Authors: Solène Debuysère1, Nicolas Trouvé1, Nathan Letheule1, Elise Colin1, Georgia Channing2
Affiliations:
1 ONERA – The French Aerospace Lab
2 Hugging Face
Satellite imagery has transformed the best way we observe our planet. More often than not, these images come from optical sensors, which capture the world in visible light, identical to our eyes. But there’s one other approach to observe the planet: Synthetic Aperture Radar (SAR). SAR uses microwaves as a substitute of visible light and may capture images at any time of day, even through clouds or bad weather.
We curated raw Umbra SAR acquisitions to create the SARLO-80 (Slant SAR Language Optic, 80 cm) dataset, a structured, high-resolution multimodal resource optimized for AI and machine learning applications. By pairing SAR imagery with geometrically aligned optical data and natural-language descriptions, it bridges radar and vision–language domains.
Before outlining the processing steps, it’s helpful to briefly recall how SAR differs from conventional optical sensing.
Dataset repo: ONERA/SARLO-80
Optics vs Radar: Two Different Views of Earth
Optical and radar imaging provide two fundamentally alternative ways of observing the Earth’s surface. While optical imagery resembles natural photographs formed by visible light, Synthetic Aperture Radar (SAR) imagery is constructed from microwave echoes that interact with the physical and electromagnetic properties of the terrain. This difference affects every aspect of image acquisition, resolution, geometry, and interpretation.
1. Lively and Passive Sensing
Unlike optical sensors that rely upon sunlight and clear skies, SAR actively emits microwaves and may image the Earth even through clouds — a key advantage when over 60% of the planet is roofed by clouds at any given time.
|
Figure 1: Example of Capella Image where Sequoia satellite of Brazil demonstrates how our high resolution SAR (left) can provide a transparent view of deforestation, even when clouds obscure optical images (right). |
2. Image Formation Principles
An optical image is a direct projection of sunshine through a lens onto a sensor array. Radar imagery, in contrast, is reconstructed computationally from a sequence of radar echoes collected because the satellite moves along its orbit. By combining measurements over time, the system synthesizes a big “virtual” antenna — the synthetic aperture — which enables fantastic spatial resolution (see Figure 3).
In optical systems, spatial resolution depends totally on the aperture size of the lens. In radar systems, it depends as a substitute on signal frequency, bandwidth, and the gap traveled by the sensor during data acquisition. This distinction allows SAR satellites to attain high resolution even with relatively compact antennas. This resolution is encoded in the dimensions of the brilliant points, with each point corresponding roughly to the smallest distinguishable feature the radar can resolve.
3. Radar Geometry and Distortions
Optical and radar sensors observe the Earth from fundamentally different geometries. Optical systems capture images in a ground-projected plane (green plane in Figure 2), where each pixel corresponds directly to some extent on the surface. In contrast, Synthetic Aperture Radar (SAR) acquires data in slant range geometry (orange plane in Figure 2), measuring distances along the radar’s line of sight. To make SAR and optical images geometrically comparable, one in every of them have to be reprojected into the geometry of the opposite—or each into a standard reference geometry—to attain approximate geometric superposability, since perfect geometric superposition is physically not possible as a consequence of their distinct viewing geometries.
Figure 2: SAR geometry acquisition with slant-range and ground-range planes.
Moreover, this oblique acquisition causes elevated terrain and tall structures to seem displaced toward the sensor, introducing geometric distortions reminiscent of:
- Layover – Tall structures, reminiscent of mountains or buildings, appear to lean toward the radar because their upper parts return signals before their bases.
- Foreshortening – Slopes facing the radar appear compressed because their top and bottom are illuminated almost concurrently.
- Shadowing – Areas hidden from the radar beam appear dark or unmeasured.
Figure 3: Comparison of optical vs. SAR image formation and distortions.
These effects are inherent to radar imaging and carry useful details about surface topography and orientation.
Figure 4: Example of layover in Copenhagen.
Figure 5: Example of volcano foreshortening.
4. Coherence and Speckle Characteristics
SAR sensors record not only the amplitude of the backscattered signal but in addition its phase — the precise timing of the returned wave. This property makes radar data coherent, enabling advanced techniques reminiscent of polarimetry and interferometry (InSAR).
Coherence also produces a characteristic speckle pattern, visible as granular texture in SAR images. Speckle results from the constructive and destructive interference of radar signals scattered by multiple small targets inside a single resolution cell. Even though it may resemble noise, speckle is a deterministic phenomenon that accommodates information in regards to the surface’s physical structure and scattering behavior.
5. Interpretation and Applications
Interpreting SAR imagery requires understanding that brightness corresponds to backscattering intensity reasonably than optical brightness or color. Highly reflective surfaces (e.g., rough terrain or metallic structures) appear vibrant, while smooth surfaces (e.g., calm water or flat soil) appear dark. Despite its more abstract appearance, SAR provides unique observational capabilities that complement optical data:
- Surface deformation monitoring using interferometry
- Mapping of soil moisture, vegetation, and ice dynamics
- Detection of infrastructure, ships, and flood extents
Together, optical and radar observations form a comprehensive view of the Earth — optical systems providing intuitive visual context, and radar systems revealing structural, dynamic, and geophysical properties invisible to the human eye.
Creating the Umbra Dataset
Figure 6: Worldwide map of Umbra data.
Open data source: Umbra Open Data
Although radar offers remarkable sensing capabilities, it stays difficult to process. To make this data more accessible, we curated and transformed the open-source radar imagery collected by the Umbra satellite constellation right into a machine-learning–ready format.
We began from around 2,500 Umbra SICD images acquired across the globe. These SAR scenes, captured in complex format and VV or HH polarization, span resolutions from 20 cm to 2 m and incidence angles between 10° and 70°. To standardize them, we refocused the spectrum and resampled all data to 80 cm × 80 cm in slant-range geometry, then split each large scene into overlapping 1,024 × 1,024 pixel patches.
To make the dataset multimodal, each SAR patch was paired with a high-resolution optical image projected into the radar’s slant-range geometry. This ensures pixel-level alignment between radar and optical imagery, though the optical projection may show geometric distortions.
Figure 7: Example of optical and SAR pair (each in slant-range plane).
Finally, to increase the dataset to vision–language research, we generated three natural-language captions for every optical image (SHORT, MID, and LONG) using CogVLM2, refined and cleaned with Qwen LLM. For instance, in Figure 7, the captions are:
-
SHORT:
A satellite image of a dense forest with a winding road, multiple water bodies, and several other buildings.
-
MID:
A satellite image of a dense forested area featuring a winding road, multiple water bodies, and several other structures, likely docks or industrial facilities.
-
LONG:
A satellite image of a dense forested landscape with a winding road, quite a few water bodies including an extended canal with locks, and several other buildings or facilities adjoining to the waterway.
The resulting collection accommodates about 119,566 triplets — each composed of a SAR crop, a co-registered optical crop, and text descriptions — forming a foundation for training multimodal models that jointly understand radar, optical, and language data.
The dataset is accessible on Hugging Face under:
ONERA/SARLO-80
Applications of SAR and AI
The Umbra SAR Dataset brings together SAR, optical, and textual data in a standardized, multimodal format, opening recent possibilities for AI applications reminiscent of:
- Classification
- Segmentation
- Change detection
- Generative modeling
By combining radar’s all-weather, structural insights with optical imagery’s intuitive visual information, the dataset supports research across diverse domains — from monitoring crop health and soil moisture in agriculture, to rapid disaster assessment, urban growth tracking, and environmental studies like deforestation and glacier movement. This complementary approach enables AI models to learn richer, more resilient representations of the Earth, demonstrating how radar and optical imagery together provide a deeper understanding of our planet.
Conclusion
The Umbra SAR Dataset was built with one goal: to make radar more accessible for AI. By aligning high-resolution SAR with optical imagery and natural-language descriptions, it provides a foundation for brand new models that may interpret radar’s unique perspective and connect it to human-understandable concepts.
Acknowledgments
This work was carried out as a part of the PhD of Solène Debuysère at DEMR-ONERA – Université de Paris Saclay, under the supervision of Nicolas Trouvé, Nathan Letheule, and Elise Colin. We gratefully acknowledge ONERA, and particularly DEMR-SEM and Olivier Lévêque Team for providing computational and research resources, Umbra for the SAR data collections and open-access initiatives enabling research use (https://umbra.space/open-data/), and Hugging Face, particularly Georgia Channing, for her assistance on this project.
Contacts
If you could have any questions or would really like to contribute, don’t hesitate to contact us:
