Modality Translation (GUSO): PSNR ≈ 20.69, SSIM ↑.
SAR Target Detection (ARTDet) / SARDet-100K: mAP ≈ 55.40.
Building Detection (UBC-V2): mAP ≈ 26.80.
Building Height Estimation (DFC23-T2): Δ (height) ≈ 54.04.
Change Detection (WHU-CD, optical): IoU ≈ 87.02.
Road Extraction (DeepGlobe, optical): IoU ≈ 68.44.
Damage Assessment (DFC25-T2): IoU ≈ 41.49.

Qualitative visualizations for four downstream tasks

Figure 5. Qualitative results over multiple VHR downstream tasks.

Feature Activations

MaRS produces sharper activations along object boundaries and shows more consistent activation regions across RGB/SAR, indicating improved modality-invariant representation and fine-grained detail modeling.

Figure 7. Feature heatmaps comparing SwinV2 vs. MaRS.

BibTeX

@inproceedings{yang2026mars,
  title={MaRS: A Multi-Modality Very-High-Resolution Remote Sensing Foundation Model with Cross-Granularity Meta-Modality Learning},
  author={Ruoyu Yang and Yinhe Liu and Heng Yan and Yiheng Zhou and Yihan Fu and Han Luo and Yanfei Zhong},
  booktitle={AAAI Conference on Artificial Intelligence},
  year={2026}
}