README.md

metadata

datasets:
  - nesteo-prototype
language:
  - en
size_categories:
  - 10M<n<100M
multilinguality: monolingual
tags:
  - ai4eo
  - earth-observation
  - remote-sensing
  - multimodal
  - grids
license: cc-by-4.0
task_categories:
  - image-segmentation
  - image-classification
pretty_name: NestEO Prototype
configs:
  - config_name: grids_selected_1200m
    default: true
    data_files:
      - split: train
        path: grids/grids_selected/selected_1200m_grid.parquet

NestEO: Modular and Hierarchical EO Dataset Framework

NestEO is a hierarchical, resolution-aligned, UTM-based nested grid dataset framework supporting general-purpose, multi-scale multimodal Earth Observation workflows. Built from diverse EO sources and enriched with metadata for landcover, climate zones, and population, it enables scalable, representative and progressive sampling for AI4EO.

Grid Levels: 120000m, 12000m, 2400m, 1200m, 600m, 300m, 150m
Grid Metadata: ESA WorldCover proportions, GHSL, Köppen Climate
Current Sample Datasets: Wyvren Hyperspectral, Satellogic Newsat, Sentinel-2, Sentinel-1, Sentinel-3 Zones: UTM 1N–60N, 1S–60S and Polar North/South Formats: Parquet, GeoParquet, Zarr (planned) License: CC-BY-4.0
More Info: See paper, GitHub.

Directory Structure Scaffold

NestEO/
├── grids/                  # UTM-aligned hierarchical grid Parquet files
├── metadata_current/       # Precomputed proportions: landcover, planned (climate, region, biomes, ghsl)
├── datasets_EO/            # Clipped EO imagery tiles (Newsat, Sentinels, planned(Landsat, HLS, MODIS, etc.))
├── datasets_AUX/           # Auxiliary datasets planned(DEM, landcover, OSM)
├── embeddings/             # Precomputed model embeddings planned (e.g., DINOv2, SigLIP, SAM, SAM2)
├── index_structure/        # tile-to-super-tile, Source-to-tile and tile-ID index maps
└── versions/               # Information about snapshots for version control

Dataset Structure

• Each tile includes a unique tile_id, and spatial geometry.
• Grouped by zone directories: grid_2400m/grid_37N_2400.parquet

Code

Github code provides the grid creation logic, ESA Worldcover proportion calculations, notebooks for processing sample datasets, and fetching/merging grids with datasets

How to Use

• Install from Github and create your own Nested grid structure
• Load a grid from grids/ (e.g., grid_1200m.parquet).
• Join with metadata (e.g., metadata_lc.parquet) on tile_id.
• Apply spatial or semantic filtering (e.g., only urban, with specific dates).
• Select and load paired EO imagery and sources from folders.

Current Status

This is a prototype release of the NestEO dataset framework, showcasing the core design principles, including hierarchical UTM-aligned grids, modular directory layout, structured GeoParquet-based metadata, and compatibility with scalable Earth Observation (EO) data ingestion. This release includes a limited set of EO source tiles and metadata layers meant to demonstrate framework usability, alignment strategies, and pipeline extensibility. Metadata-driven filtering, source-pairing, and tile-to-super-tile indexing are fully implemented.

Limitations

• This version provides a demonstration of the framework vision. Only few selected EO modalities (e.g., Sentinel-2, Sentinel-3, Satellogic, Wyvren) are currently included at sample scale.
• Further EO content layers (imagery, land cover, auxillary layers etc.) to be added in later updates for NestEOv1.
• Cross-modal pairings, auxiliary layers (e.g., DEM, OSM), and full-resolution coverage across grid levels are under active development.
• Not all grid levels are currently populated with imagery; some grid layers are provided primarily to support alignment, pairing and demonstration purposes.

Roadmap and Ongoing Work

A comprehensive v1 release is under development. This upcoming version will include:

• 12+ EO modalities (optical, SAR, hyperspectral, atmospheric, thermal)
• Over 250,000 geospatially distributed tiles across multiple grid levels
• Paired imagery samples at selected locations, across grid levels, supporting multimodal learning
• Expanded metadata layers including ESA WorldCover, Copernicus DEM, Köppen-Geiger zones, and GHSL-derived population classes
• Precomputed model embeddings (e.g., DINOv2, SigLIP, SAM2) at tile-level granularity
• Full compatibility with cloud-based filtering, lazy loading, and Hugging Face Datasets

All v1 expansions will preserve the NestEO prototype structure and align with FAIR dataset development practices. The prototype release remains stable as a referential snapshot of the NestEO design framework, while v1 will extend data completeness, modeling readiness and potential foundation model development.

Contributing

We welcome contributions on:

• Region-specific or resolution-specific imagery
• Auxiliary or annotation layers
• Grid-level metadata enrichment
• Benchmarks and model evaluations

Citation

TBD – upon official release.