SARLO-80: Worldwide Slant SAR Language Optic Dataset 80cm

Researchers released SARLO-80, a large-scale dataset combining very-high-resolution synthetic aperture radar (SAR) imagery, aligned optical images, and natural-language descriptions across 2,500 worldwide scenes. The dataset addresses a critical gap in multimodal AI training by preserving complex-valued SAR measurements and native acquisition geometry, enabling more physically grounded foundation models for Earth observation applications.

The advancement of multimodal foundation models has relied heavily on optical imaging benchmarks, leaving SAR imaging—a complementary technology essential for all-weather Earth monitoring—significantly underrepresented in large-scale datasets. SARLO-80 bridges this gap by introducing 119,566 aligned triplets spanning 257 locations across 72 countries, standardized to 80cm resolution and processed to maintain physical integrity. This represents a qualitative leap from existing SAR-optical datasets that typically use low-resolution, intensity-only products that lose critical information about SAR's complex-valued nature.

The dataset's technical rigor distinguishes it from prior work. By preserving Sensor Independent Complex Data (SICD) format and using band-limited FFT resampling, the researchers maintain physical fidelity essential for scientific applications. Pixel-level alignment between SAR and optical imagery, combined with three caption variants per sample, creates a resource suitable for cross-modal retrieval and conditional generation tasks—directly supporting the training of next-generation multimodal models.

The Earth observation and AI communities stand to benefit substantially. SAR's all-weather capabilities make it invaluable for agriculture, disaster response, and infrastructure monitoring, yet current AI tools lack the training data to exploit these advantages effectively. By releasing reproducible preprocessing code and standardized train-validation-test splits on Hugging Face, the researchers enable benchmarking and comparative studies that have been impossible at this scale. This democratization of VHR SAR data accelerates development cycles and lowers barriers to entry for researchers and commercial applications relying on robust multimodal understanding of planetary surface conditions.

• →SARLO-80 contains 119,566 SAR-optical-text triplets covering 72 countries, addressing a critical shortage of large-scale SAR datasets for multimodal AI training.
• →Dataset preserves complex-valued SAR measurements and native acquisition geometry rather than low-resolution intensity products, enabling more physically grounded learning.
• →Pixel-level alignment between SAR and optical imagery with three caption variants supports vision-language models and cross-modal retrieval applications.
• →Open release on Hugging Face with full preprocessing code enables reproducible benchmarks and accelerates development of Earth observation AI tools.
• →All-weather SAR capabilities combined with optical fusion unlock new possibilities for agriculture, disaster response, and infrastructure monitoring applications.