How AlphaEarth Foundations Redefines Planetary Mapping with AI

Background

Satellites generate petabytes of multimodal observations—including optical imagery, synthetic‑aperture radar, 3‑D lidar, and climate model outputs. Integrating these heterogeneous sources, handling data overload, and producing a consistent representation are major challenges for global Earth monitoring.

AlphaEarth Foundations Overview

AlphaEarth Foundations is an AI system developed by DeepMind that learns a unified embedding for every 10 m × 10 m grid cell on Earth. The embedding enables near‑real‑time, high‑precision mapping of land and coastal waters.

Data Ingestion

Public data streams incorporated include optical satellites (e.g., Sentinel‑2, Landsat), radar (Sentinel‑1 SAR), lidar (ICESat‑2), and climate reanalysis products (e.g., ERA5).

All inputs are reprojected onto a common 10 m grid and timestamped before training.

Model Architecture

The system relies on three technical innovations:

Adaptive decoding architecture : an implicit decoder f(t, s, x, y) that treats observation time t and sensor parameters s as continuous variables, producing a 64‑dimensional embedding for location (x, y). A correlation loss enforces consistency across overlapping observations from different sensors.

Spatially dense temporal bottleneck : features from multiple timestamps are aggregated with a time‑conditioned attention mechanism, compressing temporal information while preserving salient changes.

Geotext alignment : pixel‑level embeddings are jointly trained with geographic metadata (vector maps, land‑cover labels) using a cross‑entropy alignment term, ensuring that the learned space respects real‑world semantics.

Embedding and Compression

Each grid cell stores a 64‑dimensional float32 vector (≈256 bytes). After quantization and delta encoding, the per‑cell file size is roughly 1/16 of comparable AI mapping products, allowing planet‑scale storage at about 0.5 TB for a full‑year global dataset.

Training Procedure

Training is performed on TPU v4 pods using a distributed pipeline. Example command line:

gsutil cp gs://deepmind-alphaearth/data/*.tfrecord .
python train.py \
  --model=adaptive_decoder \
  --grid_resolution=10 \
  --embedding_dim=64 \
  --batch_size=2048 \
  --learning_rate=1e-4 \
  --loss=corr+align \
  --epochs=30

The loss combines reconstruction error, a correlation term between overlapping sensors, and an alignment term with existing land‑cover maps.

Performance Evaluation

Benchmarks on held‑out regions (e.g., Amazon basin, Sahara) show:

Mean absolute error on land‑cover classification reduced by 24 % compared with prior state‑of‑the‑art models.

When only 10 % of pixels have ground‑truth labels, error increase is less than 5 %.

Temporal consistency measured by structural similarity across consecutive days exceeds 0.92.

Visualization

Three embedding dimensions are mapped to RGB channels to produce cloud‑penetrating visualizations. Different growth stages of agricultural fields appear as distinct colors, enabling intuitive inspection of temporal dynamics.

Applications

Researchers have leveraged the embedding dataset to create custom products such as:

Food‑security indicators (crop‑type maps, yield forecasts).

Deforestation detection with sub‑monthly latency.

Urban expansion monitoring at 10 m resolution.

Water‑resource mapping in coastal zones.

Access

The dataset is available through Google Earth Engine and can be downloaded via the Cloud Storage bucket:

gsutil cp gs://deepmind-alphaearth/embeddings/*.tfrecord /local/path/

Technical report (PDF): https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphaearth-foundations-helps-map-our-planet-in-unprecedented-detail/alphaearth-foundations.pdf

Future Work

Planned extensions include expanding temporal coverage, increasing embedding dimensionality, and integrating the embeddings with large multimodal models such as Gemini to enable natural‑language querying of Earth observations.

System Architecture Illustration