Remote Sensing · ONERA
SARLO-80
119K paired SAR/optical images with text captions at 80cm resolution for multimodal learning.
What SARLO-80 is
It is designed to support multimodal earth observation and multimodal learning tasks where radar and visual spectrum imagery are analysed alongside descriptive natural language.
The radar data are sourced from Umbra satellite acquisitions resampled to an 80 cm by 80 cm azimuth and range resolution in slant-range geometry. Rather than supplying pre-rendered optical raster files, the collection bundles comprehensive Sensor Independent Complex Data (SICD) metadata and geographic coordinates that allow matching optical satellite tiles to be fetched and projected into the exact slant-range frame of the SAR acquisition.
Each entry also contains three levels of text descriptions generated from optical observations and refined with language models. By pairing complex-valued and amplitude radar data with reproducible optical alignment coordinates and captions, SARLO-80 provides a standardised benchmark for multimodal representation learning and cross-modal retrieval.
How it was made
SAR data are derived from Umbra SICD products resampled to a uniform 80 cm by 80 cm spacing in azimuth and range while preserving the native slant-range acquisition geometry. For each sample, the dataset supplies an 80 cm amplitude image, a complex-valued NumPy array containing single-look complex data, and the raw SICD XML metadata defining the sensor geometry and projection parameters.
Optical imagery is paired dynamically: metadata files record the WGS84 bounding coordinates, optical tile source, and zoom level. To project optical data into the SAR frame, control points from the SICD metadata are projected to WGS84, mapped to optical pixel coordinates, and used to calculate an affine transformation that warps the optical tile into the SAR crop frame. Natural language captions are generated from the optical tiles using CogVLM2 and subsequently cleaned and post-processed using the Qwen large language model into short, medium, and long text descriptions.
Key facts
- Spatial resolution
- 80 cm × 80 cm (azimuth × range)
- SAR geometry
- Slant-range geometry
- SAR data source
- Umbra SICD products
- File formats
- WebDataset shards containing PNG, NPY, XML, and JSON
- Optical data delivery
- On-demand reconstruction via metadata and tile sources
- Text annotations
- Three captions per sample (SHORT, MID, LONG)
What you can use it for
- Multimodal SAR-Optical Representation Learning
- Train vision models to learn joint representations between high-resolution optical imagery and slant-range synthetic aperture radar acquisitions.
- Vision-Language Alignment
- Develop and evaluate cross-modal text-to-image retrieval and captioning models using the multi-tiered short, medium, and long descriptions.
- Complex-Valued SAR Analysis
- Perform radar signal analysis and phase-sensitive model training using the included complex-valued SAR arrays rather than amplitude-only imagery.
Explore the data
No licence declared
The authors have not published a licence, which means no permission to redistribute the data has been granted to anyone. The rows are not served here; read them on the original dataset page.Strengths and limits
Good at
- Provides both single-look complex data arrays and calibrated amplitude images at 80 cm resolution.
- Includes full SICD XML files and geometric parameters to allow reproducible mathematical projection between ground coordinates and slant-range radar space.
- Supplies three distinct lengths of text descriptions per sample to support diverse vision-language objectives.
Watch out for
- Optical image files are not pre-packaged and require researchers to download external tiles and perform projection transforms locally.
- Text descriptions are machine-generated via CogVLM2 and Qwen rather than manually created by human annotators.
- Radar imagery is retained in slant-range geometry rather than projected onto an orthorectified ground plane.
Questions
Can I look inside SARLO-80 without downloading it?
Not here: it declares no licence, so no permission to redistribute its rows has been granted. Its description and statistics are still shown.
What licence is SARLO-80 under?
Its authors have not declared a licence, which means no permission to reuse or redistribute it has been granted.
Why are optical images not bundled as raster files?
Optical images are omitted from direct distribution due to licensing constraints. They are instead reconstructed on demand using the tile provider metadata and SICD projection parameters included in each sample.
What SAR image formats are provided in each WebDataset shard?
Each sample includes an 80 cm amplitude image in PNG format, a complex-valued SAR array in NPY format, an associated SICD XML metadata file, and a JSON metadata file.
How are the natural language captions created?
Captions are automatically generated from optical imagery using CogVLM2 and subsequently filtered and post-processed using the Qwen large language model into short, medium, and long formats.
In what coordinate system or geometry is the SAR data provided?
SAR images are provided in their native slant-range geometry resampled to an 80 cm by 80 cm resolution in azimuth and range.
How to cite it
arXiv:2606.20523
Written from SARLO-80's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.