MajorTOM:Elliot Pretrain: 279K multi-sensor, multi-temporal tiles for Earth observation models
HildeKuehne · x · 2026-10-17
A new arXiv dataset paper, MajorTOM:Elliot Pretrain, targets pretraining of generalist Earth observation foundation models with a unified open dataset combining sensors, temporal sampling and text:
- Scale: 279,166 tiles of 10 km sampled from the Major-TOM grid by clustering AlphaEarth embeddings.
- Each tile stacks Sentinel-2 L1C, Landsat-8/9, Sentinel-1 RTC, Copernicus DEM and ESA WorldCover on a shared 1056x1056-pixel, 10 m grid, plus cloud masks, OpenStreetMap entities and ERA5 weather.
- Three temporal regimes: Monotemporal (250,000 tiles, one acquisition), Monthly (12,500 tiles, twelve acquisitions over the annual cycle), Burst (16,666 tiles, six acquisitions within 30 days).
- Distribution: 6.3 TB of Cloud Optimized GeoTIFFs indexed by TACO v3 Parquet metadata, enabling cloud streaming and pre-read filtering; a companion library derives sensor-tagged fact sheets, captions and 16 families of vision-language tasks at load time.
- Validation: short pretraining runs of OlmoEarth-Base (effective even on a 25% subset) test the design choices.
A directly usable resource for anyone training remote-sensing foundation models or geo multimodal datasets.
More from Research
- $10M XPRIZE on Interspecies Communication launches next month, betting AI can talk to animals — anselm · 2026-10-17
- Bypassing RAG: Injecting 30,000-Token Tensor Cartridges at Layer 3 of Frozen Qwen2.5-72B on One A100 — Nearby_Indication474 · 2026-10-17
- Distillation angle overrated: frontier reasoning gains mostly come from internal RL, argues analyst — teortaxesTex · 2026-10-17
- World model trained from just hours of play data predicts world state given robot actions — micoolcho · 2026-10-17
- micro1 launches LegalReasonBench, an attorney-authored long-horizon legal reasoning benchmark — Exp_Mark · 2026-10-17
- Gen-Verse Releases DiscoveryRSI-27B for Recursive Scientific Hypothesis Development — jacek2023 · 2026-10-17