Multimodal foundation models have achieved remarkable progress, thanks in large part to vast optical benchmarks. However, according to the research paper introducing SARLO-80, comparable resources for synthetic aperture radar (SAR) remain limited. Existing SAR-optical datasets largely rely on low-resolution, intensity-only Ground Range Detected (GRD) products that do not preserve complex-valued SAR measurements or native acquisition geometry. This restricts physically grounded multimodal learning, particularly for very-high-resolution (VHR) SAR data. SARLO-80 addresses this gap by providing a public dataset that combines VHR SAR imagery, aligned optical patches, and natural-language descriptions.
The Gap in SAR-Optical-Text Data
Before SARLO-80, large-scale public datasets combining VHR SAR Single Look Complex (SLC) data, aligned optical imagery, and text were lacking. The authors note that existing resources typically use GRD products, which discard phase information and do not maintain the native slant-range geometry. This limits the ability to train models that can leverage the full richness of SAR signals, such as complex amplitudes and polarimetric information. SARLO-80 is designed to fill this void by preserving complex-valued SAR measurements and using a consistent slant-range grid.
Building SARLO-80: Data and Methodology
SARLO-80 is built from open-access Umbra spotlight acquisitions distributed as Sensor Independent Complex Data (SICD). According to the paper, the team started with around 2,500 worldwide scenes covering VV and HH polarizations, with native resolutions ranging from 20cm to 2m. To ensure uniformity, all SAR data were standardized to an 80cm slant-range grid using band-limited FFT resampling. The imagery was then tiled into 1024×1024 patches.
For each SAR patch, the authors retrieved a high-resolution optical tile from a corresponding source and warped it into the SAR grid using local coordinate correspondences to achieve pixel-level alignment. Additionally, three caption variants—SHORT, MID, and LONG—were generated per sample, supporting diverse vision–language training and evaluation tasks.
Dataset Composition and Specifications
The resulting dataset comprises 119,566 triplets: each triplet includes a complex and amplitude slant-range SAR patch, an aligned optical patch, and a natural-language description. The data covers 257 locations across 72 countries, spanning a broad range of land types and infrastructures. The table below summarizes key statistics:
| Feature | Value |
|---|---|
| Total triplets | 119,566 |
| SAR resolution | 80cm slant-range |
| Original native resolution | 20cm – 2m |
| SAR polarizations | VV, HH |
| Patch size | 1024×1024 pixels |
| Number of locations | 257 |
| Countries covered | 72 |
| Caption variants per sample | 3 (SHORT, MID, LONG) |
| Data format | Complex and amplitude SAR patch + optical patch + text |
The dataset is split into fixed train, validation, and test partitions to enable reproducible benchmarks for multimodal alignment tasks such as cross-modal retrieval and conditional generation.
Availability and Reproducibility
SARLO-80 is publicly available on the Hugging Face Hub. Alongside the dataset, the authors release the full preprocessing and baseline code, allowing researchers to reproduce the results and build upon the work. The dataset uses a slant-range geometry that preserves native acquisition characteristics, which is expected to benefit physically grounded multimodal learning.
Implications for Enterprise AI
For enterprise technology leaders in sectors like logistics and infrastructure monitoring, the availability of SARLO-80 opens possibilities for training AI models that combine radar satellite imagery with optical data and text. Such models could improve automated detection of infrastructure changes, support supply chain visibility through satellite-based monitoring, and enhance geospatial intelligence. While the dataset is primarily research-oriented, its structure aligns with practical use cases in transportation, agriculture, and defense.
The paper is authored by Debuysère, Solène, Trouvé, Nicolas, Letheule, Nathan, Colin, Elise, Channing, and Georgia. It was published on arXiv and is accessible at https://arxiv.org/abs/2606.20523.