MINDSET

Annual Earth-embedding layers from Google DeepMind's AlphaEarth Foundations model: for every 10 m square of land, 64 numbers that summarise what the satellite record (optical, radar, elevation, climate and more) shows there over the year. Places that look and behave alike have similar embeddings, so they power similarity search (“find places like this”), land-cover mapping and change detection without hand-labelled training data. Coverage: global land and coastal waters, one layer per year from 2017 to 2025, in UTM tiles of 8192 × 8192 pixels.

TerraMesh A planetary‑scale, multimodal analysis‑ready dataset for Earth‑Observation foundation models: TerraMesh merges data from Sentinel‑1 SAR, Sentinel‑2 optical, Copernicus DEM, NDVI, and land‑cover sources into more than 9 million co‑registered patches ready for large‑scale representation learning. You find more information about the data sampling and preprocessing in our paper: TerraMesh: A Planetary Mosaic of Multimodal Earth Observation Data. Samples from the TerraMesh… See the full description on the dataset page: https://huggingface.co/datasets/ibm-esa-geospatial/TerraMesh.

Sentinel-2 LeJEPA Preset-Biased (Small) A small, preset-biased Sentinel-2 L2A chip dataset curated for self-supervised pretraining of a LeJEPA ResNet-18 encoder. Built as a reproducibility artifact for the Sentinel Change Explorer proof-of-concept foundation-model change-detection feature. This is a proof of concept, not a general-purpose EO pretraining corpus. It is intentionally tiny (~thousands of chips) and biased toward the five demo AOIs the Sentinel Change Explorer app… See the full description on the dataset page: https://huggingface.co/datasets/falafel-hockey/sentinel2-lejepa-global-diverse-256.

CONUS H3 Land Cover Density (NLCD 2024) A continental United States land cover dataset aggregated to Uber's H3 hexagonal grid at resolution 10 (~120m cell edge length). Each cell contains continuous per-channel coverage fractions derived from the USGS National Land Cover Database (NLCD) 2024. Unlike NLCD's native discrete classification, this dataset expresses land cover as continuous multi-channel density values — a cell at a forest-wetland boundary is represented as 0.6 forest… See the full description on the dataset page: https://huggingface.co/datasets/Timmahw/conus-h3-land-cover.

Dataset Card for Multi-Temporal Crop Classification Dataset Summary This dataset contains temporal Harmonized Landsat-Sentinel imagery of diverse land cover and crop type classes across the Contiguous United States for the year 2022. The target labels are derived from USDA's Crop Data Layer (CDL). It's primary purpose is for training segmentation geospatial machine learning models. Dataset Structure TIFF Files Each tiff file covers a… See the full description on the dataset page: https://huggingface.co/datasets/ibm-nasa-geospatial/multi-temporal-crop-classification.

Meta and Columbia University - CIESIN provide the High Resolution Settlement Layer as the world's most accurate population datasets. https://ai.meta.com/ai-for-good/datasets/high-resolution-population-density-maps/. This dataset for India, updated in 2026, relies on an updated methodology that uses the Overture Building Footprint dataset. These maps are the distribution of human population spanning India. Each of the 7 TIFF files is a 10 x 10 degree tile (the lower latitude coordinate and longitude coordinates are in the file name). A VRT file is also included.