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

Problem Statement

Multimodal foundation models have advanced significantly due to large optical benchmarks, but comparable resources for Synthetic Aperture Radar (SAR) remain scarce. Existing SAR–optical datasets primarily rely on low-resolution, intensity-only Ground Range Detected (GRD) products (e.g., Sentinel-1). These datasets discard complex-valued measurements (phase and amplitude) and native acquisition geometry, thereby losing critical physical information such as layover, shadow, and foreshortening. Consequently, current resources offer limited support for learning representations that fully exploit SAR-specific physical properties. There is a distinct lack of large-scale, public datasets that combine very-high-resolution (VHR) SAR Single Look Complex (SLC) imagery, aligned optical observations, and natural-language descriptions in native slant-range geometry.

Methodology

The authors present SARLO-80, a multimodal dataset constructed from open-access spotlight acquisitions of the Umbra satellite constellation. The dataset creation pipeline involves the following key steps:

1. Source Data Selection: The authors selected 2,565 Sensor Independent Complex Data (SICD) scenes from Umbra, covering 72 countries and diverse environments (urban, rural, coastal, mountainous). These scenes feature VV or HH polarization, incidence angles from 10° to 70°, and native resolutions ranging from 20 cm to 2 m.
2. Frequency-Domain Resampling: To standardize the data while preserving the complex signal structure, all scenes were resampled to a fixed 80 cm slant-range grid. This was achieved via 2D Fast Fourier Transform (FFT) operations: cropping the central band for downsampling or zero-padding for upsampling, followed by an inverse FFT. This ensures Shannon–Nyquist-consistent resampling without geometric distortions.
3. Geolocation and Tiling: A dense geolocation grid was computed for each scene, mapping SAR pixels to Earth-Centered, Earth-Fixed (ECEF) and WGS84 (latitude, longitude, altitude) coordinates. The standardized scenes were then tiled into overlapping 1024 × 1024 pixel patches (stride 512).
4. Optical Alignment: High-resolution optical tiles were retrieved (Google Satellite TMS, zoom 19) and warped onto the SAR slant-range grid. This was done using a first-order 2D affine transformation estimated from corner correspondences between the SAR crop and the optical image, followed by bilinear interpolation. This approach maintains the SAR in its native geometry while providing pixel-aligned optical counterparts.
5. Caption Generation: Three caption variants (SHORT, MID, LONG) were generated for each optical patch using the CogVLM2 model. Prompts were designed to describe structural features while avoiding color terms and speculation. A Large Language Model (LLM) was subsequently used to post-process and normalize the captions for consistency.
6. Data Storage: The dataset is released as a WebDataset (~1.42 TB) containing triplets: complex SAR arrays, amplitude previews, and text descriptions. Optical images are not redistributed; instead, metadata is provided to reconstruct them on demand, adhering to third-party licensing constraints.

Key Contributions

• First Large-Scale VHR SLC SAR-Optical-Text Dataset: SARLO-80 contains 119,566 triplets covering 257 locations across 72 countries. It is the first public resource to combine VHR SAR SLC data (complex-valued), aligned optical imagery, and natural language at sub-meter resolution.
• Native Slant-Range Geometry: Unlike previous datasets that reproject SAR to ground range, this dataset preserves the native slant-range geometry and complex-valued signals. This allows for theoretically sound post-hoc resampling and retains physical phenomena like layover and shadow without geometric warping artifacts.
• Standardized 80 cm Resolution: The dataset provides a uniform 80 cm slant-range representation, bridging the gap between coarse GRD products (typically 10–30 m) and raw sensor data.
• Reproducible Benchmarking: The authors release fixed train/validation/test splits (disjoint by satellite pass to prevent data leakage), full preprocessing code, and baseline evaluation code.

Experimental Results

The authors conducted baseline experiments to demonstrate the dataset's utility for cross-modal retrieval and conditional generation, without claiming state-of-the-art performance.

• Text-to-SAR Generation: Using a fine-tuned latent diffusion model (SDXL), the authors evaluated different training strategies. A model trained with three caption variants and a final timestep-optimization phase (M_C3_TO) achieved the lowest training loss and the lowest local autocorrelation (0.38 vs. 0.21 for real SAR), indicating better restoration of high-frequency speckle structure compared to models using single captions. Qualitative results showed improved coherence in harbor structures, forest textures, and river boundaries.
• Cross-Modal Retrieval: Experiments using frozen CLIP backbones showed poor performance (R@10 < 10%), highlighting the significant modality gap between pre-trained optical models and SAR backscatter. However, full fine-tuning significantly improved retrieval performance, with the ViT-L/14 backbone achieving an R@10 of 28.47% (Image-to-Text) and 28.86% (Text-to-Image). This confirms that domain adaptation is necessary to align VHR SLC SAR with natural language.

Significance and Claims

The paper positions SARLO-80 as a foundational resource for training multimodal SAR-aware models, including Vision-Language Models (VLMs) and generative models. The authors emphasize that:

• The dataset enables research into multimodal alignment and conditional generation specifically in native SAR geometry, a capability previously unavailable at scale.
• Preserving complex-valued data allows for the study of SAR-specific physical effects (e.g., layover, non-Gaussian statistics) that are lost in intensity-only datasets.
• The dataset supports diverse tasks such as super-resolution, semantic understanding, and multimodal reasoning.
• The authors modestly frame the baseline experiments as a "proof-of-concept" to illustrate the dataset's potential rather than to establish new state-of-the-art models, noting that modeling SAR at very high resolution remains a challenging open research problem due to speckle noise and complex scattering physics.

The dataset is publicly available on the Hugging Face Hub at https://huggingface.co/datasets/ONERA/SARLO-80.