Abstract

Synthetic Aperture Radar (SAR) and optical imagery provide complementary strengths that constitute the critical foundation for transcending single-modality constraints and facilitating cross-modal collaborative processing and intelligent interpretation. However, existing benchmark datasets often suffer from limitations such as single spatial resolution, insufficient data scale, and low alignment accuracy, making them inadequate for supporting the training and generalization of multi-scale foundation models. To address these challenges, we introduce SOMA-1M (SAR-Optical Multi-resolution Alignment), a pixel-level aligned dataset containing over 1.3 million pairs of georeferenced images with a specification of 512 × 512 pixels. This dataset integrates SAR imagery from Sentinel-1, PIESAT-1, and Capella Space, spanning native spatial resolutions from 0.5 m to 10 m, with corresponding optical imagery. It provides a coarse classification into 12 typical scene categories, offering a descriptive view of scene diversity. To address multimodal projection deformation and massive data registration, we designed a rigorous coarse-to-fine image matching framework to refine SAR–optical geometric correspondence. Based on this dataset, we established comprehensive evaluation benchmarks for four hierarchical vision tasks, including image matching, image fusion, SAR-assisted synthetic cloud removal, and cross-modal translation, involving over 30 mainstream algorithms. Experiments conducted with the SOMA-0.1M training subset provide empirical evidence of the dataset's utility across four representative multimodal remote sensing tasks. For image matching, fine-tuning on SOMA-0.1M consistently improves multiple matching architectures under the adopted protocol, with generalization gains also observed on external datasets. SOMA-1M serves as a foundational resource for robust multimodal algorithms and remote sensing foundation models. The dataset will be made publicly available in accordance with the applicable data licensing agreements.

Fig. 1. Overview of the SOMA-1M dataset and examples of its multi-task applications. The two leftmost columns display the original SAR and optical input images. The remaining columns illustrate representative results generated by models trained on this dataset: (a) Image Matching; (b) Image Fusion; (c) SAR-Assisted Cloud Removal; and (d) SAR-to-Optical Translation.

SOMA-1M Dataset

Fig. 2. Global geographic distribution of SOMA-1M sampling points.

SOMA-1M is a large-scale SAR–optical multimodal remote sensing dataset spanning 1,466 geographic sampling locations worldwide. It integrates multi-resolution imagery from Sentinel-1, PIESAT-1, Capella Space, and Google Earth, with original scenes ranging from 8,000 to 35,000 pixels and spatial resolutions from 0.5 m to 10 m. After rigorous coarse-to-fine registration and quality cleaning, the dataset provides 1,300,954 pixel-level aligned SAR–optical image pairs, each standardized to $512 \times 512$ pixels, covering 12 representative land-cover categories and supporting scalable benchmarks for image matching, fusion, cloud removal, and cross-modal translation.

The test dataset can be obtained at following links:
(1) google drive https://drive.google.com/file/d/1IFCk6YSKrctfvB9LdeLq8ouvzK7stuK6/view?usp=drive_link
(2) baidu drive https://pan.baidu.com/s/1WKll9nzauD2M57OLXgkM4A Extract code: k656.
SOMA-1M dataset resources will be released progressively through this repository in accordance with applicable third-party data licenses and copyright requirements. More subsets of SOMA-1M are being prepared for public release.

Citation

If you find our work useful in your research, please consider giving a star ⭐ and a citation

@article{wu2026soma,
  title={SOMA-1M: A Large-Scale SAR-Optical Multi-resolution Alignment Dataset for Multi-Task Remote Sensing},
  author={Wu, Peihao and Yao, Yongxiang and Wan, Yi and Zhang, Wenfei and Zhao, Ruipeng and Li, Jiayuan and Zhang, Yongjun},
  journal={arXiv preprint arXiv:2602.05480},
  year={2026}
}