Introduction
While satellite imagery has revolutionized how we track everything from deforestation to urban growth, most machine learning (ML) research relies heavily on optical data (like Sentinel-2). Optical sensors are intuitive and rich in spectral information, but they have a major flaw: they cannot see through clouds or operate at night.
Synthetic Aperture Radar (SAR) offers a robust solution. By using microwave pulses to "illuminate" the ground, SAR works in any weather or lighting condition. However, SAR is notoriously difficult to utilize. It involves complex numbers and phase information, requiring specialized domain knowledge to process into advanced products like "interferograms" (which measure surface displacement) or "coherence maps" (which measure surface stability).
Before M3LEO, there was no large-scale, multi-continental dataset that provided these advanced SAR products in a format ready for standard ML pipelines. Most existing datasets were either too small, limited to a single region, or only provided basic radar amplitude. M3LEO bridges this gap by providing 17 million "ML-ready" data chips across six continents, enabling researchers to seamlessly combine all-weather SAR with traditional multispectral imagery for more robust global monitoring.
Methods and Pipeline
The authors developed a comprehensive pipeline to harmonize several distinct data sources into a unified, tiled format suitable for deep learning.
- Spatial Tiling: The Earth is divided into 4.48 x 4.48 km "tiles," with data provided as "chips" (the content of a specific data product for that tile).
- SAR Processing: Moving beyond simple backscatter (amplitude), the dataset includes interferometry (measuring millimeter-scale height changes) and coherence (correlation between radar acquisitions).
- Software Framework: The release includes a modular PyTorch Lightning framework using Hydra for configuration management. This allows researchers to easily swap models or data modalities.
- Tooling: The authors released
geetilesandsartiles, open-source tools that allow users to download and process any data from Google Earth Engine (GEE) into the M3LEO-compatible format.
The M3LEO Dataset
M3LEO is effectively a "dataset of datasets," integrating multiple modalities to provide a comprehensive view of the Earth's surface:
- Sentinel-1 (SAR): Polarimetric amplitude (VV/VH), Interferometric SAR (InSAR), and Coherence.
- Sentinel-2 (Multispectral): Surface reflectance across multiple bands (RGB, NIR, SWIR).
- Auxiliary Data: Five additional datasets describing terrain properties:
- ESA World Cover: 11 land-use classes.
- ESA CCI Biomass: Above-ground biomass measurements.
- MODIS Vegetation Cover: Percentage of tree/non-tree cover.
- GHS Built Surface: Maps of human-made structures.
- SRTM: Digital elevation models.
The final dataset spans 1.05 million unique tiles and roughly 17.2 million data chips, covering 14.1% of the Earth's land surface across six major Areas of Interest (AOIs).
Key Insights and Results
By abstracting away the most difficult parts of radar science—such as pre-selecting date-pairs for coherence and providing "unwrapped" interferograms—M3LEO democratizes access to high-value radar signals for general ML researchers.
The Power of Fusion The paper demonstrates that while SAR and optical data perform similarly on their own for tasks like land-cover classification, fusing them consistently leads to superior results. Radar provides structural and textural information that optical sensors miss. The authors ran supervised baselines using a U-Net architecture to predict land cover, biomass, and built surfaces:
- Land Cover (mIoU): Fusing Sentinel-2 with Sentinel-1 (VV+VH) achieved a score of 0.4634, significantly outperforming Sentinel-2 alone (0.4094).
- Biomass (RMSE): The fusion model reduced error to 25.137 Mg ha⁻¹, compared to 28.689 for optical-only models.
The "Appearance Shift" One of the most striking findings is that the "look" of a landscape in radar data changes significantly across geographic regions, even when the terrain type is the same. For example, a "highly vegetated" area in South America produces different self-supervised embeddings than a similar area in China, highlighting the need for geographically diverse training data.
Limitations and Future Work
While M3LEO is currently the largest ML-ready polarimetric and interferometric SAR dataset available, it does have limitations:
- Temporal Gaps: Some auxiliary labels (like ESA World Cover) are only available for 2020, making it difficult to study how landscapes change over the full 2018–2020 window of the satellite data.
- Coverage Constraints: The dataset is limited to regions where Sentinel-1 operates in "Interferometric Wide" mode with dual polarization, meaning it doesn't yet cover the entire globe (e.g., parts of the deep rainforest or poles).
- Change Detection: While the dataset contains the building blocks for change detection, the initial release doesn't include specific "event labels" for things like floods or earthquakes. The authors aim to include these in future updates.
Summary updated